- Microwave Remote Sensing of Water in the Soil Plant System MCHNES by MASSACHUSETTS INSTITUTE OF TECHNOLOGY Alexandra Georges Konings DEC 0 9 2015 S.B., Massachusetts Institute of Technology (2009) M.S., Duke University (2011) LIBRARIES Submitted to the Department of Civil and Environmental Engineering in partial fulfillment of the requirements for the degree of Doctor of Philosphy at the MASSACHUSETTS INSTITUTE OF TECHNOLOGY September 2015 Massachusetts Institute of Technology 2015. All rights reserved. Signature redacted Author ....................... Department of Civil and Environmental Engineering August 13, 2015 Certified by.............Signature redacted>..... Dara Entekhabi Bacardi and Stockholm Water Foundations Professor Signature redactedThesis Supervisor A ccepted by ................... ................. Heidi M. Nepf Donald and Martha Harleman Professor of Civil and Environmental Engineering Chair, Department Committee for Graduate Students 2 Microwave Remote Sensing of Water in the Soil - Plant System by Alexandra Georges Konings Submitted to the Department of Civil and Environmental Engineering on August 13, 2015, in partial fulfillment of the requirements for the degree of Doctor of Philosphy Abstract Remotely sensed measurements made by radars or radiometers in the low microwave frequency range are sensitive to soil moisture, soil roughness, and vegetation water content. Measurements made at multiple polarizations can be used to determine additional ancillary parameters alongside the primary variable of interest. However, if an attempt is made to retrieve too many parameters from too few measurements, the resulting retrievals will contain high levels of noise. In this thesis, I introduce a framework to determine an upper bound on the number of geophysical parameters that can be retrieved from remotely sensed measurements such as those made by microwave instruments. The principles behind this framework, as well as the framework itself, are then applied to derive two new ecohydrological variables: a) soil moisture profiles across much of the root-zone and b) vegetation optical depth, which is proportional to vegetation water content. For P-band observations, it is shown that soil moisture variations with depth must be accounted for to prevent large forward modeling - and thus retrieval - errors. A Tikhonov regularization approach is then introduced to allow retrieval of soil moisture in several profile layers by using statistics on the expected co-variation between soil moisture at different depths. The algorithm is tested using observations from the NASA Airborne Microwave Observatory of Subcanopy and Subsurface (AirMOSS) Mission over the Harvard Forest in Western Massachusetts. Additionally, at L-band, a multi-temporal algorithm is introduced to determine vegetation optical depth (VOD) alongside soil moisture. The multi-temporal approach used reduces the chance of compensating errors between the two retrieved parameters (soil moisture and vegetation optical depth), caused by small amounts of measurement noise. In several dry tropical ecosystems, the resulting VOD dataset is shown to have opposite temporal behavior to coincident cross-polarized backscattering coefficients, an active microwave indicator of vegetation water content and scattering. This possibly shows dry season bud-break or enduring litter presence in these regions. Lastly, cross-polarized backscattering coefficients are used to test the hypothesis that vegetation water refilling slows down under drought even at the ecosystem scale. Evidence for this hypothesis is only found in the driest location tested. 3 Thesis Supervisor: Dara Entekhabi Title: Bacardi and Stockholm Water Foundations Professor 4 Acknowledgments This thesis was funded by an NSF Graduate Research Fellowship, a NASA Earth and Space Science Fellowship, and the NASA AirMOSS and SMAP missions. I am grateful to both NSF and NASA for funding my research. Both this dissertation and much of my skill as a scientist and engineer is owed a great debt to my advisor, Dara Entekhabi. I am exceedingly grateful for his technical guidance and countless insightful comments and creative ideas. He has taught me far more than is contained within these pages, and I consider it a proud moment each time I catch myself thinking like he would (or adopting his speech mannerisms accidentally). Aside from being a great guide technically, I am also grateful for his general mentorship throughout these early steps of my academic career. Even when our technical viewpoints were diametrically opposed, his support was unwavering. His guidance was multiplied through countless travel opportunities enabling me to meet and learn from others in the field. He is a source of inspiration both as a researcher and as a human being, and it has been a joy to work with him on a personal level. I am grateful to my committee members Richard Cuenca, Charlie Harvey, Mahta Moghaddam, and Sassan Saatchi for their helpful comments, questions, and suggestions at various times during the completion of this work, as well as their patience with the technical challenges and travel made necessary by a committee consisting of such wide-flung members. The Entekhabi Lab members (Kaighin McColl, Hamed Alemohammad, Dave Whittleston, Parag Narvekar, Aldrich Castillo, and Siggi Magnusson) have contributed to an unusually self-reliant, cohesive, and supportive group for which I am very grateful. Special thanks are due to Kaighin McColl and Hamed Alemohammad, with whom I have enjoyed working on a variety of projects and exchanging countless conversations on the academic journey and the field of hydrology. Thanks also to the many temporary visitors of the Entekhabi lab, especially Thomas Jagdhuber, Maria Piles, and Mariette Vreugdenhil. It has been a great joy to work in such a collaborative environment. 5 Gaby Katul from Duke University continues to be a wonderful mentor, whose technical prowess and kind are equally inspiring. My thanks also to Amilcare Porporato and recent members of the Katul, Porporato, and Oren labs at Duke. The community at the Parsons lab is truly unique. The inevitable ups and downs of the research process over the last few years were transformed into a far more exciting and enjoyable adventure because of the great community at the lab. Buoyed by a joint resistance to leaks, repairs, and other facilities inconveniences, nothing can stop the lab's community. This community could not have grown and survived without the tireless work of Sheila Frankel and Jim Long. Sheila's tireless listening ear and countless hours of advice (and entertainment) kept me from despair more times than I like to acknowledge. Thanks to my Parsons friends and their partners for making me smile every day - Ruby Fu, Alison Hoyt, Irene Hu, Kyle Delwiche, Kelsey Boulanger, Dave Ridley, Jen Nguyen, Ben Scandella, Kaighin McColl, Dave Whittleston, Fatima Hussain, Jenn Apell, Anthony Carrasquillo, and so many other past and present Parsonites. Vicki Murphy deserves thanks for happily and speedily dealing with all my financial concerns no matter how complicated, including many interesting conversations and invaluable insights into the byzantine world of the research grant system along the way. Thanks also to Joe Abel, Adam Bouland, and Ryan Lewis for being great friends and for understanding the Parsons life through their partners. MIT S.B. holders have a special outlook on the world, so thanks to Denise Ichinco, Apoorva Murarka, and Knight Fu for always putting things in perspective for me and for the joy of their camaraderie. Regularly going climbing kept me sane during the last few years of my PhD - thanks to Ruby Fu, Ryan Lewis, Zach Shepherd, Kyle Peet, Kelly Daumit, Jess Bryant, and other climbing partners for allowing me into the wonderful and addictive world of climbing (and for not dropping me on belay!). Last but not least, I would not have made been able to get anywhere near completing this journey without the endless support and love of my family. Ik hou van jullie. 6 Contents 1 23 Introduction 1.1 The role of plant and soil water content in the global water, carbon, 23 1.2 Motivation for using microwave remote sensing . . . . . . . . . . . . 26 1.3 Introduction to microwave remote sensing . . . . . . . . . . . . . . 27 1.4 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 30 . . How Many Parameters Can Be Maximally Estimated from a Set of 33 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.2 Degrees of information . . . . . . . . . . . . . . . . . . . . . . . . . 35 2.2.1 Definition of Degrees of Information DoI . . . . . . . . . . . 36 2.2.2 Dependence on the bin size parameter . . . . . . . . . . . . 39 . . . . . . . . . . . . . . . . . . . . . . . 41 . . . . Measurements? Example Dol calculations 2.4 Applications to particular remote sensing observations . . . . . . . 44 2.5 C onclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 . . 2.3 . 2 . . and energy cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 The Effect of Variable Soil Moisture Profiles on P-band Backscatter 47 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 3.2 Forward M odel . . . .. 50 . 3.1 . . . . . . . . . . . . . . . . . . . . . . . .. 3.2.1 Hydrologic modeling .... 3.2.2 Backscattering Coefficient Model ... 3.2.3 Application of Scattering Model at Vaira Ranch . . . . . . . ....................... 50 ................ . 54 7 57 Behavior of Multi-layer Scattering . . . . . . . . . . . . . . . . . . . 59 3.4 Forward Modeling Error at Vaira Ranch 66 . Layering Approaches . . . . . . . . . . . . . . . . . . . . . . 66 3.4.2 R esults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 . . 3.4.1 Discussion and Conclusions . . . . . . . . . . . . . . . . . . . . . . 71 The Use of Regularization for Improving Profile Soil Moisture Re- 77 4.1 Introduction . . . . . . . . . . . . . . . . . 77 4.2 Regularization Approach . . . . . . . . . . 81 4.2.1 Cost functions . ... . . . . . . . . . 81 4.2.2 Illustrative example . . . . . . . . . 82 4.5 . . . 4.3.1 Model description . . . . . . . . . . 86 4.3.2 Experiments performed . . . . . . . 87 4.3.3 Results . . . . . . . . . . . . . . . . . 86 . . . . . . . . . Application to AirMOSS observations at Harvard Forest 90 94 Derivation of regularization parameters using an OSSE 96 4.4.2 Comparison to in situ data . . . . . . . . . . . . . . . . . . . 97 4.4.3 Comparison with landscape characteristics . . . . . . . . . . 99 . . . . . . . 107 Discussion and Conclusions . . 4.4.1 . . . . . . . . . . . . . . . . 4.4 Observing System Simulation Experiment at Vaira Ranch. . 4.3 . trievals from P-Band Radar Measurements . 4 . . . . . . . . . . . . . . . . 3.5 . 3.3 5 Vegetation Optical Depth and Albedo Retrieval using Time Series of Dual-polarized L-band Radiometer Observations 109 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 5.2 Algorithm Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 5.3 5.2.1 Classical retrieval approach 5.2.2 Timeseries motivation . . . . . . . . . . . . . . . . . . . 112 . . . . . . . . . . . . . . . . . . . . . . 114 Algorithm Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 5.3.1 Moving window timeseries design . . . . . . . . . . . . . . . . 117 5.3.2 Albedo retrieval . . . . . . . . . . . . . . . . . . . . . . . . . . 119 8 121 5.4 M ethods . . . . . . . . . . . . . . . . . . . . . . . . 121 5.5 Datasets used . . . . . . .. . . . . . . . . . . . . . 124 . NCEP land surface temperatures and flags . 125 5.5.3 . . 5.5.2 MODIS IGBP land cover . . . . . . . . . . . 125 5.5.4 MERRA-Land observation-corrected global pr 126 5.5.5 Water fraction . . . . . . . . . . . . . . . . . 126 R esults . . . . . . . . . . . . . . . . . . . . . . . . . 126 5.6.1 VOD retrievals . . . . . . . . . . . . . . . . 126 5.6.2 Albedo retrievals . . . . . . . . . . . . . . . 130 5.6.3 k retrievals 132 . . . . . . 124 Discussion and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 139 6.1 Introduction . . . . . . . . . . . . . .. . . . . . . . . 139 6.2 VOD and . 140 6.3 Possible explanations . . . . . . . . . . . . . . . . . 144 6.3.1 Dry season bud break and leaf flushing . . . 144 6.3.2 Litter . . . . . . . . . . . . . . . . . . . . . 147 Conclusions . . . . . . . . . . . . . . . . . . . . . . 147 are out of phase in several regions . . . YHV . On the Seasonal Behavior of Microwave Vegetation Indices 6.4 7 . . . . . . . . . . . . . 6 Aquarius Level 2 data . 5.7 5.5.1 . 5.6 Additional parameters . . . . . . . . . . . . . 5.3.3 Variations in Diurnal Canopy Water Content Refilling with Water Stress 8 149 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Derivation of AM and PM VWC 7.3 Soil moisture dependence of diurnal variability of 7.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . 7.1 . . . . . . . . . . . . . . - YHV - . . . . . . . 149 . . . . . . . 151 - . . . . . . 152 . . . . . . . 156 Conclusions and Future Work 159 8.1 159 C onclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 8.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 8.2.1 Root-zone soil moisture . . . . . . . . . . . . . . . . . . . . . . 162 8.2.2 Vegetation optical depth and water content 164 A Vegetation Parameters for Hydrologic Modeling 10 . . . . . . . . . . 169 List of Figures 2-1 Normalized total correlation C, between Aquarius Tbv and TbH as a function of the bin sizes ATbV and ATbH. For large bin sizes relative to the dynamic range of the variables, the C, suddenly drops when the number of bins is so low that even the approximate shape of the joint pmf is distorted by the wide bins. The black triangle corresponds to the bin sizes recommended by Scott's rule. . . . . . . . . . . . . . . . 2-2 40 The marginal (top) and joint (bottom) probability density functions (pdf) for observed vertically and horizontally-polarized brightness temperatures (Tbv and TbH, respectively) from the Aquarius satellite. Note that the edges of the joint pdf tails extend beyond the region shown; the figure is zoomed in for clarity. 3-1 . . . . . . . . . . . . . . . 42 Comparison between modeled soil moisture (black line) and Ameriflux soil moisture observations (red dashed line) at 5 cm (top) and 10 cm (bottom) depth over the 10 year simulation period used in this study. 3-2 52 Comparison between modeled latent heat flux (black line) and Ameriflux latent heat flux observations (red dashed line) over the 10 year simulation period used in this study. 3-3 . . . . . . . . . . . . . . . . . . 52 Average simulated soil moisture profile at Vaira Ranch, CA during the local dry season (June through November, black solid line) and wet season (December to May, red dashed line). The gray and pink shaded regions represent one standard deviation around the mean profile during the dry and wet season, respectively 11 . . . . . . . . . . . . . . . . 53 3-4 Seasonal evolution of simulated 'true' P-band backscattering coefficients at VV- (top) and HH-polarization (bottom) over the year 2001, assuming an incidence angle of 30'. The soil roughness rms height s=:0.02 m at the surface, while all subsurface layer are assumed to be smooth. 3-5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . 59 Evolution of the penetration depth (in cm) over the ten years of simulation. Tick marks labeled with each year correspond to the first day of that year. The inset shows the annual average cycle of the penetration depth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-6 61 Example soil moisture profile simulated by the SHAW model (see Section 3.2.1), which has cHH = -24.9 dB and c-vv = -16.9 dB when a 210-layer approximation to the full soil moisture profile is used in the P-band backscattering model (at a 30' incidene angle and frequency of 430 MHz). The vertical lines are the equivalent halfspace moistures that show the same HH-polarized backscattering coefficient (red, dashed line) and the same VV-polarized backscattering coefficient (blue, dash-dotted line) under the same texture, roughness, and sensing conditions. The roughness and sensing parameters used are identical to those in Section 3.2.3. . . . . . . . . . . . . . . . . . . . . 3-7 63 Backscattering coefficients associated with a linear soil moisture profile between various soil moisture values at the surface (z=0) and a value of 0.40 at 30 cm depth (pink dash-dotted line). The bottom x-axis indicates the value of soil moisture at the surface, while the top x-axis labels indicate the average soil moisture over the top 30 cm associated with each linearly varying profile. The black, red dashed, and blue dotted lines represent backscattering coefficients associated with uniform, two-layer, and three-layer approximations to the linear profile, respectively. The left plot shows the backscattering coefficients in the VV-polarization, while the right shows the HH-polarization. Roughness and measurement parameters were the same as in Fig. 3-6. 12 . . . 65 3-8 Soil moisture variation with depth for an example profile. The dashed horizontal lines represent the layer interfaces used in the equal layer (left) and moisture-dependent (right) approaches. The solid horizontal lines show the bottom of the depth the lowest layer is assumed to represent. For the equal-layer approach (left) this is the time-averaged penetration depth, while for the moisture-dependent approach (right) it is the penetration depth associated with the specific profile. .... 3-9 68 Seasonal evolution of backscattering coefficients for different multilayer representations of the soil moisture profile. The top row shows the backscatter at VV-(black line) and HH-polarization (red line) for a 210-layer representation of the soil moisture profile and is taken to represent the true backscatter. The bottom two rows show the difference between the backscattering coefficients for different multi-layer representations and the true backscattering coefficients at VV-pol (middle row) and HH-pol (bottom row). The black, red, and blue lines represent the differences for one-, three-, and five-layer representations, respectively. For each panel, the interfaces between homogeneous layers are spaced at equal intervals in the left column, and are spaced depending on the locations of maximum soil moisture gradients in the right column. A measurement frequency of 430 MHz, incidence angle of 300, and roughness rms-height of 2 cm are assumed. . . . . . . . . 69 3-10 RMSE (left) and bias (right) for soil moisture profile representations with different numbers of layers. Solid lines represent the error at VVpolarization, while dashed lines represent the error at HH-polarization. Black lines refer to the error when layers are distributed so as to have equal thickness, while red lines are for simulations in which the layer thicknesses depend on the soil moisture profile. Throughout, a measurement frequency of 430 MHz, incidence angle of 300, and roughness rms-height of 2 cm are assumed. . . . . . . . . . . . . . . . . . . . . . 13 70 4-1 Sample soil moisture profile derived from in situ measurements at Harvard Forest on February 4th, 2012 used to calculate the 'true' forward backscatter of Fig. 4-2. The dashed line represents the average soil moisture over two layers of 14 cm each; the optimal solution of a twolayer retrieval algorithm . 4-2 . . . . . . . . . . . . . . . . . . . . . . . . 84 Different. cost function components as a function of the top-layer and bottom-layer soil moisture. In the top left panel, the objective function without regularization (minimizing the squared normalized difference between measurements and expected values) is shown, where the measurements are those expected for the profile shown in Figure 4-1. In the bottom left panel, a pure regularization term is shown. The right panel shows the sum of the two, with the regularization term weighted by parameter A = 0.01. In each panel, the white square represents the global minimum of the displayed quantity, while the pink square represents the optimal solution for the profile used to generate the simulated m easurem ents. 4-3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 Timeseries of 'true' average soil moisture simulated by SHAW for each of the four layers retrieved by the algorithms, i.e. the top six cm (black solid line), 6-12 cm (red dotted line), 12-18 cm (blue dash-dotted line), 18-24 cm (green solid line). 4-4 . . . . . . . . . . . . . . . . . . . . . . . 88 Mean profile (left column) and inverse covariance matrix (right column) for the forward (top row) and alternative (bottom row) hydrologic m odel results. 4-5 . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 Root-mean-square error of one year of simulated regularized retrievals as a function of the regularization weight A at Vaira Ranch . . . . . . 14 91 4-6 Timeseries of the top-layer (0-6 cm) retrieved soil moisture using different retrieval algorithms. The red dotted line is the retrieved value using a regularized cost function with four layers and the value of A that corresponds to the global minimum. The blue dash-dotted line is the top layer value for unregularized retrievals with the same number of layers. The black line us the true average soil moisture over the top six cm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-7 Timeseries of the true (black solid line) and retrieved soil moisture (red dashed line) using regularization in each of the four retrieval layers. 4-8 4-9 92 . 93 used in the regularization for Harvard Forest . . . . . . . . . . . . . . 97 Mean profile (left column) and inverse covariance matrix (right column) Root-mean-square error of one year of simulated regularized retrievals as a function of the regularization weight A at Harvard Forest . . . . 98 4-10 Average 7-cm soil moisture values across interpolated soil moisture profiles (left) used as ideal retrievals for the OSSE at Harvard Forest and (right) values retrieved by the regularization algorithm with A = 10-2.1. 98 4-11 Top-layer (0-7 cm) regularized retrievals (red triangles) compared to the range of in situ measurements (0-5 cm) across each AirMOSS pixel with in situ observations. The black dot is the average of the in situ measurements. Data from the October 15th, 2012 flight are shown on the left, while the right figure shows data from October 18th, 2012. . 100 4-12 Cumulative distribution of sand fraction, clay fraction, topographic moisture index, elevation, and retrieved biomass for the entire Harvard Forest flight path (black line) and the 3 focus regions (red line), respectively. The bottom right shows the distribution of land cover type between the flight path and the focus regions (DF deciduous forest, EF = evergreen forest, MF = mixed forest, SL shrubland, GL = grassland, PS = pasture, and CR = cropland). . . . . . . . . . 101 15 4-13 Landscape characteristics of focus area 1 at the Harvard Forest. Top row: sand fraction (left), clay fraction (center), and elevation (right). Bottom row: land cover (left), retrieved aboveground biomass (center), and topographic moisture index (right). The bottom two rows show the retrieved soil moisture in the 0-7 cm (top left), 7-14 cm (top right), 14-21 cm (bottom left), and 21-28 cm layers, respectively. . . . . . . 103 4-14 Landscape characteristics of focus area 2 at the Harvard Forest. Top row: sand fraction (left), clay fraction (center), and elevation (right). Bottom row: land cover (left), retrieved aboveground biomass (center), and topographic moisture index (right). The bottom two rows show the retrieved soil moisture in the 0-7 cm (top left), 7-14 cm (top right), 14-21 cm (bottom left), and 21-28 cm layers, respectively. . . . . . . 104 4-15 Landscape characteristics of focus area 3 at the Harvard Forest. Top row: sand fraction (left), clay fraction (center), and elevation (right). Bottom row: land cover (left), retrieved aboveground biomass (center), and topographic moisture index (right). The bottom two rows show the retrieved soil moisture in the 0-7 cm (top left), 7-14 cm (top right), 14-21 cm (bottom left), and 21-28 cm layers, respectively. . . . . . . 105 4-16 Boxplot of average 0 - 7 cm layer soil moisture retrieval for different sand fraction classes. . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 4-17 Boxplot of average 0 - 7 cm layer soil moisture retrieval for different topographic moisture index classes (in m). . . . . . . . . . . . . . . . 106 16 5-1 Cost function J as a function of VOD and k for a sample set of observations (July 16th, 2012, for a pixel centered at 19.48'N, 103.531W in Central Mexico). The 'true' solution of the cost function (without noise added) is shown by a black dot. A small amount of simulated noise is added to the observations, 0.005 for the H-pol and -0.002 for the V-pol. The contours of the resulting noisy cost function are shown as black lines. The noisy solution of the resulting cost function is shown by a red triangle and is far away from the true solution. . . . . . . . . 5-2 Retrieval ratio of degrees of freedom for the different land uses and varying number of dual-polarized observations . . . . . . . . . . . . . 5-3 116 119 Relative contribution of the vegetation canopy to the total brightness temperature emitted at H-polarization, T""P"/T-H as a function of albedo w and VOD. A value of k = 20 is assumed. Results at V- polarization are qualitatively similar (not shown). . . . . . . . . . . . 5-4 120 Global maps of mean MT-DCA (left) and LPRM (right) VOD retrievals for the three year period of this study. . . . . . . . . . . . . . 128 5-5 Time series of weekly mean MT-DCA VOD, LPRM VOD, and precipitation over focus pixels. Note the different axes scale for the Amazon series. 5-6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 Global maps of standard deviation of MT-DCA (left) and LPRM (right) VOD retrievals for the three year period of this study. In both cases, a 5-week moving average is first removed from the timeseries for each pixel, so that the standard deviation primarily reflects high-frequency variability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-7 129 Joint density of Aquarius radiometer-derived vegetation optical depth vs. scatterometer c-HV in linear units (left) and radar vegetation index (right). All available combinations of active and passive measurements (e.g. one at each location and time) were used. . . . . . . . . . . . . . 130 5-8 Global map of retrieved albedo . . . . . . . . . . . . . . . . . . . . . 132 5-9 Mean difference between the two sets of k retrievals . . . . . . . . . . 133 17 5-10 Global maps of temporal mean k retrieval for Aquarius (left) and SMOS (right) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 5-11 Time series of weekly mean MT-DCA k, LPRM k, and precipitation over focus pixels. Note the different axes scale for the Amazon series. 6-1 Pearson correlation coefficient between 134 -HV and VOD for all global pixels for which there are at least 50 valid weekly VOD retrievals in three years. 6-2 Pearson correlation coefficient between with 6-3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 RUHV-VOD < -0.5. and precipitation for pixels . . . . . . . . . . . . . . . . . . . . . . . . . 142 Pearson correlation coefficient between VOD and precipitation for pixels with RHV-VOD < -0.5. . Black symbols represent the locations of the pixels shown in Figure 6-4. 6-4 cTHV . . . . . . . . . . . . . . . . . . . . . 142 Timeseries of rainfall and vegetation indices for P1 (top), P2 (center), and P3 (bottom), respectively. The P1-P3 locations are marked in Figure 6-3. For each timeseries, the average weekly VOD (blue line), aTHV in units of power (green line), EVI (red line), and rainfall (histogram) are show n. 7-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 Locations with same-day morning and afternoon observations whose footprints are less than 0.05' apart. 18 . . . . . . . . . . . . . . . . . . 153 7-2 Difference between AM and PM UHV (in dB) vs. radiometer-derived AM soil moisture for the pixels in Figure 7-1 after filtering for dew and interception. Errorbars on the differences are based on the Aquarius Kp uncertainty estimates of the backscattering cross-sections, where additive errors are assumed to be distributed normally with a standard deviation of Ko-PQ. The inset of each pixel shows the location (country, center coordinates), dominant land-cover, Gini-Simpson index of land cover, and annual average rainfall, respectively. Pixels are arranged in order of decreasing averaging rainfall. Pixels where less than three samples remained after dew and interception filtering, or where more than 0.05% of the covered land surface area was water, are removed from this analysis. 7-3 . . . . . . . . . . . . . . . . . . . . . . . 155 Mean diurnal difference of JHV (AM - PM) across days without interception or dew formation (in dB) as a function of average annual rainfall for each pixel. . . . . . . . . . . . . . . . . . . . . . . . . . . 156 19 20 List of Tables 2.1 Dol for several examples . . . . . . . . . . . . . . . . . . . . . . . . . 43 3.1 Vegetation scattering parameters used . . . . . . . . . . . . . . . . . 58 4.1 Soil parameters for different SHAW runs . . . . . . . . . . . . . . . . 89 4.2 Root-mean-square retrieval error for different algorithms . . . . . . . 93 5.1 Target areas: name, location, dominant IGBP land cover type, and coefficient of determination R 2 between time series of mean Aquarius and SM OS k retrievals. . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 122 Land cover variability of retrieved albedo w. Parameters for SMAP w are obtained from O'Neill et al. (2012) . . . . . . . . . . . . . . . . . A.1 Vegetation parameters used for hydrologic modeling 21 131 . . . . . . . . . 170 22 Chapter 1 Introduction 1.1 The role of plant and soil water content in the global water, carbon, and energy cycles Water on earth exists in a variety of states and storage components. Of these, by far the largest by volume is the ocean, followed at some distance by water in ice caps and glaciers (Bras, 1989). Together, these two components make up 97.35 % of the water on earth. Water stored in the soil and vegetation water makes up less than 1% of the remainder (Bras, 1989). Although the total volume of soil and vegetation water is small, it has an important influence on the highly dynamic global fluxes that enable life on earth. Large-scale estimation of soil moisture has a variety of societal applications, including the potential for better drought prediction, flood forecasting, wildfire risk management, irrigiation management, and predictions of disease transmission (e.g. Entekhabi et al., 2010a). The outsized influence of soil moisture is partially at- tributable to its role in modulating the fluxes of water, carbon, and energy at the land surface. Soil moisture influences recharge rates to groundwater (National Research Council, 2004), a critical component in our ability to determine at what rates water can be withdrawn sustainabily from groundwater aquifers (or more likely, how unsustainable current withdrawal rates are). Soil moisture influences the partitioning 23 of rainfall into infiltration and runoff, which can influence erosion and transport nutrients away from a site (Kinnell, 2010; Fiener et al., 2011). Soil moisture also exerts a first-order control over the rates of transpiration and bare soil evaporation. Because of the high latent heat of water, the evaporation of water from soil or plants uses a significant fraction of the energy that arrives at the land surface. As a result, soil moisture influences the partitioning of net radiation into the latent heat flux associated with evapotranspiration, and sensible heat fluxes into the atmosphere and into the ground. As a result, soil moisture influences the local weather evolution and, if large-scale free atmospheric conditions are favorable, can influence the development of convective rainfall (e.g. Konings et al., 2010). Because of the strong memory of soil moisture, this influence can extend to seasonal and even interannual timescales (Koster et al., 2000; Koster, 2004). By modulating the amount of trnaspiration, soil moisture anso has a key influence on carbon fluxes and plant health. About 65 % of global evapotranspiration fluxes occur through plant transpiration (Jasechko et al., 2013; Good et al., 2015). Plant transpiration is modulated by stomatal closure which depends not only on soil moisture but also on atmospheric demand (e.g. atmospheric humidity and temperature and available radiation energy). Plant behavior falls along a spectrum from anisohydric behavior - significant stomatal closure in response to soil moisture limitations - to isohydric behavior, in which no stomatal closure occurs under soil moisture limitations (Klein, 2014; Skelton et al., 2015). Limitations in available water not only reduce instantaneous transpiration rates, but in the longer term can also lead to mortality . Since droughts are expected to become more common under climate change (Ciais et al., 2013) and the habitats of many plant species are already shifting (Kelly and Goulden, 2008; Loarie et al., 2009), questions about plant responses to drought and plant mortality are more pressing than ever. However, the mechanisms underlying long-term vegetation response to drought are still poorly understood (van der Molen et al., 2011; Breshears et al., 2009). Depending on the type of drought, plants are at risk of xylem embolism if sufficiently large gradients of water potential cannot be maintained, leading to hydraulic failure, or, if stomatal closure is too aggressive, carbon starvation due to the 24 reduced uptake of carbon when stomata are closed (Sperry and Love, 2015; Meir et al., 2015; McDowell et al., 2008, 2011; Anderegg et al., 2012). Carbon starvation may also increase plant vulnerability to insect attacks (Anderegg et al., 2015). Predicting plant vulnerability to drought requires being able to predict transpiration across a drought or drydown period. Most current models of transpiration are based on an analogy with electrical resistors in which stomatal closure affects the resistance or (its inverse) conductance of the plant. The transpiration fluxes are driven by the difference in vapor pressure between saturated interstitial spaces in the stomata and the unsaturated humidity in the atmosphere, analogous to a voltage difference across a resistor (Campbell, 1985). The degree of isohydricity of the plant species is then generally incorporated in the maximum stomatal resistance of the plant. In doing so, these models' ability to represent variations in plant behavior is limited to those cases where average variations in maximum stomatal conductance have been measured and specified a priori. This limits the ability of models to represent species whose parameters are not well-studied. Furthermore, variations in stomatal density, size, and closure amounts due to plant height (Novick et al., 2004), nutrient availability (Ewers et al., 2001), and even origin (Bourne et al., 2015) are routinely neglected. Such models also neglect the storage and hydraulic transport of water within the plant, which would correspond to a capacitor in the electric circuit analogy. Nevertheless, plant (an)isohydry, and thus the response of transpiration to water limitations, have been shown to be related to hydraulic architecture (Schultz, 2003; Martinez- Vilalta et al., 2014). Improved representation of hydraulic architecture in transpiration and assimilation models (Bohrer et al., 2005) can lead to improved prediction of species-specific responses of transpiration to disturbances including drought and improved representation of the diurnal cycle of transpiration and hysteresis (Ewers et al., 2007; Thomsen et al., 2013; Zhang et al., 2014; Matheny et al., 2014; Kumagai, 2010). Progress in understanding how the processes that govern the flow and storage of water in the plant are related to their drought response is limited by available knowledge and observations of relevant plant traits (Kattge et al., 2011), and by the difficulty of observing plant water flow without destructive mea25 surements (as further detailed in Chapter 7). The availability of large-scale records of vegetation water content would help to understand plant water dynamics. The root-zone depth across which plant water uptake occurs depends on climate and vegetation cover and often extends to one meter or more (Schenk and Jackson, 2002; Kleidon, 2004), or even to beyond the water table in the case of phreatophytes (Orellana et al., 2012). Across this range, soil moisture varies with depth. These depth variations can have significant effects on the total evapotranspiration ( Teuling et al., 2006). Variations of soil moisture with depth are the result of the local inputs of flow into and out of a profile, including infiltration of precipitation, root water uptake profiles, subsurface lateral flow, and percolation below the zone of interest. The resulting profile shape can vary dramatically depending on conditions and does not have a consistent shape. While some correlation between depths and surface soil moisture can be assimilated into a soil water balance model to provide information about root-zone soil moisture (Galantowicz et al., 1999), the direct measurement of root-zone soil moisture is preferred for studies of plant-water interactions 1.2 Motivation for using microwave remote sensing Aside from depth-variations, soil moisture is also highly spatially variable laterally (Famiglietti et al., 2008; Crow et al., 2012). Similarly, plant water content can vary significantly within the plant canopy, and between plants. Indeed, the variables influencing plant water content can span 11 orders of magnitude or more (Katul et al., 2007). As a result, large numbers of in situ measurements are needed to be able to average over small scale variability and draw meaningful conclusions at the stand- or ecosystem scale. Remote sensing of plant and soil water content naturally smoothes over variability below these scales and allows orders of magnitude more measurements to be taken than can be made by hand. Additionally, political and other logistical issues related to the ease of setting up in situ measurements cause in situ observing networks to be highly biased towards temperate climates (Robock et al., 2000; Schimel et al., 2015), even though tropical and boreal ecosystems contribute a large percent26 age of global water and carbon fluxes (Schimel et al., 2015). Satellite-based remote sensing measurements from low-earth orbits take global measurements and do not suffer the same biases (although there might still be biases in parameter development and validation efforts). Remotely sensed observations thus provide a promising avenue for the study of plant and soil moisture behavior across climates and ecosystem types. Remotely sensed observations of the land surface can be hindered by the presence of cloud cover. Optical and near-infrared observations cannot observe the land surface on cloudy days, potentially causing large seasonal biases and requiring aggregation of multiple days of observation in order to ensure the availability of clear-sky observations. By contrast, measurements at microwave frequencies can penetrate through the atmosphere with relatively limited sensitivity to cloud cover or atmospheric water vapor profiles (Ulaby et al., 1986a). However, they are indirect measurements and require a 'retrieval' to determine the variables of interest, i.e. soil moisture or plant water content, from the observations. 1.3 Introduction to microwave remote sensing Depending on the instrument used, microwave observations can be either actively measured (as when a radar sends out electromagnetic waves and measures the amount of scattering) or passively measured (as when a radiometer measures the natural graybody emission of the land surface) (Ulaby and Long, 2014). The two measurements are related by the conservation of power, er rzl, (1.1) where ep is the emissivity of the land surface (measured by passive measurements) and rp is the reflectivity (measured by active measurements). Each depends on the polarization p - the direction the electromagnetic wave is oscillating in, and the incidence angle Oi (Staelin et al., 1998). The ep and rp also depend on the dielectric 27 constant E of the emitting or reflecting surface. For non-homogeneous materials such as soils or plants, the dielectric constant of the material is a weighted mixture of the dielectric constant of the material's components. The dielectric constant of water is much higher than that of other substances in the soil. This enables the potential inversion of a measured dielectric constant for the relative amount of water, creating the potential for remote sensing of soil moisture ( Ulaby et al., 1986a). The effective dielectric constant of soil depends not only on the amount of water but also on the soil texture - the soil particles also influence how tightly bound different components of the soil water content are, and thus how they respond to an electromagnetic wave. If soil textural properties are known, the dielectric constant at a given frequency can be predicted from soil moisture using a dielectric mixing model ( Wang and Schmugge, 1980; Peplinski et al., 1995; Mironov et al., 2004). The calculation of ep and rp for a smooth surface interface between the atmosphere and a soil with effective dielectric constant is relatively well understood. However, soil roughness and vegetation also affect the measurements. Soil roughness refers to the small-scale (< cm) variations in the height of the soil surface. The effect of vegetation cover on microwave observations depends on both the aboveground water content in the plant and on its physical structure, i.e. the vegetation allometry. Although both active and passive measurements have a similar sensitivity to soil moisture (Du et al., 2000), they differ in how they are affected by soil roughness and vegetation. Radar measurements generally require a more detailed characterization of roughness and vegetation structural effects in order to allow accurate inversion for soil moisture or vegetation water content. However, relative to radiometer measurements, they can be made at much higher spatial resolution through the use of a synthetic aperture (Van Zyl and Kim, 2011). Soil roughness is generally modeled by assuming the height of the soil surface is characterized (after detrending any larger-scale topographic slopes) by a random process with a distribution based on the root-mean-square height and the lateral autocorrelation length ( Verhoest et al., 2008), although the latter is usually neglected in the context of radiometry (Ulaby and Long, 2014). Vegetation affects microwave re28 mote sensing measurements through both a direct scattering contribution and through so-called double-bounce scattering mechanisms in which waves scattered or emitted by vegetation are re-scattered by the ground surface and vice versa. The total measured backscattering coefficient aPQ is typically modeled as the sum of surface, doublebounce, and volume scattering components, aOrP = oisurface + PQPQ double +-volume ,1.2) +pQ(12 +PQ vertical) and Q refers to the receiving polarization. For most natural surfaces, UHV UVH, so there are three possible polarizations, UHH, uvv, and UHV. ' where the subscript P refers to the transmitted polarization (either horizontal or A variety of models exist for c-pQ, ranging from the simple and semi-empirical (e.g. Attema and Ulaby, 1978; Oh et al., 1992) to complex representations of the scattering relying on tens of parameters (e.g. Durden et al., 1989; Burgin et al., 2012). For passive measurements, most retrieval approaches use the same first-order solution to the radiative transfer equations, known as the tau-omega model: TB, = TS(1 -- rP)7 + TcI - WO -- 7)( + rp7). (1.3) Above, TB, is the brightness temperature at polarization p, which is either horizontal (H) or vertical (V), T, and T, are the effective land surface and canopy temperatures, respectively. The r, is the rough surface reflectivity, -y is the vegetation transmissivity, and w is the vegetation single-scattering albedo. Thus, the effect of plant water content and scattering is accounted for through the two parameters W and '-. Numerous field campaigns and theoretical exercises have shown that -y can be represented as (Jackson and Schmugge, 1991; Van De Griend and Wigneron, 2004) y = exp bVWC) Cos 0 (1.4) where b is a canopy-specific constant of proportionality, VWC is the total aboveground vegetation water content, and 0 is the incidence angle. Thus, from radiometric 29 measurements, the total vegetation water content VWC is a direct parameter of the equations, while the total VWC is not always a direct parameter of radar scattering models. 