This file was created by scanning the printed publication.
Errors identified by the software have been corrected;
however, some errors may remain.
Spatial Data Infrastructure and
Geostatistical Analysis of Forest CanopyHydrologic Interactions, at the Fraser
Experimental Forest, Colorado, USA 1
H. Todd Mowrer2
Charles A. Troendle3
Gerhard Hunner4
Abstract-Development of a geospatial infrastructure, including
2-meter topography, digital orthophotos, satellite imagery, and
global positioning system (GPS) locations of research plots has
allowed the application of geostatistical methodology in research at
the Fraser Experimental Forest. GPS has allowed field navigation
to truly randomly located plots, georeferencing of long-term
sample locations, and navigation to new sampling locations selected through spatial analyses in the office. This precise location of
sample plots, in turn, allows the application of geostatistical techniques to interpolate point data on forest canopy conditions, integrating ancillary GIS and remote sensing information that is
measured throughout the watershed to improve overall precision.
The goal of this research is to use this spatially interpolated canopy
information to improve predictions of water yield from forested
ecosystems.
The Fraser Experimental Forest, located approximately
200 km northwest of Denver , Colorado, USA, is an 11,000 ha
research area maintained by the USDA Forest Service,
Rocky Mountain Research Station. Established in 1937 to
study the integration of water, vegetation, fish, and wildlife
in high elevation subalpine coniferous forests, much of the
work has centered on determining how forest canopy conditions affect water yield. In recent years, research has indicated that the main component in moisture loss prior to
snowmelt in the spring is canopy interception and subsequent sublimation (Troendle et al. 1993). Thus, the crown
area profile of a forested stand is an important predictor
variable in estimating snow interception in the canopy, and
deposition processes under the forest canopy, and subsequent stream flow in the spring. In the current study, efforts
are focused on spatially correlating forest canopy conditions,
physiography, and other continuously measured GIS and
Ipaper presented at the North American Science Symposium: Toward a
Unified Framework for Inventorying and Monitoring Forest Ecosystem
Resources, Guadalajara, Mexico, November 1-6,1998.
2 H. Todd Mowrer is a Principal Mensurationist at the Rocky Mountain
Research Station, 240 W. Prospect, Fort Collins, Colorado 80526-2098 USA,
tel: +970498-1255, email: tmowrer/rmrs@fs.fed.us
3 Gerhard Hunner is a Doctoral Candidate, Department of Forest Science,
Colorado State University, Fort Collins, Colorado 80523 USA, email:
gerhard@Cnr.colostate.edu
4 Charles A. Troendle is a Principal Hydrologist at the Rocky Mountain
Research Station, 240 W. Prospect, Fort Collins, Colorado 80526-2098 USA,
tel: +970498-121250, email:ctroendlelrmrs@fs.fed.us
USDA Forest Service Proceedings RMRS-P-12. 1999
remote sensing information, with snowpack accumulation
in the spring, and potentially with stream flow. This capability will provide an important research cornerstone to
measuring, monitoring and predicting the influence of overstory vegetation components on hydrologic behavior of the
landscape represented by the Experimental Forest.
Background _ _ _ _ _ _ _ __
Prior to 1995, georeferencing of research projects in the
Experimental Forest was on an ad hoc basis, and was
generally dependent on United States Geological Survey
paper maps. These maps were at a relatively coarse grain
(small cartographic scale) for the work done at Fraser
Experimental Forest, and inconsistencies between map information and actual hypsography were often encountered.
In 1995, a contract was let to (1) provide aerial photography of the entire forest, (2) develop 2-meter contour maps,
(3) provide geographic information system (GIS) thematic
layers, and (4) provide co-registered digital orthophotos at a
2-meter pixel resolution. Simultaneously, 30-meter pixel
Landsat Thematic Mapper (TM) satellite imagery was obtained, and the process of georeferencing key research
locations on the Experimental Forest through the use of
global positioning systems (GPS) was initiated.
The overall plan was to provide geospatial information
at the finest level of resolution possible, given budgetary
restrictions, and the spatial extent of the Experimental
Forest. Two-meter contour information, coupled with 2-meter
pixel black-and-white orthophotos, provided the basis for
this geospatial infrastructure. The assumption was that
information could be aggregated to a coarser level, whereas
the reverse was problematic: fine-resolution interactions
could not always be reliably disaggregated from coarse-level
measurements.
