Content uploaded by Susan L. Ustin
Author content
All content in this area was uploaded by Susan L. Ustin
Content may be subject to copyright.
Landscape Ecology 13: 79–92, 1998. 79
c1998 Kluwer Academic Publishers. Printed in the Netherlands.
Geostatistical scaling of canopy water content in a California salt marsh
E.W. Sanderson1, M. Zhang1, S.L. Ustin1and E. Rejmankova2
1Department of Land, Air, and Water Resources, University of California, Davis, CA 95616, U.S.A.; 2Division of
Environmental Studies, University of California, Davis, CA 95616,USA
(Received and accepted 2 July 1997)
Key words: geostatistics, scaling, grain, extent, canopy water content, salt marsh, remote sensing, Airborne
Visible/Infrared Imaging Spectrometer (AVIRIS)
Abstract
Remote sensing data are typically collected at a scale which is larger in both grain and extent than traditional
ecological measurements. To compare with remotely sensed data on a one-to-one basis, field measurements
frequently must be rescaled to match the grain of image data. Once a one-to-one correspondence is established,
it may be possible to extrapolate site based relationships over a wider extent. This paper presents a methodology
for rescaling the grain of ecological field data to match the grain of remotely sensed data and gives an example of
the method in verification of remote sensingestimates of canopy water content in a tidal salt marsh. We measured
canopy water content at 169 pointson a semi-regular grid in the Petaluma Marsh, CA. A variogram describing the
spatial correlation structure of the canopy water content was calculated and modeled. Ordinary kriging estimates
of the canopy water content were calculated over blocks corresponding to image pixels acquired by the Airborne
Visible/InfraredImaging Spectrometer (AVIRIS). A water content index was determined from the reflectance data
by calculating the area of a water absorption feature at 970 nm. A regression developed between the blocks and
the pixels at the site was extrapolated over the image to obtain an estimate of canopy water content for the entire
marsh. The patterns of canopy water content at the site and landscape levels suggest that different processes are
important for determining patterns of canopy water content at different spatial extents. The errors involved in the
rescaling procedures and the remote sensing interpretation are discussed.
Introduction
Remote sensing is an important tool for ecolo-
gists attempting to understand ecological patterns and
processes at landscape scales (Ustin et al., 1993; Wess-
man, 1992; Quattrochi and Pelletier, 1991) and is com-
monly used in landscape studies(O’Neill et al., 1996;
Barkhadle, et al., 1994; Iverson et al., 1994; Knight
et al., 1994; Haines-Young, 1992), however remote
sensing provides informationat a scale larger in grain
and extent (sensu O’Neill et al., 1986) and of different
kind (reflected light vs. direct measures of ecological
phenomena) than most other ecological observations.
Because of these differences in scale and kind, it is
often difficult to separate the uncertainties due to the
remote sensing interpretation from uncertainties due to
mismatches in scale when image data and field data are
compared (Verstraete et al, 1996; Raffey, 1994a,b).
To separate these uncertainties, it is necessary to
adopt some kind of scaling methodology to aggregate
thefielddata,collectedatasmaller grain, to the grain of
theremotelysensedpixel,andestimatetheaggregation
error. Separate comparison of the rescaled ground data
to the interpreted value of the pixel allows verification
of the remote sensing interpretation. If a satisfactory
relationship is found, another rescaling operation may
be required, extrapolating the relationship from the
extent of the original field site (typically smaller than
the image) to the extent of the landscape observed by
the image. Further verification (sensu Mankin et al.,
1975) is usually required.
Many landscape ecologists face the dilemma of
80
how to compare datasets collected at different scales,
which has led to many investigations of the effect of
scale on sampling (Fuhlendorf and Smeins, 1996; Qi
and Wu, 1996; McNaughton and Jarvis, 1991; Musick
and Grover, 1991; Turner et al., 1991; Woodcockand
Strahler, 1987) and of ecological scaling methodolo-
gies (Moody and Woodcock, 1995; Allen et al., 1993;
Rastetter et al., 1992; Vitousek, 1991; Waring, 1991;
Weins, 1989; O’Neillet al., 1986; Gardneret al., 1982).
Most of these studies have found that the spatial scale
ofsamplinghas an importantinfluence on the observed
pattern,buthaveyettofindany general scaling rulesfor
ecological phenomena (Jelinski and Wu, 1996; Levin,
1992).