1.4 Outline of the thesis From a single set of multi-polarized measurements, either three (in the case of active data) or two (in the case of passive data) measurements are made. However, in order to account for the effects of soil roughness and vegetation as well as soil moisture, there may be far more than three or two unknown parameters in the retrieval problem. If there are more unknowns than observations, the retrieval will be overly sensitive to noise. When more than a single parameter is retrieved, even small amounts of measurement noise or model imperfections can cause large and opposite errors ('compensating errors') in multiple retrieval parameters. Chapter 2 of this thesis introduces a framework to calculate the maximum number of parameters that can be independently estimated from a set of data even in the presence of duplicate information or correlation between the observations. Two different approaches are then used to retrieve new datasets from microwave observations by changing the balance of unknown parameters and 'degrees of information' in the measurements. In Chapter 3, I show that in order to accurately retrieve root-zone soil moisture, it is necessary to account for the shape of the soil moisture profile, rather than retrieving only a single average value. In Chapter 4, a regularization method is introduced to allow the retrieval of soil moisture profiles from P-band radar data by providing additional information about the relationship between profile layers based on prior hydrologic expectations. A second new dataset of water at the land surface is determined in Chapter 5, which uses L-band radiometer data to retrieve vegetation water constant along with the effective scattering albedo and soil dielectric constant. In Chapter 6, this dataset is compared to a coincident active vegetation index cTHv to determine the differences and possible trade-offs between active and passive microwave vegetation measurements. Lastly, Chapter 7 uses OHV data to study the response of the diurnal 30 variability of plant water content to seasonal drought. 31 32 Chapter 2 How Many Parameters Can Be Maximally Estimated from a Set of Measurements? 2.1 Introduction Remotely sensed measurements using visible, microwave, or other spectral observations of geophysical parameters are generally not a direct observation of the quantity of interest; the raw observations need to be converted to the geophysical variables in the so-called retrieval process. In most applications, multiple geophysical parameters influence the observations. These additional parameters may not be known. If so, it is often advantageous to retrieve multiple parameters at once during a single inversion. If multiple parameters are to be retrieved, however, additional measurements may be needed. These may be obtained by increasing the types of measurements made, e.g. using additional electromagnetic frequencies (spectral channels), or incidence angles and polarizations in the case of radar or radiometers. The same measurement type can also be repeated and combined, e.g. by using multiple observations over the same pixel or multiple nearby pixels. Whatever the source of the additional data, the multiple observations are rarely completely independent. This is demonstrated by the 33 success of dimensionality-reduction methods in various areas of remote sensing, e.g. Renard and Bourennane (2009); Licciardi et al. (2012); Guanter et al. (2012). It is not possible to retrieve more unknown parameters than the number of measurements. If the set of measurements are strongly correlated, a simple integer count of the number of measurements may be over-counting the number of unknowns that can be retrieved. It is therefore necessary to be able to derive the (possibly fractional) degrees of freedom that can be obtained by using a certain set of measurements, accounting for the duplicate information. Especially in the atmospheric sounding community, this is commonly done by decomposing the signal into a fractional 'degrees of freedom of the signal' and 'degrees of freedom of the noise' using a method due to Rodgers (2000). This method assumes the measurements are linearly related to the retrieval parameters. For many non-linear retrieval processes, a single linearization may not be appropriate for use in designing an algorithm that is expected to be applied to large regions, or even globally. Furthermore, Rodgers' method assumes that all errors are additive and have a Gaussian distribution. Not all sources of error are additive and Gaussian (for example, speckle noise in radar measurements is multiplicative, or non-Gaussian when transformed to additive dB units), and error magnitudes may depend on ancillary variables whose global distribution is not Gaussian. The construct of Rodgers' method limits it to Gaussian variables (since only the covariance is used to characterize their probability densities) and an extension to the non-Gaussian case is not possible. In this letter, we present an alternative methodology for those cases when Rodgers' method is not appropriate. We introduce a framework that calculates the maximum number of fractional degrees of freedom (here termed the degrees of information) in a set of measurements. The calculation depends on the full probability density function of each of the contributing measurements rather than just their covariance, and is thus expected to better capture the total amount of information in the measurements. To do this, we propose the use of the normalized total correlation, a generalization of the normalized mutual information. These information theoretic measures are discussed in Section II. Information theoretic concepts have found a wide variety of applications in remote 34 -PRM 10 sensing, e.g. (Gueguen et at., 2010; Erten et al., 2012; Cerra and Datcu, 2008; Cariou et al., 2011; Hossain et al., 2014). Here, mutual information and its generalization is used for a specific application in model selection: determining how many unknown parameters can be maximally retrieved from a given dataset. The proposed method is independent of, and does not attempt to influence, the exact choice of parameters to be retrieved, only the number of parameters. Although similar information theoretic concepts (e.g. appropriately chosen combinations of joint and/or conditional entropies) could be used to determine which parameters the observations provide the most information about, the choice of retrieval parameters may be driven by diverse scientific questions or other design factors. The framework presented therefore determines the maximum degrees of information in the data independently of which particular parameters are to be retrieved or any particular retrieval algorithm. Indeed, it may not always be practical to introduce as many parameters as there are degrees of information, but the method determines an upper bound. This letter is organized as follows. In Section 2.2.1, the normalized total correlation is introduced and shown to be confined between 0 and N-1. The fractional 'degrees of information' contained in an N-dimensional observation set is linked to the normalized mutual information. The degrees of information are dependent on the precision of the measurements through a bin size parameter A, whose derivation is discussed in Section 2.2.2. This derivation is illustrated with an an example using microwave radiometer observations. Additional examples are described in Section 2.3. Lastly, in Section 2.4, the application of the metric to a number of different common measurement types in remote sensing is discussed. This latter section is meant to be illustrative rather than exhaustive. 2.2 Degrees of information If measurements are made of two independent random variables X and Y, they can be used to retrieve two unknowns. The set of measurements can be said to contain two 'degrees of information'. If a third measurement Z is added that can be perfectly 35 predicted from one of the other two random variables, the measurements still contain only two degrees of information. In reality, it is more likely that X and Y are related but not completely independent, and Z is similarly correlated to some degree with either X and Y. Depending on how closely related the three variables are, there could be enough information in the correlated random variables to retrieve either one or two unknowns. Some measure of the total amount of information contained in a set of measurements is therefore needed. The measure should be independent of the nature of the relationship between the variables (i.e. not restricted to linear relationships) and generalizable to an arbitrary number of dimensions (number of measurement channels). The Degrees of Information DoI provides such a measure and is introduced in Section 2.2.1. Section 2.2.2 discusses the bin size parameter necessary to calculate DoI. 2.2.1 Definition of Degrees of Information Dol The Shannon entropy, one of the central tenets of information theory, is the expected value of the information content derived from a single observation of a discrete random variable X. It can also be interpreted as the uncertainty of a variable ( Cover and Thomas, 1991). The Shannon entropy can be expressed as H (X) = Zp(x) log p(x), (2.1) where p(x) is the probability mass function (pmf) of X. If the random variable has a narrow distribution, an observation will, on average, provide less information than if it has a very broad distribution. The H(X) of a discrete random variable is non-negative. For multiple variables Xi, the joint entropy is H(X1,7 ... ,XN) ---..E1 --, XN) 109 P(X1i, ..,N). (2.2) X1 XN The individual p(X ) are referred to as the 'marginal probability mass functions' and the individual H(X ) as the 'marginal entropy(ies)' of each product. 36 The mutual information is a well-known measure of the reduction in uncertainty between independent and joint measurements of two random variables X and Y. Mathematically, this can be written as, I(X; Y) = E p(x, y) log p(x) . (2.3) By comparing the joint and marginal probability distributions, the mutual information quantifies the degree to which simultaneous consideration of the two variables changes their distribution.- That is, it quantifies non-linearly how dependent the two variables are. When X and Y are independent, I(X; Y) = 0. The I(X; Y) is maximized when X and Y are dependent (i.e. perfectly correlated). From the definitions, it can easily be shown that I(X; Y) = H(X) + H(Y) - H(X, Y). (2.4) For proofs of these and other information theoretic properties used in this section, the reader is referred to an introductory information theory textbook, such as ( Cover and Thomas, 1991). Although a number of different generalizations of the mutual information exist, the total correlation C (Watanabe, 1960) captures the amount of information shared between any of the measurements in a set. Like the mutual information, the total correlation is the Kullback-Leibler divergence between the joint and the marginal entropies, 1, --- , -J, - N , --, XN) dN... dxl ogP( 109Nl --P(XN XX .. (2.5) ) C(X1,X 2, ... , XN) = N H(Xi) - H(X1 , . . , XN) = i=r1 37 (2.6) We further define the normalized total correlation C,(X 1 , ... , XN) as N E H(Xj) - H(X1,7...,IXN) Cn .. X1i~n -C(X1, ... XN) H(X1, ... , XN) -i=1(27 H(Xl, ... , XN) To prove that Cn takes a value between 0 and N - 1, we use the basic property that N H(Xl,..., XN) H (Xi) => NH (X1,..., XN)Cn (2.8) 0. Since max(H(Xi)) < H(X1 , ... , XN), multiplying by N gives, i Furthermore, since N max(H(Xi)) < NH(X1,,..., XN) (2.9) H(X,) < N max H(X,) by definition, -N ZH(Xi) N H(X,) < NH(X1,..XN) H(X1, ... , XN) - <=N By inserting this into Equation (2.7), it becomes clear that Cn < N (2.10) - 1. The Cn therefore takes a value between 0 and N-1. When the Xi are independent, C" = 0. When they have a one-to-one relationship, Cn = N - 1. The higher the normalized total correlation between the measurements, the less information they contain. The total degrees of information between the Xi is then given by DoI = N - Cn(X1, ... ,XN) (2.11) Since additional measurements cannot remove information from a first one, DoI > 1. Since, as mentioned above, mutual information and entropy are non-negative, DoI < N. Thus, Dol E [1, N], as expected. In the limit where C(X 1 , ... , XN) is maximized, H(X 1 , ... , XN) = H(Xi) for all i. Thus, it is possible to derive an alternative normalization using the minimum H(Xi). The normalization with H(X1, ... XN) used here is chosen because it is more 38 conservative, since min(H(Xi)) < H(X 1 , ... , XN). Note that in two dimensions, Le Hegarat-Mascle et al. (1997) calculated the mutual information between two remote sensing images, but these were normalized by the entropy of one of the two images, so that the resulting measure is not symmetric. The properties above were derived based on the assumption that the Xi are discrete variables. In remote sensing, many measurements are continuous rather than discrete. For continuous variables, several of the above lemmas are false and mu- tual information does not have an effective upper bound. Nevertheless, while remote sensing measurements may appear to be continuous by taking on an arbitrarily large number of values, the number of possible measurements is in practice limited by the finite accuracy or precision of the instruments. That is, small fluctuations in measurements below some accuracy threshold do not provide any physical information. For a certain bin size A, the continuous measurements can be binned into discrete classes by rounding them to the nearest interval of A. The resulting constant-bin histograms can be used directly to estimate the pmf's necessary to evaluate C". 2.2.2 Dependence on the bin size parameter Using an inappropriate bin size A may introduce errors in the estimation of the probability mass functions, and thus in the C, and Dol. If the bin size used is too small, the frequency counts in the bins will be sensitive to noise fluctuations in the dataset. If the bin size used is too large, the estimated marginal and joint pmfs may mischaracterize (or even miss all together) certain peaks in the distribution. Several different approaches have been proposed in the statistical literature to determine the optimal bin width to accurately estimate the pmf with a finite sample. Among these, Sturges' rule for calculating a bin width based on the range of the data and the number of points is the oldest and the most common. It has been shown to work well for applications of mutual information-based image registration (Legg et al., 2007) and feature selection (Hacine-Gharbi et al., 2012). However, it is known to lead to overly large bin size estimates that over-smooth the histogram, particularly for large samples sizes (which are expected in remote sensing) ( Wand, 1997). It can also be 39 sensitive to outliers. A better approach is the so-called Scott's rule, which calculates the bin size Ar? from the data's standard deviation ax, of the data instead of its range: Axi = (2.12) 3,, where n is the number of points in the sample. The dependence on ni/1 3 has been shown to be optimal for minimizing LP error norms ( Wand, 1997). The use of Scott's rule is illustrated using a two-dimensional example for ease of visualization. Horizontally and vertically polarized measurements of L-band brightness temperatures (Tbv and TbH, respectively) from the Aquarius satellite are used (Le Vine et al., 2007). The data span the period September 1, 2011 to August 31, 2012 over land and across the globe. Aquarius has three beams with three different incidence angles; only the middle beam is used here. Figure 2-1 shows the bin sizedependence of the Ca(Tbv, TbH). Since the range and shape of the distribution is similar between the two variables, it is not surprising that the dependence on ATbv and ATbH is approximately symmetric. Applying Scott's rule to each of the TbV and TbH separately leads to two different bin sizes that can be used to determine C,. C (Tbv, TbH) 0.3 0.2 0 0.1 0) -2 -2 1 0 -1 0 ) log, (A Tb Figure 2-1: Normalized total correlation C, between Aquarius Tbv and TbH as a function of the bin sizes ATbV and ATbH. For large bin sizes relative to the dynamic range of the variables, the C, suddenly drops when the number of bins is so low that even the approximate shape of the joint pmf is distorted by the wide bins. The black triangle corresponds to the bin sizes recommended by Scott's rule. 40 The Cn(Tbv, TbH) = 1.13/7.87 = 0.14 at the optimum bin size. This is much lower than the Pearson's correlation coefficient between the values, r = 0.92. The joint pmf shown at the bottom of Figure 2-2 illustrates why. Although the shapes of the marginal distributions are similar, the long tail in the joint pmfs adds a significant amount of uncertainty between the two polarizations. By contrast, it reduces the Pearson correlation coefficient relatively little because most points fall on or near the diagonal line. This demonstrates the value of non-parametric measures of the degrees of freedom in measurements rather than relying on potentially misleading Gaussian assumptions. The resulting value of DoI = 2 - 0.14 = 1.86 allows calculation of the number of overpasses that must be combined to calculate a certain number of parameters from a multi-temporal timeseries using these data. Algorithms using dual-pol radiometric data at L-band from N overpasses can retrieve a maximum of 1.86 x N parameters. Any dependent information between observations at different times is due to autocorrelation in the physical properties to be retrieved, which is generally neglected in the retrieval process. Thus, the Dof from a single set of dual-polarized measurements is multiplied by N. For example, combining data from two overpasses leads to DOI2-pass = 1.86(2) = 3.72, which is only enough information to robustly retrieve 3 parameters, even if 4 measurements are used (two polarizations on two overpasses each). Indeed, a two-overpass timeseries algorithm can be applied to these data to robustly retrieve three parameters for each pixel: a single constant vegetation optical depth and the dielectric constant during both overpasses. Additional combinations of overpass numbers and retrieved parameters are also possible, as also discussed in Chapter 5. 2.3 Example Dol calculations In this section, the degrees of information (DoI) calculation is illustrated for several additional measurements. Table 2.1 shows the Cn(X1 , ... , XN) for several data sources and compares different re-arrangements of the same timeseries. The examples of Table 41 0.03 .Tbv . ... T b H H . 0.021 CL 0.01 C 250 200 Tb 300 P p(TbvTbH) x 10- 8 30 Y I 0/ 6 28 -Q 07 H 4 ) 26 2 24 220 240 Tb 260 280 Figure 2-2: The marginal (top) and joint (bottom) probability density functions (pdf) for observed vertically and horizontally-polarized brightness temperatures (Tbv and TbH, respectively) from the Aquarius satellite. Note that the edges of the joint pdf tails extend beyond the region shown; the figure is zoomed in for clarity. 42 2.1 are discussed one-by-one below. Noisy linear relationship: We first consider the case of two linearly related timeseries of unit slope, e.g. Y = X, both distributed normally around 0 with standard deviation 1. The X and Y are jointly sampled, but are subject to independent normally distributed noise with standard deviation 0.1 to produce series x and y. The mutual information between them measures the respective dependence of variables based on their joint distributions. The addition of independent noise to all values strongly reduces the amount of redundancy between the final measurements; Cn(X, Y) = 0.21. Aquarius multi-polarization backscattering data 0-HHc-vv, and UHV: Aquarius makes coincident radar and radiometric measurements. There is a higher normalized total correlations between pairs of two co-polarized backscattering coefficients (Cn(UHH, UVV) = tering coefficients 0.28) than between a combination of co- and cross-polarized backscat(Cn(UHH, HV) = 0.19). This can be understood by noting that the cross-polarized backscatter is essentially independent of the soil moisture, unlike the co-polarized backscatter. Some total correlation remains because both the coand cross-polarized backscatters are sensitive to vegetation and soil roughness. A set of cross and co-polarized data thus carries more information than two different co-polarizations, as reflected in the higher Do. When adding a third polarization the Dol increases by less than one, as expected from the non-zero Cn between all pairs of polarizations. The total Cn increases when combining all three polarizations, suggesting the mutual information between different pairs of polarizations is in different Table 2.1: Dol for several examples Datasets -OPt Noisy linear (0.10, 0.10) (0.20 dB, 0.20 (0.20 dB, 0.30 (0.20 dB, 0.30 (0.20 dB, 0.20 (UHH, UVV) (OHH, UHV) (UVV, UHV) (OHH, JVV, OHV) (UVV, Tbv) (TbV, TbH) dB) dB) dB) dB, 0.30 dB) (0.20 dB,0.87 K) (0.87 K, 0.81K) 43 N 2 2 2 2 3 Cn 0.21 0.28 0.19 0.18 0.40 Dol 1.79 1.72 1.81 1.82 2.60 2 2 0.03 0.14 1.97 1.86 parts of the pmf (e.g. different spatial regions or seasons). The total DoI is 2.60. Aquarius multi-instrument data orvv and Tbv: Because radar and radiometric measurements are affected differently by soil and vegetation scattering, the C, between coincident brightness temperature Tb, and backscattering coefficient -vv data is low, Cn(Tbv, -vv) = 0.03. Other combinations of backscatter and brightness temperatures had even lower total correlation, and thus contain more degrees of information. As for the total correlation between Tbv and TbH, the total correlation is much lower than the Pearson correlation coefficient between o-pQ and Tbp (e.g. Piles et al. (2015)), because it is sensitive to the entire distribution. 2.4 Applications to particular remote sensing observations The degrees of information framework can be applied to a variety of remote sensing observations and used to determine how many geophysical parameters can be maximally retrieved. Note that in hyperspectral imagery, determination of the number of parameters that can be retrieved from unmixing algorithms is known as the 'intrinsic dimensionality' problem and has been well-studied (e.g. (Hasanlou and Samadzadegan, 2012; Cawse-Nicholson et al., 2013; Heylen and Scheunders, 2013)). The high number of dimensions in these images (generally more than 100) makes total correlation computationally expensive to calculate for such images. Instead, the primarily application of this method is to monospectral, multispectral, and lidar data, as outlined below and shown by example in Section III. Microwave Radiometry: Radiometric measurements are made at a certain incidence angle, frequency, and polarization. For a given incidence angle and frequency then, DoI < 2 (Dol < 4 if the radar is fully polarimetric). Additional information can be obtained by measuring the same pixel at multiple incidence angles. This concept is used by the soil moisture retrieval algorithm of the European Space Agency (ESA)'s Soil Moisture Ocean Salinity (SMOS) satellite (Kerr et al., 2012), among 44 others. The degrees of information can provide a framework to calculate how many geophysical and biophysical variables can be determined from a collection of correlated multi-angular measurements. Similar principles apply for multi-temporal retrieval algorithms, which combine measurements made at different times under the assumptions that at least one of the retrieval parameters is constant over the time period between the observations (as in 5) or for multi-frequency algorithms. Radar: Whether the data is obtained using a real or synthetic aperture, the return from radar systems can generally be described by a maximum of eight parameters - the phase and amplitude of the backscattered waves in two possible transmit polarizations and two possible receive polarizations. (Radar altimetry applications, which are based on the signal return time, provide an exception.) Thus, the number of degrees of information in a single set of measurements can be no more than eight - even though radar scattering is sometimes expressed in a 16-element Mueller matrix. As in passive microwave applications, polarimetric, multi-incidence angle and multi-temporal methods (Kim et al., 2012) can be used to increase the number of geophysical variables that can be retrieved. The DoI can be used to determine how many polarizations, angles, or temporal samples are needed. Lidar: The DoI framework may not be as useful for discrete pulse lidars as for other measurement types because different returns view different parts of the canopy. However, degrees of information can be informative when applied to waveform-recording lidars, whether used to retrieve canopy biophysical parameters or atmospheric composition information. Unlike in radar systems, the incidence angle does not vary and multi-incidence angles cannot be used to increase the DoI in the system. Instead, lidar observations at multiple wavelengths and depolarizations could be used to infer multiple properties. 2.5 Conclusion When designing retrieval algorithms, the first choice to be made is the number of parameters to be retrieved from the measurements. The degrees of information frame45 work presented in this paper provides a method for estimating how many parameters can maximally be retrieved depending on the amount of duplicate information present in the joint pmf of the measurements. Use of the entire joint pmf allows accounting for the fact that less-commonly occurring measurements add a lot of uncertainty to the retrieval, and leads to a better estimate of the uncertainty in the data. The degrees of information in the measurements are independent of the type of retrieval algorithm, whether it is statistical, physical, or some combination thereof. Once the DoI is obtained, the maximum number of independent parameters that can be retrieved is given by the floor of the DoI. Generally, the presence of noise implies that not all the information in the measurements can be used for parameter retrieval. The true information content of a set of measurement is thus below the degrees of information. Most retrieval algorithms retrieve each parameter independently (i.e. all combinations of parameter values are possible solutions). In this case, the degrees of freedom needed for the retrieval is exactly equal to the number of parameters to be retrieved. @2015 IEEE. Reprinted, with permission, from A.G. Konings, K. A. McColl, M. Piles, and D. Entekhabi, How Many Parameters Can be Maximally Estimated From a Set of Measurements? IEEE Geoscience and Remote Sensing Letters, May 2015. 46 VFPPWMWMWR NRNI R 01 NO Ilop10 P1 I R'11.1 Chapter 3 The Effect of Variable Soil Moisture Profiles on P-band Backscatter 3.1 Introduction Soil moisture is important for climate modeling (Seneviratne et al., 2010), ecological modeling (Churkina et al., 1999), and understanding vadose zone hydrology (Vereecken et al., 2008). It is highly variable at a multitude of scales (Famiglietti et al., 2008). As a result, point-scale in situ soil moisture measurements are of limited use for applications spanning local to regional scales and larger. Remotely sensed measurements naturally provide soil moisture estimates over large areas. In particular, synthetic aperture radar can make measurements synoptically at relatively high resolution. Existing microwave measurements of soil moisture have generally been made at L-band (e.g. the UAVSAR program (Hensley et al., 2008), the current passive L-band interferometric Soil Moisture and Ocean Salinity (SMOS) satellite mission (Kerr et al., 2010) or the upcoming Soil Moisture Active Passive (SMAP) satellite mission (Entekhabi et al., 2010b)) or higher frequencies (e.g. the RADARSAT satellites (van der Sanden, 2004) and ASAR (Loew et al., 2006)). However, at these frequencies, the signal does not penetrate deep into the soil; the penetration depth is generally less than about 5 cm at L-band and even less at C- or X-band ( Ulaby et al., 1986a). These instruments only measure surface soil moisture and root-zorje soil 47 moisture can only be obtained after assimilation into a hydrological model (Hoeben and Troch, 2000; Draper et al., 2012). However, the assimilation of surface data provides limited skill when the coupling between the surface and the subsurface weakens (Walker et al., 2002; Kumar et al., 2009), necessitating direct measurements of subsurface or root-zone soil moisture. At lower frequencies such as P-band, the radar signal can penetrate deeper into the soil. The penetration depth varies significantly with soil texture and moisture, but often reaches at least several tens of centimeters even at moderately high soil moisture (Moghaddam et al., 2007). Because the measurement depth at L- and C-bands is so small, variations in soil moisture with depth over this range are often also relatively small. As a result, several studies have found that the emissive (Raju et al., 1995; Escorihuela et al., 2010) and reflective (Le Morvan et al., 2008) behavior of the soil can be described as originating from a soil with a single, uniform soil moisture value extending over an infinite halfspace, as long as the associated averaging depth to obtain that uniform value is chosen correctly. The error in predicted backscattering is proportional to the difference between the averaging depth and the penetration depth at these frequencies (Zribi et al., 2014). However, over the larger depth range covered by P-band measurements, soil moisture variation is expected to be significant. As a result, it may no longer be possible to describe the soil profile using a uniform halfspace without incurring significant retrieval error. Furthermore, scattering due to shifts in soil moisture at depth will be phase-shifted relative to the scattering from the land surface, complicating interpretation of the backscattering coefficient measurements (see Section 3.3). Additionally, in cases where the surface and subsurface are decoupled - precisely those cases where the lower-frequency measurements at P-band provide the most value - retrieving only a single equivalent soil moisture value may make it difficult to disentangle the soil moisture levels at larger depths. To account for soil moisture variability with depth, multi-layer slab models can be used. Such models approximate the variable profile as consisting of a set of layered slabs, each with a given dielectric constant. In the limit of a sufficiently high number of thin (with respect to wavelength) layers, the backscattering coefficient signal of 48 a multi-layer slab model is expected to be the same as that associated with the variable profile. The multi-layer slab model can therefore be used as a forward model (relating the measured backscattering coefficient to a soil moisture profile) in soil moisture retrieval algorithms. Using multiple layers increases the number of unknowns in the inversion problem and may cause it to become ill-posed. If observations at multiple polarizations, frequencies, and/or incidence angles are available, they can be combined to retrieve the soil moisture at different depths (and other unknown variables) directly. If not enough observations are available, it may be possible to use information about expected profile variation from land surface models or in situ measurements to improve the conditioning of the retrieval problem. For any retrieval algorithm to be successful, the accuracy of the embedded forward model is crucial. In this paper, we study the effect of representing variable soil profiles using a small number of discrete layers, including a single homogeneous halfspace, on the accuracy of the simulated backscattering coefficient retrieval at P-band. The goal of this paper is not to build or test any specific retrieval algorithm, but rather to examine how the sensitivity of modeled backscattering coefficient to the profile representation can lead to design criteria for retrieval algorithms. Since testing specific retrieval algorithms is outside the scope of this study, only the forward modeling errors associated with different profile representations are compared. To do so, we consider a case study of the grassland at Vaira Ranch, CA. A hydrologic model is used to determine the full profile of soil moisture at this site over a period of ten years. These profiles are coupled to a multi-layer soil scattering model and a vegetation scattering model to simulate the 'true' scattering behavior of the site over the study period. The models used are described in Section 3.2. Section 3.3 illustrates several aspects of the behavior of multi-layer representations of soil moisture profiles and the pitfalls in their interpretation through both an example profile at Vaira Ranch and idealized case studies. Placement strategies for retrieval algorithms representing the extremes of complexity and use of prior information are described in Section 3.4.1. Section 3.4.2 presents the forward modeling errors associated with each of these layer placement strategies and with different numbers of layers at Vaira 49 Ranch. 3.2 3.2.1 Forward Model Hydrologic modeling Variations between soil moisture profiles depend on weather history, land cover and soil texture conditions. The Vaira Ranch Ameriflux site near Ione, CA (Baldocchi et al., 2004) is studied here as an example. This site will be imaged at P-band as part of the NASA Airborne Microwave Observatory of Subcanopy and Subsurface (AirMOSS) program, a polarimetric P-band airborne radar mission (Chapin et al., 2012). The site is located near the Sierra Nevada foothills and has a Mediterranean climate. There is strong seasonal variability in precipitation, with a rainy season that lasts principally from November to February. The site is covered by an annual grassland that dies completely during the hot, dry summer. To derive the time-evolution of the soil moisture profiles, the Simultaneous Heat and Water Model (SHAW) is used. SHAW simulates the physics of the coupled movement of energy and water in the soil and the surface layer of the atmosphere, as driven by meteorological forcings and plant parameters; the details of the model can be found in Flerchinger et al. (Flerchinger and Pierson, 1991, 1997). Transpiration is dependent on both the flow and uptake of water through (and by) roots and on a detailed energy balance at the land surface. Moisture flow in the soil is governed by Richards equation (corrected to account for water extracted by roots) and coupled to the flow of heat in the soil. Ten years of meteorological forcing data (precipitation, humidity, air temperature, windspeed, and solar radiation) obtained from the Ameriflux site at Vaira Ranch are used to drive the SHAW model. Wherever possible, site-specific above-ground vegetation parameters, such as those describing stomatal conductance, are obtained from the literature. The full list of parameters used is given in the Appendix. The soil is represented using 28 layers over the top 40 cm and 9 deeper layers to 50 provide appropriate boundary conditions. The layer depths are staggered so as to be more shallow near the surface. The resulting modeled soil moisture is compared to measured soil moisture from Ameriflux at two different depths in Fig. 3-1. The model correctly captures the strong seasonal cycle of soil moisture. The average RMSE is 0.083 cm 3 /cm 3 at 5 cm and 0.062 cm 3 /cm 3 at 10 cm depth. Fig. 3-2 compares the modeled latent heat flux to that observed by the eddy covariance tower at the site. In the wet spring, the modeled transpiration is lower than that observed. The reduction in transpiration during the dry summer is also slightly slower for the modeled evapo-transpiration fluxes than for the observed ones, despite the fact that the 10 cm modeled soil moisture during this season is drier than that observed. Parameter changes that allowed the modeled latent heat flux to more closely match the observed led to reductions in the quality of the soil moisture simulations (not shown) and were not adopted. The strong seasonal cycle shown in Fig. 3-1 affects the entire soil moisture profile, as illustrated in Fig. 3-3, which shows the average profile during each of the dry and wet seasons, as well as the standard deviation around the mean profile. Although the average wet season profile is relatively uniform, few instantaneous profiles show such little variability. The sign and magnitude of the slope of the profile at different depths depends on whether an infiltration front is moving through the soil column or not. Over the entire wet season, profile variability averages out to be approximately independent of depth. By contrast, because of the small number of rain events during the dry season, soil moisture conditions almost always show a strong gradient from the dry surface (where evaporation is highest) to the wetter subsurface, as reflected in the average soil moisture and the lower variability during the dry season. 51 o0. E 0.55 U 0.5 co 0.45 E 0.4 %(A .35 E 0.3 0 0.25 o 0.2 T- 0.15 ' 0.1 0.05 E 0 co 2002 2003 2004 2005 2006 2007 2008 2009 2010 Time -Model --- Observ. I ' ' c- 0.45 E 0.4. .2L 0.35 0.3 E b 0.25 to 0. 2 ...0.15 ( 0.1 E 0.05 0 U ' e -' '* 2002 2003 2004 2005 2006 2007 2008 2009 2010 Time Figure 3-1: Comparison between modeled soil moisture (black line) and Ameriflux soil moisture observations (red dashed line) at 5 cm (top) and 10 cm (bottom) depth over the 10 year simulation period used in this study. 140 I I I I I tII -Model ---Observed 120N100- I, I L I, - k 80 6040 20 0 w 2002 5 U ii 2003 11 .1 ES.' ii P~T I BIll, 2004 2005 2006 Time I 2007 2UU5 2009 2010 Figure 3-2: Comparison between modeled latent heat flux (black line) and Ameriflux latent heat flux observations (red dashed line) over the 10 year simulation period used in this study. 