Without the advent of affordable, accurate global positioning system (GPS) equipment, it would have been impossible to implement the research side of this infrastructure.
In 1995, point locations could be located with a GPS satellite
receiver at a nominal 5-meter r.m.s. accuracy (after differential correction). Currently, the same approximate investment in equipment allows differentially corrected position accuracy on the order of 1 to 3 meters. Because of the
remote location and the mountainous terrain, the potential
decimeter accuracy of current equipment can not be realized. The ability to collect information on accurately located
25
plots, and to integrate this discrete point information
with extensive contiguous information, from GIS and satellite imagery, for example, has allowed the application of
geostatistical methodology.
Geostatistical techniques have allowed us to interpolate
point estimates of forest canopy conditions across landscapes, by integrating spatially correlated predictor variables, such as slope, elevation, and aspect derived from
GIS coverages, and multispectral reflectance values obtained from satellite imagery. Geostatistical techniques used
include sequential Gaussian simulation and co-conditional
turning bands simulation. These are useful in estimating
uncertainties in GIS analyses using auto- and cross-correlations in a multivariate spatial context, and disjunctive
kriging which allows the conterminous estimation of discrete and continuous distributions of forest canopy conditions across landscapes. The overall goal is to integrate point
information on both forest canopy and snowfall accumulation with continuous data to improve the database used for
hydrologic predictions.
GPS Survey and Map Data _ __
The contract to obtain spatial data for Fraser Experimental Forest involved five main steps: survey control, aerial
photography, stereoscopy, orthorecitfication, and GIS data
formatting. Prior to any work, it was determined that the
best format for the data was Colorado State Plane coordinates in meters, on the North Zone of the North American
Datum 1983 (NAD83), using the Lambert Conformal Conic
Projection. All photographic and map products were to
provide information for an additional300-meter buffer zone
beyond the Experimental Forest boundary to ensure spatial
continuity for any analysis area. Final results were to meet
second-order, class I horizontal control survey accuracy
standards, with a relative accuracy between points of at
least one part in 50,000 (Wolf and Brinker 1994). After
completion of the GPS survey, the final minimally constrained adjustment in the network evaluation ratio was one
part in 625,000, well exceeding first order accuracy requirements of one in 100,000.
Survey Control and Aerial Photography
The Experimental Forest and surrounding boundary
was divided into nineteen rectangular tiles, 4 kilometers in
the east-west direction, and two and one-half kilometers in
the north-south direction. These tiles partitioned the GIS
information into analytically manageable sizes. Sixteen
survey control points were established throughout the Experimental Forest and immediately surrounding areas.
Precise positions for these horizontal benchmarks were
established by satellite (GPS) survey methods, and the
intermediate topography filled in photogrammetrically. In
two cases, permission had to be obtained from private
landowners before benchmarks and photo panels were installed. Two other locations required a day-long hike to set
the photo panels and obtain the GPS baselines (see Fig. 1).
Upon completion of the aerial photography, all the photo
panels were removed.
Aerial photography at a scale of 1:27,500 provided a
nominal 60 percent endlap (between adjacent photographs
on the same flight line) and a 35 percent sidelap (between
adjacent flight lines) between photographs. This resulted in
three flight lines of nine exposures each, covering the rectangular area encompassing the Experimental Forest. The
photo panels were located so that at least two and sometimes
three photo panels were visible in each photograph. Brass
Figure 1.-Aerial photography panel near Byers Peak on west boundary of the Fraser Experimental Forest.
26
USDA Forest Service Proceedings RMRS-P-12. 1999
caps, centered on the photo panels, were used to permanently benchmark the location of each control point. These
benchmarks will provide horizontal and vertical references
for future work, such as real-time differential GPS correction radio links within the Experimental Forest.
Stereoscopy and Orthorectification
The contractor used a stereoscopic plotter to develop the
digital elevation models. These, in turn, were used to create
topographic sheets with 2-meter contours. Nineteen 1:5000
scale topographic map sheets resulted, each covering a two
and one-half- by four-kilometer area. Additional mylar map
sheets were created by overlaying the contour lines on the
orthophotos, providing simultaneous visual reference for
stand delineation and topography.