Geostatistics has been used frequently in ecology
and remote sensing, both for interpolation (Rossi et
al., 1994; Atkinson et al., 1994; Van Der Meer, 1994;
Atkinson et al., 1992; Rossi et al., 1992) and scale
detection (Schlesinger et al., 1996; Hyppanen, 1996;
Jackson and Caldwell, 1993; Legendre and Fortin,
1989; Curran, 1988). This paper advances the idea
that geostatistics is also a useful scaling methodolo-
gy. Based on the statistical properties of point scale
measurements, geostatistics allows the investigator to
estimatethe value of a phenomenon at arbitrarilydeter-
mined locations within the study site, with the goal of
substantially increasing the number of observations at
the point scale which can be compared to each area
measurement. These point estimates can be aggregated
at whatever grain and extent the investigator chooses,
subject to the limits of the original grain and extent of
the point measurements. Thus the investigator deter-
mines the scale to which the point estimates will be
aggregated from a continuousrange of possible scales
and can precisely match them to the larger scale. Geo-
statistics works by estimating the spatial structure of
a phenomenon (variogram analysis), then making an
unbiased linear estimate (Issaks and Srivastava, 1989)
of the value of the phenomenon at any point within the
original study extent (kriging). The variogram guides
the estimation process by assigning weights based on
the modeled spatial autocorrelation. Further geostatis-
tical algorithms allow for quantification of expected
error in the estimates.
This paper shows how ecological field measure-
ments were rescaled to match remotely sensed obser-
vations by applying geostatistics to data from a tidal
salt marsh. Point measurements of canopy water con-
tent made over one sampling site were rescaled to the
grain of image data (pixels from the Airborne Vis-
ible/Infrared Imaging Spectrometer (AVIRIS)) using
variogram analysis and kriging. Once a relationship
was established over the extent of the study site, the
relationship was rescaled a second time, by extrapo-
lating from the extent of the site to the extent of the
landscape. This second rescaling was verified at a sec-
ond, independent reference site.
Methods
Study area
The observations for this study were made in a tidal
saltmarsh (Petaluma Marsh) along the Petaluma River,
about 8 km from the mouth of the river in San Pablo
Bay, CA (Figure 1). This marsh is part of a semi-
contiguous network of tidal wetlands that stretch along
the northern shore of San Pablo Bay and is part of the
greater San Francisco Bay Estuary (Josselyn, 1983).
Petaluma Marsh is constrained by levees along the
Petaluma River to the east and south, and by rising
topography on its western margin.
The vegetation of Petaluma Marsh is dominated
by Salicornia virginica, a succulent halophyte with
high canopy water content. Two other species grow in
zones parallel to the Petaluma River: Spartina foliosa,
a halophytic cordgrass (Mahall and Park, 1976), and
Scirpus robustus, a meso-halophytic bulrush (Ustin et
al, 1981). Both of these are less succulent than Salicor-
nia virginica and have lower water content (Zhang et
al.,1997).Asuiteof other speciesgrowalong the banks
of natural channels in the marsh. These species include
Frankenia salina,Jaumea carnosa,Grindelia cunefo-
lia,andLepidiumlatifolium. Through the remainder of
the paper, each species will be referred to by its gener-
ic name. All plant names follow the nomenclature of
Hickman (1993).
Two sites were studied in the Petaluma Marsh. The
first site (hereafter the River Site) was intensively sam-
pledwith the intention of developing a scaling relation-
ship between field and remotely sensed measurements.
The River Site covered a rectangular area (385 m by
175 m) parallel to the Petaluma River, including stands
of Spartina and Scirpus, as well as large areas of Sal-
icornia, and a small channel network (Figure 2). A
second site, the Pond Site, was located in the interior
of the marsh, approximately 1.2 km from the Petaluma
River, beside several salt ponds orpans. The vegetation
at the Pond Site was dominated by Salicornia. This site
was used to verify extrapolated predictions of canopy
water content based on the River Site results.
81
Figure 1. Study location in the Petaluma River watershed, the San Francisco Bay Estuary, and California.
82
Figure 2. Vegetation distributions of Salicornia virginica, Spartina foliosa, Scirpus robustus, and Frankenia salina at the River Site, Petaluma
Marsh. Cover estimates made over approximately 7.5 m by 7.5 m areas in the field.
