Detection, Mapping, and Monitoring of  Sudden Oak Death Using Hyperspectral Imagery    Final Report

advertisement
 Detection, Mapping, and Monitoring of Sudden Oak Death Using Hyperspectral Imagery Final Report April 2008 Institute for Technology Development Bldg. 1103, Suite 118 1401 Regency Dr. E., Suite B Stennis Space Center, MS 39529 Savoy, IL 61874 Acknowledgements This work was supported by NASA through a grant from the U.S. Forest Service in support of their Threat Assessment Program. The Institute for Technology Development’s Healthy Forest Project Team: Ken Copenhaver, Program Director Jon Fridgen, Project Manager Karly Hellrung, Image Analyst Shilpa Venkataraman, Image Analyst Additionally, the project team wishes to thank the following for their support and collaboration on this work: Jerry Beatty, WWETAC – U.S. Forest Service Terry Shaw, WWETAC – U.S. Forest Service Alan Kanaskie, Oregon Dept. of Forestry Mike McWilliams, Oregon Dept. of Forestry Chris Lee, Univ. of California Cooperative Extension Yana Valachovic, Univ. of California Cooperative Extension Rodney McKellip, NASA‐Stennis Space Center ii
Contents EXECUTIVE SUMMARY..........................................................................................................................v 1.0 BACKGROUND ............................................................................................................................. 1 2.0 GOALS AND OBJECTIVES.......................................................................................................... 3 3.0 LOCATION..................................................................................................................................... 3 3.1 Oregon Site Description .................................................................................................................. 4 3.2 California Site Description.............................................................................................................. 5 4.0 METHODOLOGY.......................................................................................................................... 6 4.1 Image Data Collection ..................................................................................................................... 6 4.1.1 Hyperspectral Imagery ...…………………………………………………………………6 4.1.2 Satellite Imagery …………………………………………………………………………...7 4.2 Ground Data Collection .................................................................................................................. 8 4.3 Image Processing.............................................................................................................................. 8 4.3.1 Hyperspectral Imagery............................................................................................................. 8 4.3.1.1 Conversion to At‐Sensor Radiance……………………………………………………..8 4.3.1.2 Atmospheric Calibration………………………………………………………………...8 4.3.1.3 Image Geometric Correction ...…………………………………………………………9 4.3.2 Satellite Imagery...................................................................................................................... 10 4.3.2.1 ASTER Data……………………………………………………………………………. 10 4.3.2.2 MODIS Data……………………………………………………………………………..11 4.4 Data Analysis................................................................................................................................. .12 4.4.1 Hyperspectral Imagery……………………………………………………………………....12 4.4.2 Satellite Imagery……………………………………………………………………………...14 5.0 RESULTS AND DISCUSSION ................................................................................................... 14 5.1 Ground Data ................................................................................................................................... 14 5.2 Statistical Analysis ......................................................................................................................... 15 5.2.1. October 2006 Imagery............................................................................................................ 16 5.2.2 July 2007 Imagery.................................................................................................................... 17 5.2.3 October 2006 and July 2007 Differences............................................................................... 19 iii
5.2.4 Temporal MODIS Datasets………...………………………………………………………..19 5.3 Image Analysis ............................................................................................................................... 22 5.3.1 Forest Land Cover Classification .......................................................................................... 22 5.3.2 Identification of SOD .............................................................................................................. 24 5.3.3 ASTER Imagery………………………………………………………………………………27 6.0 CONCLUSIONS........................................................................................................................... 30 7.0 REFERENCES ............................................................................................................................... 31 iv
EXECUTIVE SUMMARY This report summarizes the analysis and results of a project implemented to demonstrate the use of remotely sensed imagery for improving early warning system capabilities as it relates to the identification and mapping of infectious plant diseases in a forest ecosystem. Two candidate study areas were identified: one in Humboldt County California and the other in Curry County Oregon both of which have had significant amounts of Sudden Oak Death Syndrome (SOD) over the past years. Hyperspectal imagery was acquired on two occasions over both study sites and the data analyzed for its ability to discern SOD from the surrounding forest canopy. Satellite imagery was also obtained from the ASTER and MODIS sensors and analyzed in a multi‐tier/multi‐resolution approach for mapping the extent of SOD. Results indicate that significant differences do exist between healthy and diseased tanoak trees, but only in the later stages of disease development. No significant differences were observed between healthy tanoak and tanoak exhibiting the early symptoms of SOD infection. Hyperspectral imagery proved especially useful for mapping the locations of some of the SOD host species (i.e., tanoak, California Bay Laurel) as very good classification accuracies were obtained. The ASTER datasets also worked well for discriminating between different forest land cover types, especially at the Oregon study site. The reduced spatial resolution of the MODIS and ASTER datasets was found to be a limiting factor not only for accurate land cover mapping in California but also for the identification of SOD at both study sites.