Orthophotos are corrected to a uniform scale: perspective
effects due to terrain relief and photo tip and tilt have been
removed. These photographic images were co-registered
with each of the nineteen topographic map sheets. The
digital orthophotos were provided in ERDAS format with a
two-meter pixel resolution in 8-bit gray scale. They have
proved invaluable for ''heads-up'' digitizing from the computer screen, and for stand delineation and location for
subsequent field measurement, as discussed below.
The orthoimage is corrected for relief perspective so that
there is a constant scale across the image. However, the
outward radial lean of objects with vertical relief displacement, such as cliffs or trees, is still a visible distortion, which
increases as one moves outward from the center of the
photograph. Because several aerial photographs had to be
used to develop the orthophoto for each tile, and the edges of
the aerial photographs did not match the edges of the map
tiles; there are locations on the orthophotos where corners of
four aerial photos come together. Such a juncture is visible
in the upper center of Figure 2. Care must be taken that
junctures such as this do not cause misinterpretation of
stand composition and texture when these orthophotos are
used for stand delineation. .
Figure 3.-The same area of Lexen Creek watershed as in Figure 2,
showing topography. White lines are 50-meter contours, derived from
initial 2-meter contours. Deriving slope and elevation surfaces is a
straightforward process from the digital elevation model.
GIS Data Formatting
In addition to the digital orthophotos described above,
thematic layers were provided for GIS use. Ten map themes
were provided in separate ArclInfo files for each of the
nineteen tiles. These thematic layers included hypsography
(Fig. 3), hydrography, structures, roads and trails, tree-line,
barriers (fences), drainage (bridges, culverts, etc.), map
sheet lines, and control points. The hypsography provides
the basis for interpolation of a digi tal elevation model (D EM).
In addition to providing direct information on elevation,
digital elevation models provide the basis for calculation of
continuous slope and aspect information. Hydrography
(permanent streams, lakes, and wetlands) provides a useful
reference for delineation of riparian vegetation. Roads and
trails, as well as hydrographic information, provide useful
visual references for orienting an observer to specific locations. Map sheet lines are invaluable for georeferencing
point, line, and polygon information from other sources.
Control point locations are useful to determine the closest
benchmark to an existing or proposed research site or for
satellite classification (Fig. 4).
Spatially Referenced Experimental
Data
In addition to the spatial information obtained under
contract, described above, additional information has been
collected on the Experimental Forest, both specific to this
particular project and historically as part of long-term research on the Experimental Forest.
Plot-Level Forest Canopy Information
Figure 2.-A portion of a digital orhtophoto showing forested areas of
the Lexen Creek watershed.
USDA Forest Service Proceedings RMRS-P-12. 1999
Since 1992, a summer field crew has been measuring
forest plots, centered in and surrounding the Lexen Creek
watershed of the Experimental Forest. As a first step,
27
.' y. •... :'
Figure 4.-Unsupervised classification of entire Lexen Creek watershed using Landsat Thematic Mapper satellite imagery showing (from
lightest to darkest grey): open ground/bare rock; widely spaced lodgepole pine (pinus contorta); mixed lodgepole pine, Englemann spruce
(picea enge/manil) , and subalpine fir (abies /asiocarpa); and dense
Engelmann spruce/ subalpine fir (darkest grey).
previously measured temporary point sample locations
were reestablished, measured, and monumented. In subsequent seasons, these systematically located sample points
were augmented by randomly located plots to ensure complete sampling across the 121-hectare first-order watershed
(see Fig. 5).
Navigation to Plots Using GPS
GPS location has been critical to this research, first to
provide georeferencing of inventory points using differential
correction, and, more recently, to navigate to randomly
located plots using a Precise Positioning Service P(Y) code
GPS receiver. These P(Y) code GPS receivers are made
available to federal agencies through the Department of
Defense, and receive cryptographically coded signals,
which broadcast position information without the additional errors introduced to civilian C(A) code GPS signals.
The overall effect is to provide "real-time" location information that is accurate to within a 30-meter Landsat TM pixel.
This is particularly useful in remote mountainous areas
where real-time broadcast of differential correction information is unavailable or impractical.