52 0 -0.2[ - ' -0.4 S-0.6 -0.8- -6.1 Jun-Nov Dec-May 0 0.2 0.1 Soil moisture 0.3 0.4 [cm 3/cm 3I Figure 3-3: Average simulated soil moisture profile at Vaira Ranch, CA during the local dry season (June through November, black solid line) and wet season (December to May, red dashed line). The gray and pink shaded regions represent one standard deviation around the mean profile during the dry and wet season, respectively 53 3.2.2 Backscattering Coefficient Model Multi-layer soil scattering model The majority of early soil scattering models assume that the soil moisture and other properties are constant everywhere with depth, allowing the soil to be treated as a homogeneous halfspace described by a single dielectric constant and a single rough interface with air. Fung et al. (Fung et al., 1996), however, modified the well- known integral equation model (IEM) and provided a closed-form solution for the backscattering coefficient associated with a drying profile of soil moisture rather than a constant value, where the profile gradient is controlled by a transition rate factor. However, such -an approach assumes the soil moisture profile maintains a certain shape, which is not always true. Alternatively, the backscattering coefficient from an arbitrary profile may be approximated by discretizing the profile into a finite number of homogeneous layers. Changes or transitions in soil moisture between layers can generate scattered waves, which combine with the waves scattered from the top of the land surface to form the total backscattering coefficients. The relative phase (compared to that of the surface scattering) of these waves depends on kz, where k is the depth-dependent wavenumber inside the soil layer and z is the depth of the layer. The phase thus depends on both the soil moisture in the upper medium (which influences k) and the depth at which transitions occur. For a simple two-layer system, a plot of the total backscattering coefficient versus soil moisture in one of the layers does not grow monotonically, but rather oscillates around the value expected for a halfspace with the top-layer soil moisture (Boisvert et al., 1997; Ulaby et al., 1986a), because the phase is a periodic function. For real profiles, continuous variations in the soil mois- ture profile can generate a large number of reflected waves that sum with different phases. In addition, small scale heterogeneities occur due to the presence of roots, macropores, organic matter, rock inclusions, and heterogeneities in soil texture. These elements contribute volume scattering. Some efforts have been made to model the volume scattering associated with soil heterogeneity (England, 1975; Onier et al., 54 MIRROR I'M I 'W'' 11. F 2011; Duan and Moghaddam, 2011). However, such an approach requires modeling scatterers with specific dielectric constant and dimensions (or distributions thereof) occurring in a homogeneous background medium. In this paper, the soil dielectric profile is treated as corresponding to a soil moisture profile that consists of a number of arbitrarily small homogeneous layers. In the limit of an infinitely large num- ber of infinitely small layers, a multi-layer slab model can accurately represent the backscattering coefficient associated with a continuously varying profile. The specific slab model used in this paper is described below. If the interfaces between different layers are smooth, the modified Fresnel reflection coefficient at the top of a layered medium is a function of the properties of the different layers below, e.g. (Kong, 2008). Several studies have used this Fresnel reflection coefficient in combination with the IEM to find the backscattering coefficient from a two- or three-layer soil moisture profile (Song et al., 2010; Le Morvan et al., 2008). Several other multi-layer models exist, including ones in which the backscattering coefficient is found from the Green's functions ( Yarovoy et al., 2000), using the geometric optics approximation (Pinel et al., 2011), or using recursive transfer matrices and the small-perturbation model (SPM) (Imperatore et al., 2009). In this paper, we use the soil backscattering model described by Tabatabaeenejad and Moghaddam (Tabatabaeenejad and Moghaddam, 2006), which uses a first order SPM to describe the scattering from the (potentially) rough interfaces between each layer. The model calculates the effective up- and down-going waves representing the sum of the infinite reflections at each of the interfaces between layers by simultaneously solving Maxwell's equations at each interface. Because the resulting equations are arranged in matrix form, the model is generalizable to an arbitrary number of layers by inverting a matrix with larger dimensions. However, the computational cost grows quickly with the number of layers. For first-order solutions of the small perturbation model, including the one used here, depolarization for backscattering is neglected ( Ulaby et al., 1986a; Kong, 2008). As a result, the cross-polarized backscattering coefficients (HV and VH) are zero. Only co-polarized backscattering coefficients are considered throughout this paper. The surface roughness planar correlation is assumed to follow 55 an exponential distribution. Vegetation scattering model When vegetation elements are longer than the wavelength, the phases of the different components of the scattered wave are random and the scattering components can be added incoherently to a good approximation (Van Zyl and Kim, 2011). At P-band, this assumption becomes less valid for vegetated surfaces, and a model based on wave theory is necessary to accurately describe the scattering due to the vegetation covering the land surface. The vegetation scattering model used here is the one described in Burgin et al. (Burgin et al., 2012). The model is based on that of Durden et al. (Durden et al., 1989), but adapted for use in multispecies environments (though parameterized here to represent only one species, for simplicity). While it is primarily designed for forests, it can be adapted to represent other land cover types. Vegetation is assumed to consist of trunk-like and leaf-like structures, described as either finite cylinders or cylinders and disks, respectively (for deciduous forests). Each vegetation component (trunks, large branches, small branches, and leaves) is governed by parameters or distributions for their physical dimensions, orientation, and dielectric structure. The Stokes matrix associated with each component is calculated based on its geometric approximation as cylinders or disks with these parameters, and modified to account for wave attenuation when passing through the other components of the total plant structure. The Stokes matrices can then be added to provide the total Stokes matrix representing scattering from all sources in the scene (Durden et al., 1989). Crown components (branches and leaves) are assumed to be sources of volume scattering. Double-bounce scattering between the branches and the ground and between the trunks and the ground is also accounted for in the forward model. The soil scattering contribution to the double-bounce terms is represented by multiplying the Fresnel reflection coefficient associated with the multi-layer soil moisture profile by a factor accounting for the roughness at the top of the soil (Durden et al., 1989). Higher order multiple scattering is assumed to contribute only a negligible amount to the total amount of backscatter. Similarly, direct volume scattering from trunks 56 is assumed to be small. The total Stokes matrix is therefore, Mtot = Mr + Mbg + Mtg + Mvol (3.1) where Mr, Mbg, Mtg, and M,01 are the Stokes matrices for ground scattering, branchground double-bounce scattering, trunk-ground double-bounce scattering, and crown volume scattering, respectively. The resulting model is highly sensitive to the vegetation canopy parameters, which generally have to be approximated based on ground sampling and use of speciesspecific assumptions. The scattering behavior and its sensitivity to soil moisture profiles can vary widely depending on the parameters used. 3.2.3 Application of Scattering Model at Vaira Ranch The soil moisture profiles derived from the hydrological modeling are used as inputs for the backscattering model. To model the 'true' backscattering coefficient associated with the fully variably profile, 200 soil layers with 5 mm thickness are used, representing the profile over the top 1 m. Ten additional layers represent variations in soil moisture at larger depths. A constant incidence angle of 300 and measurement frequency of 430 MHz are assumed, and the equations of Peplinski et al. (Peplinski et al., 1995) are used to convert the soil moisture to the associated dielectric constant. A surface roughness rms height of 2 cm is assumed, while all subsurface layers are assumed to be smooth. The scattering effects of the grassland cover were described using the parameters in Table 3.1. The parameters were obtained from a generic grassland allometric relationship (Burgin, 2012), and modified slightly for physical realism and computational simplicity. Because dead biomass has relatively little effect on the scattering, the vegetation scattering parameters are scaled according to a repeating annual cycle that idealizes observed variations in leaf area index (LAI) (Baldocchi et al., 2004). The vegetation dielectric constant and cylinder length increase approximately quadratically between day of year (DOY) 350 and 110 (with a plateau between the start of the year and DOY 50) and decrease sharply to zero 57 Table 3.1: Vegetation scattering parameters used Value Canopy height [ml 0.30 Trunks length [cm] 5 radius [cm] 0.05 Leaves density 1#/m 3] 15000 dielectric constant length [cm] avg. radius [cm] 12-j3 20 0.05 density [#/m 3] 10000 avg. orientation dielectric constant 100 12-j3 after day 100. Thus, for most of the dry season, no live vegetation affecting the scattering is present. Because the sparse grassy vegetation primarily acts by attenuating the signal and contributes very little double-bounce or volume scattering, the results are not expected to be sensitive to the vegetation parameters used. The P-band backscattering coefficient is evaluated every six hours over the entire 10 year period, and is shown in Fig. 3-4. For simplicity, only a single year of data is shown. Sharp peaks correspond to rain events. As expected, the backscattering coefficient is much lower in the dry season between June and November than in the wet season (except for the extra scattering resulting from a single storm around DOY 175). The small oscillations occurring during the dry season show that even with reflections adding from 200 shallow profile layers, the backscattering coefficient is sensitive to the details of the modeled soil moisture profile. The backscattering coefficient measured at VV-polarization (VV-pol) generally decreases faster after a rainfall event than the backscattering coefficient at HH-polarization (HH-pol), as can be seen after the rain storms (peaks in scatter) around DOY 70, 110, and 285. This occurs because the HH-pol is generally more sensitive to the subsurface soil moisture, which stays wetter longer after a rainfall event. 58 -15-- - -20 -25 0 410 80 I I 40 80 120 120 160 200 160 200 240 280 240 280 320 360 320 360 Day of year -15 -20 I I -25 -30 -35 -40 I Day of year Figure 3-4: Seasonal evolution of simulated 'true' P-band backscattering coefficients at VV- (top) and HH-polarization (bottom) over the year 2001, assuming an incidence angle of 300. The soil roughness rms height s=0.02 m at the surface, while all subsurface layer are assumed to be smooth. 3.3 Behavior of Multi-layer Scattering In order to better interpret the forward modeling errors of the Vaira Ranch case study (described in Section 3.4), we first study the behavior of multi-layer scattering models. Scattering originating from all depths contributes to the total measured backscattering coefficient. However, because soil is a lossy medium, the electromagnetic wave attenuates with depth. In the absence of volume scattering, the power of the incident wave P(z) attenuates as ( Ulaby et al., 1986a), p(z) = P(z = 0)e- J 2a(z')dz' 59 (3.2) where z increases downwards. The factor of two signifies two-way attenuation. The attenuation coefficient a is given by, a(z) = koIm e(z)) , (3.3) where ko is the wavenumber in free space, Im denotes the imaginary component, and E(z) is the soil dielectric constant profile. Because the wave attenuates with depth, soil moisture at larger depths contributes less to the total backscattering coefficient than soil moisture near the surface. Since the contributions to the total scattering gradually decrease with depth, there is no unique way to define the sensing depth. One common approximation is given by the penetration depth, p. It is defined as the depth at which the signal power has been attenuated to 1/e times its original level (Ulaby et al., 1986a). The p is therefore determined by the equation 1 j 2koIm e(z)) dz. (3.4) Note that p is an imperfect measure of the depth contributing to the scattering, because the variation of the intensity of the wave with depth is not necessarily proportional to the contribution of that depth to the total amount of scattering. Additionally, if the signal-to-noise ratio of the radar measurement is sufficiently high, even variations in power caused by scattering at depths below the penetration depth could be detected. Since the rate of wave attenuation depends on the soil moisture profile, the measured depth can vary depending on conditions. This is shown in Figure 3-5, which shows the variation of the e-folding penetration depth p at Vaira Ranch over the ten years of simulation, as well as its average annual cycle. As the soil dries out during the dry season in the middle of the year, p increases. The annual amplitude in the penetration depth is significant. Since depths close to the surface contribute more to the total backscattering coefficient than larger depths, the equivalent constant soil moisture of a layer is generally different from the average soil moisture in that layer. Therefore, approximating a variable soil moisture profile using a multi-layer model introduces errors. Such errors 60 I I " 1 1, 1, WOR IFFINIMMIMM Imp mpg 40 35,_," E ; 25 -0 c 20 0 4. ~15 a) CL 30 20 1 Or 10 5F 0 0 100 200 DOY 300 ' ' I I I I I I 2002 2003 2004 2005 2006 2007 2008 2009 2010 Time Figure 3-5: Evolution of the penetration depth (in cm) over the ten years of simulation. Tick marks labeled with each year correspond to the first day of that year. The inset shows the annual average cycle of the penetration depth 61 are essentially unavoidable in practical soil moisture retrieval settings. This occurs because determining the exact contributions of soil moisture at different depths for an arbitrarily varying profile is mathematically challenging; by definition, the vertical variation of soil moisture is not known at a higher resolution than that of the backscattering model used in the retrieval process. Perhaps more significantly, a retrieved soil moisture that cannot be interpreted as representing an average value over some layer would be of little use to the community. Land surface models themselves typically represent soil moisture profiles through several layers with an associated average moisture value. While treating retrieved soil moisture as representing the average wetness over a certain depth is technically inaccurate, it is by far the most practical interpretation. When the soil moisture profile is represented using only a single layer model, the forward modeling error can be quite large. This is reflected in the difference between the average soil moisture over the penetration depth (or any other estimate of the sensed depth) and the equivalent soil moisture for which a homogeneous halfspace with that moisture value leads to the same backscattering coefficient. Indeed, the equivalent soil moisture may be outside of the range of the profile all together. Fig. 3-6 shows an example of such a scenario for an instantaneous profile in January 2001 at Vaira Ranch. For both polarizations, the equivalent homogeneous soil moisture is much higher than the soil moisture at any point in the profile. The equivalent homogeneous soil moisture is the uniform profile value (constant with depth) that in a halfspace forward backscattering model has the same UHH and xvv as the variable soil moisture profile used in a many-layer 'true' backscattering model. If a retrieval algorithm only estimates a single value of soil moisture, the estimated value in this case will be at least 0.10 cm 3 /cm 3 higher than the maximum soil moisture in the profile, let alone the average value. The reflections from deeper layers, which are phase-shifted relative to those at the surface, increase the total amount of backscattering, causing the equivalent soil moisture to be greater than the maximum value. Although the phase shift over each shallow layer is relatively small, the total phase shift over the measured depth significantly increases the associated backscattering, 62 ''I pi, In -1 causing the surprising result shown in Fig. 3-6. Section 3.4.2 studies the backscattering coefficient's behavior for the entire 10-year simulation period at Vaira Ranch to understand how commonly severe errors such as those shown here occur among likely observations. In more than 25% of cases studied, the equivalent homogeneous soil moisture value is outside the range of the actual profile. Similar behavior has also been observed for L-band backscattering coefficients from less-realistic linearly varying soil moisture profiles simulated using a finite element method (Khankhoje et al., 2013). V 0 = -12.5 dB, c = -15.3 dB -0.2- S-0.3 -0.4 -0.5 -Full profile --- Eqvlnt, W-pol -- Eqvlnt, HH-pol 0 0.1 0.3 0.2 0 [cm3/cm 3] 0.4 0.5 Figure 3-6: Example soil moisture profile simulated by the SHAW model (see Section 3.2.1), which has cHH = -24.9 dB and -vv = -16.9 dB when a 210-layer approximation to the full soil moisture profile is used in the P-band backscattering model (at a 300 incidene angle and frequency of 430 MHz). The vertical lines are the equivalent halfspace moistures that show the same HH-polarized backscattering coefficient (red, dashed line) and the same VV-polarized backscattering coefficient (blue, dash-dotted line) under the same texture, roughness, and sensing conditions. The roughness and sensing parameters used are identical to those in Section 3.2.3. For models with a finite number (e.g. two or three) of layers, the soil moisture profile is approximated more closely and profile representation errors are expected to be smaller than in the halfspace case. However, multi-layer models introduce an additional source of error: the large reflections occurring at the layer interfaces. 63 Unless there is a shift in soil texture, these may be unrealistically large. The behavior of errors for multi-layer scattering models can be best illustrated using the simplest possible profile, a linear one. Linear variations of soil moisture are generally unrealistic because the flow behavior that determines the profile at any given time is controlled by matric pressure, which is non-linearly related to soil moisture. Nevertheless, linear profiles are used here for illustrative purposes. Soil moisture is fixed at 0.40 cm 3 /cm 3 at 30 cm depth (near saturation). The slope of the linear profile in the top 30 cm, and thus the associated surface (z=0) soil moisture, is then varied across the dynamic range. The soil profile is modeled as extending infinitely at a value of 0.40 cm 3 /cm 3 below 30 cm. A silty loam soil is assumed, for which the signal penetration depth p is roughly 30 cm. Contributions to the total backscattering coefficient from below this depth are expected to be small. The illustrative true profile is approximated using 100 layers of 3 mm depth each. The resulting true backscattering coefficient is shown in Fig. 3-7 for both VV- and HH-polarization. An incidence angle of 300 and measurement frequency of 430 MHz are assumed. Fig. 3-7 also shows the backscattering coefficient associated with several smalllayer-number approximations of the full linear profile. In the one-layer case, the soil is assumed to consist of a uniform infinite halfspace with a moisture equal to the average soil moisture over the top 30 cm. In the two- and three-layer cases, the variable profile over the top 30 cm is approximated using two or three layers of equal size. Thus, in the two-layer case, the soil consists of a layer spanning from the surface to 15 cm with a moisture content equal to the average soil moisture over the top 15 cm, and a semi-infinite layer below it with a moisture content equal to the average soil moisture between 15 and 30 cm. Similarly, the three layer case has layers between 0 and 10 cm and 10 and 20 cm with the average moisture content across those depths, and a semi-infinite third layer below 20 cm with a soil moisture equal to the average moisture between 20 and 30 cm. Using a small number of layers creates errors of several dB in magnitude. The two-layer approximation shown in Fig. 3-7 illustrates the oscillations caused by phase-shifted reflections occurring at the subsurface layer interface. In the presence 64 Mean soil moisture [cm 3/cm3] 0.36 0.24 0.28 0.32 3 3 Mean soil moisture [cm /cm 0.24 0.28 0.32 0.36 0.4 -12 -12 -14- -14 -16- -16 0.4 V~ -18 -18 /I -20 -22 -1-layer --- 2-layer '''3-layer --- True 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 Soil moisture at z=0 [cm 3/cm 3] -20 -22 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 Soil moisture at z=0 [cm 3/cm 3] Figure 3-7: Backscattering coefficients associated with a linear soil moisture profile between various soil moisture values at the surface (z=0) and a value of 0.40 at 30 cm depth (pink dash-dotted line). The bottom x-axis indicates the value of soil moisture at the surface, while the top x-axis labels indicate the average soil moisture over the top 30 cm associated with each linearly varying profile. The black, red dashed, and blue dotted lines represent backscattering coefficients associated with uniform, twolayer, and three-layer approximations to the linear profile, respectively. The left plot shows the backscattering coefficients in the VV-polarization, while the right shows the HH-polarization. Roughness and measurement parameters were the same as in Fig. 3-6. of these oscillations, particular backscattering coefficient values no longer correspond to a unique moisture value, complicating retrieval efforts. For the 'true' profile, the combination of the many constructive and destructive phase interferences between waves scattered* from different depths results in smoothing out the oscillations. In some cases, the discrete phase shifts can counteract the gains from representing a profile more closely. For example, for surface soil moisture less than 0.06 or more than 0.26 cm 3 /cm 3 , a three-layer representation of the linear profile is actually further from the true backscattering coefficient than a two-layer representation and will lead 65 to a larger retrieval error. Although the VV-pol is more sensitive to soil moisture than the HH-pol (it has a larger range), HH-polarized reflections are more sensitive to variations in soil moisture below the surface. 3.4 Forward Modeling Error at Vaira Ranch In this section, the cumulative effect of the forward modeling errors are studied for the Vaira Ranch case study using models with different numbers of layers and different placement strategies for the layer interfaces. Throughout, the same incidence angle (300) and measurement frequency (430 MHz) are used as for the simulated 'trueAA2 backscattering coefficients. 3.4.1 Layering Approaches The error associated with discrete multi-layer representations of a continuous profile depends at least in part on where the layers are placed. The layer placement that minimizes the forward modeling error in turn varies depending on the soil moisture profile. As a result, the optimal layer placement is very difficult to predict without a priori knowledge of the shape. By combining retrieval algorithms with hydrologic modeling, data assimilation or other approaches, it may be possible to obtain an estimate of the the shape or regime of the soil moisture profile. To determine whether such an approach could benefit from improved layer placement, this paper compares the errors associated with multi-layer models based on two extreme strategies: Equal-layer approach In the absence of any prior information on soil moisture and soil texture profiles, the simplest multi-layer representation is one in which layers consist of equal depths. Although the bottom layer of a multi-layer model is assumed to extend to infinity (as with an infinite halfspace for a single-layer model), its average soil moisture still represents the average soil moisture over a finite depth. To determine that depth for the lowest layer, the total depth contributing to the backscattering coefficient must 66 be known. Here, we assume this depth is equal to the penetration depth p described in Section 3.3. Moisture-dependent approach Since multi-layer slab models assume a constant dielectric constant in each layer, layer interfaces should ideally be placed in such a way that the soil moisture profile is as uniform as possible over each layer. This can be approximated by placing layer interfaces at the depths where soil moisture is changing the most rapidly. Thus, the moisture-dependent approach tested here places layer interfaces at the depths where the modulus of the gradient of the soil moisture profile is highest. As before, the bottom layer extends from the penetration depth to infinity in the electromagnetic model; its soil moisture is equal to that of the layer above it. The layer placement scheme is illustrated for a sample profile in the right panel of Fig. 3-8, which shows the interface locations for a 4-layer model for an example profile. For contrast, the interface locations used for the equal-layer approach are shown in the left panel. In general, the layer interfaces calculated using this method tend to fall closer to the top of the profile than to the bottom, as variations in soil moisture are more pronounced near the surface. Since shallower depths contribute more to the total backscattering coefficient, this is a useful attribute. This approach can only be used where prior information on the profile shape is available, but represents a limiting scenario. 3.4.2 Results The evolution of the simulated backscattering coefficient for different layer strategies and numbers of layers is shown in Fig. 3-9 for the first year of simulation. Not surprisingly, the error generally decreases with the number of layers, although some exceptions occur. The average profile representation error is higher in the dry season than in the wet season. This occurs primarily because the sensitivity of radar backscatter to soil moisture is higher for drier soils, so that a given error in the soil moisture representation corresponds to a higher error in backscattering coefficients. 67 i 0 Moisture profile-dependent layers Equal, constant layers 0 -0.05 -0.05 -0.1 -0.1 -0.15 -0.15 01 -0---------- ---------------- -0.2- -0.2- 0.1 0.2 0.3 0.4 Soil moisture [cm 3/cm 3l 0.1 0.2 0.3 0.4 Soil moisture [cm 3/cm 3I Figure 3-8: Soil moisture variation with depth for an example profile. The dashed horizontal lines represent the layer interfaces used in the equal layer (left) and moisturedependent (right) approaches. The solid horizontal lines show the bottom of the depth the lowest layer is assumed to represent. For the equal-layer approach (left) this is the time-averaged penetration depth, while for the moisture-dependent approach (right) it is the penetration depth associated with the specific profile. Since the measured depth is higher in the dry season, the total unaccounted profile variability is also generally greater during the dry season than during interstorm periods in the wet season. The dry season error dramatically decreases when the number of layers is increased beyond one, especially at VV-polarization. Increasing the number of layers allows the model to better capture the strong gradient from the dry surface to the (somewhat) wetter subsurface, improving the accuracy of the equivalent soil moisture representation near the surface. Since the VV-pol is less sensitive to the subsurface than the HH-pol, and thus more sensitive to the soil moisture near the surface, improving the near-surface soil moisture accuracy is particularly beneficial to the VV-pol backscattering coefficient simulation accuracy. Fig. 3-9 also illustrates that using a moisture-dependent layering approach most dramatically improves the backscattering representation shortly after a rainfall event (e.g., during and right after one of the sharp peaks in the evolution of the backscattering coefficient). At such times, the rainfall infiltration fronts cause strong variations in soil moisture, the depth of which evolves quickly in time. Varying the layer depths in time allows the 68 soil moisture variability associated with the rainfall infiltration to be at least partially captured. Moisture-Dependent Approach Equal-Layer Approach . 10 0-1C 0 0 CO gr -15 0 0 100 200 Day of year 300 I 300 5 0 V-1" 0 -1u -0 -- -5 -5. 100 200 300 0 Day of year 0 100 200 Day of year 300 10 1C 5- LO 0 -5 -10-15 C 0 5 0 -5 100 200 Day of year -10 300 ( CO 200 10 5- 0 . Day of year 10 CV -- -2 2-2 5 -W-Pol ---HH-Pol S-2C -2 -25 -1 -I D o -1 -layer ---3-layer 200 --5-ayerDay of year 30 Figure 3-9: Seasonal evolution of backscattering coefficients for different multi-layer representations of the soil moisture profile. The top row shows the backscatter at VV-(black line) and HH-polarization (red line) for a 210-layer representation of the soil moisture profile and is taken to represent the true backscatter. The bottom two rows show the difference between the backscattering coefficients for different multilayer representations and the true backscattering coefficients at VV-pol (middle row) and HH-pol (bottom row). The black, red, and blue lines represent the differences for one-, three-, and five-layer representations, respectively. For each panel, the interfaces between homogeneous layers are spaced at equal intervals in the left column, and are spaced depending on the locations of maximum soil moisture gradients in the right column. A measurement frequency of 430 MHz, incidence angle of 300, and roughness rms-height of 2 cm are assumed. Fig. 3-10 summarizes the profile representation error statistics for different numbers of layers and for both layering strategies over the ten simulation years. The bias and root-mean-square-error (RMSE) of the backscattering coefficient (in dB) 69 2.5 1 2 0.8 1.5- ff 0. 6 0.5 0.2 4 3 2 Number of layers %11 5 -Equal Layers, VV --- Equal Layers, HH VV -Moisture-dependent, --- Moisture-dependent, HH 4 3 2 Number of Layers 5 Figure 3-10: RMSE (left) and bias (right) for soil moisture profile representations with different numbers of layers. Solid lines represent the error at VV-polarization, while dashed lines represent the error at HH-polarization. Black lines refer to the error when layers are distributed so as to have equal thickness, while red lines are for simulations in which the layer thicknesses depend on the soil moisture profile. Throughout, a measurement frequency of 430 MHz, incidence angle of 300, and roughness rms-height of 2 cm are assumed. are shown. The RMSE is higher at HH-pol than at VV-pol. A systematic bias b contributes to the RMSE for all numbers of layers and strategies. The bias is small when averaged over multiple seasons with different soil moisture profiles. As a result, 2 the bias-removed (random error) RMSE, calculated as uRMSE = RMSE 2 + b2 (Entekhabi et al., 2010a), rarely differs from the biased RMSE by more than 10%. As expected, using a more sophisticated moisture-dependent layering approach improves the RMSE relative to using constant, equally-sized layers. The decrease is most consistent at VV-pol. the number of layers. Both the bias and the RMSE generally decrease with Still, the HH-pol error increases, if only slightly, when the number of layers is increased from 3 to 4. 70 As was discussed in Section 3.3, this increase occurs because of the introduction of unrealistically large reflections at the layer interfaces. Because the HH-polarization is more sensitive to the subsurface than the VV-polarization, the errors from the unrealistically large reflection are larger at HH-pol, relative to the errors due to misrepresenting the profile. As a result, it is more common for HH-pol backscattering coefficient simulation to perform worse after additional layers are added, as reflected in the RMSEs shown in Fig. 3-10. In the one-layer case, the only difference between the two layering strategies is the bottom depth over which the soil moisture is averaged. For the equal-layer approach, this depth is 24 cm, while for the moisture-dependent approach, it is the penetration depth at that particular time. Given the large variability in the penetration depth with time (shown in Fig. 3-5), one would expect that the moisture-dependent approach has a much lower error than the equal-layer approach for single-layer models. Instead, Fig. 3-10 shows it is marginally higher, reinforcing the fact that the penetration depth does not actually represent the depth that contributes to the signal, even though it is often used to represent this depth in the literature. 3.5 Discussion and Conclusions The relationship between the soil moisture profile and the observed backscattering coefficient is highly complex. The sensitivity to soil moisture decreases with depth at a rate that depends on the intervening soil moisture profile. Profile variability also causes volume scattering, here interpreted as the superposition of waves multiply scattered from layer interfaces. Depending on their phase, these reflected waves may not add monotonically. Because of these and other factors, multiple soil moisture profiles can lead to very similar or identical backscatters values. As a result, there is no stable one-to-one relationship between backscattering coefficient and average profile soil moisture over a certain depth; the relationship depends on how the soil moisture varies. This is potentially problematic for efforts to estimate a single value of soil moisture using radar, as the soil moisture profile is by definition not known when it is being measured. It is therefore not the correct approach to use low-frequency 71 radar measurements to retrieve only a single value of soil moisture that represents the average profile value over the contributing depth. Indeed, the homogeneous soil moisture value that corresponds most closely to the backscattering coefficient from a variable profile may well be entirely outside the range of the profile all together, as illustrated in Fig. 3-6. For the ten years of simulated soil moisture profiles at Vaira Ranch used as a case study in this paper, this occurred in more than 25% of profiles. Profile variations can even lead to a backscattering coefficient that is outside the dynamic range of homogeneous halfspace soils all together (occurring in 5% of profiles). The predicted backscattering error induced by using the average homogeneous soil moisture was about 2 dB in a root-mean-square sense, but could be as large as 10 dB in individual cases. Although an error of 2 dB corresponds to different soil moisture retrieval errors depending on the soil moisture regime, it is clearly significant relative to the dynamic range of measured backscattering coefficients, which could be 18 dB or higher. In an operational soil moisture retrieval setting, unknown soil roughness and/or vegetation parameters have to be retrieved at the same time as soil moisture. In such cases, the profile error can lead to incorrect estimation of these ancillary parameters on top of the soil moisture estimation error. In theory, profile variability errors could be reduced by retrieving soil moisture at multiple depths simultaneously. In most soil moisture estimation applications, when each pixel is measured at only a single frequency or incidence angle, the number of measurements is often small relative to the total number of unknowns, which include not only the soil moisture at various depths but also vegetation and soil roughness parameters. Retrieving the entire profile along with the other unknowns based on single-frequency radar measurements only is thus impractical and indeed often impossible. It may be possible to increase the number of profile layers resolved by the retrieval if the measurement data are combined with a priori information about the likely variation of the soil moisture profile from hydrologic modeling (see . also Chapter 4) The forward modeling error calculated for ten years of realistic soil moisture profiles at Vaira Ranch follows similar meteorological patterns when multiple layers are 72 used as when a single halfspace is assumed. The error is highest shortly after rainfall and during extended drydowns - those times when the soil moisture profile is most variable. At VV-polarization, the forward error decreases as the number of layers is increased and the variations of soil moisture with depth are represented more accurately. The VV-pol sensitivity to the subsurface is relatively lower even at P-band, so that the reduction in error is primarily caused by the improved representation of the average soil moisture in the top few centimeters. The error decreases most quickly when switching from one to two layers, and more slowly for each additional layer thereafter. However, the error does not always continuously decrease with the number of layers. The interfaces between the additional layers produce additional scattering that is unrealistically large, partially counteracting the gain from representing the profile more accurately. Since the HH-polarization is the most sensitive to the reflections from the subsurface, HH-pol errors are more likely to increase with the number of layers than VV-pol errors. This is reflected both in backscattering models for individual profiles and in the overall error statistics for different modeling configurations. The forward modeling errors due to profile variability largely consisted of random fluctuations rather than a systematic bias. Errors in retrieved soil moisture originating from profile variability are thus similarly likely to consist mostly of random fluctuations. The forward modeling error is also influenced by the layer placement. The sensitivity to layer placement was tested by using both a naive layering scheme, the equal layer approach, and a more complex one in which the layers were placed at different depths depending on the profile and chosen to minimize variability across each layer depth, the moisture-dependent approach. Doing so reduced the RMSE of the forward error by 0.2 to 0.7 dB, depending on the polarization and the number of layers. The moisture-dependent approach is most helpful shortly after a rainfall event, when profile variability is highest and infiltration fronts may be associated with sharp inflection points. In reality, designing a layer strategy as detailed as the moisture-dependent approach is impossible for retrieval. Nevertheless, depending on the details of a given retrieval algorithm and the amount of prior information, it may be possible to devise 73 an intermediate-complexity layering strategy. The two extreme layering strategies tested here suggest such a scheme might cause a moderate reduction in the forward modeling error, and therefore also in the retrieval error. Even the representation of the fully variable profile with more than 200 homogeneous layers used here does not capture the real complexity in the field. Soil dielectric mixing models assume a perfectly homogeneous soil texture, and neglect heterogeneities in the form of roots, rocks, or organic material. Real profiles are unlikely to be as smooth as those simulated in this study. Furthermore, no errors other than those arising from soil moisture profile variability were included in this study. Radar speckle, instrument calibration errors, and errors in vegetation modeling and parameterization may contribute as much or more to the total error as errors in subsurface treatment. Furthermore, depending on the exact vegetation and roughness conditions of any given measurement location, the subsurface errors may be smaller or larger. In this study, the vegetation had only a minimal effect on the scattering (through attenuation), and thus had only a slight effect on the simulated forward modeling errors (not shown). The comparisons between scattering model simulations of various numbers of layers and soil moisture profile shapes performed in this study provide insights to guide the development of retrieval algorithms. Such algorithms need to balance the accuracy of the profile representation (through the number and positions of different layers) with the well-posedness of the inverse problem, which is constrained by the limited number of polarimetric measurements available at any point in the field. If the retrieval process is underdetermined, the retrieved solution will be strongly dependent on the measurement noise and other errors. Such noise can cause the global minimum of the retrieval cost function to shift to a local minimum of the noise-free cost function, which may be far away from the true solution. The balance between profile accuracy and profile degrees of freedom is first studied by considering the effect of representing the fully variable soil moisture profile as a set of discrete layers. Additionally, the impact of layer positioning is studied under limiting scenarios with and without prior knowledge of likely soil moisture profile shapes. The insights provided 74 in this paper will be used in follow-up studies on the design and testing of specific retrieval algorithms. At their core, robust and reliable retrieval algorithms should include a combination of inverse-problem regularization with a priori information and an adaptive selection of discrete layering representation. @2014 IEEE. Reprinted, with permission, from A.G. Konings, D. Entekhabi, M. Moghaddam and S. Saatchi, The effect of variable soil moisture profiles on P-band backscatter? IEEE Transactions on Geoscience and Remote Sensing, October 2014. 