v
1.0 BACKGROUND Sudden oak death (SOD) caused by the invasive non‐native pathogen Phytophthora ramorum, is the name given to an epidemic tree disease, first detected in 1995. It is lethal to a number of oak species: coast live oak (Quercus agrifolia), California black oak (Q. kelloggii), and Shreve oak (Q. parvula var. shrevei).and tanbark oak (Lithocarpus densiflorus) (McPherson et al., 2001). The disease is currently known to exist in the coastal ranges in California and several locations in southern Oregon (Kelly, 2001). Since first detected, over one million tanoaks, California black oaks, and coast live oaks have died due to SOD infection (California Oak Mortality Task Force, 2007a). The symptoms of P. ramorum infection vary among plant species and symptoms are not consistent within any given species. One characteristic of the infection in oak species is the sudden simultaneous leaf death on a major stem or of the entire tree, thus, the term sudden oak death. Infected trees may exhibit ‘bleeding’ of sap from apparently intact bark or through cankers that develop on the tree from near the soil level up to as high as 4 m (California Oak Mortality Task Force, 2007b). Several other organisms are often associated with P. ramorum infection, including ambrosia beetles (Monarthrum scutellare and M. dentiger) and the saprophytic fungus Hypoxylon thuorsianum (Goheen et al., 2006). Black, dome‐shaped fruiting bodies may appear near the P. ramorum canker and on other portions of the bole of the tree (Garbelotto et al., 2001). At present, the only definitive ways to diagnose a SOD infection in a tree are by culturing the pathogen or by amplifying the DNA using polymerase chain reaction (PCR) techniques (Kelly, 2001). Since it was first detected, SOD has had a significant economic impact. At a cost of more than $15 million, over 1.6 million plants were investigated by state agencies and USDA‐APHIS for SOD. In one Southern California nursery, over one million camellias valued at nearly $9 million, were destroyed because of SOD infection (Alexander, 2005). California exports nearly half its nursery production (i.e., plants and flowers) to markets in other states or countries, while Oregon exports an estimated 90% if its production. Canada is a major market for that nursery stock and initially, Canada shut its markets to most plant stock from the states of Oregon and California. Without reopened market access, Oregon nurseries alone faced losses in sales to Canada of $15 to $20 million. Costs also appear in the form of quarantines and regulations. Of the funds available to address the epidemic, California has dispersed over $1 million dollars to SOD‐infected counties for the removal of trees that pose a threat to life, property, and/or public works (Cole, 2001). In Oregon, more than $1.5 million has been spent in an attempt to eradicate the pathogen from forests. 1
There are many examples of the use of remotely sensed imagery to monitor forest health. Although a majority of the work has focused on coniferous forests, remote sensing has also been utilized to investigate the health and structure of hardwood forests (Boyer et al., 1998; Everitt et al., 1999; Gong et al., 1999). High‐resolution, multispectral imagery has been used for the identification of SOD in the Western forests (Kelly, 2001; Kelly, 2003; Kelly et al., 2004). The seasonality of image acquisition is critical for accurate monitoring of this disease, as late‐spring through summer appears to be the ideal time for data acquisition. The specific months may vary from year‐to‐year, but imagery must be acquired after complete leaf expansion in the oaks, but before California buckeye (Aesculus californica) begins its summer deciduous process. When buckeye begins to change color, it can appear very similar to Sudden Oak Death affected trees when viewed from a remote location. In some areas of California, buckeye may begin changing in late June (Kelly, 2003). Sudden Oak Death has several characteristics that make a monitoring approach that combines remote sensing with fieldwork ideal. Generally, as trees with the disease die, the entire crown changes dramatically from healthy green to light brown over a relatively short period of time. The change from green to light brown, however, may be more subtle with the canopy progressively yellowing over a period from months to years. After canopy change has occurred, the brown leaves can stay attached to the branches for months, and in areas where SOD is advanced, the affected trees display spatial clumping, with diseased trees likely to be clumped together. This pattern can result in dramatic spectral reflectance changes across broad areas (Kelly, 2001). High spatial resolution remotely sensed data has provided mixed results in mapping and monitoring SOD. Kelly (2003) observed that SOD has spectral reflectance shift across the visible to NIR range as the disease progresses. They found, however, that automated spectral classification methods alone do not provide high accuracies (i.e., 68.8%). None‐the‐less, classifications using the visible and NIR bands combined with Tasseled Cap “greenness” and “wetness” components improved overall accuracy to 81.7% Further work (Kelly et al., 2004) evaluated supervised, unsupervised, and “hybrid” classification methods for discriminating dead and dying tree crowns (as a result of SOD) from the surrounding forest mosaic. The unsupervised and supervised methods provided overall classification accuracies of 66.1‐78.8% and 73.5‐87.9%, respectively. Use of hybrid methods, which incorporate both unsupervised and supervised techniques, resulted in overall accuracies of 93‐95%. 2
The Western Wildland Environmental Threat Assessment Center (WWETAC) in Prineville, OR was designated to coordinate U.S. Forest Service (USFS) efforts in the western U.S. to reduce the likelihood of forest disturbances, mitigate the effects of disturbances, improve information access through centralization, and improve tracking of forest risk and disturbance over space and time. Another center, located in Asheville, NC has a similar directive, with a focus on threats to Eastern forests. The overarching goal of the threat assessment centers is the development of an integrated, national Early Warning System (EWS) to identify, detect, and rapidly respond to forest environmental threats. Although the operational requirements for the EWS are still evolving within the USFS, it is envisioned to be a multi‐source, multi‐scale tool that provides broad area, high frequency reconnaissance; detailed location‐specific analysis using higher‐resolution imagery and ground data; and predictive modeling of potential and emerging forest threats. 2.0 GOALS AND OBJECTIVES The overall goal of this project was to evaluate the use of hyperspectral imagery for improving early warning system capability as it relates to pre‐visual detection of Sudden Oak Death (SOD) syndrome. Specific objectives are to: 1) collect representative ground data from forest species infected and not infected with SOD; 2) perform two hyperspectral image acquisitions over the study area; 3) develop a classification product that separates tanoak and other host species (when applicable) from other forest species; 4) determine the degree of separability between SOD infected and non‐infected species and at what stage of development SOD may first be identified with hyperspectral imagery; 5) identify spectral locations most useful for identifying SOD; and 6) classify and map the extent of SOD in the study areas using multiple analysis techniques. 3.0 LOCATION The project was conducted in Curry (Oregon) and Humboldt (California) counties in the Western U.S. (Figure 1). These areas were selected because of the large number of tanoak at these locations and because Sudden Oak Death (SOD) has been observed in the area. Additionally, some ground data had previously been collected in these areas. 3