This capability allows navigation in the field to a precise
location that was previously established in the office
through computer analysis of digital spatial information, or
through the more traditional method of determining the
location of a point on a map on a digitizer tablet. This has
been extremely useful in locating sample plots in areas that
have been selected through a combination of data sources
such as satellite imagery and digital orthophotographs, in
conjunction with stereoscopic viewing of aerial photography. The coordinates of the desired plot location can either
be obtained from the cursor location on the digital imagery,
or from a digitizer table. The coordinates are entered in to the
P(Y) code receiver, and the field crew can reliably navigate
to the selected point location to collect their information.
This is a substantial improvement over traditional forest
inventory procedures, where crews relied on coarse topographic and overstory conditions to estimate their location
within a stand. Plots were then systematically located along
compass lines by pacing a predetermined distance between
plots. Thus, plot locations had only nominal certainty with
regard to their location within the stand. It was therefore
necessary to statistically base inventory information on this
"systematic sampling with multiple random starts" methodology, which is generally accepted as only suitable for predicting the mean values of typical forest conditions (Avery
and Burkhart 1994). Through GPS location of forest plot
location, spatial information can be maintained throughout
an analysis, preventing information loss through averaging
of plot information (Mowrer 1997).
Snow Course Information
Figure5.--Global Positioning System (GPS) satellite locations of 821/
125th hectare plots. Lines delineate stands from the orthophoto in
Figure 2.
28
Mountain snowpack accumulation, and subsequent collection and transport of snow melt to supply water for
agricultural and urban uses is of considerable economic
importance in Colorado and elsewhere in the western USA.
Part ofthe importance ofthe Fraser Experimental Forest, is
its long-term historic data set of snow accumulation and
consequent runoff at measured stream locations. Four "snow
courses" have been measured for more than forty years.
These consist oflarge loops, beginning and ending at stream
flow measurement points (v-notch weirs) at the base of the
watershed. Measurements of snow depth and water equivalent are taken at monumented points located 40.24 meters
(2 chains) apart throughout the loop. The on-going process of
GPS-Iocating each of these measurement posts will allow
the spatial interpolation of water equivalents throughout a
watershed using the contiguous GIS information as a function of canopy information measured at point locations.
USDA Forest Service Proceedings RMRS-P-12. 1999
Geostatistical Analysis _ _ _ __
In 1993, the purchase of the first GPS receiver on the
Experimental Forest was coincident with the measurement
of 82 inventory and randomly added plots across the Lexen
Creek watershed. Just being able to visualize the spatial
response surface for percent crown cover across this 121
hectare watershed was a great advantage. Mowrer (1994)
developed a variogram and a kriged response surface for the
watershed. The variogram (or semi-variogram) models the
spatial autocorrelation between values of a particular variable at successively greater separation distances. Kriging
uses this variance model to estimate the minimum variance
combination of weights for adjacent measured values to
estimate the value at an unmeasured location Osaaks and
Srivastava 1989)
Mowrer (1996) extended this methodology using sequential Gaussian simulation to develop multiple, equally probable realizations of input to GIS analyses. Sequential
Gaussian simulation applied the kriging algorithm to each
cell or pixel visited randomly in the area of interest. If the
cell had a measured location in it, that value was adopted for
the cell. If not, the kriging process was used with the
variogram to estimate the set of weights to apply to adjacent measurement points and previously simulated cell
values to create a minimum variance estimate for the point
in question. A random deviate was then drawn from distribution determined by the kriged mean and variance, and
assigned to that cell. This process was repeated as cells were
visited randomly until values for the entire area had been
estimated.
Mowrer (1997) again extended this approach to estimate
the potential areas for old growth across the Lexen Creek
watershed by applying the sequential Gaussian simulation
algorithm to three variables of interest: mean stand diameter, age of dominant and codominant trees, and percent
crown cover. Creating independent realizations of these
three variables, and using them as input to a simple GIS
analysis to select cells meeting a minimum threshold for
each of the three variables, created the ad hoc spatial
confidence region for potential old growth on the watershed
shown in Figure 6. Of particular relevance, this study
established that at least 500 realizations of the sequential
Gaussian simulation process was necessary to provide
stable estimates of the variance.