75 76 Chapter 4 The Use of Regularization for Improving Profile Soil Moisture Retrievals from P-Band Radar Measurements 4.1 Introduction Using lower microwave frequencies for remotely sensed measurements of soil moisture allows for better penetration through vegetation and into the soil ( Ulaby et al., 1986a). The recently launched ESA Soil Moisture Ocean Salinity (SMOS) Kerr et al. (2010) and NASA Soil Moisture Active Passive satellites (Entekhabi et al., 2010a) both operate at L-band, for which no more than about the top 5 cm of the soil contribute to the measurements (Entekhabi et al., 2010a; Ulaby et al., 1986a). Although the penetration depth varies with soil moisture and texture, it is much deeper than 5 cm for P-band. P-band synthetic aperture radar (SAR) data generally sense multiple tens of centimeters and can penetrate as deeply as 1 m (Moghaddam et al., 2007). Deeper measurements allow sensing of soil moisture over (much of) the vegetation root zone, significantly increasing the hydrologic utility of the measured data. 77 For several years, the NASA Airborne Microwave Observatory of Subcanopy and Subsurface (AirMOSS) campaign has been the first to collect large-scale P-band soil moisture measurements. Observations were made several times a year over ten North American sites representative of different biomes. The root-zone soil moisture measurements will be assimilated into ecological models to provide new estimates of North American net ecosystem exchange (Chapin et al., 2012). Surface soil moisture such as that measured by L-band frequencies and higher can be estimated using a single average value as long as the averaging depth is chosen correctly (Raju et al., 1995; Le Morvan et al., 2008). Over the range of the root-zone (0 to ~1 m), however, soil moisture may vary strongly with depth. Ignoring the variation of soil moisture and estimating only an average value can lead to large errors in P-band soil moisture retrieval (Tabatabaeenejad and Moghaddam, 2011; Konings et al., 2014). Alternatively, it is possible to approximate the continuously variable soil moisture profile as consisting of a set of homogeneous layers, each with a unique moisture content and corresponding dielectric constant. Chapter 3 studied the backscatter modeling error (or forward error) associated with treating the profile as consisting of different numbers of layers. As the number of layers used in the retrieval process increases, the estimation error of the backscattering coefficient decreases. Increasing the accuracy of the embedded forward model should improve the retrieval process. However, as the number of layers increases, so does the number of unknowns. Furthermore, vegetation and roughness parameters must routinely be determined alongside soil moisture during the inversion of the backscattering coefficient, further adding to the number of unknowns. Although multiple transmit and receive polarizations can be used to obtain additional measurements, there are only 2.5 Degrees of Information within a typical set of AirMOSS backscattering coefficients (uHH, uvv, and UHV)- Thus, the number of unknowns can easily be greater than the available information in P-band SAR measurements, leading the inversion problem to become ill-posed. The soil moisture determined by this underdetermined retrieval problem may thus be overly sensitive to measurement noise and is unlikely to reflect true conditions. The ill-posedness of the inversion can be mitigated by using additional data, such as 78 __M multi-frequency (e.g. Rao et al. (1993)) or multi-angular observations (e.g. Jagdhuber et al. (2013)) to reduce the dimensionality of problem. However, such data are not always available. In the absence of additional data, one can use relationships between the soil moisture at different depths to reduce the effective number of degrees of freedom in the inversion problem. However, the shape of a soil moisture profile can vary dramatically depending on conditions. The AirMOSS algorithms assumes that soil moisture varies quadratically with depth ( Truong-Lol et al., 2015; Tabatabaeenejadet al., 2015), based on the observation that a quadratic polynomial can represent a variety of different profile shapes. The three parameters of the quadratic function are then inverted based on the data. Truong-LoI et al. (2015) further minimize the overall variability with depth. However, real soil moisture profiles are unlikely to follow a simple polynomial approximation like the quadratic. Instantaneous soil moisture profiles depend on the flow of water in the soil and the recent history of water sources and sinks (precipita- tion, evaporation, and drainage). Soil water flow depends on the matric head in the soil, which is a function of pressure and elevation. However, since the relationship between matric pressure and soil moisture follow a power law, the resulting soil moisture profiles are highly non-linear and cannot always be well-captured by a quadratic shape. Furthermore, changes in soil type with depth lead to discontinuities in the soil moisture profile that cannot be captured with a continuous quadratic profile. Similarly, infiltration fronts after precipitation events often lead to sharp changes in soil moisture at the front interface. Lastly, the use of a single quadratic profile to repre- sent soil moisture throughout the profile, implies that the spatial variability of soil moisture has a similar magnitude at different depths. By contrast, it is well known that deeper moisture layers are significantly less spatially and temporally variable than more shallow layers. Thus, while the assumption of a quadratic shape is more flexible than a variety of other functional forms, it is contradictory to soil moisture physics. Furthermore, the flexibility of the soil moisture profile comes at a relatively high cost of retrieving three different soil moisture parameters. Since there is duplicate information between polarizations (see also Chapter 2) and the cross-polarized 79 backscattering has only low sensitivity to soil moisture, retrieving three soil moisture unknowns (potentially on top of vegetation and roughness parameters) may lead to unstable, noisy retrievals. In this chapter, I propose to use an alternative method that uses Tikhonov regular- ization (Tikhonov and Arsenin, 1977) for soil moisture retrievals from P-band radar data. Tikhonov regularization has previously been used in other remote sensing retrievals, for example to minimize spatial gradients across single-look soil moisture retrievals from X-band data (Kseneman and Gleich, 2013) or in spatial upscaling and downscaling algorithms (Qin et al., 2013; Stoy and Quaife, 2015). Akbar and Moghaddam (2015) used regularization to combine active and passive measurements of the same area for surface soil moisture retrieval. Here, the regularization is used for soil moisture profile retrieval. The regularization incorporates the statistics of soil moisture - determined a priori - to reduce the underdeterminedness of the problem. It can result in a variety of soil moisture shapes and does not assume a single functional form. The use of soil moisture statistics as an additional source of information allows for retrieving more independent layers from a single set of multi-polarization observations than might be obtained with a classical retrieval approach. An example of how regularization reduces the likelihood of retrieving unrealistic profiles is shown in Section 4.2. Section 4.3 explains the regularization method, including the use of an observing system simulation experiment to determine the optimal value of the weighting parameter between the observations and the prior statistics. The method is tested in Section 4.4 using P-band observations from AirMOSS covering the Harvard Forest in Western Massachusetts, USA. The regularized retrievals are compared with limited in situ observations made at the time of the observations. However, since these may not capture the true value of soil moisture across the larger AirMOSS pixels, the retrievals are also evaluated by comparing them to expected patterns based on variations in soil type and vegetation cover. 80 4.2 4.2.1 Regularization Approach Cost functions The retrieval of soil moisture from backscattering coefficients is usually accomplished by trying to find the set of conditions for which the predicted backscattering coefficient most closely matches the observations. Mathematically, this entails finding the global minimum of a cost or objective function J, Ji= z (Upq(X) (4.1) q) Upq p~q where the subscripts p and q denote the incident and scattering polarizations, X is the vector of unknowns, including soil moisture at different depths and roughness and/or vegetation parameters, Upq(X) is the predicted backscattering coefficient and o-j"q represents the observed backscattering coefficient. Because the magnitude of -pq can vary significantly depending on the polarization, the cost function is often normalized by the observations. In this paper, we focus on soil moisture retrieval and assume that the vegetation and roughness parameters are known. For relatively homogeneous vegetation covers, this can be achieved by using data from in situ sampling. Alternatively, data from a timeseries of observations may be used to gain information from additional observations (Truong-LoI et al., 2015). In that case, the smooth-surface reflectivities F can be used instead of the total backscattering coefficients Up,, such that the classical cost function becomes, F p p,q ) (4.2) / J = Additional experiments using a backscattering coefficient-based cost functions (not shown) have demonstrated that the performance of the regularization method is not sensitive to the choice of observational parameters as long as the vegetation and roughness parameters are known. While accounting for vegetation and soil roughness effects can present its own challenges, it is outside the scope of this chapter. 81 In order to retrieve accurate soil moisture statistics, it is necessary to change the topology of the cost function J so that its global minimum is closer to the true solution. This can be done by adding a Tikhonov regularization term J2 that accounts for the expected covariance between the two layers. This can be written quantitatively as, s2 = (Oz T C-j -ZY) (z -i) The Oz is the long-term average value of Oz and C- (4.3) is the inverse of the covariance matrix of 6. They both depend on the local climate, land cover, and soil texture. The total cost function J is then given by + A (0 J = J + AJ2 = pq ( Z)T Cz1(0z ), (4.4) 0p~q where the A is a single weighting parameter that controls the relative contribution of the measurement and regularization terms. The added regularization term penalizes different profile deviations from the mean according to the magnitude of their expected covariance. The regularization method thus has three parameters: the mean profile, the inverse covariance matrix of the profile, and the weighting term A. The mean profile and inverse covariance matrix can be determined using either simulated soil moisture data from a hydrologic model, or statistics derived from nearby in situ measurements. Since these parameters depend on the average profile statistics, they are not expected to vary significantly in space. The optimal value of A depends on the variability of the soil moisture profiles and the measurement conditions and is difficult to determine a priori. The source of data that is used to calculate the profile statistics can also be used to generate a simulation experiment that can be used to determine the optimal value of A, as will be further discussed in Sec. 4.3. 4.2.2 Illustrative example In this illustrative example, the forward or 'true' soil moisture profile is close to continuous. It is shown in Figure 4-1. This is a sample profile based on in situ 82 measurements, which are further described in Sect 3.4.2. The observed smooth- surface reflectivities 1F associated with this forward profile are calculated by using the continuous fractions solution to Maxwell's equations for a set of smooth slabs with different dielectric constants, derived by Kong (2008). To use this solution, the continuous forward profile is approximated as consisting of 160 layers of 5 mm of depth each. In doing so, the soil moisture difference between consecutive layers becomes small, and in the limit of infinitely small layers, the resulting calculated reflectivities will equal the reflectivities of the continuous profile. It is of course infeasible to retrieve all 200 layers of soil moisture independently with polarized but single-frequency backscatter measurements. Instead, the full profile is characterized in this example by the average soil moisture values over two layers: 0-14 cm and 1428 cm. The bottom depth of 28 cm is chosen based on the approximate penetration depth (the e-folding depth of the wave power) for the region, as further discussed in Sec 4.4. Figure 4-1 also shows the average soil moisture of the continuous profile in each of the two 14 cm layers. Even if a two-layer retrieval algorithm cannot retrieve the entire profile, it would ideally retrieve these average values. A P-band sensing frequency of 430 MHz and incidence angle of 40' are assumed. Figure 4-2 shows the cost function J as a function of the two unknowns X {01, 02} where 0, is the soil moisture in layer z. The pink square represents the average soil moisture values associated with the profile (the ideal solution), while the white square shows the true global minimum of J1 . The simplification of the fully variable profile to a two-layer profile changes the topology of J such that the equivalent twolayer soil moisture in the profile is not the one that most closely matches observations. Furthermore, the function has several local minima. The bottom left panel of Fig. 4-2 shows the topology of J2 as a function of 0,. Again, the pink square shows the average soil moisture values associated with the true profile, while the white square shows the global minimum of J2 . Not surprisingly, the global minimum J2 is equal to 0;, where J2 = 0. Because Czz is symmetric, so is J2 . As shown in the right panel of Fig. 4-2, the combined cost function includes contributions from both the measurements and the a priori expectations, and has a 83 0 -0.05 E -0.1 'E-0.15 -0.2 -0.25 0 0.1 0.2 0.3 0.4 Soil moisture [cm 3/cm 3 Figure 4-1: Sample soil moisture profile derived from in situ measurements at Harvard Forest on February 4th, 2012 used to calculate the 'true' forward backscatter of Fig. 4-2. The dashed line represents the average soil moisture over two layers of 14 cm each; the optimal solution of a two-layer retrieval algorithm. global minimum (white square) close to the true solution (pink square). Additionally, the new cost function has far fewer local minima. In this example, the many local minima in the topology of J1 originate in part from the differences between the reflectivities of the fully variable profile and the two-layer simplification used in the retrieval process. In general, observations will be influenced by other sources of modeling and measurement noise that will serve to further distort the topology of J1 , such that the global minimum is even further from the true solution. As in the above example, an appropriately weighted regularization term is necessary to reduce the number of incorrect minima, where the weighting is determined by the value of A. In many data assimilation problems, the cost function to be minimized has a similar form to Eq. (4.4), but rather than having an explicit weighting parameter A, the relative weights of the observational and regularization or model term are determined by the size of the covariance terms in each. This approach has also been used for the retrieval of atmospheric temperature and humidity profiles from the inversion of satellite soundings (e.g. Eyre, 1989). In such cases, the covariance terms represent the covariances of the errors in the observation and 84 Ji 2 0.4 J +kJ2 0.0.2 2 0.5 0.1 400 0.4 15 0.3 1 J2 . 0.2 Sm 0.4 0.4 ~0.2 C14 0 sm1 20.2 200 0.2 0.4 sm1 Figure 4-2: Different cost function components as a function of the top-layer and bottom-layer soil moisture. In the top left panel, the objective function without regularization (minimizing the squared normalized difference between measurements and expected values) is shown, where the measurements are those expected for the profile shown in Figure 4-1. In the bottom left panel, a pure regularization term is shown. The right panel shows the sum of the two, with the regularization term weighted by parameter A = 0.01. In each panel, the white square represents the global minimum of the displayed quantity, while the pink square represents the optimal solution for the profile used to generate the simulated measurements. model, such that the model and observation are weighted equally if the two are equally uncertain. Here, to avoid a dependence on the particular assumptions of the hydrologic model, no such assumption is made. This also avoids the difficulty of estimating the uncertainty associated with the observations and radiative transfer models. Note that even though C, is defined based on the covariance of the soil moisture profiles themselves, rather than of the errors associated with their estimates, this is equivalent to calculating the error covariance term associated with a model assuming the soil moisture profile is always equal to its climatological mean. In this example, A = 0.01. For different values of A, J = J + J2 may not represent the true solution as a minimum. For a given site, the optimal value of A must be determine using a simulation experiment, as discussed further below. 85 4.3 Observing System Simulation Experiment at Vaira Ranch The core of the simulation system is comprised of a coupled hydrologic model and electromagnetic scattering model. These are used as a forward model to simulate a set of P-band backscatter measurements at a site in California. This part of the simulation system is identical to that used in Chapter 3, but is also described in detail below for completeness. The resulting simulated measurements can then be used as inputs to a variety of potential retrieval algorithms. In turn, the output of these retrieval algorithms can be compared to the original hydrologic simulations to study their performance. 4.3.1 Model description At a given location, the soil moisture profile is determined by the local history of precipitation entering the soil on the one hand and moisture leaving the soil through evaporation, root uptake, and diffuse recharge on the other hand. Furthermore, the flow of moisture through the soil acts to redistribute the soil water content across depths over time, as governed by Richards equation. Evapotranspiration fluxes at the land surface are limited by energetic as well as water balance equations. To capture these effects, the one-dimensional Simultaneous Heat and Water (SHAW) model is used. A detailed description of the model equations can be found elsewhere in the literature (Flerchingerand Pierson, 1991, 1997). The model is parameterized to represent the Vaira Ranch Ameriflux site near lone, CA (Baldocchi et al., 2004); the specific parameters used are given in Appendix A. Micrometeorological parameters measured at the site are used to drive the evolution of the model. The site has a strongly seasonal climate with a hot, dry summer (lasting roughly from June to November) and a cool, wet winter (lasting roughly from December to May). The simulated soil moisture profiles are then coupled to a multi-layer slab model for the surface reflectivity (Kong, 2008). The soil scattering model represents the 86 soil as a set of 'slabs' of homogeneous soil moisture, such that scattering originates from the interfaces between the layers. Here, 200 layers of 5 mm depth each are used. The electromagnetic properties of the bottom layer, representing a depth of around 1 m, extend to infinity. Since 1 m is well below the average penetration depth of 24 cm at the site, the bottom boundary condition is not expected to have an influence on the accuracy of the simulations. We assume that the coherent sum of the scattered waves from the layer interfaces accurately approximates the true backscatter from a continuously varying soil moisture profile for a sufficiently high number of sufficiently small layers. The contributions of rock inclusions, macropores, organic matter and other sources of heterogeneity to the total volume scattering are neglected. Using more than 200 layers does not change the simulated forward soil backscattering coefficient. The land surface roughness is assumed to be distributed isotropically and exponentially, with a root-mean-square height of 1.3 cm and a correlation length of 13 cm. Subsurface layers are assumed to be smooth. An incidence angle of 300 and a measurement frequency of 430 MHz are assumed. This is the center frequency used by the AirMOSS mission. 4.3.2 Experiments performed The regularized retrieval algorithm minimizes the objective function J in Equation (4.4) to retrieve an unknown soil moisture profile and unknown roughness root-meansquare height. The retrieved profile is approximated by four constant layers, each of six cm depth, consistent with the fact that the average P-band penetration depth for this area is about 24 cm (Chapter 3). The four layers provide a balance between a high enough number of layers to reduce errors due to profile variability and a sufficiently low number of degrees of freedom so that the regularized retrieval is not strongly underdetermined. Figure 4-3 shows the evolution of the four soil moisture layers over 2001. During the dry summer, the top layers are drier than the subsurface layers as they are more affected by evaporation. In the wet season, precipitation events first enter the soil at the surface, so that the shallow layers are generally wetter than deeper layers. The dynamic variability decreases with depth. The combined retrieval 87 0.4 0.350.3- 0.25- E E 0.2 S0.15 0.1 -0-6 cm --- 6-12 cm 0.05 -- 12-18 cm -18-24 cm 050 100 150 200 Day of Year 250 300 350 Figure 4-3: Timeseries of 'true' average soil moisture simulated by SHAW for each of the four layers retrieved by the algorithms, i.e. the top six cm (black solid line), 6-12 cm (red dotted line), 12-18 cm (blue dash-dotted line), 18-24 cm (green solid line). algorithm has somewhere between one and five (four soil moisture profile layers and one roughness parameter) degrees of freedom. The mean profile and inverse covariance matrix used in the regularization are determined after calculating the average soil moisture in each of the four 6-cm layers for all profiles in the ten years of hydrologic simulation. To determine how sensitive the inversions are to the value of the weighting parameter A, different values are tested. The regularized retrievals are also compared to two alternative algorithms without regularization. In one algorithm, four layers of soil moisture are retrieved without regularizing the inversion cost function, i.e. by minimizing Eq. 4.2. Additionally, a third algorithm is tested in which only a single average soil moisture value is retrieved for the entire profile. Each of these algorithms is tested on a full year of daily simulated backscatter observations representing January 1st, 2001 to December 31st, 2001. For each algorithm, errors are compared by calculating the root-mean-squareerror (RMSE) between the retrieved and true average value in each of the four 6-cm 88 Table 4.1: Soil parameters for different SHAW runs Parameter name Units Forward value Campbell's pore size distribution index 2.83 Air-entry potential cm -42.3 Saturated hydraulic conductivity mm/hr 8.3 3 Bulk density cm/m 1.09 Porosity cm 3 /cm 3 0.59 Sand % 0.30 Alternate value 5.33 -7.7 7.7 1.39 0.40 0.20 Silt % - 0.57 0.64 Clay % - 0.13 0.16 layers per profile. For the single-layer retrievals, each of the four layers is set equal to the single value resulting from the inversion of the backscatter measurements. To ensure the global minimum is found for each retrieval, the cost function minimization is performed using a datacube approach. The mean O; and covariance matrix C,, of the profile depend on the local climate, land cover, and soil texture. In this study, the calculated values were obtained by averaging all the years of daily profiles simulated by using SHAW. Operationally, the mean value of the model results used to infer the local statistics may be imperfect. Aside from potential errors in the model physics or parameterizations, the climatologies of modeled soil wetness produces often differ significantly from each other and observations (Reichle et al., 2004; Koster et al., 2009). To determine the sensitivity of the retrieval performance to the values used in regularization, values are retrieved using statistics determined from two different runs of the SHAW model. Aside from the run used in the forward model, a second run is used in which the soil type differs. Table 4.1 lists the two sets of soil parameters. The resulting mean profile and inverse covariance matrix are shown in Figure 4-4. Although the magnitude of both OZ and Cj differs widely between the two cases, the vertical structure is similar for both. 89 C-1 0 10000 E 5000 2 -10 E-5000 3 0 4 0-20 1 0.1 0.15 0.2 0 [cm3 /cm 33 0 2 3 4 1 0000 -10 2 5000 CL C) -20 3 0 4 -5000 1 0.1 0.15 0.2 2 3 4 6, [cm3/cm 3] Figure 4-4: Mean profile (left column) and inverse covariance matrix (right column) for the forward (top row) and alternative (bottom row) hydrologic model results. 4.3.3 Results Regularization Weight The value of A affects whether the true profile at any given time actually minimizes J. If A is too large, true deviations from the mean due to normal variations in weather history will be unduly punished and the retrieved profile will always be close to the mean. If A is too small, however, the retrieval reverts to the ill-posed, unregularized retrieval. The optimal value of A varies widely between profiles. Depending on how closely the profile values and layer covariance (relative to the mean profile) follow the specified profile-average mean profile inverse covariance, the optimal value of A may be lower or higher. Additionally, in the presence of measurement or modeling errors (such as the error created by retrieving only four soil moisture layers), an artificially high or low value of A may cause cancellation of errors and give the retrieval the appearance of improvement. When averaged over multiple profiles, however, there is a globally optimal value of A. This is shown in Fig. 4-5, which depicts the combined root-mean- 90 0.2r E 0.15- CD 0.1 - E - 0 I0.05 10 10 104 100 Figure 4-5: Root-mean-square error of one year of simulated regularized retrievals as a function of the regularization weight A at Vaira Ranch square error of all retrieved profiles and averaged over all depths, for different values of A. The curve has a clear minimum around A = 10-1-3. This value is used in all further retrievals with regularization. Furthermore, the error curve is relatively flat near the minimum value of A. Because all simulation studies are imperfect to some extent, the 'true' error curve may differ slightly from that shown in Fig. 4-5, and the value of A = 10". may not be optimal when applied to a real dataset. However, the flatness of the curve near the minimum suggests that using A = 10-1. rather than the true optimal value will not significantly affect the retrieval performance. Retrieval with perfect regularization statistics Fig. 4-6 shows the evolution of the retrieved soil moisture of the shallowest soil layer (0-6 cm) for the three algorithms. As small changes in the fully variable soil moisture profile and differences between measurement noise between days change the location of the local and global minima of J1, the four-layer soil moisture retrieved without regularization differs. in Fig. As a result, the associated timeseries (green line) 4-6 shows large fluctuations. If only a single average soil moisture value and roughness height are retrieved, the inversion is no longer underdetermined and 91 0.4 0.35 It. 0.3 E E 0.25 It 0.2 oil II 'I Ig' 0.15 0.1 ,111111 -True 0.05 -- Retr. w/ reg Retr. w/out reg 0 50 ? 100 150 200 Day of Year 250 300 350 Figure 4-6: Timeseries of the top-layer (0-6 cm) retrieved soil moisture using different retrieval algorithms. The red dotted line is the retrieved value using a regularized cost function with four layers and the value of A that corresponds to the global minimum. The blue dash-dotted line is the top layer value for unregularized retrievals with the same number of layers. The black line us the true average soil moisture over the top six cm. becomes significantly more stable. However, additional error is caused by neglecting the variability of the soil moisture profile with depth. This leads to particularly large errors in the one-layer retrieval at times when the soil moisture profile varies most: after rainfall events (during peaks in the timeseries), when the surface is often much wetter than the subsurface, and during the dry season (roughly from Julian day 120 to 250), when the surface is much drier than the subsurface. In each case, the average retrieved soil moisture is drier (wetter) than the value in the top-most layer, so that there is a consistent bias to the average retrieved values in each of the two periods. By contrast, the regularized retrievals are both stable and accurate. Figure 4-7 shows the timeseries of true soil moisture and regularized retrievals for all four layers. The retrieval quality is similar for all layers. After rainfall events, however, the regularization term constrains the solution too much. Shortly after a rainfall event, an infiltration front travels through the soil. While shallower depths are very wet, deeper layers are still dry if the infiltration front has not reached them 92 6-12 cm 0-6 cm 0.4 E0.3 E 0.3 0.2 E 0.2 .) E S0. S 0.4 100 200 Day of Year 12-18 cm 0.3 100 c~-' 300 0.4 -True --- Retrieved 0.3- -". 100 200 Day of Year 18-24 cm E E.20.2a. 0 1 300 E 0.3- 01 <D" . - 0.4 200 Day of Year 0 300 Day of Year 3 0 Figure 4-7: Timeseries of the true (black solid line) and retrieved soil moisture (red dashed line) using regularization in each of the four retrieval layers. Table 4.2: Root-mean-square retrieval Regularized Regularized retrieval, retrieval, alternate forward covariance covariance 0.028 0.026 01 [% vol] 0.018 0.018 02 [% vol] 0.015 0.015 03 [% vol] 0.026 0.024 04 [% volj 0.023 all [% vol] 0.021 error for different algorithms Unregularized Unregularized retrieval, 1 retrieval, 4 layers layer 0.036 0.019 0.050 0.017 0.064 0.029 0.026 0.070 0.070 0.064 yet. The regularized retrievals at these times are too wet at depth, leading to large errors there, as can be seen during the rain event near Julian day 180 or, for the deepest layer, during the wetting and drying phases near day 300. The root-mean-square errors for all layers are shown in Table 4.2. For all layers, the regularized retrievals have the smallest root-mean-square errors. No matter the retrieval algorithm, the error increases with depth due to the reduced sensitivity of the observations to soil moisture at depth. Four-layer retrievals without regularization are unstable, leading to high retrieval errors more than three times as high as those of the regularized retrieval. 93 Sensitivity to regularization statistics Table 4.2 also shows the error statistics for soil moisture retrieved using a regularization term with the alternate profile statistics. Both sets of simulated retrievals have the same optimal value of A. Not surprisingly, the performance of the regularized retrievals degrades when the regularization statistics are imperfect. However, despite the widely different magnitudes of both 0, and C- 1 , the algorithm still performs well and significantly better than alternative retrieval algorithms. 4.4 Application to AirMOSS observations at Harvard Forest Vaira Ranch (and the Tonzi Ranch nearby) is covered by one of the AirMOSS sites. However, fewer AirMOSS observations exists over Vaira Ranch than over most other AirMOSS sites due to the Ranch's location close to the Beale Air Force Radar, which often limits the ability to operate a P-band radar such as that of AirMOSS. Over much of the AirMOSS domain, the landscape is covered by a mixture of shrubs and oak trees that is highly heterogeneous and for which it is difficult to determine optimal scattering parameters describing the vegetation. Instead, the regulariation method is applied to AirMOSS data flown over the Harvard Forest region in Western Massachusetts, USA. An alternative procedure for accounting for vegetation and roughness has been applied (Truong-LoI et al., 2015), which is briefly summarized here for completeness. The method uses data from the Forest Inventory Analysis (FIA) as input structural data for a canopy backscattering model representing the various components of the vegetation (Saatchi and McDonald, 1997), which is then in turn used to derive sitespecific vegetation parameters that allow the effect of vegetation to be described as 94 a function of biomass, o-pQ =ApQW PQ cos 6 (1 - e-BPQWOPQlcosO) 3 + CPQFPQ(E, )W PQ sin 0e-BQW PQ/cos0 + SHH (E, 0) eBWP/cosO. (4.5). where W is the above-ground biomass, 0 is the local incidence angle, 6 and SPQ are an effective dielectric constant and its associated bare-soil reflectivity, respectively, and ApQ, BpQ, CPQ, ozpQ, !3pQ, and ypQ are site- and polarization-specific parameters determined from the full canopy backscattering model. The biomass and soil roughness values are then assumed constant over the multiple (usually three) flights in a campaign. The values that minimize the mismatch between the observed backscatter and the predicted backscatter from Eq. (4.5) are used. In this first minimization, a placeholder optimal effective dielectric constant is retrieved. The optimal biomass and roughness values are then used to invert the backscattering coefficients for the surface reflectivities. The Harvard Forest is owned by Harvard University and consists primarily of red oak and red maple trees. The vegetation is 50-70 years old (Goulden et al., 1996). It has been used for experimental studies since 1907 and has been making eddy covariance measurements of carbon fluxes in the region since the late 1980's ( Wofsy et al., 1993). These measurements will be used to validate the AirMOSS L4 modeled carbon fluxes derived from AirMOSS root-zone soil moisture measurements. The site contains only moderate topography and has well-drained, loamy soils. Data from a three-day AirMOSS campaign in October 2012 are used; data are available on October 15, 2012, October 18, 2012, and October 21st, 2012. The observed backscattering coefficients are further averaged over a 3 by 3 block to reduce speckle. A set of three in situ soil moisture profiles is available spanning two of the AirMOSS pixels. These will be used to determine the regularization statistics and to run an observing system simulation experiment in Section 4.4.1. A much large range of in situ surface soil moisture measurements spanning many pixels were taken 95 at the time of overpass on October 15th and October 18th. These will be used to validate regularized AirMOSS retrievals in Section 4.4.2. Because the in situ measurements have some representativeness error relative to the AirMOSS pixels and are only available for the top 5 cm, the spatial patterns of the regularized retrievals are also examined for physical consistency as an additional check in Section 4.4.3. 4.4.1 Derivation of regularization parameters using an OSSE As in Section 4.3, an observing system simulation experiment (OSSE) is used to determine the regularization statistics. In Section 4.3, the forward soil moisture profiles were derived from a hydrologic simulation of the unsaturated zone. In situ measurements are used here to show that these can also be used to determine the regularization statistics. In situ soil moisture profiles are available at 2, 5, 10, 20, 40, 60, and 80 cm depth (Cuenca and Hagimoto, 2012). The measurements were averaged across the three profiles. The soil moisture values were converted to matric head using the Van Genuchten equations (van Genuchten, 1980) and interpolated linearly in pressure units. Soil texture is determined from the Soil Survey Geographic Database (SSURGO) (Soil Survey Staff, Natural Resources Conservation Service). This matric interpolation was sampled daily at 6:00 AM to generate a record of daily soil moisture profiles. The record is used to calculate the mean soil moisture and covariance matrix across four layers of 7-cm each. The evolution of the four 7-cm layers is shown in Figure 4-10a. They are remarkably constant over time, likely due to the lack of strong rainfall seasonality and well-drained soils at the site. The four 7-cm layers add up to a total measurement depth of 28 cm, this is the average penetration depth (the e-folding depth of the wave power ( Ulaby et al., 1986a)) at the AirMOSS center frequency. the dielectric model of Mironov et al. (2004) is used throughout. The resulting mean profile and inverse covariance matrix are shown in Figure 4-8. Note that the inverse covariance matrix has a similar spatial structure to that of Vaira Ranch (Figure 4-4) but has values that are an order of magnitude higher. This reflects the fact that the soil moisture at the Harvard Forest is less temporally variable than at Vaira Ranch. 96 10 4 0 10 -5 ~-15 3 0 -20 4 -25 -5 0 0.1 0.2 3 j [cm 3/cm I 0.3 1 2 3 4 Figure 4-8: Mean profile (left column) and inverse covariance matrix (right column) used in the regularization for Harvard Forest Daily soil moisture profiles based on in situ measurements are used to simulate observed reflectivities, to which the regularized retrieval method is applied. Figure 4-9 shows the resulting RMSE between forward and retrieved profiles as a function of the regularization weight A in black. As in figure 4-4, an intermediate value of A, A = 10-2.1 provides the optimal value of the weighting parameter (the lowest simulated retrieval error). Note that this value is almost an order of magnitude less than the optimal value found in the analogous OSSE for Vaira Ranch. This is consistent with the higher magnitude of the inverse covariance matrix at Harvard Forest. If this value of A is used, the regularized soil moisture across the four 7-cm layers compares well with the average values across the interpolated layers used in the forward part of the OSSE, as shown in Figure 4-10. 4.4.2 Comparison to in situ data Over a small subset of the AirMOSS flight lines, in situ measurements of soil moisture were made during two of the campaigns. Measurements were made using a Stevens POGO portable soil sensor representing a depth of 5.8 cm. Because the support scale (a few cm) of a single in situ measurement is many orders of magnitude smaller than that of the AirMOSS measurements (Bl6schl and Sivapalan, 1995) and soil moisture 97 0.1 0.09 0.08 0.07 e;-0.06 E 0.05 - 0.04 0.030.02- - 0.01 10 Figure 4-9: Root-mean-square error of one year of simulated regularized retrievals as a function of the regularization weight A at Harvard Forest "Retrieved" soil moisture "True" soil moisture 0.4 0.4 ,-0.35- g-0.35 E C 0.3 E (00.3 0.25 0-0.25 S0.2 0.2 E 0.15 E 0.15 0.1 - 00.0 0.10.05- 0.05- 0 -0-7 cm --- 7-14 cm -- 14-21 cm -21-28 cm 100 200 100 360 DOY 200 300 DOY Figure 4-10: Average 7-cm soil moisture values across interpolated soil moisture profiles (left) used as ideal retrievals for the OSSE at Harvard Forest and (right) values retrieved by the regularization algorithm with A 10-". 98 has significant spatial variability, many measurements are needed to determine representative soil moisture across a remote sensing pixel (Famiglietti et al., 2008). It is not practical to determine the soil moisture profile across the penetration depth for many points in a single day, so only surface soil moisture measurements are made here. Three measurements were made every 50 m across a transect of 15 plots. On October 18th, an entire quadrant of such transects were measured. Because the flight paths and transects are not aligned, the number of in situ measurements per AirMOSS pixels varies between 9 and 21 on October 15th, and between 3 and 6 on October 18th. Figure 4-11 shows the range of in situ soil moisture measurement for each AirMOSS pixel, along with the soil moisture retrieved value. Because of the low number of measurements, the full range is shown instead of the more common standard deviation. Despite the high spatial variability within each pixel, the retrieved value is within or close to the range of observations in most cases. With the exception of pixel 1 on October 18th, the spatial trend along the transect is also similar between the retrievals and the observations. The remaining errors are likely a combination of errors in the regularization retrieval method, errors in the assumed vegetation and roughness parameters, noise in the observations, and representativeness error between the in situ measurements and the AirMOSS observations. Overall, there is a small positive bias in the retrievals. Since many soils increase in wetness with depth, a portion of this bias might be due to the differences between the 7 cm depth represented by the retrievals and the 5 cm depth over which the measurements were taken. However, this effect is probably minor relative to other sources of bias. 4.4.3 Comparison with landscape characteristics To (further) evaluate the profile retrievals, their spatial patterns are compared to the spatial patterns of landscape characteristics. Landscape characteristics such as vegetation, soil type, and elevation influence the fluxes of water in, through, and out of the soil, and thus the expected trends of the soil moisture fields. To be able to visually identify patterns in the retrievals, three 5 x 5 km focus regions are first selected for closer inspection. These regions are determined prior to performing the retrievals and 99 October 15th, 2012 October 18t, 2012 0-6 *POGO 0A A E A 1 203 retrieval w/ reg, 0-7 cm 05- 1 0.4 0.3 E - I 05 SGPOGOmeas, 0-5.8 cm - retrieval w/ rea, 0-7 cm neas,0-5.8cm E 02 0-1 01- 1 2 3 4 5 Unique pixel # 6 7 8 -0 0.5 1 165 2 Unique pxel 2-5 3 3,5 4 # -0 - 902 Figure 4-11: Top-layer (0-7 cm) regularized retrievals (red triangles) compared to the range of in situ measurements (0-5 cm) across each AirMOSS pixel with in situ observations. The black dot is the average of the in situ measurements. Data from the October 15th, 2012 flight are shown on the left, while the right figure shows data from October 18th, 2012. selected such that their landscape characteristics are similar to those of the entire flight area. As shown in Figure 4-12, the cumulative distribution of soil type (sand fraction and clay fraction), topography (elevation and topographic moisture index TMI, also known as the topopgraphic moisture index (Beven and Kirkby, 1979)), and vegetation (IGBP land cover type from the National Land Cover Database and aboveground biomass derived from AirMOSS timeseries according to Truong-Lol et al. (2015) ) are comparable. Note that the clay fraction and retrieved biomass have a direct influence on the soil moisture retrievals by affecting the dielectric mixing model and reflectivity retrieval approaches ( Truong-LoI et al., 2015), respectively. Thus, it can be difficult to disentangle whether a spatial pattern of soil moisture that coincides with patterns in these ancillary variables is occurring because of natural co-variability or because of direct compensation in the retrievals. This is not a concern for the other landscape characteristics. Figure 4-13 shows the landscape characteristics and soil moisture retrievals for the first focus region on October 18th, 2012. No single landscape characteristic dominates 100 Clay Fraction - ----- Sand Fraction 1 1 0.5 0.5 0 50 10 0 10 5 Elevation [m] 110 TMI 1 1 0.5 0.5 0 10 20 Retrieved biomass 0 30 0 Figure 4-12: 100 200 200 400 600 800 1000 Land cover type 0.5 00 15 1wI - - 0 DF EF MF SL GL PS CR 300 Cumulative distribution of sand fraction, clay fraction, topographic moisture index, elevation, and retrieved biomass for the entire Harvard Forest flight right path (black line) and the 3 focus regions (red line), respectively. The bottom shows the distribution of land cover type between the flight path and the focus regions deciduous forest, EF = evergreen forest, MF = mixed forest, SL = shrubland, (DF grassland, PS = pasture, and CR = cropland). GL 101 the pattern of the soil moisture in the retrievals, but their influence on overall soil moisture spatial variation can still be seen. The diagonal 'line' of pixels with a high sand fraction corresponds to significantly lower soil moisture. This is consistent with physical expectations. Relatively sandy soils generally have a higher hydraulic conductivity and lower porosity than more clay-like soils, and thus drain faster. For relatively wet conditions such as those shown in the retrievals, we expect this effect to be especially significant near the wetter surface compared to deeper layers, as is observed in Figure 4-15. In addition, the lower right corner of the domain, where biomass is generally lower than in the rest of the domain, corresponds to slightly wetter retrievals. The lower biomass is likely to have lower transpiration and root water uptake rates. Figure 4-14 shows the second focus area on the same day. Soil moisture is generally drier in the regions with deciduous forests than in regions with evergreen forests, perhaps reflecting higher transpiration rates in deciduous forests. No clear corre- spondence between soil moisture patterns and landscape characteristics is observed in focus region 3 (Figure 4-15). Multiple landscape characteristics with different spatial patterns affect soil moisture, and the ancillary data may itself have errors. It is not clear if the lack of systematic patterns in Figure 4-15 is due to the quality of the retrievals or the limited predictive power of the comparison with landscape characteristics. To better disentangle the multiple effects that might be influencing the soil moisture retrievals, Figure 4-16 shows a boxplot of the retrieved top-layer soil moisture as a function of the soil sand fraction (note that this is not used directly in the Mironov dielectric mixing model for the retrievals, unlike the clay fraction). As expected, sandier soils generally have drier soil moisture. Similarly, as shown in Figure 4-17 soil moisture is driest in locations with the lowest topographic moisture index, reflecting expected patterns in subsurface lateral flow. The topographic moisture index (TMI = log AD (Beven and Kirkby, 1979), where A, is the cumulative drainage area and 0 is the local topographic slope) captures the effect of drainage on subsurface wetness and is expected to scale with soil moisture. 