OR
Curry County
Humboldt
County
CA
NV
Figure 1. Location of the study areas in Northern California and Southern Oregon. The counties of interest are highlighted in blue. 3.1 Oregon Site Description The Oregon study site was located in southwest Oregon near the city of Brookings in the Siskiyou National Forest. The area resides within the SOD quarantine zone established in Curry County by the Oregon Department of Forestry. For the primary hyperspectral data acquisition that occurred in July, a total of 12 flight lines were established (Figure 2). 4
Figure 2. The SOD study area in southwest Oregon. The 12 flight lines (in yellow) are shown on an ASTER scene acquired on 16 Jun. 2007. 3.2 California Site Description The California study site was located in southern Humboldt County near the small towns of Garberville and Redway. The area has known wide‐spread SOD infestations and includes areas along the Eel River and the Humboldt Redwoods State Park. For the hyperspectral data acquisitions at this location, a total of 18 flight lines were established (Figure 3). 5
Figure 3. The SOD study area in northern California. The 18 flight lines (in yellow) are shown on an ASTER scene acquired on 04 Jul. 2007. 4.0 METHODOLOGY 4.1 Image Data Collection 4.1.1 Hyperspectral imagery At each study site, hyperspectral imagery was collected using the Institute for Technology Development’s (ITD) RDACS‐H4 hyperspectral sensor. The RDACS‐H4 is updated version of the sensor described by Mao (2000) and collects data from 294 to 1173 nm in 6‐nm bandwidths 6
for a total of 150 bands. Imagery was acquired at a spatial resolution of one (1) meter and the image footprint from the sensor covers and area of approximately 1600 × 5500 meters. Two airborne image acquisitions were performed over each study site: the first occurring in late October 2006 and the second occurring in mid‐ to late‐July 2007. The timing and extent of the image acquisitions were largely determined by the project collaborators with input from ITD. For the Oregon site in particular, the image acquisition area was shifted slightly south of the originally planned site as additional SOD infestations were identified via ground and aerial surveys. For the initial data acquisition that occurred in October 2006, imagery was acquired over eight flight lines in Oregon and 18 flight lines in California. For the primary data acquisition that occurred in July 2007, the same 18 flight lines were utilized in California and 12 flight lines slightly to the south and west of the original site were used in Oregon. 4.1.2 Satellite Imagery In an effort to evaluate the multi‐tier/multi‐resolution (MT/MR) approach of the EWS image data was also obtained from satellite data sources. Specifically, imagery was obtained from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) and the Moderate Resolution Imaging Spectrometer (MODIS) sensor. The ASTER data (visible and NIR) was collected at a spatial resolution of 15 m and the MODIS NDVI products utilized in the study were obtained at a spatial resolution of 250 m. ASTER imagery was acquired on multiple dates corresponding to the primary hyperspectral data collection missions in California and Oregon (Table 1) and was downloaded via FTP from the Land Processes Distributed Active Archive Center (LP‐DAAC). The MODIS data products used in the project were provided by NASA collaborator Science Systems and Applications, Incorporated (SSAI) and were acquired from 2000 through 2006. Table 1. ASTER image acquisition dates. Location Acquisition Date Oregon 16 Jun. 2007 California 04 Jul. 2007 7
4.2 Ground Data Collection For both study areas, ground data collection was accomplished using a combination of differentially‐corrected global positioning systems (DGPS) and hand‐held GPS units. When points were collected, a description of the feature being referenced was also included (e.g., tanoak, SOD, California Bay Laurel). All data was stored as ArcView shapefiles and was used later in the data analysis process for the extraction of image spectra from the hyperspectral datasets. In addition to the ground data collected by ITD and the project collaborators, historic SOD infestation data was obtained from the OakMapper online database (www.oakmapper.org) developed and maintained by the University of California‐Berkeley. The database contains the spatial location of every positive confirmation of SOD within the state of California going back to 2000. The OakMapper data was obtained for the purpose of evaluating coarser resolution satellite imagery for mapping and monitoring SOD. 4.3 Image Processing 4.3.1 Hyperspectral Imagery 4.3.1.1 Conversion to At‐Sensor Radiance The first processing step performed after image acquisition is the conversion of the raw image digital numbers (DNs) to at‐sensor radiance. Prior to data acquisition, the RDACS‐H4 camera system was radiometrically calibrated at the NASA Sensor Calibration Lab at the Stennis Space Center. The calibration process involved characterizing every detector on the camera’s CCD array such that an empirical relationship was developed between a noise‐removed DN pixel value and the radiance values measured within the integrating sphere in the calibration lab. NASA collaborator SSAI performed the necessary sensor characterization and provided an executable Matlab program that subtracts the dark‐current values from the raw image DNs and applies the necessary regression equation to convert the DNs to at‐sensor radiance units of [mW/(cm2∙um∙sr)]. 4.3.1.2 Atmosphric Calibration After converting the hyperspectral imagery to at‐sensor radiance, the data must then be corrected for atmospheric effects and converted to relative reflectance. Atmospheric correction of the RDACS‐H4 hyperspectral imagery was performed using the Tafkaa_6S atmospheric model (Montes and Gao, 2004). Tafkaa_6S is based the ATREM 4.0 model (Gao et al., 1993) and 8