In the 1997 study, sequential Gaussian realizations were
created based on the assumption of spatial independence
between variables. In 1998, Mowrer (in press) extended the
study through adaptation ofthe FORTRAN program COSIM
(Carr and Myers 1985). The co-conditional simulation algorithm in this program includes spatial correlations through
cross-variograms, which provide a model of the spatial
covariance between two variables as a function ofseparation
distance. Thus, the co-variogram can be used with the
variograms for the individual variables in co-kriging. The
Carr and Myers (1985) program uses the "turning bands"
algorithm (Journel and Ruijbregts 1978) to create random
fields that are subsequently conditioned using co-kriging.
The results of the ad hoc spatial confidence region generated
using this algorithm are shown in Figure 7. The conclusion
from this study was that the symmetrical and circular
artifacts of the process we impossible to justify ecologically.
USDA Forest Service Proceedings RMRS-P-12. 1999
Figure S.-Percentiles based on the histogram of 500 realizations
from the sequential Gaussian simulation geostatistical uncertainty
assessment procedure. Areas above the 90th percentile are shown in
medium gray, 95th percentile in dark gray, and 99th percentile in black,
representing successively higher levels of certainty for old growth
locations.
For this, and other reasons, Mowrer recommended not
using this algorithm in generating multiple realizations of
forest stand conditions.
An ongoing controversy in the field of GIS, is how to
represent different spatial phenomena that may vary continuously in some areas, yet have discrete boundaries in
others. Runner, et al. (in press) investigated the use of the
GIS and remote sensing information described above, to
explore different geostatistical techniques such as cokriging
and disjunctive kriging to interpolate these types of phenomena accurately. Cokriging, described above, has the
advantage of using continuously measured information,
:*
tI!,.
...... ~
.•
Figure 7.-Ad hoc spatial confidence region, from the same GIS
analysis in Figure 6, but using the co-conditional simulation algorithm
to create spatially cross-correlated realizations for age and diameter.
Results are considered inferior to sequential Gaussian simulation,
despite the explicit spatial cross-correlation.
29
such as that available from GIS and remote sensing, that is
spatially cross-correlated with other information measured
only at particular point locations. Thus, the more extensive
data can be used to improve the estimation of the sparser
variables of interest across the entire area. Disjunctive
kriging (Rivoirard 1994) provides a non-linear unbiased
estimate of conditional probabilities that some value exceeds a cutoff value.
Forest Canopy Effects on
Hydrology
Our objective in the work described above is to improve
spatial estimates of snowpack water equivalent on the
ground on April 1st of each year. This measure provides the
basis for estimation of runoff later that spring. Ninety-five
percent of the surface water exiting watersheds on the
experimental forest originates as melting snowpack.
Simple empirical models indicate that the amount of water
in the snowpack on April 1 explains 70-95 percent of the
variability in annual runoff (Troendle et aI., 1998). Canopy
interception and subsequent sublimation of snow , which can
approach 35 percent of gross precipitation before it reaches
the ground, is a major manipulable source of water loss in
forested stands (Troendle et al. 1993). Winter interception
losses vary with aspect, stand density, species composition,
and, to some extent, precipitation amount. Interception
losses are therefore quite variable spatially (ibid.). On
north-facing slopes, peak water equivalent or the amount of
water in the snowpack on April 1s t, can be increase by 50
percent through forest removal. On a similar south-facing
slope, the increase would only be 20 percent (ibid.). Because
of the almost direct relationship between vegetation density
and snowpack accumulation, and between snowpack accumulation and streamflow, reliable predictions of steamflow
from subalpine forest environments require spatially correct estimations of vegetation and physiography.
Dubrasich et al. (1997) dev.eloped estimates of betweentree porosity, based on intensive individual tree sampling
in five structurally complex and three structurally simple
forest stands in the Pacific Northwest of the USA. Their
estimate of porosity treated each tree as a solid object and
expanded crown areas from individual plots to volumes per
hectare. Between-tree porosity was calculated as the percentage of the total volume space in the stand occupied by
stand crown volume.
Conclusions ------------------------------Obvious extensions of existing research may improve predictive capabilities for forest canopy hydrologic interactions.