102 I EWAbanlml 5.0 Clay Ffacetan Sand FRacoan Land coWr Type Moad Forgat Evd Forat Topo MoNu"r ndx AG Soma" j1g") reno 7-14 cm layer 0-7 cm layer 0.5 0.4 - m 0.3 14-21 21-28 cm layer cm layer 0.2 0.1 0 Figure 4-13: Landscape characteristics of focus area 1 at the Harvard Forest. Top row: sand fraction (left), clay fraction (center), and elevation (right). Bottom row: land cover (left), retrieved aboveground biomass (center), and topographic moisture index (right). The bottom two rows show the retrieved soil moisture in the 0-7 cm (top left), 7-14 cm (top right), 14-21 cm (bottom left), and 21-28 cm layers, respectively. 103 I. E Clay Frmcbon Sand FrM*W m) pW - Co "and AG om. Type E ai "M Top (OMSfa) Molm"u hodx F~ 7-14 cm layer 0-7 cm layer 0.5 0.3 21-28 cm layer 14-21 cm layer 0.2 0.1 I I. m ." 0 Figure 4-14: Landscape characteristics of focus area 2 at the Harvard Forest. Top row: sand fraction (left), clay fraction (center), and elevation (right). Bottom row: land cover (left), retrieved aboveground biomass (center), and topographic moisture index (right). The bottom two rows show the retrieved soil moisture in the 0-7 cm (top left), 7-14 cm (top right), 14-21 cm (bottom left), and 21-28 cm layers, respectively. 104 Se~ E*vataonlml COy Froa~n Fr.c?~ leamo AG Sf*mas Lane Co*W Type Topo (AI EvdFore D*W Forest 2 ~out 10 200 'U' i o Z5 S u yend 0-7 cm layei 7-14 cm layer 0.5 0.4 0.3 14-21 cm layer 21-28 cm layer 0.2 0.1 0 Figure 4-15: Landscape characteristics of focus area 3 at the Harvard Forest. Top row: sand fraction (left), clay fraction (center), and elevation (right). Bottom row: land cover (left), retrieved aboveground biomass (center), and topographic moisture index (right). The bottom two rows show the retrieved soil moisture in the 0-7 cm (top left), 7-14 cm (top right), 14-21 cm (bottom left), and 21-28 cm layers, respectively. 105 0.50.450.4E - 0.35 E 0.25 E 0.2 - 0.15 0.1- 60-70 50-60 90-100 80-90 70-80 Sand Fraction 7 cm layer soil moisture retrieval for different - Figure 4-16: Boxplot of average 0 sand fraction classes. 0.5k E 0,4&i IB B S E 0.4 0.35 *; - *' 0.3 I E 0.25 0.2 E 6 0.15 I I I S -J 5-6 6-7 8-9 7-8 I 9-10 >10 A tanG Figure 4-17: Boxplot of average 0 - 7 cm layer soil topographic moisture index classes (in m). 106 moisture retrieval for different 4.5 Discussion and Conclusions The backscattering coefficient associated with a variable soil moisture profile can differ significantly from the backscattering coefficient associated with a uniform profile with soil moisture equal to the mean profile value. Thus, for low-frequency microwave SAR measurements of soil moisture that measure a significant depth into the soil (such as P-band measurements or L-band measurements over very dry areas), it is necessary to invert the backscatter model for different values of soil moisture at multiple depths in order to obtain in accurate retrieval. However, in the absence of multi-angular or multi-frequency data, the associated inversion is generally underdetermined and therefore ill-posed. This can be mitigated by introducing additional information based on hydrologic linkages within the profile to reduce the degrees of freedom in the inversion process. Specifically, one can regularize the cost function to include an additional term that penalizes profiles that deviate significantly from the expected covariance structure. In this paper, this approach was tested using a year of simulated daily SAR measurements, including simulated speckle, at P-band. The simulated data were parameterized to reflect the measurement frequency of the current NASA AirMOSS mission to measure root-zone soil moisture using P-band SAR. Retrievals from an unregularized, underdetermined retrieval are highly sensitive to noise and will fluctuate excessively, both between measurements and between depths, due to overfitting. The retrieval problem may be made well-posed by inverting only a single average soil moisture, such that two unknowns (soil moisture and roughness) are determined by two measurements. However, errors will result from missing the soil moisture variability, particularly shortly after rain events or during extended drydowns, when the soil moisture difference between layers is especially high. Additional errors result from phase effects at the transition between layers and the reducing soil moisture sensitivity with depth, which combine to make the total backscattering coefficient a highly variable function of the soil moisture profile (see also Chapter 3). Regularization avoids both of these problems and leads to stable and accurate retrievals, as indicated both by the visual quality of the retrievals. 107 Further improvements to the regularized retrieval algorithm are possible. For example, for applications in which timeseries data are available, observations from multiple days could be combined to estimate the soil moisture roughness height, making the assumption that it is constant in time (Kim et al., 2012). A host of possible improvements are possible if vegetation parameters must be retrieved, including regularization of multiple vegetation properties based on allometric relationships or the use of an additional regularization relationship between vegetation water content and soil moisture. However, the aim of this chapter is not to advocate for or create a single optimal retrieval algorithm for use in particular mission, but rather to test the use of regularization more generally. The experiments performed here show regularization can be an effective method to enable the retrieval of multi-layer soil moisture profiles without overfitting even from single-angle and single-frequency data. The dependence on the structure of the covariance matrix sometimes causes soil moisture profiles to be overly constrained. That is, although it prevents the retrieval of highly variable profiles due to overfitting, it also occasionally unfairly penalizes natural variability of the profiles that is not reflected in the covariance structure. For example, while surface and deeper soil moisture values are often positively correlated, there may be little correlation shortly after a rainfall event, when the infiltration front has not yet reached the deeper soil moisture layers. Regularization-retrieved soil moisture values will lead to a retrieval of profiles that are too wet at depth, as the weaker signal from the deeper dry layers is not enough to overcome the penalty from statistics that reflect the fact that soil moisture at different layers usually shows a large correlation. This problem could be avoided by making the structure of the covariance matrix state-dependent, as commonly done in the meteorological literature (known as flowdependence) (e.g. Bannister, 2008). Future work should study the potential improved in soil moisture retrieval from using state-dependent regularization statistics. 108 Chapter 5 Vegetation Optical Depth and Albedo Retrieval using Time Series of Dual-polarized L-band Radiometer Observations 5.1 Introduction Our ability to close the Earth's carbon budget and predict feedbacks in a changing climate depends on knowing where, when and how much carbon dioxide and water vapor is exchanged between the land surface and the atmosphere. Both these fluxes are intimately tied to vegetation: roughly 60% of global land evapotranspiration fluxes occur through plant-mediated transpiration (Schlesinger and Jasechko, 2014), and vegetation photosynthesis response to increasing CO 2 concentrations is the biggest carbon cycle feedback in climate models (Ciais et al., 2013; Schimel et al., 2014). Microwave radiometric data at L-band are sensitive to both vegetation characteristics and soil moisture. In particular, radiometric observations are sensitive to vegetation optical depth (VOD). Passive soil moisture retrieving satellites at L-band like the NASA Soil Moisture Active Passive (SMAP) (Entekhabi et al., 2010b), the ESA Soil 109 Moisture and Ocean Salinity (SMOS) (Kerr et al., 2012), and the NASA/CONAE Aquarius-SAC/D (Le Vine et al., 2007) must properly account for the effect of VOD on observations in order to accurately retrieve soil moisture. Furthermore, microwave VOD estimates have previously been shown to be useful indicators for understanding vegetation state and variability, complementing the information provided by optical indices (Andela et al., 2013; Poulter et al., 2014; Zhou et al., 2014). VOD is also a potentially useful tool for crop monitoring (Patton and Hornbuckle, 2013) that could help detect crop water stress before optical sensors can ( Van Emmerik et al., 2014). VOD is directly proportional to total vegetation water content (VWC), with a constant of proportionality that is dependent on frequency and canopy structure. Since total VWC is related to biomass (it influences the amount of available storage for water), VOD has been used as an indicator of biomass in the past (e.g. Liu et al. (2015)). However, since vegetation water content also varies depending on the soil water availability (even in the absence of changes in biomass), VOD can also be interpreted as an indicator of vegetation water content useful for studying plant responses to hydrologic stress. The VOD measured by passive microwave sensors is an integrated measure of vegetation water content and structural effects. The total VOD is always less sensitive to the lower canopy layers than to the upper canopy layers, although the exact rate of attenuation of the microwave signal depends on the canopy. The rate of attenuation is also frequency-dependent ( Ulaby et al., 1986b), although few studies have been done comparing the effect of frequency on the measured VOD. If differences in canopy penetration between observations at different frequencies are ignored, different satellites can be combined into a single long-term record of VOD (Owe et al., 2008; Liu et al., 2011). Such an existing record has been used as a vegetation indicator complementary to optical indices (Shi et al., 2008; Andela et al., 2013). VOD retrievals from recently launched L-band radiometers such as SMOS and SMAP could be used to extend long-term multi-frequency VOD records (van der Schalie et al., 2015). Additionally, vegetation water content, and thus the amount of plant stress inferred by measuring vegetation water content, generally varies throughout the canopy (e.g. 110 (Hellkvist et al., 1974; Bohrer et al., 2005; Janott et al., 2011)). Studies of vegetation water content based on remote sensing may thus be better served by using VOD from lower frequencies such as L-band, which attenuate less quickly and are more sensitive to lower canopy layers. Furthermore, the development of VOD datasets and of joint VOD and soil moisture retrieval algorithms at L-band is of interest because of the greater soil sensing depth of these frequencies. Several approaches exist for the simultaneous retrieval of vegetation optical depth and soil moisture that is necessary at L-band. Both variables can be simultaneously derived from a snapshot of measurements by using information from observations at both horizontal and vertical polarizations (Jackson et al., 2002; Meesters et al., 2005). However, because the two polarizations are closely correlated, such a retrieval is sensitive to noise, as will be further explained in Section 5.2. If multi-angular data are available, such as in the case of SMOS, these can be used to further constrain the retrievals (Cui et al., 2015). Alternatively, observations from multiple overpasses can be combined into a single retrieval. Such a multi-temporal approach rests on the assumption that vegetation state as reflected in VOD is likely to change more slowly than soil moisture, and is constant over adjacent overpasses. The use of a timeseries approach also allows for the retrieval of the single-scattering albedo, the amount of power scattered by the vegetation cover. The value of albedo is often assumed to be independent of polarization and constant as a function of land cover (Van de Griend and Owe, 1994; O'Neill et al., 2012; Kerr et al., 2011). Its values are often close to zero ( Wigneron et al., 2004). A correctly chosen effective value of the single-scattering albedo allows accounting for higher-order scattering effects, which are especially important over moderate to dense vegetation cover (Kurum et al., 2012a). Many of the land-cover dependent values used in the literature are therefore in some sense fitting-parameters ( Wigneron et al., 2004; Kurum, 2013). However, a land-cover dependent assignment is sensitive to errors in the land cover classifications, as well as to variations in albedo within a certain land cover type. A sensitivity study has shown that errors in assumed albedo adds more uncertainty to single-incidence angle VOD and soil moisture retrievals than errors in soil and canopy temperature, 111 soil roughness, or bias or noise in observed brightness temperature (Davenport et al., 2005). The ability to retrieve albedo directly rather than relying on assumptions about its value may therefore significantly improve both VOD and soil moisture retrievals. In this chapter, I introduce a new multi-temporal algorithm for simultaneous retrieval of vegetation optical depth, single-scattering albedo, and soil dielectric constant using dual-polarized single incidence-angle observations at L-band frequencies. The method is referred to as the Multi-Temporal Dual Channel Algorithm (MT-DCA) and tested using three years of L-band passive observations from the Aquarius sensor. The paper is organized as follows. Section 5.2 motivates the need for a timeseries algorithm to avoid compensating errors when retrieving multiple parameters from a snapshot of dual-polarized observations. Section 5.3 describes the algorithm design. The testing methodology and datasets used in this paper are described in Sections 5.4 and 5.5, respectively. Retrieval results are shown in Section 5.6 and discussed in Section 5.7. 5.2 Algorithm Motivation 5.2.1 Classical retrieval approach Almost all radiometric soil moisture retrieval approaches are based on the so-called T-W model, a zeroth-order solution of the radiative transfer equations describing the emission of the land surface TP= Tffi' + T 'nopy = TS (1 - rp) - + T(1 - W) (1 - Y) (1 + r,7y) (5.1) The TB, is the brightness temperature at polarization p, which is either horizontal (H) or vertical (V), T, is the effective land surface temperature, rp is the rough surface reflectivity, and T, is the canopy temperature. The quantity -y is the vegetation transmissivity, T, is the canopy temperature, and w is the vegetation single-scattering 112 albedo, the fractional power scattered by the vegetation. The vegetation transmissivity -y accounts for attenuation of the emission through the vegetation layer. It is related to the vegetation optical depth, y = exp (VOD\ cos) (5.2) (_Cos 0 where 0 is the measurement incidence angle. When the VOD equals 0, there is no vegetation attenuation on the microwave emission from the soil and the corresponding -y is 1. The VOD increases with vegetation density; over dense vegetation, -y approaches 0 and the microwave emission is dominated by vegetation. VOD is commonly assumed to be linearly proportional to vegetation water content (Jackson and Schmugge, 1991; Van De Griend and Wigneron, 2004), (5.3) VOD = b -VWC, where the constant of proportionality b depends on the vegetation structure. The rough surface reflectivity can be decomposed as rp = r* exp(-hcos(O)n), where r* is the reflectivity of the flat (smooth) soil, h is the roughness parameter, which is assumed to be linearly related to the root-mean-square surface height of the soil surface, and n is an angular value (Ulaby and Long, 2014). The Fresnel equations relate r* to the complex dielectric constant k of the soils, which is in turn governed by soil moisture and soil texture. Most soil moisture retrieval algorithms rely on the same (or an equivalent) mathematical problem. In order to determine the vector of unknown parameters X from a set of observations, the mismatch between the observed (Tvbs) and modeled brightness temperatures (Tmo'dI (X)) is minimized, X = min J = E (T obs - Tmodel 2 ,(5.4) p=H,V where p represents the polarization. There are a variety of algorithms that differ in how many observations are com113 bined - whether the sum over polarization is included or whether additional summations are made over different incidence angles, frequencies, or overpasses - and in how many unknowns are included in X. The cost function in (4) can also incorporate additional terms to account for a priori information on the unknown parameters and its associated uncertainty (Piles et al., 2010), which is the solution adopted for the SMOS L2 processor (Kerr et al., 2011). Soil moisture is a key unknown and is always retrieved. By contrast, additional parameters such as h, w, and VOD can be either assigned dependent on some ancillary information or retrieved alongside soil moisture. 5.2.2 Timeseries motivation In order for the retrieved values to be stable - that is, insensitive to measurement noise - the algorithm cannot have more unknowns than the degrees of freedom provided by the measurements considered. If this requirement is not met, the global minimum of the cost function in Eq. (5.4) will be overly sensitive to measurement noise or small imperfections in the radiative transfer model. For the SMOS satellite, multiple incidence angles are used to obtain additional degrees of freedom (Kerr et al., 2012). For data where only a single frequency and incidence angle is available, either a so-called single-channel algorithm using a single polarization (e.g. (Bindlish et al., 2015)) or a dual-channel algorithm (DCA) using both the horizontal and vertical polarizations (e.g. (Jackson et al., 2002)) can be used. A commonly used variant of the traditional dual channel algorithm is the Land Parameter Retrieval Model or LPRM (Owe et al., 2008). It uses only the H-polarization in the cost function, but also uses the V-polarization as an additional piece of information by algebraically rearranging the tau-omega model to provide a direct relationship between k and VOD (Meesters et al., 2005) that is a function of the multi-polarization difference index (Owe et al., 2001). For both the traditional DCA and LPRM, two polarizations are used to retrieve two unknowns. Because the horizontally and vertically polarized brightness temperatures are highly correlated, there is duplicate information in the two measurements. This duplicate information reduces the ability of a DCA or similar algorithms to robustly 114 retrieve two parameters in the presence of measurement or modeling noise and adds errors to the retrievals (Konings et al., 2015). This is illustrated in Figure 5-1. The background colors show the cost function, as a function of VOD and k, for a sample set of observations. The perfect retrievals would be those leading to the minimum of the cost function, indicated by a black dot. Small amounts of noise (AeH = 0.005, Ae, = -0.002) are added to the 'observed emissivities' to simulate observational or model noise. The cost function contours of the noisy observations are overlaid as black lines and the new solution (and associated retrievals) is shown as a red triangle. This example shows that even small amounts of noise cause large shifts in the observed solution due to compensation between VOD and k along the diagonal curvature of the cost function. For only a single polarization the cost function moves up and down by a far smaller amount than the distance between the true and noisy solutions (not shown), but retrieving two variables simultaneously allows compensation between the two. This leads to large errors in retrieved VOD and k. Konings et al. (2011) performed an observing system simulation experiment (OSSE) in which the errors associated with different retrieval algorithms were tested over an area representing the United States by using known truth conditions and simulating observations with realistic parameterization, model, and observational errors. Both the bias and random errors of retrieved soil moisture increased several fold for a dualchannel algorithm relative to a single-channel algorithm due to compensating errors. This is consistent with the effect of observational errors tested in a smaller-scale OSSE (Crow et al., 2005), where dual-channel algorithm errors were also significantly higher than single-channel algorithm errors. The problem of compensating errors can be reduced by using additional observations to increase the 'Degrees of Information' (the fractional degrees of freedom) (Konings et al., 2015) in the data used. The use of additional observations makes the retrieval problem less sensitive to noise. For sensors like SMAP where only a single incidence angle and frequency is available, this can be achieved by combining measurements from different overpasses. If the time between different overpasses is sufficiently short, vegetation properties can be assumed constant across the differ115 1. 0.8 0.2 0 10 20 30 k Figure 5-1: Cost function J as a function of VOD and k for a sample set of observa0 tions (July 16th, 2012, for a pixel centered at 19.48 N, 103.53"W in Central Mexico). The 'true' solution of the cost function (without noise added) is shown by a black dot. A small amount of simulated noise is added to the observations, 0.005 for the H-pol and -0.002 for the V-pol. The contours of the resulting noisy cost function are shown as black lines. The noisy solution of the resulting cost function is shown by a red triangle and is far away from the true solution. 116 ent overpasses. This assumption has also been used to improve multi-angular soil moisture retrievals from SMOS ( Wigneron et al., 2000; Kerr et al., 2011). The soil dielectric constant varies much more rapidly than vegetation and must be retrieved separately for each overpass. Adding each additional overpass therefore increases the number of (correlated) observations by two, but the number of unknowns by only one (by assuming the same VOD, the only extra unknown is the new k). 5.3 5.3.1 Algorithm Design Moving window timeseries design For each retrieval, the timeseries algorithm proposed in this work combines all observations within a moving window and retrieves a single value of VOD along with N different values of the dielectric constant k, where N is equal to the number of overpasses within the moving window. Thus, the retrieval is the solution to, N 2 (e0 od -oel(X)) . (5.5) mi X=VODlk1,....kN J(X) = e t=1 p=H,V The MT-DCA algorithm retrieves N+ 1 independent parameters (1 x VOD and N x k) with 2 x N observations (H x N and V x N). Increasing N increases the number of measurements available for the retrieval, but also increases the possible errors from changes in VOD over the time period spanning the observations (e.g. violations of the assumption that VOD is constant across the N overpasses). The optimal choice of N is thus the minimum value such that the 2 x N observations provide enough information to determine N + 1 parameters. Because the H- and V-polarized emissivities at any given pixel and time are correlated, however, they contain duplicate information and do not provide 2 full degrees of information for the retrieval. Instead, the measurements provide some fractional number of 'Degrees of Information' (Dol). The Dol is less than two by an amount depending on the non-linear correlation between the polarizations. It can be estimated using the normalized mutual information between 117 the Aquarius-based polarized brightness temperatures. Here, the H- and V-polarized data together contain 1.86 Degrees of Information (Konings et al., 2015). Using measurements at independent days provides 1.86 x N DoL. The Dol provide an upper bound on the number of parameters that can be estimated robustly from a given set of observations. Depending on the forward model and algorithm implementation, this bound may or may not be reached. To find the minimum N that allows robust retrieval, the ratio of the total degrees of information divided by the total number of unknowns can be used. That is, the retrieval ratio RR = 1.86. N/(N + 1). If RR is greater than 1, the retrieval algorithm is expected to be robust to noise. Fig. 5-2 plots RR for several values of N. The Dol and the resulting RR are also separately calculated for each land cover type. Because the non-linear relationship between the polarizations varies between land cover classes, the relationship between the two is weaker when data from multiple land cover types are combined. The DoI and RR therefore increase when calculated across all pixels rather than only those of a single land cover type. A choice of N = 2 is enough to get robust retrievals (RR > 1), so two overpasses are combined for this application. At each pixel, two different values of VOD are retrieved for each overpass: once when the current overpass is the first of two in the moving window, and once when it is the second. Similarly, k is retrieved twice for each overpass and pixel depending on the location of the moving window. In each case, the two possible window positions are averaged to provide a single dataset of VOD and k. The two possible retrievals for each k (estimated from different multitemporal windows associated with each overpass) will later be compared as a consistency check. Since Aquarius has a revisit time of 7 days (see Section 5.5.1), retrievals are only performed when there are at least two coincident observations in 14 days. This filters out times when missing data might otherwise affect the validity of the assumption that VOD is constant between data takes. 118 1.2 rr * . .* . . . . -------- - - - - - - - ---- - * *Z . . . * 0 * 0 0 * 1.2 0 * * 1.4- * 1.6 c 0.8 0.4 - N = 1 overpasses - N = 2 overpasses - N = 3 overpasses 0.2 - N = 4 overpasses - N = 5 overpasses 'O 'k Figure 5-2: Retrieval ratio of degrees of freedom for the different land uses and varying number of dual-polarized observations 5.3.2 Albedo retrieval Since the albedo is sensitive to canopy architecture and influences the retrieved values of VOD and k (Davenport et al., 2005), it is beneficial to retrieve its value directly instead of using an assumed dependence on land cover type. Such an assumption is sensitive both to errors in land cover classification (mostly based on optical data) and to vegetation variability within land cover types. It is possible to set up a retrieval approach wherein three overpasses are combined to retrieve 3 values of kt, a constant VOD, and a constant w. Such an approach has an RR > 1. However, the results illustrate the fact that the DoI only provides an upper bound that is not always reached - VOD and albedo compensate for each other significantly, leading to temporal fluctuations in retrieved albedo that are unrealistically large relative to its dynamic range and to its spatial variations (results not shown). The reason for such compensation can be understood by examining Eq. (5.1). The total bright- ness temperature can he separated into two components Ti" = T'(1 - 'r(k))y and is small relaTW noy = Tc(1 - w)(1 - -y)(1 + rp(k) 1 ). If the contribution from TW" tive to the total and T" nf" dominates, the functional form of the effect of both w 119 and -y (which is a monotonic function of VOD) on the observed TB is the same. It becomes impossible to distinguish between w and VOD, causing large fluctuations in each. Figure 5-3 shows the relative contribution of the TW"nOPY to the total brightness temperature under different w and VOD. The contribution of TW"I to the total TB is often small, especially over wet and heavily vegetated soils. This explains why allowing albedo to vary leads to unrealistically large temporal variations in both w and VOD. 0.2 0.18 0.9 0.16 0.8 0.14 0.7 0.12 0.6 a 0.1 0.5 0.08 0.4 0.06 0.3 0.04 0.2 0.02 0.1 0 0.8 0.4 1.2 0 Figure 5-3: Relative contribution of the vegetation canopy to the total brightness temperature emitted at H-polarization, T"7Hy IT-H as a function of albedo W and VOD. A value of k = 20 is assumed. Results at V-polarization are qualitatively similar (not shown). Instead, albedo is assumed constant ( Van de Griend and Owe, 1994; Wigneron et al., 2004) and retrieved separately across the full record of observations (and alongside time-varying VOD and k). The retrieval of albedo is robust for a given pixel if the total Dol across the M VOD - k retrieval pairs throughout the timeseries (M - DoI- N) is greater than the total number of unknowns (1 + M(N + 1)). If two consecutive overpasses per VOD - k retrieval pair are used (N = 2), this requirement is met if more than two VOD - k retrievals are available (M > 2). If the total number of available retrieval pairs M < 2, no retrievals are attempted. Otherwise, all available observations for a given pixel are combined to find w. The constant value of albedo 120 is chosen that minimizes the sum of the optimal cost function for each retrieval pair, = i fmin J = min -p=1 b - minkt,k2 tvoDt Todel (X)) t=1 p=H,V _p=l (5.6) 5.3.3 Additional parameters The land surface temperature is determined from ancillary data as described in Section 5.5. In this paper, we further retrieve k from the rough-surface reflectivity for validation of the overall algorithm. The roughness parameter h is assumed to be equal to 0.13, the average of the different land-cover dependent values assumed by SMAP (O'Neill et al., 2012). 5.4 Methods The MT-DCA algorithm is applied to three years of data from the NASA Aquarius sensor. While an alternative version of MT-DCA could be build that retrieves soil moisture directly instead of the soil dielectric constant k, here we retrieve soil dielectric constant in order to estimate parameters that are entirely independent of ancillary data (such as soil texture), which might contain errors. A similar approach was previously used by de Jeu et al. (2014). The search space for k is limited between 2.5 and 35, values that were chosen based on the Mironov dielectric mixing model (Mironov et al., 2004) for a range of soil types. For VOD, the search space is limited to values between 0 and 1.3 nepers. Validation of the resulting VOD retrievals using direct ground-based measurements of vegetation water content is difficult, as no regional monitoring networks exists at the spatial scale of the Aquarius satellite. Vegetation water content is highly spatially variable, so that any in situ measurements that cover only a small fraction of the total Aquarius instrument field-of-view scale cannot be considered representative. It is therefore difficult to directly validate VOD data. Instead, the spatial patterns of the retrieved datasets are examined for physical 121 realism. Additionally, several focus pixels are chosen that represent relatively homogeneous land cover conditions (measured using the Gini-Simpson index (Simpson, 1949) on the discrete land cover classes, as proposed in Piles et al. (2015) and a wide variety of climatic and land cover conditions. One of the SMAP Core Cal/Val sites (SMAPEx)is also chosen as a focus pixel. The specific location and dominant land cover type of each focus pixel are included in Table 5.1. The temporal dynamics of each of these focus pixels are compared to the temporal dynamics of precipitation over the same area, which is expected to have a strong influence on vegetation water content in many regions. The MT-DCA VOD retrievals are also compared to those from the commonly used LPRM algorithm (Owe et al., 2001; Meesters et al., 2005; Owe et al., 2008). In order to be able to distinguish algorithm differences from differences in frequency or sensor characteristics, the LPRM algorithm is implemented and applied to the Aquarius observations as the MT-DCA. For LPRM, w = 0.06 is assumed, based on the value in Owe et al. (2001). Table 5.1: Target areas: name, location, dominant IGBP land cover type, and coefficient of determination R 2 between time series of mean Aquarius and SMOS k retrievals. Site name SMAPEx Amazon Latitude 34.70 OS 2.23 0S Longitude 145.73 0 E 66.00 W Land cover Open Shrubland Evergreen Broadleaf Forest Nordeste 7.30 42.63 OW Savanna Pampas 33.82 60.17 OW Cropland East Africa Central Asia West Africa 5.49 0S 45.27 ON 9.68 0 N 34.50 OE 66.30 OE 6.37 OE Woody Savanna Grassland Natural Vegetation Mosaic 0S 0S Since Aquarius provides radar and radiometer collocated observations, this study also explores the relationship between Aquarius MT-DCA VOD and two alternative active vegetation indices. Because soil scattering generally leads to negligible depolarization (Van Zyl and Kim, 2011), the cross-polarized backscattering coefficient can be used as an index of vegetation scattering intensity and water content. Alternatively, Arii et al. (2010) defined the Radar Vegetation Index (RVI), which is a measure of the randomness of canopy elements and vegetation scattering. 122 Although the retrieved VOD dataset is the primary focus of this study, the MTDCA retrievals are further evaluated by analyzing the retrieved values of w and k. The spatial and temporal patterns of the retrievals of k are also compared to those of LPRM. To avoid contamination from differences in assumed soil texture and dielectric mixing, the retrievals of the dielectric constant k are compared rather than the soil moisture estimates. The two are monotonically related. Lastly, a consistency check is performed on the k retrievals. For each date and location, two retrievals of k are obtained - one when the current date is the first in the two-overpass window, and one when it is the second. The two sets of retrievals are compared to test the robustness of the retrievals. Soil roughness is often assumed to depend on land cover type, as in the SMAP retrieval algorithm (O'Neill et al., 2012) or retrieved from additional information, as done by SMOS (Kerr et al., 2011). In this study, we use a globally constant soil roughness value of h = 0.13 and n = 2, the average of the land-cover dependent values used in the SMAP retrieval algorithm (O'Neill et al., 2012). Using a globally constant value allows the algorithm retrieval test to be independent of any possible errors in ancillary land cover data. Thus, any spatial patterns in Figures 5-4 and 5-6 are a direct test result of the data retrievals and not of ancillary data. Sensitivity tests showed that the exact value of soil roughness used had only a minor effect on the retrieved VOD and albedo values (not shown). To isolate the effects of the unique albedo and VOD retrieval assumptions of the MT-DCA, the same roughness assumptions are used for both the MT-DCA and LPRM implementations. The different datasets used for these analyses are described in Section 5.5. All datasets are converted to the same gridding scheme and spatial resolution, which is chosen to match the Aquarius observations. Since Aquarius measurements do not exactly overlap over time, the first 7 days of observations are used to set up the grid. Subsequent overlapping footprints with centers less than 0.15 degrees from a grid center are included in that grid cell, otherwise they are excluded. More detailed information on the gridding strategy can be found in McColl et al. (2014) and Piles et al. (2015). To enable spatial and temporal consistency, all data sets used in this 123 work have been resampled to the Aquarius footprint grid: land-cover classification data (used in interpreting the results only) is resampled using the most common land cover class, while ancillary precipitation and temperature data are resampled using linear averaging. When converting datasets with a higher spatial resolution to the Aquarius gridding scheme, a circular orbital footprint is assumed, with a radius dependent on latitude. Note that the land cover data are only used in the analysis of the results, not in the retrieval algorithm itself. 5.5 5.5.1 Datasets used Aquarius Level 2 data The Aquarius/SAC-D mission, launched in June 2011, is a joint U.S.-Argentinian mission to map the surface salinity field of the oceans from space. It has equatorial crossing times of 6 A.M. (descending) and 6 P.M. (ascending) and a 7-day repeat cycle. Its payload includes the NASA Aquarius sensor, the first combined active/passive polarimetric L-band microwave instrument in space. It consists of three L-band radiometers and a scatterometer, which image the Earth in a pushbroom fashion at 29.360 (inner beam), 38.49' (middle beam), and 46.29' (outer beam) incidence angles, with 3 dB footprints of 76 x 94 km, 84 x 120 km and 96 x 156 km. (Le Vine et al., 2007). The present study uses three years of global Aquarius Level 2 data (version 2.0), covering the period from September 1st, 2011 to August 31st, 2014. Dual-polarized brightness temperatures (TBH and TBv) from the middle beam acquired during morning (descending) overpasses are used for joint VOD, k, and w retrievals. Coincident cross-polarized backscattering coefficients (UHV) are also selected to explore their relationship with retrieved VOD. Only data from morning overpasses are used to ensure the vegetation and near-surface soil are in thermal equilibrium. Data from the central beam is chosen since, out of the three available beams, the greatest amount of independent information can be obtained from the center-most angle at 38.49', which 124 is also the closest to SMAP's incidence angle. Radar and radiometer data have been screened for orbital maneuver times and Radio Frequency Interference (RFI) Le Vine et al. (2014). In addition, data over ocean, land-sea transitions, Antarctica, Greenland and non-vegetated surfaces (water, urban and barren land covers) have been masked out. 5.5.2 NCEP land surface temperatures and flags The land surface temperature T provided as auxiliary information with Aquarius data is used as an input retrieval parameter in the present study. They are obtained from the National Centers for Environmental Prediction (NCEP) Global Data Assimilation System (GDAS) and interpolated from the daily 0.25' product to the exact time and location of the Aquarius observations. Pixels with land surface temperatures less than 00 C were assumed to have frozen soils and masked out of the analysis. Similarly, pixels where the observed emissivity was greater than one were assumed to have an erroneous land surface temperature and masked. Lastly, locations and times where NCEP data suggest the presence of snow or ice cover were also removed from the analysis. 5.5.3 MODIS IGBP land cover The 2005 MODIS MCD12Q1 International Geosphere-Biosphere Programme (IGBP) collection 5 landcover product has been used in this study to characterize the dominant land cover within each Aquarius footprint. The MODIS IGBP land cover is a world-wide product at 500-m spatial resolution that encloses 17 distinctive land cover classes. MODIS products are freely distributed by the U.S. Land Processed Dis- tributed Active Archive Center (www.lpdaac.usgs.gov). Note that land cover data are only used to interpret the results and not within the retrieval algorithm. 125 5.5.4 MERRA-Land observation-corrected global precipitation - Global daily precipitation from the Modern Era Retrospective Analysis (MERRA) Land run (Reichle et al., 2011) have been used in this study, with additional corrections applied to match the data from the Global Precipitation Climatology (GPCP) project and the NOAA Climate Prediction Center (CPC) (Reichle and Liu, 2014). 5.5.5 Water fraction The observed brightness temperatures from Aquarius are corrected for the effect of emission from surface water bodies. The NCEP Land Surface Temperature is assumed to be equal to the temperature of any water bodies in the pixel, whose brightness temperature Tter is calculated using the model of Klein and Swift (1977) with an assumed salinity of 0.5 parts per thousand. bodies in the pixel f, The fractional coverage of water is then used to separate the land emission and water emission contributions to the Aquarius observations. It is assumed that Tbs = fw 7 at'e + f.)Tland, which can be re-arranged to solve for Tj1lnd. The static water fraction (1- fw is determined by calculating what fraction of the high-resolution 250 m land cover data from the MODIS MOD44W dataset are classified as water or land. The data are first aggregated to the 3km EASE grid used in the SMAP Testbed and then converted to the Aquarius footprint grid. Pixels with more than 10% static water cover were removed from the analysis entirely. 