uses the Second Simulation of the Satellite Signal in the Solar Spectrum (6S) for the atmospheric scattering calculations. The copy of Tafkaa_6S used for this project was requested through the Naval Research Laboratory (NRL) in Washington, D.C. In preparation for running the model, the at‐sensor radiance data is first scaled by 1000 and converted to 16‐bit integers and stored in a binary band interleaved by line format. To run the model, the user must specify the latitude and longitude of the center of the image, the date and time at which the image was acquired, the average elevation of the ground surface for which imagery was collected, and the altitude at which the imagery was acquired. Additionally, the user must specify the atmospheric ozone content (available at http://toms.gsfc.nasa.gov/ozone/ozone.html) for the general area in which the imagery was collected. Using the specified input parameters, the model uses the atmospheric absorption features in the hyperspectral data (i.e., 940 nm or 820 nm) to compute atmospheric water content on a line‐by‐line basis in the imagery. After removal of atmospheric effects, the data are transformed to reflectance (scaled by a value of 10000) and output to a binary band‐interleaved by line file as 16‐bit integers. 4.3.1.3 Image Geometric Correction The final step in the processing of the RDACS‐H4 hyperspectral imagery is geometric correction. To help compensate for the roll, pitch, and yaw of the aircraft, the RDACS‐H4 camera system is mounted in a Zeiss T‐AS gyro‐stabilized mount. Additionally, the camera system is equipped with an inertial navigation system (INS) GPS that measures the roll, pitch, and yaw of the aircraft. The INS measurements are then used in a customized software program that removes any remaining distortion in the imagery caused by the movement of the aircraft. The software operates by creating a simulated pushbroom camera using a three‐axis rotation model with the pitch represented by the rotation of the x‐axis, the roll represented by the rotation of the y‐axis, and the yaw represented by the rotation of the z‐axis (Figure 4). The z‐axis also represents the direction of flight or image acquisition. 9
Yaw
Roll
Direction of flight
Pitch
Figure 4. Representation of the three‐axis model utilized by the distortion removal software. Although the software described above applies a rudimentary geometric correction to the image datasets, it is necessary to perform an additional geometric correction to obtain the desired registration accuracy (i.e., ≤ 0.5 pixel). ENVI (ITT Visual Information Solutions, Boulder, CO) was used to perform a geometric correction based on a second‐order polynomial model and nearest neighbor resampling. The build the appropriate geo‐rectification model, points of discernable features were selected from both the hyperspectral image and a suitable reference image (typically a DOQQ) until the desired registration accuracy was obtained. In some instances, up to 100 points were selected throughout the image obtain the desired accuracy. All images were output in the North American Datum (NAD) of 1983 using the Universal Transverse Mercator (UTM) coordinate system zone 10‐north. 4.3.2 Satellite Imagery 4.3.2.1 ASTER Data After acquisition, ASTER imagery was downloaded from the LP‐DAAC as Level‐1B (L1B) calibrated, byte‐scaled, at‐sensor radiance data. The files are delivered in the HDF file format and consist of multiple bands with wavelengths in the visible, NIR, short‐wave infrared (SWIR), and thermal infrared (TIR) portions of the spectrum. The first step in preparation of the ASTER data is conversion of the byte‐scaled L1B data to floating point radiance values. This is performed with the ASTER calibration utility in ENVI. The utility extracts the necessary 10
conversion/calibration factors from the HDF file and applies them to the dataset (ITT Visual Information Solutions, 2007). Although the L1B datasets are delivered with geolocation information, the accuracy of the product is not adequate for analysis. Thus, the datasets are subjected to an additional geometric correction similar to that used in the preparation of the hyperspectral imagery. Points of discernable features were selected from both the ASTER image and a DOQQ reference image until the desired registration accuracy was obtained (≤ 0.5 pixel). After the selecting the points, the images were warped using a first‐order polynomial model and nearest neighbor resampling. 4.3.2.2 MODIS Data The MODIS datasets were provided to ITD by NASA collaborator SSAI and largely consist of NDVI and phenological datasets derived from the MOD13 (Terra) and MYD13 (Aqua) data products. SSAI developed customized tools using Matlab to process the temporal MODIS datasets and derive phenological parameters. To create the datasets, the Time Series Product Tool (TSPT) was used to calculate NDVI values from MODIS radiance or reflectance data (gridded or swath). After performing the NDVI computations, the output data are subjected to a QA/QC process using the metadata provided with the raw products. Next, TSPT fuses images from the Aqua and Terra satellites and then applies an outlier detection algorithm to remove suspect pixels. After some additional processing to remove noise from the time series, the datasets are layer‐stacked into 3D band sequential (BSQ) files for use in various image processing programs. Once the filtered time series data has been produced by TSPT, the Phenological Parameters Estimation Tool (PPET) is used to generate season‐specific phenological information on a pixel‐
by‐pixel basis. The software extracts the time series data, identifies the growing seasons within each year and then locates specific points within the growing season (Figure 5). 11
Figure 5. Phenology parameters identified by the PPET software using the time series NDVI data from the MODIS sensor. In addition to the computed NDVI values at the growing season points of interest, PPET also determines the day of year (DOY) on which the computed value was observed. The NDVI values along with the corresponding DOY parameters are output in BSQ files for use in image processing software. Data points obtained from the OakMapper online database along with the data points provided by the project collaborators were overlaid on the phenology datasets and the pixel values were extracted. 4.4 Data Analysis 4.4.1 Hyperspectral Imagery To address the project’s objectives, a number of statistical and image processing techniques were utilized. All analyses efforts focused on techniques that may be integrated into the operational environment of the EWS. Because of the nature of the work being performed, the analysis that was conducted relied on a tiered approach. First, analysis was performed to separate the species of interest (i.e., tanoak) from other forest land cover. Second, analysis was 12
performed to identify SOD within the tanoak species. In addition to identifying the tanoak species, the hyperspectral data is also in the process of being analyzed for its ability to identify and map the locations of other SOD host species such as California Bay Laurel. To determine if statistical differences existed between the spectral reflectance data for infected and non‐infected trees and other native land cover, a mixed model (Littell et al., 1996) was utilized. The mixed model analysis was used to characterize the spectral differences between forest land cover/species types as well as the various symptoms expressed by SOD. For each combination of land cover types and/or vegetation health, comparisons were made using the ESTIMATE statements in the MIXED procedure in the SAS System Version 9.0 (SAS Institute, Cary, NC). The analysis was performed on each image date as well as on the reflectance differences that were observed between the two dates. For the image classifications, a majority of the analysis was performed using the Iterative Self‐