Sampling design
Canopy water content was selected for study because
it has a strong influence on the spectral reflectance of
vegetation (Lillesand and Kiefer, 1987) and because it
ishighlycorrelatedwithstandingbiomass(Zhangetal,
1997), and therefore, related to productivity. Moreover
since all plants have a canopy water content, it is a con-
tinuous phenomenon which crosses spatial discontinu-
ities in species distributions. Although the species are
distributed in distinct zones, the study was designed to
estimate canopy water content across the entiremarsh,
withoutpreviousknowledgeof species locations, so all
comparisons were made withoutrespect to zonation.
At the River Site sample points were placed on
an quasi-regular grid at approximately 15 m intervals
(average nearest point-to-point distance was 12.3 m).
The sample spacing was chosen to underestimate the
grainofAVIRIS data which nominally has 20 m square
pixels. Canopy reflectance measurements and cover
estimates by species were made at 169 sampling points
on the grid. Canopy water determinations by destruc-
tive harvest were made at a subset of 38 points, strati-
fied approximately equally in different species zones,
following the other measurements. See Zhang et al.
(1997) for methods. At the Pond Site fifteen sam-
pling points were destructively harvested to determine
canopy water content. All sampling at both sites was
completed between May 15, 1994 and June 7, 1994 to
coincide approximately with an AVIRIS overflight on
May 21, 1994.
The location of each sampling point was deter-
mined as an average of ten measurements using a
Pathfinder Plus Global Positioning System (GPS)
83
(Trimble Navigation, Sunnyvale, CA). In our study,
GPS location measurements had a precision of 1–
2 m (based on standard deviation of ten simultaneous
measurements) and an accuracy of 3–10 m (based on
repeated sampling at a later time.) Because of difficul-
ties in processing the GPS measurements and satellite
availability, twenty-two data locations were not mea-
sured directly, but had their position estimated relative
to the other points. The accuracy for these points is
somewhat less, but no worse than 15 m. The loca-
tion of the natural channels and mosquito ditches were
determined by mapping relativeto the known sampling
points.
Canopy spectral measurements and analysis
Upwelling canopy radiance was measured using an
Analytical Spectral Devices Personal Spectrometer II
(Analytical Spectral Devices, Boulder, CO) for the
345–1072 nm wavelength interval and with an approx-
imately 2 nm spectral bandwidth. The fiber-optic head
was suspended one meter above the canopy and ori-
ented nadir. An 18 view restrictor was mounted on
the optic to restrict the field of view to a circle of
approximately 42 cm diameter (0.126 m2). At each
sampling point, radiance measurements were collect-
ed in the visible and near-infrared regions separate-
ly and with different integration times, to maximize
the signal-to-noise ratio of each measurement. Imme-
diately before acquisition, incident radiation on the
canopy was observed by measuring the upwellingradi-
ance from a Spectralon white reflectance panel (Lab-
sphere, Inc., North Sutton, NH) which is approximate-
ly 100% reflective for visible and near-infrared radi-
ation. In post-processing the canopy radiance spectra
were divided by the Spectralon panel radiance spec-
tra to determine canopy reflectance (as a percentage
of incident light). All radiance measurements were
made between 1100 and 1500 hours PST. Reflectance
spectra from each integration time were matched by
making a least squares fit over the overlapping regions
(approximately 850 to 900 nm), weighted by a run-
ning standard deviation five bands wide based on the
shorter integration time spectra. This weighting func-
tion was used because the shorter integration time was
optimized for measurement of upwelling radiance in
the visible, which resulted in poor signal-to-noise in
the near-infrared.
Canopy reflectance spectra were truncated to 400
to 1050 nm, averaged to 10 nm bands to correspond
approximately to AVIRIS bands and normalized by
dividing by the square root of the sum of squared
reflectance values for each spectra. While normal-
ization minimized albedo variations in the reflectance
spectra, it produced more consistency in the shape of
the spectra, which is the focus of the remote sens-
ing method used. Normalization over the entire spec-
trum has been suggested recently (Price, 1994; Pinzon,
1995) and is analogous to using NDVI (Normalized
Difference Vegetation Index) instead of a simple ratio
(Tucker, 1979).
Canopy water content was approximated from the
reflectance spectra using a technique described as
continuum removal (Clark and Roushe, 1984). This
semi-empirical technique assumes a linear relationship
betweenthe area of an absorptionfeatureandthechem-
ical content that causes that absorption. A linear con-
tinuum is calculated over the wavelength interval of
the feature to approximate a hypothetical reflectance
in the absence of the feature. The area between the
hypothetical reflectance and the measured reflectance
is determined and compared to chemical content (in
this case, canopy water content) to derive an empiri-
cal (linear regression) relationship(Zhang et al., 1997).