5.6 5.6.1 Results VOD retrievals A global map of three-year time-average Aquarius VOD retrievals using the MT-DCA is shown in Fig. 5-4. The spatial patterns of VOD retrievals follow global vegetation distributions, with the highest vegetation optical depth in tropical and boreal forests and low VOD in arid climates. Across the Sahel, there is a gradient of increasing average VOD from North to South. Fig. 5-4 also shows the mean VOD obtained 126 by applying the LPRM algorithm. The spatial patterns of the two VOD temporal means are generally consistent except over densely forested areas. The difference over densely forested areas may partially be occurring because valid LPRM retrievals are so rare in these regions that the time-average VOD of LPRM is capturing a different subset of the seasonal cycle than the annual average of MT-DCA. Across the globe, LPRM predicts a negative VOD for 11% of retrievals, which is physically impossible. These occur predominantly over dry, lowly vegetated ecosystems. Indeed, for regions where MT-DCA predicts a VOD less than 0.1, more than 50% of all LPRM VOD retrievals are negative. These unphysical values occur because the LPRM assumes a fixed relationship between k and VOD based on an exact equality of the T - W model in both polarizations. In reality, noise and model error (including, for example, an imperfect w specification) may mean that there is no perfect solution to both equations, so after a least-error value of k is found the accompanying VOD may not be physically realistic. For a relatively coarse resolution (90 km footprint) satellite like Aquarius, this may be especially common. The occurrence of negative VOD retrievals at L-band is consistent with a previous C-band application of LPRM in which switching from soil moisture retrievals with possibly noisy soil texture values to soil dielectric constant retrievals increased the number of valid VOD retrievals by as much as 200 days a year, again predominantly over dry areas (de Jeu et al., 2014). In this paper, LPRM-retrievals that predict a negative VOD are removed from the comparison. For an additional 10% of observations with valid MT-DCA retrievals, including many over the Amazon and Congo river basins, LPRM retrievals are not made because they have a MPDI of less than 0.01 (Meesters et al., 2005). To gain further insight into the behavior of the MT-DCA retrievals, their temporal dynamics are compared to those of LPRM VOD retrievals and of precipitation (an expected strong predictor of VOD in several areas) for several focus pixels in Fig. 5-5. All datasets are shown at a weekly temporal resolution, equal to the average revisit time of the Aquarius satellite. In cases where there is significant seasonal variability in precipitation, the retrieved VOD is responsive to accumulated precipitation and consistent with expected seasonal changes in vegetation water content. For example, 127 1.2 0.8 0.6 -IL 0.4 0.2 Figure 5-4: Global maps of mean MT-DCA (left) and LPRM (right) VOD retrievals for the three year period of this study. a clear seasonal cycle is evident in the VOD retrievals over West Africa, Nordeste, and East Africa, where VOD shows a steady decline after the end of the rainy season. In each of these sites, there is a lag between the end of the rainy season and the minimum value of VOD. This suggests that VOD is sensitive to changes in vegetation water content accompanying the plant response to water stress and/or to changes in leaf biomass - the response of vegetation to a reduction in (stochastic) rainfall is not instantaneous. For most target regions, the LPRM VOD shows more high-frequency variability than the MT-DCA VOD. Although it is possible that the constant VOD assumption in the MT-DCA slightly dampens natural variability, the near-oscillatory behavior of many of the high-frequency LPRM retrievals suggest that they are due to retrieval noise rather than due to true variability in the signal. Such differences in temporal behavior occur over much of the globe, as shown in Figure 5-6. For both MT-DCA and LPRM, the standard deviation is shown after the 5-week moving window mean is removed. This moving window subtraction acts as a high-pass filter, and the remaining variability is more likely to be retrieval noise than true temporal variability in the signal. The MT-DCA high-frequency variability is significantly lower than that of LPRM for much of the globe. The high-frequency variability of the L2 VOD from SMOS (which uses multi-angular measurements and a prior estimate of VOD based on leaf area index) (Kerr et al., 2011) is about halfway between that of MT-DCA and that of LPRM (not shown). The availability of both active and passive data from Aquarius allows a preliminary 128 cEntral 0 2 -12 30- Asia ~~~MRain~fa11--LPRM VOD -- TDAV 0.5 25- 0.4 E20-0 .0 :06> 9 1510 !04 M16 10. 1 - M04 0 01114 01113 01112 0 0.30 1 01)13011 0112 Afr NordsteEast 20 20r '_15 E .301 010- 01/13 01112 0.4 02 01)12 v01/12 01/14 01/13 01114 SMAPEx Pampas 20 S 0.4 -0.5 W 05 15 15 -0.4 E -0.30 0 1 10 30 I.. 1/1NI3 01)12 01/12 01/13 .2 .9 0. AQ I .210> >ii b 0113011 01)1 01/13 01112 01114 01114 West Atma 3j 0.5 f1 L1J5 04 10(1 0.I L % 0 12 01/12 0 i - 01/13 W/12 01/14 Figure 5-5: Time series of weekly mean MT-DCA VOD, LPRM VOD, and precipitation over focus pixels. Note the different axes scale for the Amazon series. 0.1 p0 0.08 0.06 p 40 0.04 M46I 0.02 4A Figure 5-6: Global maps of standard deviation of MT-DCA (left) and LPRM (right) VOD retrievals for the three year period of this study. In both cases, a 5-week moving average is first removed from the timeseries for each pixel, so that the standard deviation primarily reflects high-frequency variability 129 0 investigation of active and passive vegetation indices. Figure 5-7 compares Aquarius MT-DCA VOD data to coincident OHV and RVI observations by showing the joint density of the two for all locations and times. The VOD-RVI joint density (R 2 = 0.46) is more flat and has longer tails than that of VOD and cTHV (R2 = 0.76). That is, there is more scatter in the relationship between VOD and RVI than in that between VOD and OHV. This suggests the latter may hold more promise as an active-microwave-based predictor of VOD for single-channel soil moisture retrieval approaches. However, some saturation in the 0 HV may be occurring for densely vegetated sites. 200 1.22 8 1.2 180 1 7 160 1 140 0.8 0.8 120 0 0 100 >0.6 > 0.6 80 0.4 0.2 - 3 60 0.4 .40 0.2- 20 0 0.02 0.04 0.06 0 00 0.2 0.6 0.4 0.8 10 RVI cHV Figure 5-7: Joint density of Aquarius radiometer-derived vegetation optical depth vs. scatterometer (THV in linear units (left) and radar vegetation index (right). All available combinations of active and passive measurements (e.g. one at each location and time) were used. 5.6.2 Albedo retrievals Table 5.2 shows the average retrieved albedo values for each land cover type, as well as the average parameters assumed in the SMAP passive-only retrieval algorithm. The retrieved values are generally lower than the land-cover dependent values used by SMAP as well as the globally constant value of 0.06 assumed by LPRM (Owe et al., 2008), consistent with theoretical findings that higher-order scattering reduces the 130 effective albedo values used in the tau-omega model (Kurum, 2013). Albedo values are generally higher for vegetation covers with significant woody components, such as forests and woody savannas. However, there is significant variability in retrieved albedo values across and within land cover class. Grasslands, croplands, open shrublands, and savannas show particularly large amounts of variability within each class relative to the class average. For open shrublands, this variability appears to be due to differences between the tundra regions and shrublands in less densely vegetated areas. Table 5.2: Land cover variability of retrieved albedo w. Parameters for SMAP W are obtained from O'Neill et al. (2012) Land cover type Evergreen Needleleaf Forest Evergreen Broadleaf Forest Deciduous Needleleaf Forest Deciduous Broadleaf Forest Mixed Forest Closed Shrublands Open Shrublands Woody Savannas Savannas Grasslands Croplands Cropland/Natural Veg. Mosaic SMAP w 0.12 0.12 0.12 0.12 0.10 0.05 0.05 0.12 0.08 0.05 0.05 0.065 Retrieved w : mean (std. dev.) 0.05 0.05 0.06 0.03 0.05 0.03 0.05 0.04 0.02 0.03 0.04 0.02 (0.02) (0.03) (0.02) (0.03) (0.03) (0.04) (0.05) (0.03) (0.03) (0.05) (0.04) (0.03) A global map of the retrieved constant albedo values is shown in Figure 5-8. Not surprisingly, the spatial patterns of albedo roughly follow those of average VOD and of expected vegetation cover, although there are a few more noisy high-albedo outliers. In general, the transition between low and high vegetation regions is more rapid for albedo than for VOD, as can be seen for example in Northern Africa. This is consistent with the apparent sensitivity of albedo to woody biomass; average VOD trends may be capturing smaller-scale spatial variations in leaf cover that albedo is insensitive to. 131 0.2 0.15 0.1 * 0.05 0 Figure 5-8: Global map of retrieved albedo 5.6.3 k retrievals For each date and location, k can be retrieved at either the start or the end of the moving window. A global map of the time-average of the instantaneous difference between the two sets of k retrievals is shown in Fig. 5-9. The difference is generally small, as confirmed by the figure inset, which shows the overall distribution of differences. The standard deviation between the differences is only 1.54 and there is no significant bias (mean Ak = 0.10). The similarity between the two sets of k confirms the robustness of the algorithm for soil moisture, VOD, and albedo retrievals. Figure 5-10 shows the mean retrieved k for both MT-DCA and LPRM. The MT-DCA dielectric constants are slightly higher (wetter soils) than those of LPRM for much of the globe. In regions such as southeastern China, Russia, Scandinavia and Bolivia/Matto Grosso, there is significantly more spatial variability in the mean LPRM retrievals than in the MT-DCA ones. Over much of the Amazon and Congo basins, LPRM retrievals are not valid because the MPDI is not sufficiently large (Meesters et al., 2005). There are a few pixels in the Amazon where the average LPRM k is much higher than for MT-DCA, but this is probably a mixture of higher LPRM retrievals and the fact that LPRM retrievals are not valid during many times of the year, so that the two averages may represent different seasonal cycles. When 132 I 3 2 1 0 I 0.5 -1 -2 0- -3 Figure 5-9: Mean difference between the two sets of k retrievals the LPRM algorithm is applied to Aquarius with albedo values from MT-DCA instead of the LPRM assumed value of 0.06 (Owe et al., 2008), the difference in average k reduces to near-zero values for the vast majority of pixels (not shown), suggesting that the ability to retrieve albedo is an important component of the MT-DCA algorithm for soil moisture retrievals. The effect of the albedo assumption is smaller for VOD retrievals than for k retrievals, but also significant. I I130 20 10 0 Figure 5-10: Global maps of temporal mean k retrieval for Aquarius (left) and SMOS (right) Time series of MT-DCA and LPRM k retrievals over the focus pixels for the study period are shown in Fig. 5-11. The temporal dynamics of retrievals from both algorithms are very similar, though there is a slight bias between the two in many cases. In the Amazon pixel, the seasonal cycle in soil dielectric constant is 133 considerably larger than that of VOD (Figure 5-5), consistent with the fact that forests in this region access deep stores of groundwater (Baker et al., 2009). Long drydowns in West Africa and Nordeste occur more slowly for MT-DCA than for LPRM, though it is difficult to say which is more accurate at these large scales. --Amazon - - Central Asia -- 4 ____20 UMRanWaI-.LPRMA k -(T Ok 3015 30 20 0 110- 110 - 10 5i1 10-20 10 01112 3O 10 20 30 15 20- 2 10 for 01/14 01/13 01/12 scale over focus-pixels.-Notethe-dWest Afit 20 402,-4 01/14 01/13 01/14 01/13 01/12 30 220- -40 20SMAPEx .4 15 ----- 15L 01/14 01/13 -Pampas 20 - 20 3 0_ 01/14 01/13 01/12 -40 1530 01/12 20 0114 01/13 01/12 20 r 40O - Amao-s-r4 15 s 30 10 0 0112 01/13 01/14 Figure 5-11: Time series of weekly mean MT-DCA k, LPRM k, and precipitation over focus pixels. Note the different axes scale for the Amazon series. 5.7 Discussion and Conclusions A new method, the Multi-Temporal Dual Channel Algorithm (MT-DCA), is proposed to retrieve microwave vegetation optical depth (VOD), the effective single-scattering albedo w, and soil dielectric constant (monotonically related to soil moisture) from time-series of dual-polarized L-band radiometer observations. It is applied to three years of Aquarius data at L-band. The algorithm relies on the premise that vegetation 134 changes more slowly than soil moisture. A moving average window is used to combine observations from two overpasses while retrieving only a single constant value of VOD alongside the dielectric constant for each of the two overpasses. Single-scattering albedo is assumed to be constant in time and optimized separately across the full record of observations. Note that when soil moisture conditions are similar during the two overpasses, observations from the second overpass do not provide additional information and can lead to noisy retrievals. For Aquarius applications, using 3 overpasses or more leads to a moving window size of 21 days or more, over which the assumption of constant vegetation may not hold. For other satellites with a more frequent revisit time, using a slightly longer window can increase the chance that soil moisture conditions change significantly during the moving window time period, adding additional information to the measurements. Future work is needed to investigate whether such variations can improve the quality of MT-DCA retrievals for other satellites such as SMOS and SMAP, whose more frequent revisit time is likely to lead to improved performance of MT-DCA. Another limitation of the MTDCA algorithm is that it cannot capture sudden changes in VOD such as those that might be induced by fire or deforestation during the time when the moving window passes over the destruction event. However, such an event may still be detectable with some delay when there is a large change in VOD over a small number of moving windows. The proposed multi-temporal algorithm retrieves microwave scattering albedo alongside VOD and soil dielectric constant rather than requiring an a priori assumption on its value based on land cover as is commonly done. This not only improves the quality of the VOD and k retrievals, but also allows simultaneous retrieval of the effective single-scattering albedo. Albedo is assumed to be constant in time, consistent with prior literature and the fact that it is strongly dependent on canopy architecture. With the exception of the retrievals in Rahmoune et al. (2013), which are limited to forests, albedo has previously only been determined using airborne or tower-based field campaigns with limited spatial coverage. The retrieved values of single-scattering albedo are low, but non-zero. The albedo is largest across land cover 135 conditions with significant woody components. The forested regions in South America are found to have a lower albedo than those of the Northern Hemisphere, consistent with Rahmoune et al. (2013). In almost 95% of cases, the retrieved albedo value is lower than the value that would have been assigned by the land cover-based parameterization in the SMAP ATBD (O'Neill et al., 2012). There are large variations in albedo even within a given land cover class, suggesting the common assumption of a land-cover based albedo value is rather poor. Indeed, for the closed shrublands, savannas, grasslands, croplands, and cropland / natural vegetation mosaic classes, the variability of albedo across land cover is larger than its average value. When LPRM is run with albedo values retrieved from the MT-DCA, much of the difference in dielectric constant retrievals disappears. Although it is difficult to validate the retrievals in this paper, given the likely errors in prior assumptions of albedo, the significant effect of albedo on k retrievals suggests that the ability of MT-DCA applications to retrieve it alongside other variables is a major advantage. Global VOD spatial patterns from MT-DCA correspond well to those obtained using LPRM on average, but differ significantly in dynamic range and temporal behavior. The new VOD estimates have temporal dynamics that are consistent with precipitation and prolonged dry-downs, and both canopy water retention and biomass drying and wilting processes are reflected in the VOD retrievals. This confirms that VOD is a vegetation indicator that can be complementary to well known visible infrared indices such as NDVI or LAI. The MT-DCA Aquarius retrievals of VOD show significantly less high-frequency temporal variability due to noise than those from LPRM applied to Aquarius brightness temperature data. Additionally, because of the fixed relationship between VOD and k assumed by LPRM in the presence of noisy observations, LPRM retrievals lead to unphysical predictions of negative VOD or invalid retrievals in 16% of cases. The MT-DCA does not suffer from this problem. At higher frequencies, brightness temperatures are less sensitive to soil moisture and have less penetration through the canopy than at L-band. Nevertheless, retrieval algorithms using temporal snapshots are still prone to measurement noise. Using a method with timeseries such as the one proposed in this paper may also lead to 136 some improvements in VOD retrieval accuracy at C, X and Ku-bands. The VOD represents an integrated value over the canopy, weighted depending on the rate of attenuation of the signal through the canopy. This attenuation rate is dependent on both the frequency band of the measurement and the canopy properties themselves. As a result, the canopy properties and height ranges that dominate the signal at a given frequency band vary by location. This complicates interpretation of VOD data, as water stress in both branches and leaves as well as canopy structure (which affects the b-parameter) both vary with canopy height. Additional work is needed to better understand the effect of measurement frequency on VOD retrievals and interpretation. Overall, VOD estimates obtained using the proposed algorithm have temporal dynamics that are generally consistent with precipitation and prolonged dry-downs. Both canopy water retention and biomass drying and wilting processes are reflected in the VOD retrievals. This suggests optical-infrared indices such as NDVI or LAI, which are not directly sensitive to vegetation water content, are a poor basis for VOD in soil moisture retrieval algorithms. Indeed, other studies of VOD (Jones et al., 2011; Lawrence et al., 2014) have found that the end-of-season VOD variability is distinct from the signal found in optical-infrared vegetation indices. By contrast, a comparison of MT-DCA VOD with collocated radar observations (JHV) shows there is a strong relationship between the two microwave measures of vegetation. The UHV is a better predictor of VOD than RVI (Figure 5-7). This may be relevant to the design of multi-resolution active-passive vegetation retrievals from upcoming SMAP data. SMAP plans to use Normalized Difference Vegetation Index (NDVI) climatology from MODIS as ancillary dataset for VOD in its baseline soil moisture retrieval algorithm. With the proposed algorithm, SMAP independent VOD retrievals could be used to improve SMAP soil moisture retrievals. This study further suggests that cross-polarized backscatter signals, which will be gathered at higher spatial resolution than radiometric measurements by SMAP, can carry information that may be helpful in providing effective VOD over the SMAP radiometer footprint based on information from the higher spatial resolution radar. The proposed method could be applied to SMOS and SMAP L-band data to better quantify the contribution of the vegetation 137 to total emissivity and therefore improve soil moisture retrievals. 138 Chapter 6 On the Seasonal Behavior of Microwave Vegetation Indices 6.1 Introduction Active (radar) and passive (radiometer) measurements are sensitive to both soil moisture and vegetation. In both cases, vegetation scattering and attenuation is sensitive to both the amount of water contained within the vegetation and to structural factors (e.g. shape and size distributions of different components). In past studies employing microwave vegetation indices, active cross-polarized backscatter measurements have generally been interpreted as sensitive to biomass (Le Toan et al., 1992; Mitchard et al., 2009; Englhart et al., 2011) or to both biomass and water content (Asefi-Najafabady and Saatchi, 2013; Saatchi et al., 2012; Woodhouse et al., 2012). Passive VOD measurements have been previously interpreted as sensitive to either total biomass (Liu et al., 2015; Guan et al., 2014) or plant water content (Zhou et al., 2014; Andela et al., 2013; Poulter et al., 2014). The two are related - as aboveground biomass increases, total aboveground vegetation water content does too. Under a given vegetation cover, both active and passive measurements are about equally sensitive to soil moisture (e.g. Du et al., 2000). The differential sensitivity of active and passive measurements to the vegetation properties themselves is less well-studied. Radar backscattering coefficients increase with biomass initially but 139 saturate at high biomass values. At least over forests, model simulations have shown that the emissivity signal saturates less rapidly than the backscattering coefficients do relative to biomass (for some assumed deciduous forest structural/allometric set of properties) (Ferrazzoli and Guerriero, 1996). By contrast, based on case studies across a 200 latitude transect and using clustering approaches for land cover detection, Prigent et al. (2001) concluded that passive measurements are less sensitive to vegetation than active measurements (without the ability to specify relative to what exact parameter). A better understanding of the similarities and differences between the behavior of active and passive measurements is needed to be able to determine which measurement type is preferable for different ecological applications. The derivation of a new VOD dataset from Aquarius in Chapter 5 creates, for the first time, coincident active and passive vegetation indices. This enables a direct comparison of vegetation indices derived from each measurement type without contamination from sensor, frequency, or orbital differences between different active-only or passive-only satellites. In this chapter, the temporal dynamics of active and passive vegetation indices from Aquarius are compared to test whether any systematic differences occur. Based on the results in Chapter 5, OHV is used as an active vegetation index rather than the Radar Vegetation Index, which is susceptible to contamination from soil moisture. 6.2 VOD and c-HV are out of phase in several regions As shown in Figure 5.7, when compared across all locations and times, VOD and UHV are positively correlated. However, when using global scatterplots such as in Figure 5.7 (and such as also often made between VOD and optical indices, e.g. Liu et al. (2011)) it is difficult to separate measurement noise from true differences in dynamic behavior between related but substantially different factors (e.g. leaf greenness compared to biomass). Here, we explicitly compare the temporal dynamics of the two indices. Figure 6-1 shows the correlation between temporal fluctuations of VOD and O-HV at each pixel. As expected, there are several regions of high temporal 140 1 50 0.5 0 0 - -50- -150 -100 50 0 -50 100 150 Figure 6-1: Pearson correlation coefficient between uHV and VOD for all global pixels for which there are at least 50 valid weekly VOD retrievals in three years. correlation. These are particularly common in highly water-limited regions such as the Sahel, Southwestern US, Australia, and Southern Africa. This is consistent with expectations - in more water-limited regions, the vegetation behavior is more likely to respond to seasonal variations in water availability, creating a stronger seasonal cycle that is detected in both VOD and uHV. Perhaps most striking are the re- gions of highly negative correlation near Angola and Zambia and in parts of Brazil. The high magnitude of the negative correlation there suggests that noise alone is unlikely to be responsible for the differences in behavior between the two microwave vegetation indices in these regions - they must be sensing at least partially different canopy properties. Figures 6-2 and 6-3 show the correlation between each of OHV or VOD and precipitation for areas where gHV and VOD show very different behavior (RUHV-VOD < -0.5). In most of these regions, 9HV approximately follows the sea- sonal cycle of precipitation, but VOD does not. Within the abovementioned regions of coherent spatial patterns, two types of behavior occur. In Africa and Eastern Brazil, oHV closely tracks precipitation while VOD does not. In the western part of the Brazilian region of interest, neither microwave vegetation index tracks precipitation 141 particularly closely, but neither are they strongly out of phase with precipitation. 1 - ~M 50 0.5 r. 0 A V -50k -150 -100 -50 0 Figure 6-2: Pearson correlation coefficient between with R.HV-VOD < -0.5. 50 0 100 -0.5 -1 150 -HV and precipitation for pixels P1 S-P2 P3 SOP50-_ 1 0.5 -1 0 0 -0.5 -50k -150 -100 U OU I1UU I OU -1 Figure 6-3: Pearson correlation coefficient between VOD and precipitation for pixels with RHV-VOD < -0.5. . Black symbols represent the locations of the pixels shown in Figure 6-4. 142 Sample timeseries for one representative pixel in each of the three regions mentioned above are shown in Figure 6-4. The pixel locations are marked in Figure 6-3. To further understand the differences in signal between the two microwave indices, they are also compared to timeseries of the Enhanced Vegetation Index (EVI) (Huete et al., 1994). The EVI is a version of the commonly used Normalized Difference Vegetation Index (NDVI), modified to reduce atmospheric corrections and to avoid the saturation evident in NDVI datasets. The NDVI is a normalized ratio of the nearinfrared and red spectral-band reflectances ( Tucker, 1979). This ratio is proportional to chlorophyll abundance in the plant canopy (Sellers, 1985; Myneni et al., 1995) and thus correlates well with photosynthetic capacity. Like all optical data sources, EVI is susceptible to the presence of clouds. Widely distributed EVI products from MODIS represent 8-day periods to ensure the occurrence of cloud-free days and reduce noise in the measurements. At the coarse resolutions of Aquarius, it is very rare for an entire 0.900 footprint to be covered by clouds. Furthermore, the native MODIS resolution (250 m) is several orders of magnitude higher than that of Aquarius. Instead of using the standard product with high spatial resolution and coarse temporal resolution, the raw MODIS reflectances are used to calculate a daily EVI at coarse spatial resolution and produced at the same footprint gridding system used for the Aquarius data. For P3, EVI data are often missing due to the high frequency of cloudy conditions during the wet season and are not included in Figure 6-4. The timeseries are consistent with the patterns expected based on the correlation coefficients in Figures 6-1-6-3. In each of the regions, there is a strong dry season, which is most pronounced for P1 (top row). For each of the three pixels, approximately the same seasonal pattern occurs across the three year record, although the start of the increase in VOD occurs a little later in the wet season for the second year of P3 than for the other two years. Near the transition from wet to dry season the VOD increases, coming to a maximum value within the dry season before declining again after the start of the wet season. For P1, the VOD starts to decrease at the very start of the wet season, during the first weeks of low rain. For P2 and P3, it does not decrease until later in the wet season. Generally, the seasonal increase in VOD is 143 relatively rapid at first and then increases more slowly. By contrast, the aHV starts to decline as the wet season transitions to the dry season, and does not increase again until the rainfall increases again. For P1, the decline at the beginning of the dry season gives way to a period of near-constant orHV until the subsequent wet season, even as EVI continues to decline. For P2 (center), both EVI and -HV decline slightly over the entire duration of the dry season. For P3, rainfall, VOD, and OHV are all out of phase with each other, and the timeseries (especially that of JHV) appear to be a lot noisier than the analogous ones for P1 and P2. This is consistent with a rapid response to rainfall in the dry season, given the fact that the P3 dry season has more rain events than those of P2 and P1. 6.3 Possible explanations The large differences between the seasonal cycle of VOD and that of OrHV and precipitation over several regions is surprising. While there might be differences in the sensitivities of active and passive measurements that cause differences in their dynamics, there are no clear reasons why VOD should show related but opposite behavior to precipitation, aHV, or EVI. Several hypotheses to describe these observations are described below. 6.3.1 Dry season bud break and leaf flushing For many tropical dry ecosystems, the emergence of new leaves and shoots has been observed during the dry season (e.g. Singh and Singh, 1992; Eamus, 1999; Hutyra et al., 2007; Holbrook et al., 1995). The VOD measurements may be sensitive to this dry season biomass growth, as also previously theorized about Ku-band backscatter in African evergreen forests near the Equator (Guan et al., 2013). The difference in behavior between VOD and a-HV might be explained by their differing sensitive to leafy and woody biomass. Using the MIMICS radiative transfer model ( Ulaby et al., 1990) parameterized to represent trembling aspen (a deciduous species) at L-band, Steele-Dunne et al. (2012) found that aHV was about equally sensitive to 144 -0.06 0.6 a 0.40 ,-0.04 .- -0.02 0.2- 0 Sep11 Sep12 Jan12 Jan13 ,O Sep13 Mrainfall Jan14 -VOD 0.61 0.04 -EVI 0.03 - cHV Y r= 0.4 0 0.02 0 7 0.2 -0.01 .1 .M - QL Sep11 1.' > Jan12 I Sep12 Jan13 Sep 1 3 I Jan14 - [I 0 0 0 0.07 0.065 C 0.06 0 0.055 -S 2Se p12 Sep13 Jan1 3 Jan14 0.05 Figure 6-4: Timeseries of rainfall and vegetation indices for P1 (top), P2 (center), and P3 (bottom), respectively. The P1-P3 locations are marked in Figure 6-3. For each timeseries, the average weekly VOD (blue line), JHV in units of power (green line), EVI (red line), and rainfall (histogram) are shown. 145 branch gravimetric water content as to leaf gravimetric water content (their figure 9f). For passive measurements, however, using a field theory based model, Ferrazzoli and Guerriero (1996) found that the effect of leaves on the total emission is relatively small at L-band. Additionally, cylinder-like shapes are predicted to be primarily absorbers, increasing VOD, while more disc-like shapes like leaves can decrease emissivity and VOD (Ferrazzoli and Guerriero, 1996) (consistent with observations over agricultural fields at higher frequencies (Ferrazzoli et al., 1995). These findings are consistent with a dry season loss of leaves (which are more disc-like) and leaf flushing during the wet season, along with the occurrence of the growth of new shoots or buds (more cylindrical) during the dry season. For passive measurements, the effect of the leaf loss is relatively small but shoot growth and bud break increase VOD during the dry season. Since aHV is more sensitive to leaves, the loss of leaves during the dry season may dominate its overall seasonal variability, explaining why c-HV decreases during the dry season. The attribution of the increase in VOD to bud break and shoot growth are also consistent with the dry season decrease in EVI, which is primarily sensitive to chlorophyll not visible until leaves open. There are some unanswered questions about this possible explanation. The expected timing of bud break and leaf fall are only partially consistent with the seasonal dynamics of VOD - each of P1, P2, and P3, shows a decline in VOD at the start of the wet season, but no such early leaf-fall has been observed using ground-based measurements. For P1, the timing of the VOD increase and decrease is particularly closely linked with the wet/dry season transition. Of the 9 pixel-years shown in Figure 6-4, the dry season in 2013 for P3 is the only one where VOD increases the most rapidly in the middle of the dry season rather than at the very beginning. By contrast, many of the species that flower or show bud break during the dry season do so in the middle of the dry season (Holbrook et al., 1995). Similarly, leaf abscission is generally observed during the dry season, while declines in observed VOD in the pixels of Figure 6-4 and other similar locations tend to start during the wet season only. Attributing seasonal fluctuations in VOD and O-HV to changes in biomass only is 146 also at odds with the fact that the largest dry season increase in VOD occurs in P1 (difference of about 0.3 nepers instead of 0.1-0.2 nepers), both in absolute terms and relative to its average VOD value. One would generally expect to see the greatest seasonal variability in forested areas and less so in more grassy areas. While P2 and P3 are evergreen forests and woody savannas, respectively, the land cover in P1 is more grassy - based on optical data, the woody fraction of the P1 pixel (a savanna) is only around 0.5 (Guan et al., 2014). This hypothesis is therefore most likely to apply to P2 and P3 - and the regions they represent - than to P1, whose seasonal cycle also shows the clearest correlation with rainfall occurrence (and thus the most conflict with mid-season changes in phenology observed in the ecological literature). 6.3.2 Litter Additional contributions could be coming from wet litter in the dry season. In temperate regions, litter has previously been shown to increase L-band emissivity (Kurum et al., 2012b; Grant et al., 2009) and decrease L-band backscattering (De Roo et al., 1991). This suggests that litter is present for much of the dry season in these regions and might therefore be decomposing relatively slowly. Such an observation is consistent with previous ground-based studies that have observed reductions in litterfall decomposition rates during the dry season ( Wieder and Wright, 1995) in Panama, but would be the first large-scale observation of this effect across large areas of dry tropical regions. 6.4 Conclusions As shown in Chapter 6, YHV and VOD are positively related across much of the globe. The Pearson correlation coefficient was calculated for each pixel in order to compare the temporal dynamics of VOD and UHV. The regions with the highest correlations between VOD and UHV are often semi-arid, consistent with the fact that semi-arid climates are expected to have the strongest seasonal cycle in vegetation water content. In several dry tropical regions, the seasonal dynamics of VOD and JHV show 147 a strong negative correlation. This is counter to expectations, but the strength of the negative correlation suggests it cannot be attributed to noise and may provide new information about ecosystem behavior in these regions. In particular, in the Angola/Zambia region and in the southwestern regions of Brazil, VOD increases during the dry season but aHV does not. This may be due to the effect of budbreak and shoot growth during the dry season that is causing the increase in VOD, coupled with a loss of leaves causing the decrease in CHV. The occurrence of dry season litterfall is also consistent with this seasonal behavior. Neither of these two behaviors has previously been observed at large scales. They are largely consistent with ground-based observations, although questions remain about why the increase in VOD starts closer to the start of the dry season rather than towards the middle, which is usually observed in the ecological literature. While these are possible explanations for the observed signal in the South American regions, they cannot provide the sole explanation for the observed signals in Zambia, which has less woody vegetation. Additional research is needed to validate the hypotheses of dry season bud break and litterfall as explanations for the differing seasonal dynamics of VOD and UHV in the South American regions, and to establish the relative contributions of each of the two hypothesis to the total signal in regions where they are out of phase. Additional research is also needed to determine why the behavior shown in these regions is not shown elsewhere in woody regions with significant dry seasons. Nevertheless, the magnitude of the behavior shown in Figure 6-4 strongly suggests that active and passive vegetation indices can provide complementary information that could lead to new insights about the seasonal ecological variations of tropical dry ecosystems. 148 Chapter 7 Variations in Diurnal Canopy Water Content Refilling with Water Stress 7.1 Introduction In Chapter 5, it was assumed that the vegetation properties that affect microwavefrequency measurements (vegetation water content and canopy structure) are approximately constant over the time span of two overpasses, i.e. over an 8 day window for the Aquarius satellite. The validity of this assumption depends on the fact that each of the satellite overpasses occurs at approximately the same local time of day. Vegetation water content follows a strong diurnal cycle (e.g. Brodribb and Holbrook (2004)) that is much larger than the expected day-to-day variability at a given time of day over the span of a few days. This variability occurs because water is lost at the leaves, but replenished through the roots - the water needs time to move from the roots to the canopy. Thus, the diurnal cycle of the stored vegetation water content represents the integral of the difference between transpiration and root uptake rates. The canopy water storage is maximized in the early morning and minimized in the early afternoon. The process of replenishing water lost to transpiration is often termed refilling. Several studies have found that the refilling time seems to increase with water stress. That is, under relatively drier conditions, water moves more slowly across the vegetation. For example, in dry oak trees at the Tonzi Ranch (near the Vaira Ranch site 149 studied in Chapters 3 and 4, Fisher et al. (2007) found that during the dry summer season the proportion of nighttime sap flux increased. Similar patterns have been observed in a freestanding C 3 - CAM tree species (for which CAM photosynthesis is induced by drought) (Herrera et al., 2008), and in laboratory measurements of young oak trees (Ehrenberger et al., 2012). Insofar as nocturnal refilling (like daytime refilling) is driven by gradients in water potential between the roots and the leaves (e.g. (Cavender-Bares et al., 2007)), low soil moisture may reduce these gradients. Additionally, drought stress often leads to reductions in xylem hydraulic conductivity due to embolism (Sperry and Tyree, 1988; Tyree and Sperry, 1989; Sperry et al., 2002). However, it is not clear how often these effects are significant, how widespread this phenomenon is, or how much it varies at the stand level or larger scales. Observationdriven studies are challenging because direct measurements of total vegetation water content are difficult. In analogy with gravimetric measurements for measuring soil moisture, plant samples weighed before and after drying can allow calculation of component water content, but this requires destructive sampling. Dendrometers can be used to measure stem water content through radius changes (Zweifel et al., 2001, 2006) and turgor pressure probes can be used to measure leaf water potential (a function of leaf water content) (Zimmermann et al., 2008), but each of these requires speciesspecific calibration, and neither captures variations throughout the canopy (Bohrer et al., 2005). While sap flux measurements are comparatively easy, they require a complicated hydraulic flow model to scale across the plant, requiring many speciesspecific parameters. Although there has been a concerted effort to create databases of plant hydraulic traits (Kattge et al., 2011; Medlyn et al., 2011) and multi-species meta-analyses are becoming more common (e.g. (Manzoni et al., 2011)), knowledge of these parameters is still limited to a comparatively small number of species. A larger observational dataset of diurnal variability in vegetation water content may help increase understanding of variations in refilling time and size. Given sufficient care, refilling time can be studied by comparing vegetation water content measured by microwave remote sensors between morning and evening overpasses. Friesen et al. (2012) observed diurnal differences in C-band backscatter from the European 150 Remote Sensing (ERS) ERS-1 and ERS-2 satellites that could not be attributable to soil moisture or azimuthal difference over several regions in the globe. Steele-Dunne et al. (2012) studied these differences over a single pixel in West Africa and noted that the diurnal difference increased as the dry season progressed, eventually becoming smaller again near the end of the dry season and disappearing during the rainy season. Since the low-earth orbit of ERS causes morning and evening overpasses to occur at approximately the same local time, such an increase (decrease) in diurnal variability could imply an increase (decrease) in refilling speed, since drier conditions would normally lead to lower transpirational fluxes and a reduced refilling amount. However, in both studies, diurnal differences were assessed by aggregating all morning and evening overpasses in a given month and location and comparing their distributions. Since morning and evening overpasses occurred on different dates, this provides only a limited capacity to probe the drivers of diurnal variability in backscattering coefficients. Using the Michigan Microwave Canopy Scattering Model (MIMICS) ( Ulaby et al., 1990), Steele-Dunne et al. (2012) also showed that cross-polarized backscattering coefficients at L-band are primarily sensitive to leaf vegetation water content. In this chapter, we compare same-day morning and evening cross-polarized backscattering coefficients measured by the L-band Aquarius satellite in order to determine whether a) diurnal differences in vegetation water content can be observed at L-band and b) whether the refilling of vegetation water content depends on water availability. The morning overpasses (at 6:00 AM) is interpreted as occurring after refilling is complete but before the onset of transpiration causes a loss in aboveground canopy water content, while refilling may only be partially complete by the time of the 6:00 PM overpass. 