Organizing Data Analysis Technique (ISODATA) unsupervised classification algorithm (Tou and Gonzalez, 1974). The ISODATA algorithm was utilized to break the imagery into a large number of classes (at least 60 initial classes). After the initial cluster development, the classes were labeled and merged to identify tanoak and/or other species with SOD using statistical techniques and ground data collected by the project team. If shadows were prevalent in the data, they were labeled as such using unsupervised techniques. Accuracy assessments were performed using a portion of the ground data that was set aside prior to the start of analysis. Because of the less‐diverse species matrix found in Oregon, a majority of the analyses have focused on that study site. It is hoped that some of the spectral signatures developed from that study site will be useful in identifying some of the species of interest as well as SOD in California. Due to the stray‐light issues that were observed in the hyperspectral data, all analyses have been performed with a 75‐band subset of the 150 bands that were originally collected by the RDACS‐H4 camera system. The 75‐band spectral subset used in all analyses covered a spectral range of 400 to 850 nm. During the latter part of the project, staff from SSAI and ITD were able to effectively characterize the stray‐light effects and developed a post‐acquisition software program to remove the second‐order effects from the data. However, because the project focus shifted toward the MT/MR approach, only a limited amount of analysis was performed on the corrected data. 13
4.4.2 Satellite Imagery Analysis of the MODIS phenological parameters utilized an approach similar to that described by Lange et al. (1992) in which disease progression was modeled over time (i.e., a longitudinal study). Based on the characteristics of the SOD pathogen in tanoak, one would expect the NDVI values to decline after the infection has occurred (i.e., indicating stress). Thus, the MIXED procedure in SAS was used to analyze the temporal trend in the phenological parameters following SOD infection. For each subject (i.e., ground data point) in the analysis, phenology parameters recorded prior to the onset of SOD infestation were used as baseline values. After the SOD infestation was observed, MIXED essentially modeled a new slope for the trend of the phenological parameters. The statistical significance of the post‐infection slope parameter was then used to determine if the imagery was able to distinguish the presence of the disturbance. Similar to the hyperspectral imagery, the ASTER datasets were evaluated for their ability to: 1) differentiate different forest species; and 2) identify and map SOD. Classifications were performed using the supervised Maximum Likelihood classification algorithm (Jensen, 1996) as the number of image bands (3 in ASTER versus 75 in the hyperspectral) was no longer a limiting factor for supervised classification. 5.0 RESULTS AND DISCUSSION 5.1 Ground Data Ground data points were collected by the project cooperators and by ITD staff during visits to the study sites. The points that were collected not only recorded the locations of SOD, but also recorded the locations of other species (e.g., alder, douglas fir, madrone). At the California study site (Figure 6) a total of 321 points were collected; while at the Oregon site, a total of 281 points were collected. 14
Legend
Flight line
Figure 6. Ground data points shown with the location of the flight lines in Humboldt County, CA. Additional ground points were obtained from the OakMapper online database of known SOD infections in coastal California. From 2001 to 2007, nearly 1000 points of known SOD infections (in tanoak only) were obtained and utilized in the analysis of the temporal MODIS datasets prepared by SSAI. 5.2 Statistical Analysis Statistical analysis was performed on data collected in October 2006, July 2007, and on the reflectance differences between the two dates of imagery using the SAS System and the MIXED procedure. Using the ground data points and polygons, spectra were extracted from the hyperspectral data sets and subjected to a mixed model analysis of variance (ANOVA). All results presented are from the Oregon study site. 15
5.2.1. October 2006 Imagery For the dataset acquired in late‐October 2006, little to no significant differences were observed between tanoak and the other forest species. It is unclear at this time if that was a result of the environmental conditions at the time of image acquisition or the phonological characteristics of the species of interest. With the large number SOD infections that were identified in early 2007 by the Oregon project cooperators, comparisons were made between tanoak and trees that were later discovered to be infected with SOD. It was thought that there may be subtle differences that would be observed in the October imagery indicating the onset of the SOD pathogen. The results shown in Figure 7 indicate no significant difference in reflectance between known healthy tanoak and the trees that were later discovered to be infected with SOD. Two conclusions can thus be drawn from this observation: 1) either the imagery was collected prior to the onset of detectable symptoms; or 2) in its early stages, SOD infections are not easily distinguishable from healthy tanoak. Figure 7. Plot of reflectance differences (or lack thereof) between healthy tanoak and trees that were later identified to be infected with SOD. 16
5.2.2 July 2007 Imagery For the data acquired in July 2007, a similar analysis was performed with objective of trying to identify portions of the spectrum in which statistically significant differences were occurring between not only the species of interest, but also between the healthy and diseased tanoak. The spectral differences between the primary forest tree species (i.e., alder, douglas fir, and tanoak) are shown in Figure 8. The results indicate that all three species are clearly separable in the near‐infrared (NIR) portion of the spectrum. In the visible portion of the spectrum, significant differences occur between alder and tanoak as well as between douglas fir and tanoak. It’s also interesting to note that no significant differences occur between alder and douglas fir in the visible portion of the spectrum. Figure 8. Plot of the difference in reflectance versus wavelength for the primary forest species in Oregon. 17