Thisregressionrelationshipwas used to predict canopy
water content at the 169 sampling points at the River
Site.
Variogram calculation and modeling and kriging
estimation
Observed variograms were calculated and modeled for
the continuum removed area (CORA). A lag distance
of 15 m and a width tolerance of 7.5 m, corresponding
to the approximate spacing of the sample grid, gave
the best resolved variogram, with each point on the
variogram representing between 407 and 1321 data
pairs(Figure 3). The searchneighborhoodwasoriented
roughly alongthe north-south axis, extending 0–100m
in the east-west direction and 0–200 m in the north-
south direction to accommodate the rectangular area
of the River Site. Directional variograms were also
examined for evidence of anisotropy, but none was
found.
We also calculated cross-variograms between
canopywater content and CORA,butbecause of sparse
water content sampling (38 points), those variograms
had poorly defined structure. As a result, co-kriging
approaches, which estimate water content directly, as
suggested by Atkinson et al. (1992), gave poor estima-
tions for our data set.
84
Figure 3. Observed variogram and exponential variogram model
for canopy water content, based on spectral measurement (CORA),
River Site, Petaluma Marsh.
The CORA observational variogram was modeled
using an exponential model and no nugget(Figure 3).
Therangeof the omnidirectionalvariogrammodel was
100mwithasillequivalenttoacanopywater content of
0.71 kg/m2. The nugget is definedas the semivariance
at a lag distance of zero. Strictly the semivariance at
lag zero should be zero; a nonzero nugget causes a
discontinuityin the variogramwhich restricts the range
of weighting values used in the estimation (Isaaks and
Srivastava, 1989). Nonzero nuggets are often found in
observed variograms and may indicate measurement
error or short scale variability. In this study we attribute
the nugget effect in the observed variogram to noise
in the CORA estimates of canopy water content, so
we did not include a nugget in the variogram model.
See Atkinson et al., 1996, for further discussion on
handling nugget effects in variograms calculated from
remotely sensed data.
Ordinary kriging algorithms were used with the
variogram model to estimate the canopy water con-
tent at a density of nine points per pixel, which was a
marked increase over the original field data (an aver-
age 0.76 points per pixel, for the combined CORA
and destructive harvests, or 0.14 points per pixel, for
the destructive harvests only). Point estimates from
the geostatistics were averaged (or “blocked”) to cor-
respond to areas of the same size and location as the
AVIRIS pixels.
Variogram modeling and kriging estimations were
calculated using GEOPACK (Yates and Yates, 1989)
and GEO-EAS (Englund and Sparks, 1988) software
packages.
Image acquisition and processing
A hyperspectral, visible and near-infrared image
was acquired over the Petaluma River watershed by
the Airborne Visible/Infrared Imaging Spectrometer
(AVIRIS) (Vane et al., 1993) on May 21, 1994 at 13:
19 PST. AVIRIS images are 512 pixels wide and 614
pixels long, with reflectance measured in 224, approx-
imately 10 nm bands, covering the spectral range
from 400–2500 nm. The instantaneous field of view
of AVIRIS pixels is 20 m, but pixels overlap approxi-
mately 2.5 m on each edge, so that resolved pixels are
approximately 17.5 m on a side. For 17.5 m pixels,
the extent of the image is 8.96 km wide and 10.75 km
long.
The image was calibrated to apparent surface
reflectance using the Atmosphere Removal Program
(ATREM, version 1.0) calibration algorithm (Gao et
al., 1993). The calibration was optimized by adjusting
the atmospheric calibration parameters and repetitive-
ly comparing calibrated pixel output to known ground
spectra.
ThecalibratedimagewasviewedusingtheSpectral
Image Processing System (SIPS) (Kruse et al., 1993)
for analysis and interpretation. Using SIPS, we identi-
fied 16 bands in the near-infrared, from 918–1062nm,
which contained the 970 nm water absorption feature.
These band reflectances were extracted from the image
data cube to calculate the continuum removed area,
using the same technique (Clark and Roushe,1984) as
described above for the canopy spectra. These steps
were implemented using the Image Processing Work-
bench (Frew, 1990). The resulting one band image was
imported in ARC/INFO and georeferenced using GPS
acquired ground points at six road intersections on the
image. The rectification errorwas less than 17.5meters
(i.e. one pixel) in both north-south and east-west direc-
tions.