7.2 Derivation of AM and PM VWC As mentioned in Chapter 5, the Aquarius orbit is not an exactly repeating orbit. Roughly every 7 days, the Aquarius orbit passes over a previously imaged area, but 151 the footprint centers are usually around 0.100 - 0.25' apart. This is significant relative to the 0.80' radius of the footprint. Both refilling amounts (e.g. average evapotranspiration and uptake rates under a given set of conditions) and other factors may vary in space, so this chapter departs from the gridding system previously used in Chapter 5 and 6 (and originated in McColl et al. (2014) and Piles et al. (2015)). Pairs of coincident AM and PM observations are formed if their footprint centers are less than 0.05' apart. This is a conservative requirement. Assuming a footprint radius of 0.90' (the exact radius is latitude-dependent), this implies that 92% of the area covered by the two footprints overlaps. Thus, in order for spatial variability alone to cause a diurnal difference in backscatter AUHV > 0.5 dB, there would have to be a difference in the diurnal average c-pQ of the non-overlapping areas at least 3.33 dB in size. However, such a constraint means that pairs from different days rarely get close to overlapping. An additional constraint of day-to-day drift less than 0.30' is used, implying that the center 0.25' by 0.25' of a grid cell is covered. Unfortunately, coincident AM and PM observations occur in only a relatively small set of latitude bands in the Aquarius orbit. Using the restrictions above, only a handful of footprints remain. They all occur at a latitude of 11.37' N. Their location is shown in Figure 7-1. Fortunately, the region represents a natural gradient of land cover and precipitation conditions. 7.3 Soil moisture dependence of diurnal variability of 6THV Refilling is not the only factor that can create differences between morning and evening values of UHV. Ascending and descending overpasses view the same location from a different azimuth angle. Although many flat natural surfaces have azimuthal symmetry, azimuthal effects can be important in areas of significant topography where azimuthal angles can influence the apparent incidence angle induced by the surface slope or whether a surface is shaded (e.g. (Schuler et al., 1999). Faraday rotation 152 Figure 7-1: Locations with same-day morning and afternoon observations whose footprints are less than 0.050 apart. may vary in time, and oHV is particularly sensitive to the quality of the applied Faraday rotation correction (Freeman and Saatchi, 2004). Although dew has previously been found to have a negligible effect on L-band observations over wheat (Gillespie et al., 1990), it may be more significant for other land cover types. The presence of intercepted water on leaves shortly after rainfall may also obscure the signal. Lastly, although the UHV is expected to be almost independent of soil moisture in the crosspolarization (Steele-Dunne et al., 2012; Ulaby and Long, 2014), some small influence may remain. There is at least a small correlation between soil moisture and UHV over much of the globe, although it is fundamentally impossible to determine whether this is due to a direct scattering influence of soil moisture or due to a correlation about soil moisture and overall vegetation water content without additional datasets. In order to better understand whether the diurnal AcHV can be attributed to changes in refilling, backscatter differences are compared to morning soil moisture derived from coincident Aquarius radiometer measurements (Bindlish et al., 2015). If changes in refilling speed consistent with (Steele-Dunne et al., 2012) are responsible, AUHV is expected to decrease with soil moisture. Days when the occurrence of dew is a possibility are conservatively filtered by removing days when the 6:00 AM 153 relative humidity is greater than 98% (calculated based on 2-m humidity and surface temperature data from the Modern Era Retrospective Reanalysis - Land (MERRALand) (Reichle et al., 2011). Similarly, combinations where rainfall between the two overpasses or in the six hours before the morning overpass exceeded 1 mm are also filtered. The rainfall data used are from the SMAP 'Nature Run', which is also based on MERRA-Land but has additional precipitation corrections applied (Reichle and Liu, 2014). All ancillary datasets are converted to the same grid. After such conservative filtering, the remaining pixels occur predominantly in the dry season. Locations with more than 5% open water content or significant topography are removed from the analysis entirely. Figure 7-2 shows the average diurnal variability as a function of soil moisture for the remaining pixels. The scatter in each of the subplots of Figure 7-2 is likely due to day-to-day variations in evaporative demand (e.g. humidity and net radiation) or other factors. For most pixels, AUHV is positive (i.e. the backscattering coefficients are higher in the morning) - parts of the Kp-based uncertainty range are negative for only a small minority of points. This is consistent with the theory of an incomplete refilling amount by the 6:00 PM time of the evening overpass. However, for most pixels, if any trend is detectable at all, the diurnal difference generally increases with soil moisture. This is counter to the observation in Steele-Dunne et al. (2012) that backscattering coefficients increased over the course of the dry season, and counter to the explanation that this may be happening due to reduced refilling speeds. Alternatively, trends in diurnal differences of vegetation water content may be dominated by the effect of diurnal differences in soil moisture and their direct influence on CHV. Unlike vegetation water refilling, the diurnal difference in soil moisture is expected to increase with the morning soil moisture value - wetter soils have higher drainage and evapotranspiration rates. Interestingly, the only pixel that shows the expected negative relationship between morning soil moisture and AUHV is also the driest pixel. It has an average precipitation over the three-year study period of 348 mm/yr and is located in Northeastern Nigeria. The location studied by (Steele-Dunne et al., 2012) is further North and even 154 I 1 I 4 0.5 0.51 2 C - ----- --- ---------------- 0 - - 0 NGsnea(11T7N. 1062W) Doianat LC: Sawarma, GSU 0.55 A Rit 88.1 nun 0 0.1 0.3 0.2 Soi moisture SE Clwd(1137N, 2W63E) DomrWnu LC: Savana, GSI: 0.75 Arxef Ran 501 nun SMUn {11.37N,.7.33hY D0mrnrt LC: Sawama, GSt 0.54 A'u 04 -05 RUr:59 .O8nmm ' -0.5 ------------ 0 0.1 0 0.3 0.2 Soil moisture 0.4 -0.5 0 0.1 0.2 0.3 0.4 Soil moisture 0.5 #f 0 --------------------- 'I 0.5 1 U 0.5 1 1-111p f -41 ----------S Sudan (1M7, 31.11E) DonaiMLC Crqand, GS. 0.71 AnmuO -0.51 0 ---------------- RaW. 45a4.8 m 0.1 0.3 02 Sod moisture SE Niger (t7,3.14W) SWChrad411 37h, 17.12E) Da*MT LC: Mosaic, GSL 0.6 Amusm Rat Sea6nun j 0.4 0 0.1 0.2 0.3 Sod moisture DWint LC: Maic, GSI: 0.64 Annum Ran: 353.77 mn 0.4 -0.5. 0 0.1 0.2 0.3 Soil moisture 0.4 15 1 0.5 f 0----------------- NENIgera(1t37%.13i3E) Donant LC MOLac, GSL 0.2 Annum Ra31 347 78uam -0.5 0 0.1 0.2 0.3 0.4 SoA moisture Figure 7-2: Difference between AM and PM UHV (in dB) vs. radiometer-derived AM soil moisture for the pixels in Figure 7-1 after filtering for dew and interception. Errorbars on the differences are based on the Aquarius Kp uncertainty estimates of the backscattering cross-sections, where additive errors are assumed to be distributed normally with a standard deviation of KUpaQ. The inset of each pixel shows the location (country, center coordinates), dominant land-cover, Gini-Simpson index of land cover, and annual average rainfall, respectively. Pixels are arranged in order of decreasing averaging rainfall. Pixels where less than three samples remained after dew and interception filtering, or where more than 0.05% of the covered land surface area was water, are removed from this analysis. 155 drier (average precipitation of about 250 mm/yr). This is consistent with evidence from Ehrenberger et al. (2012) that the slow-down in refilling increases super-linearly with drought stress - e.g. as the drought stress becomes more extreme, the (negative) derivative of the refilling rate increases in magnitude. However, it must be emphasized that there are only a handful pixels tested in this analysis, and that the Northeastern Nigeria pixel itself has only a small number of coincident diurnal pairs remaining after filtering. Furthermore, the non-linear relationship between soil moisture and UHV may also influence these results. R2 = -0.73 0.8- Ug0.7 0.60.5- 0.4 - 0 60.30.2 ' 0.1 _0 200 400 600 Avg annual precip [mm] 800 1000 Figure 7-3: Mean diurnal difference of GHV (AM - PM) across days without interception or dew formation (in dB) as a function of average annual rainfall for each pixel. 7.4 Conclusions Across the gradient of locations shown in Figure 7-1, only limited evidence was found for the dependence of diurnal variability on soil water availability observed by (SteeleDunne et al., 2012). While there was a diurnal decrease in cHv detected in the dry season across a gradient of pixels in super-equitorial Africa, this gradient increased with soil moisture for most of the pixels, suggesting that either the behavior of UHV 156 was influenced by factors other than vegetation water content (e.g. diurnal soil moisture variability) or relative refilling over a 12-hour daytime period actually increases with increasing water stress. That is, the effect of reduced transpiration rates under increasingly dry conditions (causing less water to need to be refilled) may outweigh the reduced speed in refilling, if such a reduction takes place. For the driest location studied, the trend between -HV and soil moisture was reversed, suggesting that water stress slows down the refilling process only in the driest of environments. However, note that the difference in average annual precipitation (calculated over the three years of the Aquarius record) between the driest and second-driest pixel in Figure 7-2 is less than 10 mm, suggesting these results should still be viewed with caution. All in all, it is difficult to draw conclusive interpretations from the small number of points studied across the Aquarius gradient. Certainly, seven pixels cannot provide conclusive evidence, no matter what they show. However, with the recent launch of the Soil Moisture Active Passive satellite (Entekhabi et al., 2010a) on January 31st, 2015, a new L-band radar dataset has become available. Because of the much higher swath width and better integration of the instrument integration time with the orbit, SMAP radar data do have diurnally coincident observations over most of the globe. Furthermore, at a spatial resolution of 3 km instead of -90 km, the spatial resolution is almost two orders of magnitude higher for SMAP than for Aquarius. As of the time of this writing, the SMAP radar record was interrupted after about 2.5 months (likely due to the high ionospheric activity affecting the instrument) and may only be relatively short overall. Nevertheless, given the high spatial resolution, it is expected that the SMAP record can lead to insight even over a short timespan. The multiple interpretations of the results in Section 7.2 suggest the limitations of working with o-HV directly. Ideally, a radar scattering model would be used to calculate some measure of vegetation water content directly. However, most radar scattering models, including those of Chapters 3 and 4 (e.g. (Burgin et al., 2012; Truong-LoI et al., 2015)) are highly sensitive to parameters and assumptions describing the structure and density of the vegetation canopy and its components. This makes them very difficult to parameterize, especially across large spatial areas that 157 may contain a variety of vegetation covers. Perhaps the simplest scattering model is the so-called 'cloud model', which represents vegetation as a single layer of water droplets (Attema and Ulaby, 1978). However, even this model relies on three different parameters beyond that describing vegetation water content, as well as additional soil-specific parameters. Alternatively, a new data-driven method for soil moisture retrieval from radar backscattering coefficients without ancillary parameters has recently been derived (Narvekar et al., 2015; Bruscantini et al., 2015). It may be possible to combine this method with a cloud model and timeseries-based approach to derive relative vegetation water content amounts over time. 158 Chapter 8 Conclusions and Future Work 8.1 Conclusions Microwave measurements are sensitive to water content in the soil and plants. As such, they can provide information relevant for monitoring and understanding changes to the distribution and availability of water at the land surface (including both within the soil and within plants). Such information is relevant for predicting changes to vegetation function and mortality under changing climate and land cover. This thesis has discussed the development and use of two new datasets from microwave remote sensing: root-zone soil moisture profiles derived from P-band radar observations and vegetation optical depth derived from L-band radiometer observations. Microwave observations include contributions from both the moist soil and vegetation scattering or emission. Using the microwave observations requires that the soil moisture and vegetation contributions are disentangled through the retrieval process. In doing so, multiple observations can be combined from different incidence angles, frequencies, polarizations, or observation times. Many of these combinations of data are highly correlated and thus contain some degree of duplicate information. As such, additional measurements do not always lead to a full additional degree of freedom in the retrieval process. Chapter 2 has introduced a framework to determine the fractional degrees of freedom contained within a set of measurements, termed the degrees of information. The degrees of information in a set of measurements depends on the 159 shape of its joint probability distribution - duplicate information reduces the spread of the joint probability distribution. The degrees of information can be calculated based on an appropriately normalized version of the information theoretic total correlation measure. It is applicable to data with an arbitrary number of dimensions and sensitive to the amount of duplicate information in the entire distribution without requiring assumptions on the shape of the data. The degrees of information provide an upper bound on the number of parameters that can be robustly retrieved from a set of data without being overly sensitive to small amounts of measurement noise. Chapters 3 and 4 discuss the retrieval of soil moisture profiles from radar observations in the P-band frequency range. Unlike observations made at higher electromagnetic frequencies, they are sensitive to soil moisture across much of the root zone. An observing system simulation experiment (OSSE) is used to test the effect of profile representation on the simulated backscatter. The OSSE consists of a year of hydrologically simulated soil moisture profiles and associated backscattering coefficients representing the Vaira Ranch near lone, California. The soil moisture profile is represented using several homogeneous slab layers. As the number of layers used increases, the forward error from misrepresenting the continuously varying soil moisture profile decreases, although using a more sophisticated moisture-dependent layering system does not significantly reduce the error relative to using a simpler layering systems in which all layers have a constant (in time) and equal depth. The error decreases as more layers are used to represent the profile. Such forward modeling errors affect the cost function of the soil moisture profile retrieval problem and are thus likely to propagate to retrieval errors. For some profiles, if a homogeneous halfspace is used to represent the profile, the equivalent halfspace soil moisture corresponding most closely to the 'observed' backscattering coefficients may be outside the range of the profile all together. This is a result of the phase shift induced by subsurface reflections. Chapter 3 shows that retrieving depth-dependent soil moisture profiles improves the accuracy of the soil moisture retrieval. Thus, representing the depth-variability of the soil moisture profile increases not only the hydrologic utility of the retrieved data, but also reduces the forward modeling error within the retrieval. 160 Using multiple layers to represent the soil moisture profile increases the number of unknowns above the degrees of information contained in P-band observations. Chapter 4 introduces a Tikhonov regularization method to increase the stability of the retrieved soil moisture values. The regularization adds a term to the retrieval cost function that penalizes possible solution profiles that deviate significantly from the expected profile, based on the mean profile and inverse covariance between different layers. This penalizes solutions that are hydrologically unrealistic. The mean profile and inverse covariance matrix can be determined from either simulations or from ground-based observations. The second dataset derived in this thesis is vegetation optical depth (VOD), a measure of canopy attenuation sensitive to vegetation water content. In Chapter 5, multi-temporal observations are used to increase the degrees of information above those contained in a single snapshot of H- and V-polarized brightness temperatures. The multi-temporal approach relies on the premise that, when observed at the same time each morning, the temporal dynamics of vegetation are sufficiently slow that VOD can be assumed constant between two consecutive overpasses. This assumption also allows the retrieval of the effective single-scattering canopy albedo, avoiding retrieval errors due to incorrect prior specifications of this parameter. The method is applied to L-band frequency radiometric observations from the Aquarius satellite. The resulting retrievals show significantly less high-frequency temporal variability (the high frequency variability is likely attributable to noise) than state-of-the art snapshot retrieval methods. The VOD data derived in this thesis provide coincident active and passive measurements from the same observing platform. These were used to compare the seasonal dynamics of passive vegetation indices like VOD and an active index like aHV, both of which are expected to be sensitive to vegetation water content. The highest correlations between VOD and 6-HV occur in regions with semi-arid climates, where the strongest seasonal cycle in vegetation water content is expected. In several dry tropical regions, including the savannas south of Central Africa and several regions in Southwestern Brazil, the two microwave vegetation indices had opposite seasonal 161 dynamics. This may be attributable to dry season bud break and shoot growth occurring simultaneously with or shortly after leaf fall. If so, this would be the first large-scale observation of these processes. Additional research is needed to test this theory. Chapter 7 shows that microwave vegetation indicators can also be used to study the diurnal variability of vegetation water content and refilling. The diurnal variability of Aquarius UHV was compared to soil moisture across a gradient of land cover and precipitation in Africa to test the hypothesis (based on laboratory studies) that refilling speed decreases during times of water stress. Evidence for this was only found at the driest pixel studied, which has an annual precipitation of less than 250 mm/yr. 8.2 8.2.1 Future Work Root-zone soil moisture Chapter 4 introduced a regularization method using the mean and covariance of different profile layers in the retrieval cost function to avoid the retrieval of hydrologically unrealistic profiles. The method was applied to observations over the Harvard Forest in Western Massachusetts and compared to in situ measurements of surface soil moisture and spatial patterns. P-band observations are available at 9 other sites, and the regularization method should be tested on some of these to ensure it is applicable under a wide variety of conditions. In particular, AirMOSS observations over the MOISST site in Oklahoma contain multiple strips that observe the same area using different headings and incidence angles. Regularized retrievals from two same-day observations of the overlapping pixels could be used to test the stability of the retrievals. Additionally, in August 2015, near-simultaneous observations of the Walnut Gulch region in Arizona will be made at both L- and P-band. The top layer soil moisture obtained from regularized retrievals from the P-band data could be compared to L-band retrievals as a further check on the regularization method. Alternatively, the regularization method could be adapted to ingest both L- and P-band observations. 162 This may increase the vertical resolution with which the soil moisture profile can be identified. The method described in Chapter 4 uses smooth-surface reflectivities rather than the backscatter coefficients that are observed by the radar. The smooth-surface reflectivities are derived from the radar observations using a separate method that relies on parameters obtained from running a detailed vegetation-structure-based canopy backscattering model. Such a method is sensitive to the availability of ancillary data and cannot be applied globally. Furthermore, although it allows a dynamic amount of biomass that changes from pixel to pixel, the vegetation-structural model applied is constant across pixels (Truong-LoI et al., 2015). This method uses only the HH-, VVand HV-polarized backscattering coefficients. However, AirMOSS data provide fully polarimetric observations, which include not only the magnitude of the backscattering but also its phase. Polarimetric models (Lee and Pottier, 2009; Yamaguchi et al., 2005) could be used as an alternative approach to account for the effect of roughness and vegetation and retrieve smooth-surface reflectivities. Such models often treat vegetation as consisting of some distribution of (randomly or not) oriented ellipsoids. This reduces the number of unknowns in the vegetation model, allowing them to be retrieved directly from the polarimetric observations even in the absence of ancillary information. The Degrees of Information framework of Chapter 2 can be used to test whether the polarimetric observations contain enough information for such a retrieval. The development of such a retrieval method is already under way at MIT (Alemohammad, 2015). The large scale maps of soil moisture produced by applying the regularized retrieval methods can be used to determine the relative controls of landscape characteristics such as topography, land cover, and soil texture on the spatial variability of soil moisture at different depths. Knowledge of the expected controls on spatial soil moisture patterns can be used in the development and validation of land surface models - testing model performance based on physical realism and spatial patterns rather than simply on their ability to match observations at a given point will reduce equifinality and other issues (Beven, 1993; Refsgaard, 2000). For the purposes of 163 evaluating the regularization algorithm, Figures 4-16 and 4-17 simply correlated sand fraction and elevation with soil moisture. In order to make additional progress in understanding the controls on soil moisture fields at different depths, explicit consideration should be made of the actual spatial patterns in soil moisture. For example, the connectivity of soil moisture patterns (the probability that two points at a given distance apart belong to the same cluster, where clusters can be determined using thresholding) (Western et al., 2001; James and Roulet, 2007) can be compared to the connectivity of landscape characteristics. The consideration of multiple points simultaneously is less sensitive to the effects of lateral flow than the analysis of figures 4-16 and 4-17, and the use of thresholding allows for studying the relationships between soil moisture and texture, land cover, and topography under different regimes. 8.2.2 Vegetation optical depth and water content Chapter 5 developed a multi-temporal algorithm to retrieve vegetation optical depth and single-scattering albedo, reducing errors from the need to specific albedo a priori. The method was tested on radiometric observations from the Aquarius satellite in order to allow comparison with co-incident radar data in Chapter 6. However, the Aquarius observations have a relatively coarse special resolution ( 0.900) and have shown significant calibration bias and drift (Dinnat et al., 2012; Piepmeier et al., 2015). Furthermore, the assumption of constant vegetation optical depth across two consecutive overpasses is likely to be impacted by the relatively long Aquarius revisit time (7-8 days) and the fact that Aquarius does not have a perfectly repeating orbit. The method could also be applied to data from the SMOS (if adapted for multi-angular data) or SMAP satellites, which would avoid many of the above disadvantages and create a less noisy dataset. Furthermore, when the soil moisture does not change over two consecutive overpasses (such as might happen during an extended drydown in the absence of rainfall), the two sets of measurements in the moving window are identical and the second set of observations does not provide any additional information. Under such conditions, it might be useful to dynamically adapt the moving average window length to include three overpasses, increasing the 164 I .. .11, 1 - . 1. _ a .1 - I . . - . 1-~ -1.1 - 1a__-&-&-a" -k"_&. '_ J- - : 11 . " wuwk . - . 1 .1 1 1 1. 111 chance of a change in soil moisture between overpasses. While this is a possibility for the 3-day revisit orbit of SMAP (9 days total in moving window), the constant VOD assumption is questionable for three consecutive overpasses (21 days) of Aquarius. In theory, the VOD dataset is proportional to vegetation water content. Nevertheless, it may not equally reflect vegetation water content in all canopy layers due to attenuation across the canopy. The algorithm of chapter 5 should be applied to radiometric observations at higher frequencies to generate the data necessary to study the effect of frequency on VOD measurements. Data analysis should be supplemented by theoretical studies of expected extinction in multi-level canopies (as in Ferrazzoli et al. (1995), but for a wider variety of canopy types). Plant water stress varies across the canopy, so lower frequency observations that integrate across the canopy may be more representative than higher frequencies. They are also less sensitive to canopy heterogeneity (Matzler, 1994). Higher frequencies, however, may be less sensitive to litter or stem water content, which may be difficult to disentangle from drought-response reductions in vegetation water content, as in Chapter 5. Additional studies are needed to help elucidate which frequency ranges are preferred for studies of vegetation water content behavior. Note that the algorithm developed in Chapter 5 can easily be applied to observations made at higher frequencies. A dataset of VOD observations at higher frequencies is already available (Owe et al., 2008; Liu et al., 2011), but it uses the LPRM retrieval algorithm, which is sensitive to errors in the assumed effective single-scattering albedo and to observational noise. As shown in Chapter 6, there are significant differences in the behavior of L-band active and passive vegetation indices. If VOD is more sensitive to litter than aHV, an active-based vegetation index may be more suitable for studies of vegetation water content, while VOD can possibly provide complementary information. However, the use of active vegetation indices for studies of vegetation water content is limited by the difficulty of interpreting gHV, whose relationship to vegetation water content may be non-linear. Relatively simple models such as the cloud model (Attema and Ulaby, 1978) show the most promise in enabling the retrieval of vegetation water content from active measurements. This can be combined with recently developed parsimo165 - -- 1--__'-' -- .1 1 1 nious models to estimate the soil scattering contribution to measured backscattering coefficients (Narvekar et al., 2015; Bruscantini et al., 2015) to derive an effective VOD from UHV. Additional work is needed to validate whether the possible explanations in Chapter 6 explains the out of phase behavior of VOD and UHV in certain dry tropical regions. Ground-based observations in the locations of Figure 6-4 are necessary to determine the exact seasonal cycle of bud break, leaf flushing, and leaf fall in these regions. In validating satellite observations with ground-based observations, there is always a risk that the locations of the ground-based observations are not representative of the scale of the satellite pixel. The future availability of a full annual cycle of VOD data from higher-resolution SMAP records ( 36 km instead of 80 km) should help with this issue. Furthermore, the spatial homogeneity of the VOD seasonal behavior patterns in these regions suggests that representativeness error will be relatively low and thus unlikely to be a major concern. Additionally, radiative transfer simulations should be run to determine how much different canopy components contribute to the total. VOD and UHV in these regions. This is likely to depend on land cover type and vegetation characteristics, so it is imperative that region-specific parameters are used in these simulations. Remote sensing datasets such as lidar-based canopy height can be used to help this parameterization, and may be supplemented by groundbased observations where available. Lastly, a comparison of the seasonal behavior of VOD at L-band and at higher frequencies should help to validate these results, as the sensitivity of passive emissivity to leaves is highly dependent on the frequency used (Ferrazzoli and Guerriero, 1996). In Chapter 7, the diurnal variability of cTHV, interpreted as related to vegetation water content, was used to study the speed of vegetation water content refilling. Variations in the size of the difference in oHV between morning and afternoon were interpreted as sensitive to variations in refilling speed, but may also be sensitive to variations in refilling amount. Additional studies should take this into account. Rather than assuming the amount of necessary diurnal refilling is constant in time and changes in the diurnal variability are only reflective of changes in refilling speed, the 166 amount of expected refilling should be calculated explicitly based on transpiration estimates. Although modeled transpiration estimates may be imperfect, they are likely to capture the approximate seasonal cycle of refilling amount, providing a firstorder correction for this effect. Additional insights into this question may be gained from tower-based studies that can make continuous microwave measurements over the course of the diurnal cycle. The strong diurnal cycle of vegetation water content is superimposed upon slower seasonal variability. An easy first step to explore this seasonal variability is to focus on regions with a distinct dry season. In many of these regions, VOD dries down after the end of the rainy season. The size and timing of the VOD dry-down depends on plant water strategies. Their spatial variability can be studied and compared to a range of other rainfall (mean annual amount, intensity and duration of seasonality), energy (net radiation, cloud cover), and known plant properties (canopy height, plant functional type, woody fraction, etc) to find dominant controls. These can be used to test theoretical predictions that suggest that the seasonality in VOD may be highly sensitive to the duration of the wet season (Feng et al., 2012). Previous investigations have shown there is as much variability in photosynthetic parameters within different vegetation covers of the same plant functional type as there is between plant functional type groups themselves (Groenendijk et al., 2011). Given the strong coupling between plant water loss through transpiration (which has a first-order effect on vegetation water content) and photosynthetic behavior (e.g. Katul et al., 2010), similar findings are expected for the drydown behavior of VOD. Models of photosynthesis and transpiration that include vegetation water content as an explicitly modeled component are becoming more common (e.g. Matheny et al., 2014). When they are run on a larger scale, VOD data could be assimilated into such models. Comparing the day-to-day variability of vegetation water content to soil moisture at different times may allow the development of an indicator of canopylevel plant isohydry that can be derived across the globe, whose controls can then be studied. A key challenge in such an effort will be the conversion of VOD to leaf water potential. This challenge could be overcome, for example by using a machine 167 learning approach (as has been developed for snow (Forman and Reichle, 2015)) or if need be by assuming a transfer function dependent on land cover type. The resulting dataset can be useful for predicting drought sensitivity in different regions under current and future conditions. VOD-based indices of drought sensitivity can also be used to evaluate theoretical predictions that taller trees are more prone to hydraulic failure (McDowell et al., 2008). Remotely sensed observations can provide valuable new information about the behavior of the global hydrologic cycle - they enable observations at scales unimaginable using in situ measurements. Because they are indirect measurements and not all model parameters can be directly inferred from remote sensing, remotely sensed observations are often interpreted only in the context of other remotely sensed observables or in the context of large-scale models. The gap to the literature based on ground observations is rarely bridged unless it is in the explicit context of algorithm validation. Whenever possible, studies should be designed so as to explicitly test theories developed at smaller scales (such as, for example, the leaf level in the context of ecohydrology). This not only serves to enrich the interpretation of remotely sensed observations, but may also provide valuable lessons for the development of hydrologic models. 168 Appendix A Vegetation Parameters for Hydrologic Modeling The vegetation parameters used for hydrologic modeling of the Vaira Ranch site are detailed in Table A.1. The stomatal resistance r. model takes the form (Campbell, 1985; Flerchingerand Pierson, 1997), rs = (A.1) rso (1 + (4'i/4c)n) Because several of these model-specific parameters were not available in the literature, they were manually optimized to produce the best fit between modeled soil moisture and the Ameriflux measurements. Changing the parameters did not cause a significant change in the modeled soil moisture. The variation of root biomass with depth was assumed to be distributed exponentially, p = 1 - f', where p is the cumulative fraction of roots located above depth z in centimeters (Jackson et al., 1996). For Vaira Ranch, # = 0.94 (Baldocchi et al., 2004). The model is spun-up for 32 years. The soil is assumed to be a silty loam, consisting of 30% sand fraction and 13% clay fraction, with a saturated hydraulic conductivity of 200 mm/day (Miller et al., 2007). 169 Table A.1: Vegetation parameters used for hydrologic modeling Source Value Parameter Ameriflux site description 10 cm Plant height Assumed 0.2 cm Characteristic dimension Baldocchi et al. (2004), avg. of fig. 1 0.33 Leaf Area Index 2 24 g/m LAI/0.0135, Montaldo et al. (2005) Dry biomass Baldocchi et al. (2004), fig. 5 Canopy albedo 0.15 Temp. of respiration shutdown 50 C Assumed Minimum stomatal resistance ro Critical leaf water potential I' Stomatal resistance parameter n 70.0 s/m -150 m 5.0 Xu and Baldocchi (2003), fig. 2 Optimized Flerchinger and Pierson (1997) Leaf resistance Root resistance 106 m 3 s kg- 1 2 x 106 m 3 s kg-' Optimized Optimized Surface roughness 1 cm 0.1 (plant height) 170 Bibliography Akbar, R., and M. Moghaddam (2015), A Combined Active AA; Passive Soil Moisture Estimation Algorithm With Adaptive Regularization in Support of SMAP, IEEE Trans. Geosci. Remote Sens., 53(6), 3312-3324. Alemohammad, S. H. (2015), Personal Communication. Andela, N., Y. Y. Liu, A. I. J. M. van Dijk, R. A. M. de Jeu, and T. R. McVicar (2013), Global changes in dryland vegetation dynamics (1988 - 2008) assessed by satellite remote sensing: comparing a new passive microwave vegetation density record with reflective greenness data, Biogeosciences, 10(10), 6657-6676, doi:10.5194/bg-106657-2013. Anderegg, W. R. L., J. M. Kane, and L. D. L. Anderegg (2012), Consequences of widespread tree mortality triggered by drought and temperature stress, Nat. Clim. Chang., 3(September 2012), doi:10.1038/nclimatel635. Anderegg, W. R. L., et al. (2015), Tree mortality from drought, insects, and their interactions in a changing climate., New Phytol., doi:10.1111/nph.13477. Arii, M., J. J. V. Zyl, Y. Kim, S. Member, and A. Current (2010), A General Characterization for Polarimetric Scattering From Vegetation Canopies, IEEE Trans. Geosci. Remote Sens., 48(9), 3349-3357. Asefi-Najafabady, S., and S. Saatchi (2013), Response of African humid tropical forests to recent rainfall anomalies, Philos. Trans. R. Soc. B Biol. Sci., 368(1625), Attema, E. P. W., and F. T. Ulaby (1978), Vegetation modeled as a water cloud, Radio Sci., 13(2), 357-364. Baker, I. T., L. Prihodko, a. S. Denning, M. Goulden, S. Miller, and H. R. Da Rocha (2009), Seasonal drought stress in the amazon: Reconciling models and observations, J. Geophys. Res. Biogeosciences, 114 (1), 1-10, doi:10.1029/2007JG000644. Baldocchi, D. D., L. Xu, and N. Kiang (2004), How plant functional-type, weather, seasonal drought, and soil physical properties alter water and energy fluxes of an oak-grass savanna and an annual grassland, Agr. For. Meteorol., 123(1-2), 13-39. 171 Bannister, R. N. (2008), A review of forecast error covariance statistics in atmospheric variational data assimilation . II : Modelling the forecast error covariance statistics, Q. J. R. Meteorol. Soc., 1996, 1971-1996, doi:10.1002/qj. Beven, K. (1993), Prophecy, reality and uncertainty in distributed hydrological modelling, Adv. Water Resour., 16(1), 41-51. Beven, K. J., and M. J. Kirkby (1979), A physically based, variable contributing area model of basin hydrology, Hydrol. Sci. Bull., 24 (1), 43-69, doi: 10.1080/02626667909491834. Bindlish, R., T. Jackson, M. Cosh, T. Zhao, and P. O'Neill (2015), Global Soil Moisture From the Aquarius/SAC-D Satellite: Description and Initial Assessment, IEEE Geosci. Remote Sens. Lett., 12(5), 923-927. B16schl, G., and M. Sivapalan (1995), Scale issues in hydrological modelling: A review, Hydrol. Process., 9(3-4), 251-290. Bohrer, G., H. Mourad, T. A. Laursen, D. Drewry, R. Avissar, D. Poggi, R. Oren, and G. G. Katul (2005), Finite element tree crown hydrodynamics model (FETCH) using porous media flow within branching elements: A new representation of tree hydrodynamics, Water Resour. Res., 41 (11), n/a--n/a, doi:10.1029/2005WR004181. Boisvert, J. B., Q. H. J. Gwyn, A. Chanzy, D. J. Major, B. Brisco, and R. J. Brown (1997), Effect of surface soil moisture gradients on modelling radar backscattering from bare fields, Int. J. Remote Sens., 18(1), 153-170, doi: 10.1080/014311697219330. Bourne, A. E., A. M. Haigh, and D. S. Ellsworth (2015), Stomatal sensitivity to vapour pressure deficit relates to climate of origin in Eucalyptus species, Tree Physiol., 35(3), 266-278, doi:10.1093/treephys/tpvOl4. Bras, R. L. (1989), Hydrology: An Introduction to Hydrologic Science, 660 pp., Addison-Wesley. Breshears, D. D., 0. B. Myers, C. W. Meyer, F. J. Barnes, C. B. Zou, C. D. Allen, N. G. McDowell, and W. T. Pockman (2009), Research communications research communications Tree die-off in response to global change-type drought: Mortality insights from a decade of plant water potential measurements, Front. Ecol. Environ., 7(4), 185-189, doi:10.1890/080016. Brodribb, T. J., and N. M. Holbrook (2004), Diurnal depression of leaf hydraulic conductance in a tropical tree species, Plant, Cell Environ., 27(7), 820-827, doi: 10.1111/j.1365-3040.2004.01188.x. Bruscantini, C. A., A. G. Konings, P. Narvekar, K. A. McColl, and D. Entekhabi (2015), L-band radar soil moisture retrieval without ancillary information, IEEE Trans. Geosci. Remote Sens., Submitted. 172 Burgin, M. (2012), Personal Communication. . Burgin, M., D. Clewley, R. M. Lucas, and M. Moghaddam (2012), A generalized radar backscattering model based on wave theory for multilayer multispecies vegetation, IEEE Trans. Geosci. Remote Sens., 49(12), 4832-4845, doi: 10.1109/TGRS.2011.2172949. Campbell, G. S. (1985), Soil physics with BASIC: Transport models for soil-plant systems, 150 pp., Elsevier, Amsterdam. Cariou, C., K. Chehdi, and S. Le Moan (2011), BandClust: An Unsupervised Band Reduction Method for Hyperspectral Remote Sensing, IEEE Geosci. Remote Sens. Lett., 8(3), 565-569. Cavender-Bares, J., L. Sack, and J. Savage (2007), Atmospheric and soil drought reduce nocturnal conductance in live oaks., Tree Physiol., 27(4), 611-620, doi: 10. 1093/treephys/27.4.611. Cawse-Nicholson, K., S. B. Damelin, A. Robin, and M. Sears (2013), Determining the Intrinsic Dimension of a Hyperspectral Image Using Random Matrix Theory, IEEE Trans. Image Proc., 22(4), 1301-1310. Cerra, D., and M. Datcu (2008), A Model Conditioned Data Compression Based Similarity Measure, in Data Compression Conf. 2008. DCC 2008, p. 509. Chapin, E., A. Chau, J. Chen, B. Heavey, S. Hensley, Y. Lou, R. Machuzak, and M. Moghaddam (2012), AirMOSS: An Airborne P-band SAR to measure root-zone soil moisture, in Radar Conf. (RADAR), 2012 IEEE, pp. 693-698. Churkina, G., S. W. Running, and A. L. Schloss (1999), Comparing global models of terrestrial net primary productivity (NPP): the importance of water availability, Glob. Chang. Biol., 5(S1), 46-55, doi:10.1046/j.1365-2486.1999.00006.x. Ciais, P., et al. (2013), Climate Change 2013. The Physical Science Basis. Contribution of Working Group 1 to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change, chap. Carbon and, pp. 465-570, Cambridge University Press, New York. Cover, T. M., and J. A. Thomas (1991), Elements of Information Theory, John Wiley & Sons, Ltd. Crow, W. T., et al. (2005), An observing system simulation experiment for Hydros radiometer-only soil moisture products, IEEE Trans. Geosci. Remote Sens., 43(6), 1289-1303, doi:10.1109/TGRS.2005.845645. Crow, W. T., A. A. Berg, M. H. Cosh, A. Loew, B. P. Mohanty, R. Panciera, P. de Rosnay, D. Ryu, and J. P. Walker (2012), Upscaling sparse ground-based soil moisture observations for the validation of coarse-resolution satellite soil moisture products, Rev. Geophys., 50(2), RG2002--. 