Comparisons were also made between tanoak and the two classes of vegetation health that were observed (Figure 9). The early SOD class shown on this graph represents tanoak trees that were exhibiting the early symptoms of SOD (e.g., bleeding cankers), but showed no visible symptoms in the tree canopy. Incidentally, no significant differences were observed between healthy tanoak and trees exhibiting the early symptoms of the disease. Significant differences were observed, however, between healthy tanoak and those infected with SOD. The trees that were infected with SOD exhibited a significantly lower NIR reflectance values as well as higher reflectance values in the green portion of the spectrum. Figure 9. Plot of the differences between the two categories of SOD and healthy tanoak for the Oregon study site. 18
5.2.3 October 2006 and July 2007 Differences A mixed model ANOVA was also applied to the reflectance differences that were observed between the October 2006 and July 2007 datasets. The results (Figure 10) indicated that from fall to summer, little change is observed in the canopy reflectance of douglas fir. For the tanoak and SOD infested tanoak, the primary changes occurred in the NIR portion of the spectrum, with the SOD infested tanoak exhibiting a greater reduction in NIR reflectance during the summer. Interestingly, the changes observed in the visible portion of the spectrum for healthy tanoak and SOD infested tanoak are relatively similar. Figure 10. Reflectance differences observed between the October 2006 data collection and the July 2007 data collection. 5.2.4 Temporal MODIS Datasets The phenological parameters derived from the temporal MODIS datasets were analyzed using the MIXED procedure in SAS and a random coefficient model (Littell et al., 1996). The model was specified such that a new slope was modeled by MIXED after the SOD infestation was 19
observed in the ground data. The analysis was performed on the phenological parameters using infestations that were recorded in 2002, 2003, and 2004. Using data from these three years ensured that sufficient temporal data was available prior to and after the infestation. The random coefficient model analysis results for several phenological parameters (season maximum NDVI, left 20% (season start), and right 20% (season end) are shown in Figures 11‐13. The plots show mean values of the selected phenological parameters for all years data was available and provide insight into the overall trend in the parameters after SOD infection. In the three years studied for this analysis (2002, 2003, 2004) the slope parameter computed by MIXED were not statistically different from zero. The slope parameter being equal to zero indicates that the MODIS imagery was not able to sufficiently distinguish any potential changes that may have occurred in the forest as a result of SOD. Most likely, this is due to the coarse resolution of the data. 1
0.9
0.8
NDVI
0.7
0.6
0.5
0.4
Season Max
0.3
0.2
L20
0.1
R20
0
2000
2001
2002
2003
2004
2005
2006
Year
Figure 11. Plot of the season maximum NDVI, left 20% NDVI (season start), and right 20% NDVI (season end) versus the year from which they were obtained for SOD infections observed in 2002. 20
1
0.9
0.8
NDVI
0.7
0.6
0.5
0.4
Season Max
0.3
L20
0.2
R20
0.1
0
2000
2001
2002
2003
2004
2005
2006
Year
Figure 12. Plot of the season maximum NDVI, left 20% NDVI (season start), and right 20% NDVI (season end) versus the year from which they were obtained for SOD infections observed in 2003. 1
0.9
0.8
NDVI
0.7
0.6
0.5
0.4
Season Max
0.3
0.2
L20
0.1
R20
0
2000
2001
2002
2003
2004
2005
2006
Year
Figure 13. Plot of the season maximum NDVI, left 20% NDVI (season start), and right 20% NDVI (season end) versus the year from which they were obtained for SOD infections observed in 2004. 21
In general, the MODIS NDVI values appeared to be significantly higher than what would typically be expected in a forest ecosystem. For the MODIS data analyzed, season minimum NDVI values typically ranged from 0.50 to 0.92 and season maximum values ranged from 0.73 to 0.96. Given the location of the data (coastal California and Oregon) and climate of that region, the data values would be expected to be substantially lower. 5.3 Image Analysis Image analysis methods have focused primarily on unsupervised ISODATA classifications using a tiered approach. First, a classification was developed to identify tanoak from all the other forest land cover types. Second, the analysis focused on identifying SOD within the tanoak class that was created during the first step. Because the statistical analysis indicated better overall separability between the various land cover and vegetation health for the data acquired in July 2007, all classifications were performed using the imagery from that time. 5.3.1 Forest Land Cover Classification In Oregon excellent results have been obtained for the identification of the primary forest land cover types using hyperspectral imagery (Figure 14 and Table 2). For the four primary flight lines (OR‐3, OR‐9, OR‐10, OR‐11) from which a majority of the ground data was collected, the overall classification accuracy was 97%, with an accuracy of 90% for the identification of tanoak. Table 2. Classification accuracies for the identification of forest land cover types averaged over the four primary flight lines in Oregon. Class Accuracy (%) tanoak 90 alder 96 Douglas fir 93 Clearing/Burn Area 97 Sparse Vegetation 94 Overall 94 22
Legend
Figure 14. False color composite (left) and unsupervised forest land cover classification map (right) from flight line OR‐10. In California, the results of the forest land cover classification have exhibited somewhat lower accuracies than those observed in Oregon (Table 3). This is likely due to the larger number of species present and the fact that many of tanoaks are often obscured or partially obscured by larger trees such as the California Redwood. Additional ground data that was recently received from the California project collaborators will likely improve these accuracies. 23
Table 3. Forest land cover classification accuracies obtained for the California study site. Class Accuracy (%) California redwood 93 Tanoak 79 California bay laurel 81 White oak 73 Madrone 85 Coast live oak 73 Douglas fir 89 Overall 81 5.3.2 Identification of SOD For the identification of SOD, the tanoak classes that were identified from the previously discussed land cover classification were segmented into additional classes and then used to identify trees potentially infected with SOD. On several flight lines of the Oregon study site, some of the tanoak had been treated with herbicide approximately five weeks for image data collection. Thus, an attempt was also made to define a class that contained the tanoaks that were treated with herbicide. Results from a portion of flight line OR‐11 in Oregon are shown in Figure 15. The classification shows the surrounding healthy vegetation as well as the dead tanoak and the tanoak that are declining from the herbicide treatment. Accuracy assessment results (Table 3) indicate very good accuracies for the mapping of the dead tanoak and the declining tanoak. Overall accuracies were typically above 85% for the classifications identifying tanoak that died as a result of SOD infection. 24
Figure 15. Identification of dead tanoak (from SOD) and declining tanoak from a prior application of herbicide to help control the spread of SOD. Table 3. Classification accuracies for the identification of dead and declining tanoak in Oregon. Class Accuracy (%) Advanced SOD 90 Declining tanoak (herbicide treated) 83 Overall 87 25