Modeling the between-tree porosity using geostatistical
techniques such as three-dimensional kriging with external
drift, or three-dimensional cokriging are two immediate
goals. Both of these techniques require that the primary
variable of interest, i.e., between-tree porosity, be secondorder stationary with a constant mean and spatial covariances that are only dependent on the distance between
30
points. Kriging with external drift does allow for a linear or
polynomial trend model. Cokriging has the strength of
including cross-variates that are measured more intensively
than the variable of interest and improve predictive ability
for the primary variable at unmeasured locations. Possibilities for following up Pomeroy and Schmidt's 1993 work on
fractal analysis may provide predictive capabilities for
within-tree porosity, or interception. The confluence of spatial data, predictive capabilities, and computing power bode
well for progress in these areas.
Literature Cited
Avery, Thomas E. and Harold E. Burkhart. 1994. Forest Measurements. McGraw-Hill, New York. 408 pp.
Carr, James R and Donald E. Myers. 1985. COSIM: A Fortran IV
program for coconditional simulation. Computers and Geosciences 11(7): 675-705.
Dubrasich, Michael E., David W. Hann, and John C. Tappeiner II.
1997. Methods for evaluating crown area profiles offorest stands.
Canadian Journal of Forest Research 27:385-392.
Hunner, Gerhard, Robin M. Reich, and H. Todd Mowrer. In press.
Modeling forest stand structure using spatial statistics, geographic information systems, and remote sensing. In: Proceedings of the Second Southern Forestry GIS Conference, October
28-29, 1998, Athens, Georgia, USA
Isaaks, Edward H. and R Mohan Srivastava. 1989. An Introduction to Applied Geostatistics. Oxford University Press, New York.
561 pp.
Journel, Andre and Ch.J. Huijbregts. 1978. Mining Geostatistics.
Academic Press, London. 600 pp.
Mowrer, H. Todd. 1994. Spatially quantifying attribute uncertainty
in input data for propagation through raster-based GIS. In:
Proceedings of GIS '94: the Eighth Annual Symposium on Geographic Information Systems. Polaris Conferences, Vancouver,
British Columbia, Canada. Pp. 373-382.
Mowrer, H. Todd. 1996. Incorporating uncertainty into spatial
estimates of old-growth subalpine forests using geostatistics.
In: GIS Applications in Natural Resources 2, edited by M. Heit,
H. Parker, and A Shortreid. GIS World Books, Fort Collins,
Colorado, USA Pp. 178-184.
Mowrer, H. Todd. 1997. Propagating uncertainty through spatial
estimation processes for old-growth subalpine forests using sequential Gaussian simulation in GIS. Ecological Modeling
98(1997):73-86.
Mowrer, H. Todd. (In press.) Spatial interpolation of forest conditions using co-conditional geostatistical simulation. In: Integrated Tools for Natural Resources Inventories in the 21 st Century, Proceedings ofan International Conference on the Inventory
and Monitoring of Forested Ecosystems, August 16-20, 1998,
Boise, Idaho. General Technical Report, North Central Forest
Experiment Station, St. Paul, Minnesota, USA
Pomeroy, J. W. and RA Schmidt. 1993. The use of fractal geometry
in modeling intercepted snow accumulation and sublimation.
In: Proceedings of the 50 th EasternJ61 st Western Snow Conference, Quebec City, 1993. Pp. 1-10.
Rivoiard, J. 1994. Introduction to disjunctive kriging and non-linear
geostatistics. Clarendon Press, Oxford. 180 pp.
Troendle, Charles A, M.S. Wilcox, and G.S. Bevinger. The Coon
Creek water yield augmentation pilot project. In: Proceedings of
the 66 th Western Snow Conference, Snowbird, Utah, USA. Pp.
123-130.
Troendle, Charles A, RA Schmidt, and Manuel Martinez. 1993.
Partitioning the deposition of winter snowfall as a function of
aspect on forested slopes. In: Proceedings of the 50 th Easternl61 st
Western Snow Conference, Quebec City, 1993. Pp. 373-379.
Wolf, Paul R and Russell C. Brinker. 1994. Elementary Surveying, Ninth Edition. Harper Collins, New York. 760 pp.
USDA Forest Service Proceedings RMRS-P-12. 1999