173 Cuenca, R., and Y. Hagimoto (2012), Personal Communication. Cui, Q., J. Shi, J. Du, T. Zhao, and C. Xiong (2015), An Approach for Monitoring Global Vegetation Based on Multiangular Observations From SMOS, IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 8(2), 604-616. Davenport, I., J. Fernandez-Galvez, and R. Gurney (2005), A sensitivity analysis of soil moisture retrieval from the tau-omega microwave emission model, IEEE Trans. Geosci. Remote Sens., 43(6), 1304-1316, doi:10.1109/TGRS.2005.845640. de Jeu, R. a. M., T. R. H. Holmes, R. M. Parinussa, and M. Owe (2014), A spatially coherent global soil moisture product with improved temporal resolution, J. Hydrol., 516, 284-296, doi:10.1016/j.jhydrol.2014.02.015. De Roo, R., M. C. Dobson, and F. T. Ulaby (1991), Dielectric Constant of Forest Specular Reflectivity of Forest Litter Bistatic Scattering from Forest Litter, Proc. IEEE Geosci. Remote Sens. Symp., pp. 15-18. Dinnat, E. P., S. Abraham, D. M. Le Vine, P. De Matthaeis, and C. Utku (2012), Comparison of aquarius measurements over oceans with radiative transfer models at L-band, Int. Geosci. Remote Sens. Symp., pp. 382-385, doi: 10.1 109/IGARSS.2012.6351558. Draper, C. S., R. H. Reichle, G. J. M. De Lannoy, and Q. Liu (2012), Assimilation of passive and active microwave soil moisture retrievals, Geophys. Res. Lett., 39, 5 PP., doi:201210.1029/2011GL050655. Du, Y., F. T. Ulaby, and M. C. Dobson (2000), Sensitivity to soil moisture by active and passive microwave sensors, IEEE Trans. Geosci. Remote Sens., 38(1), 105-114, doi: 10.1109/36.823905. Duan, X., and M. Moghaddam (2011), Vector electromagnetic scattering from layered rough surfaces with buried discrete random media for subsurface and root-zone soil moisture sensing, in Geosci. Remote Sens. Symp. (IGARSS), 2011 IEEE Int., pp. 1227-1230. Durden, S. L., J. J. van Zyl, and H. A. Zebker (1989), Modeling and observation of the radar polarization signature of forested areas, IEEE Trans. Geosci. Remote Sens., 27(3), 290-301, doi:10.1109/36.17670. Eamus, D. (1999), Ecophysiological traits of deciduous and evergreen woody species in the seasonally dry tropics, Trends Ecol. Evol., 14 (1), 11-16, doi:10.1016/S01695347(98)01532-8. Ehrenberger, W., S. Riiger, R. Fitzke, P. Vollenweider, M. Giinthardt-Goerg, T. Kuster, U. Zimmermann, and M. Arend (2012), Concomitant dendrometer and leaf patch pressure probe measurements reveal the effect of microclimate and soil moisture on diurnal stem water and leaf turgor variations in young oak trees, Funct. Plant Biol., 39(4), 297-305. 174 England, A. W. (1975), Thermal Microwave Emission from a Scattering Layer, J. Geophys. Res., 80(32), 4484-4496, doi:10.1029/JB080i032p04484. Englhart, S., V. Keuck, and F. Siegert (2011), Aboveground biomass retrieval in tropical forests - The potential of combined X- and L-band SAR data use, Remote Sens. Environ., 115(5), 1260-1271, doi:10.1016/j.rse.2011.01.008. Entekhabi, D., R. H. Reichle, R. D. Koster, and W. T. Crow (2010a), Performance Metrics for Soil Moisture Retrievals and Application Requirements, J. Hydrometeorol., 11 (3), 832-840, doi:10.1175/2010JHM1223.1. Entekhabi, D., et al. (2010b), The soil moisture active passive (SMAP) mission, Proc. IEEE, 98(5), 704-716, doi:10.1109/JPROC.2010.2043918. Erten, E., A. Reigber, L. Ferro-Famil, and 0. Hellwich (2012), A New Coherent Similarity Measure for Temporal Multichannel Scene Characterization, IEEE Trans. Geosci. Remote Sens., 50(7), 2839-2851. Escorihuela, M. J., A. Chanzy, J. P. Wigneron, and Y. H. Kerr (2010), Effective soil moisture sampling depth of L-band radiometry: A case study, Remote Sens. Environ., 114 (5), 995-1001, doi:10.1016/j.rse.2009.12.011. Ewers, B. E., R. Oren, N. Phillips, M. Str6mgren, and S. Linder (2001), Mean canopy stomatal conductance responses to water and nutrient availabilities in Picea abies and Pinus taeda., Tree Physiol., 21 (12-13), 841-850, doi:10.1093/treephys/21.1213.841. Ewers, B. E., R. Oren, H. S. Kim, G. Bohrer, and C. T. Lai (2007), Effects of hydraulic architecture and spatial variation in light on mean stomatal conductance of tree branches and crowns, Plant, Cell Environ., 30(4), 483-496, doi:10.1111/j.13653040.2007.01636.x. Eyre, J. R. (1989), Inversion of cloudy satellite sounding radiances by nonlinear optimal estimation. I: Theory and simulation for TOVS, Q.J.R. Meteorol. Soc., 115(489), 1001-1026. Famiglietti, J. S., D. Ryu, A. A. Berg, M. Rodell, and T. J. Jackson (2008), Field observations of soil moisture variability across scales, Water Resour. Res., 44(1), 1-16, doi:10.1029/2006WR005804. Feng, X., G. Vico, and A. Porporato (2012), On the effects of seasonality on soil water balance and plant growth, Water Resour. Res., 48(5), 1-12, doi: 10.1029/2011WR011263. Ferrazzoli, P., and L. Guerriero (1996), Passive microwave remote sensing of forests: a model investigation, IEEE Trans. Geosci. Remote Sens., 34(2), doi: 10.1109/36.485121. 175 Ferrazzoli, P., L. Guerriero, S. Paloscia, and P. Pampaloni (1995), Modeling X and Ka band emission from Leafy vegetation, J. Electromagn. Waves Appl., 9(3), 393-406, doi: 10.1 163/156939395X00541. Fiener, P., K. Auerswald, and K. Van Oost (2011), Spatio-temporal patterns in land use and management affecting surface runoff response of agricultural catchments-A review, Earth-Science Rev., 106(1-2), 92-104, doi:10.1016/j.earscirev.2011.01.004. Fisher, J. B., D. D. Baldocchi, L. Misson, T. E. Dawson, and A. H. Goldstein (2007), What the towers don't see at night: nocturnal sap flow in trees and shrubs at two AmeriFlux sites in California, Tree Physiol., 27(4), 597-610. Flerchinger, G. N., and F. B. Pierson (1991), Modeling plant canopy effects on variability of soil temperature and water, Agr. For. Meteorol., 56, 227-246. Flerchinger, G. N., and F. B. Pierson (1997), Modelling plant canopy effects on variability of soil temperature and water: Model calibration and validation, J. Arid Environ., 35(4), 641-653, doi:10.1006/jare.1995.0167. Forman, B. A., and R. H. Reichle (2015), Using a support vector machine and a land surface model to estimate large-scale passive microwave temperatures over snowcovered land in North America, IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., pp. 1-24, doi:10.1109/JSTARS.2014.2325780. Freeman, A., and S. S. Saatchi (2004), On the Detection of Faraday Rotation in Linearly Polarized L-band SAR Backscatter Signatures, East, 42(8), 1607-1616. Friesen, J., S. C. Steele-Dunne, and N. Van De Giesen (2012), Diurnal differences in global ers scatterometer backscatter observations of the land surface, IEEE Trans. Geosci. Remote Sens., 50(7 PART1), 2595-2602, doi:10.1109/TGRS.2012.2193889. Fung, A. K., M. S. Dawson, K. S. Chen, A. Y. Hsu, E. T. Engman, P. 0. O'Neill, and J. Wang (1996), A modified IEM model for scattering from soil surfaces with application to soil moisture sensing, in Geosci. Remote Sens. Symp. 1996. IGARSS '96. 'Remote Sens. a Sustain. Futur. Int., vol. 2, pp. 1297-1299 vol.2. Galantowicz, J. F., D. Entekhabi, and E. G. Njoku (1999), Tests of sequential data assimilation for retrieving profile soil moisture and temperature from observed L- band radiobrightness, IEEE Trans. Geosci. Remote Sens., 37(4), 1860-1870, doi: 10.1109/36.774699. Gillespie, T., B. Brisco, R. Brown, and G. Sofko (1990), Radar detection of a dew event in wheat, Remote Sens. Environ., 33(3), 151-156, doi:10.1016/0034- 4257(90)90026-I. Good, S. P., D. Noone, and G. Bowen (2015), Hydrologic Connectivity Constrains Partitioning of Global Terrestrial Water Fluxes, Science (80-. )., 349(6244), 175- 177. 176 Goulden, M. L., J. W. Munger, S.-M. Fan, B. C. Daube, and S. C. Wofsy (1996), Measurements of carbon sequestration by long-term eddy covariance: Methods and a critical evaluation of accuracy, Glob. Chang. Biol. [Global Chang. Biol.], 2(3), 169-182, doi:10.1111/j.1365-2486.1996.tb00070.x. Grant, J. P., A. A. van de Griend, M. Schwank, and J.-P. Wigneron (2009), Observations and Modeling of a Pine Forest Floor at L-Band, IEEE Trans. Geosci. Remote Sens., 47(7), 2024-2034, doi:10.1109/TGRS.2008.2010252. Groenendijk, M., et al. (2011), As sessing parameter variability in a photosynthesis model within and betw een plant functional types using global Fluxnet eddy covariance data, Ag ric. For. Meteorol., 151(1), 22-38, doi: 10. 1016/j.agrformet.2010.08.013. Guan, K., A. Wolf, D. Medvigy, K. K. Caylor, M. Pan, and E. F. Wood (2013), Seasonal coupling of canopy structure and function in African tropical forests and its environmental controls, Ecosphere, 4 (3), art35, doi:10.1890/ES12-00232.1. Guan, K., E. F. Wood, D. Medvigy, J. Kimball, M. Pan, K. K. Caylor, J. Sheffield, X. Xu, and M. 0. Jones (2014), Terrestrial hydrological controls on land surface phenology of African savannas and woodlands, J. Geophys. Res. Biogeosciences, 119, 1652-1669, doi:10.1002/2013JG002572.Received. Guanter, L., C. Frankenberg, A. Dudhia, P. E. Lewis, J. G6mez-Dans, A. Kuze, H. Suto, and R. G. Grainger (2012), Retrieval and global assessment of terrestrial chlorophyll fluorescence from GOSAT space measurements, Remote Sens. Environ., 121(0), 236-251, doi:http://dx.doi.org/10.1016/j.rse.2012.02.006. Gueguen, L., S. Cui, G. Schwarz, and M. Datcu (2010), Multitemporal analysis of multisensor data: Information theoretical approaches, in Geosci. Remote Sens. Symp. (IGARSS), 2010 IEEE Int., pp. 2559-2562. Hacine-Gharbi, A., P. Ravier, R. Harba, and T. Mohamadi (2012), Low bias histogram-based estimation of mutual information for feadoi: 33(10), 1302-1308, Pattern Recognit. Lett., ture selection, http://dx.doi.org/10.1016/j .patree.2012.02.022. Hasanlou, M., and F. Samadzadegan (2012), Comparative Study of Intrinsic Dimensionality Estimation and Dimension Reduction Techniques on Hyperspectral Images Using {K-NN} Classifier, IEEE Geosci. Remote Sens. Lett., 9(6), 1046-1050. Hellkvist, J., R. G.P., and J. P.G. (1974), Vertical Gradients of Water Potential and Tissue Water Relations in Sitka Spruce Trees Measured with the Pressure Chamber, J. Appl. Ecol., 11(2), 637. Hensley, S., et al. (2008), The UAVSAR instrument: Description and first results, in Radar Conf. 2008. RADAR '08. IEEE, pp. 1-6. 177 Herrera, A., C. Ballestrini, and W. Tezara (2008), Nocturnal sap flow in the C3-CAM species, Clusia minor, Trees - Struct. Funct., 22(4), 491-497, doi:10.1007/s00468008-0209-8. Heylen, R., and P. Scheunders (2013), Hyperspectral Intrinsic Dimensionality Estimation With Nearest-Neighbor Distance Ratios, IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 6(2), 570-579. Hoeben, R., and P. A. Troch (2000), Assimilation of active microwave observation data for soil moisture profile estimation, Water Resour. Res., 36(10), 2805-2819. Holbrook, N. M., J. L. Whitbeck, and H. A. Mooney (1995), Drought responses of neotropical dry forest trees, in Seas. dry Trop. For., pp. 243-276, Cambridge University Press. Hossain, M. A., X. Jia, and M. Pickering (2014), Subspace Detection Using a Mutual Information Measure for Hyperspectral Image Classification, IEEE Geosci. Remote Sens. Lett., 11 (2), 424-428. Huete, A., C. Justice, and H. Liu (1994), Development of Vegetation and Soil Indices for MODIS-EOS, Remote Sens. Environ., 49, 224-234. Hutyra, L. R., J. W. Munger, S. R. Saleska, E. Gottlieb, B. C. Daube, A. L. Dunn, D. F. Amaral, P. B. de Camargo, and S. C. Wofsy (2007), Seasonal controls on the exchange of carbon and water in an Amazonian rain forest, J. Geophys. Res. Biogeosciences, 112(3), 1-16, doi:10.1029/2006JG000365. Imperatore, P., A. Iodice, and D. Riccio (2009), Electromagnetic Wave Scattering From Layered Structures With an Arbitrary Number of Rough Interfaces, IEEE Trans. Geosci. Remote Sens., 47(4), 1056-1072. Jackson, R. B., J. Canadell, J. R. Ehleringer, H. A. Mooney, 0. E. Sala, and E. D. Schulze (1996), A global analysis of root distributions for terrestrial biomes, Oecologia, 108(3), 389-411, doi:10.1007/BF00333714. Jackson, T. J., and T. J. Schmugge (1991), Vegetation effects on the microwave emission of soils, Remote Sens. Environ., 36(3), 203-212. Jackson, T. J., A. Y. Hsu, and P. E. O'Neill (2002), Surface Soil Moisture Retrieval and Mapping Using High-Frequency Microwave Satellite Observations in the {Southern Great Plains}, J. Hydrometeor, 3(6), 688-699, doi:10.1175/15257541(2002)003<0688:SSMRA M>2.0.CO;2. Jagdhuber, T., I. Hajnsek, A. Bronstert, and K. P. Papathanassiou (2013), Soil Moisture Estimation Under Low Vegetation Cover Using a Multi-Angular Polarimetric Decomposition, Geosci. Remote Sensing, IEEE Trans., 51 (4), 2201-2215. 178 James, A. L., and N. T. Roulet (2007), Investigating hydrologic connectivity and its association with threshold change in runoff response in a temperate forested watershed, Hydrol. Process., 21(25), 3391-3408. Janott, M., S. Gayler, A. Gessler, M. Javaux, C. Klier, and E. Priesack (2011), A onedimensional model of water flow in soil-plant systems based on plant architecture, Plant Soil, 341 (1-2), 233-256, doi:10.1007/s11104-010-0639-0. Jasechko, S., Z. D. Sharp, J. J. Gibson, S. J. Birks, Y. Yi, and P. J. Fawcett (2013), Terrestrial water fluxes dominated by transpiration, Nature, 496(7445), 347-350. Jones, M. 0., L. A. Jones, J. S. Kimball, and K. C. McDonald (2011), Satellite passive microwave remote sensing for monitoring global land doi: 1102-1114, Remote Sens. Environ., 115(4), surface phenology, http://dx.doi.org/10.1016/j.rse.2010.12.015. Kattge, J., et al. (2011), TRY - a global database of plant traits, Glob. Chang. Biol., 17(9), 2905-2935, doi:10.1111/j.1365-2486.2011.02451.x. Katul, G., A. Porporato, and R. -Oren (2007), Stochastic Dynamics of PlantWater Interactions, Annu. Rev. Ecol. Evol. Syst., 38(1), 767-791, doi: 10.1 146/annurev.ecolsys.38.091206.095748. Katul, G., S. Manzoni, S. Palmroth, and R. Oren (2010), A stomatal optimization theory to describe the effects of atmospheric C02 on leaf photosynthesis and transpiration, Ann. Bot., 105(3), 431-442. Kelly, A. E., and M. L. Goulden (2008), Rapid shifts in plant distribution with recent climate change., Proc. Natl. Acad. Sci. U. S. A., 105(33), 11,823-11,826, doi:10.1073/pnas.0802891105. Kerr, Y., P. P. Waldteufel, P. Richaume, L. Davenport, P. Ferrazzoli, and J.P. Wigneron (2011), Algorithm Theoretical Based Document (ATBD) for the SMOS level 2 Processor for Soil Moisture, Tech. rep., CESBIO, IPSL-Service difYiAifironomie, INRA-EPHYSE, Reading University, Tor Vergata University. Kerr, Y. H., et al. (2010), The SMOS Mission: New Tool for Monitoring Key Elements ofthe Global Water Cycle, Proc. IEEE, 98(5), 666-687, doi: 10.1109/JPROC.2010.2043032. Kerr, Y. H., et al. (2012), The SMOS Soil Moisture Retrieval Algorithm, IEEE Trans. Geosci. Remote Sens., 50(5), 1384-1403. Khankhoje, U. K., J. J. van Zyl, and T. A. Cwik (2013), Computation of Radar Scattering From Heterogeneous Rough Soil Using the Finite-Element Method, Geosci. Remote Sensing, IEEE Trans., 51(6), 3461-3469. 179 Kim, S.-B., L. Tsang, J. T. Johnson, S. Huang, J. J. Van Zyl, and E. G. Njoku (2012), Soil Moisture Retrieval Using Time-Series Radar Observations Over Bare Surfaces, Geosci. Remote Sensing, IEEE Trans., 50(5), 1853-1863. Kinnell, P. I. A. (2010), Event soil loss, runoff and the Universal Soil Loss Equation family of models: A review, J. Hydrol., 385(1-4), 384-397, doi: 10.1016/j.jhydrol.2010.01.024. Kleidon, A. (2004), Global Datasets of Rooting Zone Depth Inferred from Inverse Methods, J. Clim., 17(13), 2714-2722, doi:10.1175/15200442(2004)017<2714:GDORZD>2.0.CO;2. Klein, L., and C. Swift (1977), An improved model for the dielectric constant of sea water at microwave frequencies, IEEE Trans. Antennas Propag., 25(1), 104-111. Klein, T. (2014), The variability of stomatal sensitivity to leaf water potential across tree species indicates a continuum between isohydric and anisohydric behaviours, Funct. Ecol., pp. 1313-1320, doi:10.1111/1365-2435.12289. Kong, J. A. (2008), Electromagnetic Wave Theory, 1016 pp., EMW Publishing, Cambridge, MA. Konings, A. G., G. G. Katul, and A. Porporato (2010), The rainfall-no rainfall transition in a coupled land-convective atmosphere system, Geophys. Res. Lett., 37(14), 1-5, doi:10.1029/2010GL043967. Konings, A. G., D. Entekhabi, S. K. Chan, and E. G. Njoku (2011), Effect of Radiative Transfer Uncertainty on L-Band Radiometric Soil Moisture Retrieval, IEEE Trans. Geosci. Remote Sens., 49(7), 2686-2698. Konings, A. G., D. Entekhabi, M. Moghaddam, and S. S. Saatchi (2014), The Effect of a Variable Soil Moisture Profile on P-band Backscatter Estimation, IEEE Trans. Geosci. Remote Sens., 52(10), 6315-6325. Konings, A. G., K. A. McColl, M. Piles, and D. Entekhabi (2015), How Many Parameters Can Be Maximally Estimated From a Set of Measurements?, IEEE Geosci. Remote Sens. Lett., 12(5), 1081-1085. Koster, R. D. (2004), Regions of Strong Coupling Between Soil Moisture and Precipitation, Science (80-. )., 305(5687), 1138-1140, doi:10.1126/science.1100217. Koster, R. D., M. J. Suarez, and M. Heiser (2000), Variance and Predictability of Precipitation at Seasonal-to-Interannual Timescales, J. Hydrometeorol., 1 (1), 2646, doi:10.1175/1525-7541(2000)001<0026:VAPOPA>2.0.CO;2. Koster, R. D., Z. Guo, R. Yang, P. A. Dirmeyer, K. Mitchell, and M. J. Puma (2009), On the Nature of Soil Moisture in Land Surface Models, J. Clim., 22(16), 43224335, doi:10.1175/2009JCLI2832.1. 180 Kseneman, M., and D. Gleich (2013), Soil-moisture estimation from X-band data using tikhonov regularization and neural net, IEEE Trans. Geosci. Remote Sens., 51 (7), 3885-3898, doi:10.1109/TGRS.2012.2228486. Kumagai, T. (2010), Impact of stem water storage on diurnal estimates of whole-tree transpiration and canopy conductance from sap flow measurements in Japanese cedar and Japanese cypress trees, Hydrol. Process., 2274 (November 2008), 22672274, doi:10.1002/hyp. Kumar, S. V., R. H. Reichle, R. D. Koster, W. T. Crow, and C. D. Peters-Lidard (2009), Role of Subsurface Physics in the Assimilation of Surface Soil Moisture Observations, J. Hydrometeor., 10(6), 1534-1547, doi:10.1175/2009JHM1134.1. Kurum, M. (2013), Quantifying scattering albedo in microwave emission of vegetated terrain, Remote Sens. Environ., 129(0), 66-74. Kurum, M., P. E. O'Neill, R. H. Lang, A. T. Joseph, M. H. Cosh, and T. J. Jackson (2012a), Effective tree scattering and opacity at L-band, Remote Sens. Environ., 118(0), 1-9. Kurum, M., P. E. O'Neill, R. H. Lang, M. H. Cosh, A. T. Joseph, and T. J. Jackson (2012b), Impact of conifer forest litter on microwave emission at L-band, IEEE Trans. Geosci. Remote Sens., 50(4), 1071-1084, doi:10.1109/TGRS.2011.2166272. Lawrence, H., et al. (2014), Comparison between SMOS Vegetation Optical Depth products and MODIS vegetation indices over crop zones of the USA, Remote Sens. Environ., 140 (0), 396-406, doi:http://dx.doi.org/10.1016/j.rse.2013.07.021. Le Hegarat-Mascle, S., D. Vidal-Madjar, 0. Taconet, and Z. M. (1997), Application of shannon information theory to a comparison between L- and C-band SIR-C polarimetric data versus incidence angle, Remote Sens. Environ., 60(2), 121-130, doi:http://dx.doi.org/10.1016/S0034-4257(96)00164-2. Le Morvan, A., M. Zribi, N. Baghdadi, and A. Chanzy (2008), Soil Moisture Profile Effect on Radar Signal Measurement, Sensors, 8(1), 256-270, doi:10.3390/s8010256. Le Toan, T., A. Beaudoin, J. Riom, and D. Guyon (1992), Relating forest biomass to SAR data, IEEE Trans. Geosci. Remote Sens., 30(2), 403-411, doi: 10.1109/36.134089. Le Vine, D. M., G. S. E. Lagerloef, F. R. Colomb, S. H. Yueh, and F. A. Pellerano (2007), Aquarius: An Instrument to Monitor Sea Surface Salinity From Space, IEEE Trans. Geosci. Remote Sens., 45(7), 2040-2050, doi: 10.1109/TGRS.2007.898092. Le Vine, D. M., P. De Matthaeis, C. S. Ruf, and D. D. Chen (2014), Aquarius RFI detection and mitigation algorithm: Assessment and examples, IEEE Trans. Geosci. Remote Sens., 52(8), 4574-4584, doi:10.1109/TGRS.2013.2282595. 181 Lee, J. S., and E. Pottier (2009), Polarimetric Radar Imaging: From Basics To Applications, CRC Press. Legg, P. A., P. L. Rosin, D. Marshall, and J. E. Morgan (2007), Improving accuracy and efficiency of registration by mutual information using SturgesaA2 histogram rule, Proc. Med. Image Underst. Anal. Univ. Aberystwyth, pp. 17-18. Licciardi, G., P. R. Marpu, J. Chanussot, and J. A. Benediktsson (2012), Linear Versus Nonlinear PCA for the Classification of Hyperspectral Data Based on the Extended Morphological Profiles, IEEE Geosci. Remote Sens. Lett., 9(3), 447-451. Liu, Y. Y., R. A. M. de Jeu, M. F. McCabe, J. P. Evans, and A. 1. J. M. van Dijk (2011), Global long-term passive microwave satellite-based retrievals of vegetation optical depth, Geophys. Res. Lett., 38(18), L18,402--. Liu, Y. Y., A. I. J. M. van Dijk, R. a. M. de Jeu, J. G. Canadell, M. F. McCabe, J. P. Evans, and G. Wang (2015), Recent reversal in loss of global terrestrial biomass, Nat. Clim. Chang., (May), doi:10.1038/nclimate2581. Loarie, S. R., P. B. Duffy, H. Hamilton, G. P. Asner, C. B. Field, and D. D. Ackerly (2009), The velocity of climate change., Nature, 462(7276), 1052-1055, doi: 10.1038/nature08649. Loew, A., R. Ludwig, and W. Mauser (2006), Derivation of surface soil moisture from ENVISAT ASAR wide swath and image mode data in agricultural areas, IEEE Trans. Geosci. Remote Sens., 44 (4), 889-899. Manzoni, S., G. Vico, G. Katul, P. A. Fay, W. Polley, S. Palmroth, and A. Porporato (2011), Optimizing stomatal conductance for maximum carbon gain under water stress: a meta-analysis across plant functional types and climates, Funct. Ecol., 25(3), 456-467. Martinez-Vilalta, J., R. Poyatos, D. Aguad, J. Retana, and M. Mencuccini (2014), A new look at water transport regulation in plants, New Phytol., 204, 105-115. Matheny, A. M., et al. (2014), Characterizing the diurnal patterns of errors in the prediction of evapotranspiration by several land-surface models: An NACP analysis, J. Geophys. Res. Biogeosciences, pp. n/a-n/a, doi:10.1002/2014JG002623. Matzler, C. (1994), Microwave transmissivity of a forest canopy: Experiments made with a beech, Remote Sens. Environ., 48(2), 172-180, doi:10.1016/00344257(94)90139-2. McColl, K. A., D. Entekhabi, and M. Piles (2014), Uncertainty Analysis of Soil Moisture and Vegetation Indices Using Aquarius Scatterometer Observations, IEEE Trans. Geosci. Remote Sens., 52(7), 4259-4272. 182 4"'"Amw_ . -1 "1.-1, . "I " a _a - - - McDowell, N., et al. (2008), Mechanisms of plant survival and mortality during drought: why do some plants survive while others succumb to drought?, New Phytol., 178(4), 719-739, doi:10.1111/j.1469-8137.2008.02436.x. McDowell, N. G., D. J. Beerling, D. D. Breshears, R. a. Fisher, K. F. Raffa, and M. Stitt (2011), The interdependence of mechanisms underlying climate-driven vegetation mortality, Trends Ecol. Evol., 26(10), 523-532, doi: 10. 1016/j.tree.2011.06.003. Medlyn, B. E., et al. (2011), Reconciling the optimal and empirical approaches to modelling stomatal conductance, Glob. Chang. Biol., 17(6), 2134-2144. Meesters, A., R. A. M. de Jeu, and M. Owe (2005), Analytical derivation of the vegetation optical depth from the microwave polarization difference index, IEEE Geosci. Remote Sens. Lett., 2(2), 121-123. Meir, P., M. Mencuccini, and R. C. Dewar (2015), Drought-related tree mortality addressing the gaps in understanding and prediction, New Phytol., 207, 28-33. Miller, G. R., D. D. Baldocchi, B. E. Law, and T. Meyers (2007), An analysis of soil moisture dynamics using multi-year data from a network of micrometeorological observation sites, Adv. Water Resour., 30(5), 1065-1081, doi: 10.1016/j.advwatres.2006.10.002. Mironov, V. L., M. C. Dobson, V. H. Kaupp, S. a. Komarov, and V. N. Kleshchenko (2004), Generalized refractive mixing dielectric model for moist soils, IEEE Trans. Geosci. Remote Sens., 42(4), 773-785, doi:10.1109/TGRS.2003.823288. Mitchard, E. T. A., et al. (2009), Using satellite radar backscatter to predict aboveground woody biomass: A consistent relationship across four different African landscapes, Geophys. Res. Lett., 36(23), 1-6, doi:10.1029/2009GL040692. Moghaddam, M., Y. Rahmat-Samii, E. Rodriguez, D. Entekhabi, J. Hoffman, D. Moller, L. E. Pierce, S. Saatchi, and M. Thomson (2007), Microwave Observatory of Subcanopy and Subsurface (MOSS): A Mission Concept for Global Deep Soil Moisture Observations, IEEE Trans. Geosci. Remote Sens., 45(8), 2630-2643, doi:10.1109/TGRS.2007.898236. Montaldo, N., R. Rondena, J. D. Albertson, and M. Mancini (2005), Parsimonious modeling of vegetation dynamics for ecohydrologic studies of water-limited ecosystems, Water Resour. Res., 41(10), 1-16, doi:10.1029/2005WR004094. Myneni, R., F. Hall, P. Sellers, and A. Marshak (1995), The Interpretation of Spectral Vegetation Indexes, IEEE Trans. Geosci. Remote Sens., 33(2), 481-486. Narvekar, P. S., D. Entekhabi, S.-b. Kim, and E. G. Njoku (2015), Soil Moisture Retrieval Using L-Band Radar Observations, IEEE Trans. Geosci. Remote Sens., 53(6), 3492-3506. 183 - - , - - - - .- - I - _11- 1.1.1 I " -, National Research Council (Ed.) (2004), Groundwater fluxes across interfaces, The National Academies Press. Novick, K. a., P. C. Stoy, G. G. Katul, D. S. Ellsworth, M. B. S. Siqueira, J. Juang, and R. Oren (2004), Carbon dioxide and water vapor exchange in a warm temperate grassland., Oecologia, 138(2), 259-274, doi:10.1007/s00442-003-1388-z. Oh, Y., K. Sarabandi, and F. T. Ulaby (1992), An empirical model and an inversion technique for radar scattering from bare soil surfaces, Geosci. Remote Sensing, IEEE Trans., 30(2), 370-381. O'Neill, P., S. Chan, E. Njoku, T. Jackson, and R. Bindlish (2012), Algorithm Theoretical Basis Document SMAP Level 2 & 3 Soil Moisture (Passive), Tech. rep., Jet Propulsion Laboratory, California Institute of Technology, Initial Release, v.1. Onier, C., A. Chanzy, A. Chambarel, R. Rouveure, M. Chanet, and H. Bolvin (2011), Impact of Soil Structure on Microwave Volume Scattering Evaluated by a TwoDimensional Numerical Model, IEEE Trans. Geosci. Remote Sens., 49(1), 415425. Orellana, F., P. Verma, S. P. Loheide, and E. Daly (2012), Monitoring and modeling water-vegetation interactions in groundwater-dependent ecosystems, Rev. Geophys., 50(3), RG3003--. Owe, M., R. de Jeu, and J. Walker (2001), A methodology for surface soil moisture and vegetation optical depth retrieval using the microwave polarization difference index, Geosci. Remote Sensing, IEEE Trans., 39(8), 1643-1654. Owe, M., R. de Jeu, and T. Holmes (2008), Multisensor historical climatology of satellite-derived global land surface moisture, J. Geophys. Res. Earth Surf., 113(F1), doi:10.1029/2007JF000769. Patton, J., and B. Hornbuckle (2013), Initial Validation of SMOS Vegetation Optical Thickness in Iowa, IEEE Geosci. Remote Sens. Lett., 10(4), 647-651. Peplinski, N. R., F. T. Ulaby, and M. C. Dobson (1995), Dielectric properties of soils in the 0.3-1.3-GHz range, IEEE Trans. Geosci. Remote Sens., 33(3), 803-807. Piepmeier, J. R., S. Member, L. Hong, F. A. Pellerano, and S. Member (2015), Aquarius L-Band Microwave Radiometer : 3 Years of Radiometric Performance and Systematic Effects, IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., pp. 1-8. Piles, M., M. Vall-llossera, A. Camps, M. Talone, and A. Monerris (2010), Analysis of a Least-Squares Soil Moisture Retrieval Algorithm from L-band Passive Observations, Remote Sens., 2(1), 352-374, doi:10.3390/rs2010352. 184 Piles, M., K. A. McColl, D. Entekhabi, N. Das, and M. Pablos (2015), Sensitivity of Aquarius Active and Passive Measurements Temporal Covariability to Land Surface Characteristics, IEEE Trans. Geosci. Remote Sens., 53(8), 4700-4711. Pinel, N., J. T. Johnson, and C. Bourlier (2011), Fully polarimetric scattering from random rough layers under the geometric optics approximation: geoscience applications, Radio Sci., 46, RSOE20--. Poulter, B., et al. (2014), Contribution of semi-arid ecosystems to interannual variability of the global carbon cycle., Nature, 509(7502), 600-603, doi: 10.1038/naturel3376. Prigent, C., F. Aires, W. Rossow, and E. Matthews (2001), Joint characterization of vegetation by satellite observations from visible to microwave wavelengths: A sensitivity analysis, J. Geophys. Res., 106(D18), 20,665, doi:10.1029/2000JD900801. Qin, J., K. Yang, N. Lu, Y. Chen, L. Zhao, and M. Han (2013), Spatial upscaling of in-situ soil moisture measurements based on MODIS-derived apparent thermal inertia, Remote Sens. Environ., 138, 1-9, doi:10.1016/j.rse.2013.07.003. Rahmoune, R., P. Ferrazzoli, Y. H. Kerr, and P. Richaume (2013), SMOS level 2 retrieval algorithm over forests: Description and generation of global maps, IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 6(3), 1430-1439, doi: 10. 1109/JSTARS.2013.2256339. Raju, S., A. Chanzy, J. Wigneron, J. Calvet, Y. Kerr, and L. Laguerre (1995), Soil moisture and temperature profile effects on microwave emission at low frequencies, Remote Sens. Environ., 54(2), 85-97, doi:10.1016/0034-4257(95)00133-L. Rao, K. S., S. Raju, and J. R. Wang (1993), Estimation of soil moisture and surface roughness parameters from backscattering coefficient, Geosci. Remote Sensing, IEEE Trans., 31(5), 1094-1099. Refsgaard, J. C. (2000), Spatial Patterns in Catchment Hydrology, chap. Towards a, pp. 329-354, Cambridge University Press. Reichle, R. H., and Q. Liu (2014), Observation-Corrected Precipitation Estimates in GEOS-5, NASA/TM2014-104606, Vol. 35. Reichle, R. H., R. D. Koster, J. Dong, and A. A. Berg (2004), Global Soil Moisture from Satellite Observations, Land Surface Models, and Ground Data: Implications for Data Assimilation, J. Hydrometeor., 5(3), 430-442, doi:10.1175/15257541(2004)005<0430:GSMFSO>2.0.CO;2. Reichle, R. H., R. D. Koster, G. J. M. De Lannoy, B. A. Forman, Q. Liu, S. P. P. Mahanama, and A.. Tour6 (2011), Assessment and Enhancement of MERRA Land Surface Hydrology Estimates, J. Clim., 24(24), 6322-6338, doi:10.1175/JCLI-D10-05033.1. 185 Renard, N., and S. Bourennane (2009), Dimensionality Reduction Based on Tensor Modeling for Classification Methods, IEEE Trans. Geosci. Remote Sens., 47(4), 1123-1131. Robock, A., K. Y. Vinnikov, G. Srinivasa, J. K. Entin, S. E. Hollinger, N. A. Speranskaya, S. Liu, A. Namkhai, and G. Srinivasan (2000), The Global Soil Moisture Data Bank, Bull. Am. Meteorol. Soc., 81(6), 1281-1299, doi:10.1175/15200477(2000)081<1281:TGSMDB>2.3.CO;2. Rodgers, C. D. (2000), Inverse Methods for Atmopsheric Sounding: Theory and Practice, 238 pp., World Scientific, Singapore. Saatchi, S., S. Asefi-Najafabady, Y. Malhi, L. E. 0. C. Aragdo, L. 0. Anderson, R. B. Myneni, and R. Nemani (2012), Persistent effects of a severe drought on Amazonian forest canopy, Proc. Natl. Acad. Sci., pp. --. Saatchi, S. S., and K. C. McDonald (1997), Coherent effects in microwave backscattering models for forest canopies, IEEE Trans. Geosci. Remote Sens., 35(4), 10321044, doi:10.1109/36.602545. Schenk, H. J., and R. B. Jackson (2002), The global biogeography of roots, Ecol. Monogr., 72(3), 311-328. Schimel, D., B. B. Stephens, and J. B. Fisher (2014), Effect of increasing CO 2 on the terrestrial carbon cycle, Proc. Natl. Acad. Sci., 112(2), 436-441, doi: 10. 1073/pnas.1407302112. Schimel, D., et al. (2015), Observing terrestrial ecosystems and the carbon cycle from space, Glob. Chang. Biol., pp. n/a-n/a, doi:10.1111/gcb.12822. Schlesinger, W. H., and S. Jasechko (2014), Transpiration in the global water cycle, Agric. For. Meteorol., 189-190, 115-117, doi:10.1016/j.agrformet.2014.01.011. Schuler, D. L., J. S. Lee, and T. L. Ainsworth (1999), Compensation of terrain azimuthal slope effects in geophysical parameter studies using polarimetric SAR data, Remote Sens. Environ., 69(2), 139-155, doi:10.1016/S0034-4257(99)00017-6. Schultz, H. R. (2003), Differences in hydraulic architecture account for nearisohydric and anisohydric behaviour of two field-grown Vitis vinifera L. cultivars during drought, Plant, Cell Environ., 26(8), 1393-1405, doi:10.1046/j.13653040.2003.01064.x. Sellers, P. J. (1985), Canopy reflectance, photosynthesis and transpiration, Int. J. Remote Sens., 6(8), 1335-1372, doi:10.1080/01431168508948283. Seneviratne, S. I., T. Corti, E. L. Davin, M. Hirschi, E. B. Jaeger, I. Lehner, B. Orlowsky, and A. J. Teuling (2010), Investigating soil moisture-climate interactions in a changing climate: A review, Earth-sci Rev., 99(3aA4), 125-161, doi: 10.1016/j.earscirev.2010.02.004. 186 Shi, J., T. Jackson, J. Tao, J. Du, R. Bindlish, L. Lu, and K. S. Chen (2008), Microwave vegetation indices for short vegetation covers from satellite passive microwave sensor AMSR-E, Remote Sens. Environ., 112(12), 4285-4300, doi: http://dx.doi.org/10.1016/j.rse.2008.07.015. Simpson, E. H. (1949), Measurement of Diversity, Nature, p. 688. Singh, J., and V. K. Singh (1992), Phenology of seasonally dry tropical forest, Curr. Sci., 63(11), 684-689. Skelton, R. P., A. G. West, and T. E. Dawson (2015), Predicting plant vulnerability to drought in biodiverse regions using functional traits, Proc. Natl. Acad. Sci., 112(18), 201503,376, doi:10.1073/pnas.1503376112. Soil Survey Staff, Natural Resources Conservation Service, U. S. D. o. A. (), Web Soil Survey. Song, K., X. Zhou, and Y. Fan (2010), Retrieval of soil moisture content from microwave backscattering using a modified IEM model, Prog. Electromagn. Res. B, 26, 383-399. Sperry, J. S., and D. M. Love (2015), What plant hydraulics can tell us about responses to climate-change droughts, New Phytol., 207, 14-27. Sperry, J. S., and M. T. Tyree (1988), Mechanism of water stress-induced xylem embolism., Plant Physiol., 88(3), 581-587, doi:10.1104/pp.88.3.581. Sperry, J. S., U. G. Hacke, R. Oren, and J. P. Comstock (2002), Water deficits and hydraulic limits to leaf water supply, Plant, Cell Environ., 25(2), 251-263, doi: 10.1046/j.0016-8025.2001.00799.x. Staelin, D. H., A. W. Morgenthaler, and J. A. Kong (1998), Electromagnetic Waves, 562 pp., Prentice Hall, Upper Saddle River, NJ. Steele-Dunne, S. C., J. Friesen, and N. Van De Giesen (2012), Using diurnal variation in backscatter to detect vegetation water stress, IEEE Trans. Geosci. Remote Sens., 50(7 PARTi), 2618-2629, doi:10.1109/TGRS.2012.2194156. Stoy, P. C., and T. Quaife (2015), Probabilistic Downscaling of Remote Sensing Data with Applications for Multi-Scale Biogeochemical Flux Modeling, PLoS One, 10(6), e0128,935, doi:10.1371/journal.pone.0128935. Tabatabaeenejad, A., and M. Moghaddam (2006), Bistatic scattering from threedimensional layered rough surfaces, IEEE Trans. Geosci. Remote Sens., 44(8), 2102-2114, doi:10.1109/TGRS.2006.872140. Tabatabaeenejad, A., and M. Moghaddam (2011), Radar Retrieval of Surface and Deep Soil Moisture and Effect of Moisture Profile on Inversion Accuracy, Geosci. Remote Sens. Lett. IEEE, 8(3), 478-482, doi:10.1109/LGRS.2010.2087741. 187 Tabatabaeenejad, A., M. Burgin, X. Duan, and M. Moghaddam (2015), P-Band Radar Retrieval of Subsurface Soil Moisture Profile as a Second-Order Polynomial : First AirMOSS Results, IEEE Trans. Geosci. Remote Sens., 53(2), 645-658. Teuling, a. J., S. I. Seneviratne, C. Williams, and P. a. Troch (2006), Observed timescales of evapotranspiration response to soil moisture, Geophys. Res. Lett., 33(23), 0-4, doi:10.1029/2006GL028178. Thomsen, J. E., G. Bohrer, A. M. Matheny, V. Y. Ivanov, L. He, H. J. Renninger, and K. V. R. Schdfer (2013), Contrasting hydraulic strategies during dry soil conditions in Quercus rubra and Acer rubrum in a sandy site in Michigan, Forests, 4(4), 1106-1120, doi:10.3390/f4041106. Tikhonov, A., and V. Arsenin (1977), Solutions of Ill-Posed Problems, Halsted Press, New York. Truong-Lol, M.-L., S. Saatchi, and S. Jaruwatanadilok (2015), Soil Moisture Estimation Under Tropical Forests Using UHF Radar Polarimetry, Geosci. Remote Sensing, IEEE Trans., 53(4), 1718-1727. Tucker, C. J. (1979), Red and photographic infrared linear combinations for monitoring vegetation, Remote Sens. Environ., 150, 127-150, doi:10.1016/00344257(79)90013-0. Tyree, M. T., and J. S. Sperry (1989), Cavitation and Embolism, Water, pp. 19-38. Ulaby, F. T., and D. G. Long (2014), Microwave Radar and Radiometric Remote Sensing, University of Michigan Press, Ann Arbor. Ulaby, F. T., R. K. Moore, and A. K. Fung (1986a), Radar Remote Sensing and Surface Scattering and Emission Theory, Microwave Remote Sensing: Active and Passive, vol. II, Artech House, Norwood, MA. Ulaby, F. T., R. K. Moore, and A. K. Fung (1986b), From theory to applications, Microwave Remote Sensing: Active and Passive, vol. III, Artech House, Norwood, MA. Ulaby, F. T., K. Sarabandi, McDonald Kyle, Whitt Michael, and M. C. Dobson (1990), Michigan microwave canopy scattering model, Int. J. Remote Sens., 11(7), 1223-1253, doi:10.1080/01431169008955090. Van de Griend, A. A., and M. Owe (1994), Microwave Vegetation Optical Depth and Inverse Modeling of Soil Emissivity Using Nimbus SMMR Satellite-Observations, Meteorol. Atmos. Phys., 54(1-4), 225-239. Van De Griend, A. A., and J.-P. Wigneron (2004), The b-factor as a function of frequency and canopy type at H-polarization, IEEE Trans. Geosci. Remote Sens., 42(4), 786-794, doi:10.1109/TGRS.2003.821889. 188 van der Molen, M. K., et al. (2011), Drought and ecosystem carbon cycling, Agric. For. Meteorol., 151(7), 765-773, doi:10.1016/j.agrformet.2011.01.018. van der Sanden, J. J. (2004), Anticipated applications potential of RADARSAT-2 data, Can. J. Remote Sens., 30(3), 369-379. van der Schalie, R., R. Parinussa, L. Renzullo, A. van Dijk, C.-H. Su, and R. de Jeu (2015), SMOS soil moisture retrievals using the land parameter retrieval model: Evaluation over the Murrumbidgee Catchment, southeast Australia, Remote Sens. Environ., 163, 70-79, doi:10.1016/j.rse.2015.03.006. Van Emmerik, T., S. C. Steele-Dunne, J. Judge, and N. van de Giesen (2014), Impact of diurnal variation in vegetation water content on radar backscatter of maize during water stress, IEEE Trans. Geosci. Remote Sens., 16(7), 6070, doi: 10.1109/TGRS.2014.2386142. A Closed-form Equation for M. T. (1980), van Genuchten, Soilsi, Unsaturated of Conductivity the Hydraulic dicting 10.2136/sssaj1980.03615995004400050002x. Predoi: Van Zyl, J., and Y. Kim (2011), Synthetic Aperture Radar Polarimetry, JPL Space Science and Technology Series, 312 pp., Wiley, Hoboken, NJ. Vereecken, H., J. A. Huisman, H. Bogena, J. Vanderborght, J. A. Vrugt, and J. W. Hopmans (2008), On the value of soil moisture measurements in vadose zone hydrology: a review, Water Resour. Res., 44, 1-21, doi:10.1029/2008WR006829. Verhoest, N. E. C., H. Lievens, W. Wagner, J. Alvarez Mozos, M. S. Moran, and F. Mattia (2008), On the Soil Roughness Parameterization Problem in Soil Moisture Retrieval of Bare Surfaces from Synthetic Aperture Radar, Sensors, 8, 4213-4248, doi: 10.3390/s8074213. Walker, J. P., G. R. Willgoose, and J. D. Kalma (2002), Three-dimensional soil moisture profile retrieval by assimilation of near-surface measurements: Simplified Kalman filter covariance forecasting and field application, Water Resour. Res., 38(12), 1301--. Wand, M. P. (1997), Data-based choice of histogram bin width, Am. Stat., 51 (1), 59-64. Wang, J. R., and T. J. Schmugge (1980), An Empirical Model for the Complex Dielectric Permittivity of Soils as a Function of Water Content, IEEE Trans. Geosci. Remote Sens., GE-18(4), 106-107, doi:10.1109/TGRS.1980.350304. Watanabe, S. (1960), Information Theoretical Analysis of Multivariate Correlation, IBM J. Res. Dev., 4 (1), 66-82, doi:10.1147/rd.41.0066. 189 Western, A. W., G. Bloschl, and R. B. Grayson (2001), Toward capturing hydrologically significant connectivity in spatial patterns, Water Resour. Res., 37(1), 83-97. Wieder, R. K., and S. J. Wright (1995), Tropical Forest Litter Dynamics and Dry Season Irrigation on Barro Colorado Island, Panama, Ecology, pp. 1971-1979. Wigneron, J. P., P. Waldteufel, a. Chanzy, J. C. Calvet, and Y. Kerr (2000), Two-dimensional microwave interferometer retrieval capabilities over land surfaces (SMOS Mission), Remote Sens. Environ., 73(3), 270-282, doi:10.1016/S00344257(00)00103-6. Wigneron, J.-P., M. Parde, P. Waldteufel, A. Chanzy, Y. Kerr, S. Schmidl, and N. Skou (2004), Characterizing the Dependence of Vegetation Model Parameters on Crop Structure, Incidence Angle, and Polarization at L-Band, IEEE Trans. Geosci. Remote Sens., 42, 416--425, doi:10.1109/TGRS.2003.817976. Wofsy, S. C., M. L. Goulden, J. W. Munger, S. M. Fan, P. S. Bakwin, B. C. Daube, S. L. Bassow, and F. a. Bazzaz (1993), Net Exchange of C02 in a Mid-Latitude Forest., Science, 260(5112), 1314-1317, doi:10.1126/science.260.5112.1314. Woodhouse, I. H., E. T. A. Mitchard, M. Brolly, D. Maniatis, and C. M. Ryan (2012), Radar backscatter is not a 'direct measure' of forest biomass, Nat. Clim. Chang., 2(8), 556-557, doi:10.1038/nclimatel60l. Xu, L., and D. B. Baldocchi (2003), Seasonal trends in photosynthetic parameters and stomatal conductance of blue oak (\emph{{Q}uercus douglasii}) under prolonged summer drought and high temperature, Tree Physiol, 23(13), 865-877. Yamaguchi, Y., T. Moriyama, M. Ishido, and H. Yamada (2005), Four-component scattering model for polarimetric SAR image decomposition, IEEE Trans. Geosci. Remote Sens., 43(8), 1699-1706. Yarovoy, A. G., R. V. de Jongh, and L. P. Ligthart (2000), Scattering properties of a statistically rough interface inside a multilayered medium, Radio Sci., 35(2), 455-462. Zhang, Q., S. Manzoni, G. Katul, A. Porporato, and D. Yang (2014), The hysteretic evapotranspirationdATVapor pressure deficit relation, J. Geophys. Res. Biogeosciences, 119, 125-140, doi: 10.1002/2013JG002484.Received. Zhou, L., et al. (2014), Widespread decline of Congo rainforest greenness in the past decade., Nature, 509(7498), 86-90, doi:10.1038/naturel3265. Zimmermann, D., R. Reuss, M. Westhoff, P. Geg ner, W. Bauer, E. Bamberg, F. W. Bentrup, and U. Zimmermann (2008), A novel, non-invasive, online-monitoring, versatile and easy plant-based probe for measuring leaf water status, J. Exp. Bot., 59(11), 3157-3167, doi:10.1093/jxb/ernl7l. 190 Zribi, M., A. Gorrab, N. Baghdadi, Z. Lili-Chabaane, and B. Mougenot (2014), Influence of Radar Frequency on the Relationship Between Bare Surface Soil Moisture Vertical Profile and Radar Backscatter, Geosci. Remote Sens. Lett. IEEE, 11 (4), 848-852. Zweifel, R., H. Item, and R. Hdsler (2001), Link between diurnal stem radius changes and tree water relations, Tree Physiol., 21(12-13), 869-877. Zweifel, R., L. Zimmermann, F. Zeugin, and D. M. Newbery (2006), Intra-annual radial growth and water relations of trees: Implications towards a growth mechanism, J. Exp. Bot., 57(6), 1445-1459, doi:10.1093/jxb/erjl25. 191