For the California study site, it was expected that accuracies for the identification of SOD would be lower than those observed in Oregon. The overall accuracies for the identification of SOD were typically around 80% ‐ agreeing well with the results Kelly et al. (2004) obtained from their study. Figure 16 shows the results of an unsupervised ISODATA classification performed on flight line CA‐9 (July 2007). The zoom area shown in the map indicates two instances where SOD was identified via the ground data and in the imagery. SOD Positive
Ground Truth Point
Legend
Barren/Non-Vegetated
SOD Positive
SOD Likely
Healthy Vegetation
False Color Composite
Classification Map
Figure 16. Unuspervised classification map from flight line CA‐9 in the California study area. 26
5.3.3 ASTER Imagery For the ASTER imagery a similar tiered analysis approach was used. First, a classification was performed to identify the different forest land cover types. After the tanoak was identified from the other forest species, a classification was performed to identify SOD. Classifications were primarily conducted using the Maximum Likelihood algorithm as the amount of ground data was sufficient to train the classifier and perform an accuracy assessment. In Oregon, very good classification accuracies were obtained for discriminating the various land cover types (Figure 17 and Table 4). Overall, the classification had an accuracy of 86% with an accuracy of 84% for tanoak. The use of the coarser resolution data, however, necessitated the addition of a mixed forest class that was comprised of tanoak, alder, and Douglas fir. In some areas, the abundance (or lack thereof) of the three species was such that the resolution of the imagery could not effectively discern a single appropriate species. Figure 17. Forest land cover map produced using ASTER imagery and the Maximum Likelihood classification algorithm. 27
Table 4. Classification accuracies for the identification of forest land cover types in Oregon using ASTER imagery. Class Accuracy (%) tanoak 84 alder 94 Douglas fir 73 Clearing/Burn Area 92 Sparse Vegetation 75 Mixed Forest 80 Water 99 Overall 86 In California, however, the outcome of the forest land cover classification using ASTER imagery was poor (Table 5). The overall classification accuracy was 73% with accuracies of 52% and 63% for the identification of tanoak and California bay laurel, respectively. These accuracies were substantially lower than those observed in Oregon as well as those obtained with the hyperspectral imagery in California. The reduction in spatial resolution associated with ASTER data combined with the diverse species matrix of the California study site appear to require the greater spatial and spectral resolution provided by the hyperspectral imagery. Table 5. Classification accuracies for the identification of forest land cover types in California using ASTER imagery. Class Accuracy (%) California redwood 86 Tanoak 52 California bay laurel 63 White oak 62 Madrone 73 Coast live oak 64 Douglas fir 82 Water 99 Clearing 90 Sparse Vegetation 72 Overall 73 28
After identification of the forest land cover types, an attempt was made to utilize the ASTER datasets for mapping the extent of SOD at both study sites. Again, the supervised Maximum Likelihood algorithm was used to classify the imagery based on ground data collected by the project collaborators and ITD. In general, the resulting classification accuracies for identifying SOD were poor (Table 6). At both study sites, accuracies decreased substantially from those that were observed with the hyperspectral imagery. The accuracies for identifying tanoak with advanced symptoms of SOD were 54% and 47% for Oregon and California, respectively. Although the previously conducted statistical analysis (refer to Figure 8) indicated that the spectral resolution of multispectral imagery may be sufficient for the identification of SOD, the spatial resolution of the ASTER data is likely inadequate as SOD is often found in individual and smaller groups of tanoak. Table 6. Classification accuracies for the identification of SOD using ASTER imagery. Study Site Class Accuracy (%) Oregon Advanced SOD 54 Declining tanoak (herbicide treated) 67 Healthy 60 Overall California Advanced SOD 47 Healthy 56 Overall 60 52 29
6.0 CONCLUSIONS After the September 2006 working meeting in Portland, OR, an EWS component project was established to evaluate the use of remotely sensed imagery for the detection and mapping of SOD. With the assistance of the USFS, project collaborators were identified and two study sites selected: one in Humboldt County, California; the other in Curry County, Oregon. Both areas had and continue to have significant amounts of SOD. Hyperspectral imagery was acquired on two occasions over both study sites and ground data was collected by the project collaborators and ITD for use in all subsequent analyses. The hyperspectral datasets provided excellent results for the identification and mapping of forest land cover types, especially of those that are host species for SOD. SOD detection, however, presented more of a challenge and was only possible in the later stages of infection. From the two dates of imagery acquired (October 2006 and July 2007), the July imagery proved to be the most useful in discriminating between SOD‐infected and healthy tanoak trees (accuracies ≥ 80%). In addition to the hyperspectral datasets, ASTER and MODIS satellite data were also obtained. The MODIS datasets were processed to derive phenological parameters for the 2000‐2006 growing seasons using tools developed by NASA collaborator SSAI. ASTER datasets were obtained during June and July of 2007 which corresponded to the time where the hyperspectral datasets were acquired. The coarser spatial resolution provided by ASTER and the time‐series MODIS products proved largely ineffective in discriminating between healthy and SOD‐
infected tanoak trees (accuracies ≤ 60%). The ASTER data, however, was useful for mapping forest land cover types, especially at the Oregon study site. Despite the larger spatial extent covered by the satellite image datasets, the coarser spatial resolution of the data resulted in significantly lower accuracies than those obtained using airborne hyperspectral imagery. For this particular application, the MT/MR approach may prove difficult to implement as the amount of disturbance would need to be substantial for detection with image datasets at the provided resolution. Alternative datasets with higher spatial resolution and future sensors may improve the feasibility of this application. The classification products developed by this effort will be made available to the project collaborators for use in their ongoing efforts in mitigating the impact of SOD in coastal California and Oregon. 30
7.0 REFERENCES Alexander, J. 2005. Review of Phytophthora ramorum in European and North American nurseries. [Online]. Available at http://www.fs.fed.us/psw/publications/documents/psw_gtr196/psw_gtr196_001c_01Alexander.
pdf (verified 26 Feb. 2007). Sudden oak death 2nd Science Symposium: The state of our knowledge, Monterey, CA. Boyer, M., J. Miller, M. Belanger, and E. Hare. 1988. Senescence and spectral reflectance in leaves of northern pin oak (Quercus palustris Muenchh.). Remote Sens. of Environ., 25:71‐87. California Oak Mortality Task Force. 2007a. Sudden oak death [Online]. Available at http://nature.berkeley.edu/comtf/index.html (accessed 26 Feb. 2007; verified 01 Mar. 2007). Univ. of California, Berkeley. California Oak Mortality Task Force. 2007b. Sudden oak death: Plant symptoms [Online]. Available at http://nature.berkeley.edu/comtf/html/plant_symptoms.html (accessed 26 Feb. 2007; verified 01 Mar. 2007). Univ. of California, Berkeley. Cole, E.F. 2001. The relentless spread of sudden oak death: Tracking a mysterious killer. California Coast and Ocean 17(3):1‐4. Everitt, J.D., D. Escobar, D. Appel, W. Riggs, and M. Davis. 1999. Using airborne digital imagery for detecting oak wilt disease. Plant Dis. 83(6):502‐505. Gao, B.C., K.B. Heidelbrecht, and A.F.H. Goetz. 1993. Derivation of scaled surface reflectances from AVIRIS data. Remote Sens. Environ. 44:165‐178. Garbelotto, M., P. Svihra, and D. Rizzo. 2001. Sudden oak death syndrome fells three oak species. California Agric. 55(1):9. Goheen, E.M., E. Hansen, A. Kanaskie, N. Osterbauer, J. Parke, J. Pscheidt, and G. Chastagner. 2006. Sudden oak death and Phytophthora ramorum: A guide for forest managers, Christmas tree growers, and forest‐ree nursery operators in Oregon and Washington. Publication EM 8877. Oregon State Univ. Ext. Service, Corvallis. 31
Gong, P., X. Mei, G.S. Bignig, and Z. Zhang. 1999. Monitoring oak woodland change using digital photogrammetry. J. of Remote Sens. 3(4):285‐289. ITT Visual Information Solutions. 2007. ENVI User’s Guide. ITT Visual Information Solutions, Inc. Boulder, CO. Jensen, J.R. 1996. Introductory digital image processing: A remote sensing perspective. 2nd ed. Prentice‐Hall, Upper Saddle River, NJ. Kelly, M.N. 2001. Monitoring sudden oak death in California using high resolution imagery [Online]. Available at http://nature.berkeley.edu/comtf/pdf/Bibliography/kelly2002a.pdf (verified 26 Feb. 2007). 5th Symp. on Oak Woodlands: Oaks in California’s changing landscape, San Diego, CA. Kelly, M.N. 2003. Remote sensing of sudden oak death using ADAR imagery. In Proc. 9th Forest Service Rem. Sens. App. Conference: Rapid delivery of remote sensing products, Amer. Soc. Photogram. Remote Sens., San Diego, CA. Kelly, M.N., D. Shaari, Q. Guo, and D. Liu. 2004. A comparison of standard and hybrid classifier methods for mapping hardwood mortality in areas affected by sudden oak death. Photogram. Engr. and Remote Sens. 70(11):1229‐1239. Lange, N., B.P. Carlin, and A.E. Gelfand. 1992. Hierarchical Bayes Models for the progression of HIV infection using longitudinal CD4 T‐Cell numbers (with discussion). J. Amer. Stat. Assoc. 87:615‐632. Littell, R.C., G.A. Milliken, W.W. Stroup, and R.D. Wolfinger. 1996. SAS system for mixed models. SAS Institute, Cary, NC. Mao, C. 2000. Hyperspectral focal plane scanning: An innovative approach to airborne and laboratory pushbroom hyperspectral sensing. p. 424‐428. In Proc. Geospatial Info. in Agric. and Forestry, Vol. 1. ERIM Int’l, Ann Arbor, MI. McPherson, D.L. Wood, A.J. Storer, N. Maggi Kelly, and R.B. Standiford. 2001. Sudden oak death, a new forest disease in California. Integrated Pest Mngmt. Rev., 6:243‐246. 32
Montes, M.J. and B.C. Gao. 2004. Tafkaa_6S: An atmospheric correction algorithm for the land. Remote Sens. Div., Naval Research Laboratory, Washington, D.C. Tou, J.T. and R.C. Gonzalez. 1974. Pattern recognition principles. Addison‐Wesley, Reading, MA. 33
Download