The region of salt marsh vegetation was delineated
by creating a mask from the AVIRIS band centered at
1222 nm, which gave the maximum contrast between
the salt marsh vegetation and surrounding upland veg-
etation. This band was extracted from the data cube
using SIPS, reformatted to a pixel grayscale map and
contrast sharpened in XV, version 3.0 (John Bradley,
Bryn Mawr, PA), then used as a mask in ARC/INFO.
A second mask of the open water and channel network
was prepared using the band centered at 1591 nm.
85
Figure 4. Scatterplot and regression of block kriged estimates to image derived measurements (CORA) of canopy water content at the River
Site, Petaluma Marsh.
Comparison of ground observations to remote
observations
Block kriged estimates and image measured canopy
water content data were related by regression analysis
(Figure 4). Because of extrapolation problems, edge
pixels were not included in the regression relation-
ship, resulting in a comparison of 157 pixels. This site-
based regression relationship was applied across the
entire image to estimate canopy water content values.
Regression predictionsof water content less than zero
kg/m2wereset tozerokg/m2. In generalthe pixels with
less than zero kg/m2(8.6% of total) corresponded to
pixels lining the Petaluma River and tributaries, where
pixels were not masked out because they included both
vegetation and water elements.
Verification
The Pond Site destructiveharvests were used to test the
relationship between vegetation canopy water content
and image CORA from the River Site. The predict-
ed canopy water content map for the entire salt marsh
was laid over an ARC/INFO coverage containing the
Pond Site point sample data. Points were comparedto
their corresponding pixels. Where two points fell in
one pixel, the point valueswere averaged prior to com-
parison to the pixelvalue to avoid pseudo-replication.
The point and pixel values werecompared on a scatter-
plot and a regression line was calculated through the
values (Figure 5). Three extreme values which were
outside the range of canopy water content observed at
the River Site were removed.
Results
The main quantitative results of this study are sever-
al distributions of canopy water content measured or
estimated at different spatial grains and extents. Com-
parison of these distributions statistically and spatially
constitutes the scaling framework used to extend the
resultsfromthesitelevel to the landscape level.Exami-
nationofthestatistics associated with each comparison
allow us to estimate error for each scaling transforma-
tion.
The statistical distributions of canopy water con-
tent measured by destructive harvest and continu-
um removal (CORA) were similar but not identi-
cal. The univariate statistics of these distributions
86
Figure 5. Scatterplot and regression of measured canopy water
content to rescaled predictions of canopy water content at Pond Site,
Petaluma Marsh.
(Table 1) showed similar means, but a higher max-
imum value and higher standard deviation associated
withthedestructiveharvest data.Thedistributionofthe
destructiveharvest data wasnegativelyskewedbecause
of disproportionate sampling of Spartina, which had
a lower canopy water content, relative to Spartina’s
areal coverage at the site. Although Spartina covers
only approximately1% of the sampled area (Figure 2),
Spartina samples account for 26% of the destructive
harvest water content observations.
The results of the kriging interpolation steps com-
pared favorably with the ground measurements. The
point kriged estimates were normally distributed with
ameanidenticaltothemeanCORAestimateof canopy
water content, 2.08 kg/m2(Table 1), and with similar
standard deviations (0.61 and 0.82 kg/m2), respective-
ly; however, the overall range of the kriged distribution
was contracted. Like most estimation algorithms, krig-
ing tends to smooth the data.
The distribution of block kriged estimates was also
normal with a slight increase in the mean compared
to the point estimates, to 2.10 kg/m2(Table 1), with a
standard deviation of 0.57 kg/m2. The extreme values
werealsotruncated.In this case howeverthe smoothing
was not only a function of the kriging procedure, but
also of the block averaging. Block averaging may be
comparable to the smoothingdue to pixel averagingby
the sensor, and thereforenot necessarily undesirable.
The block kriging estimates were very similar to
the image-derivedestimates (image CORA) of canopy
water content. The block kriged estimate distribution
(Table 1) was not statistically distinguishable from the
image-derived distribution (Wilcoxon Rank Sum test
(Z = –0.2872; p = 0.7740)), though the image-derived
distribution had a greater range and was less regu-
lar. Given that the distributions were approximately
normal, we applied a Student’s t test (t = 0.033; p
= 0.8552) to show that means were also statistically
indistinguishable (Table 1).
The spatial patterns of canopy water content
through the various interpolation and averaging steps
are shownin Figure 6. The original fieldmeasurements
(Figure 6a), the estimation steps for points (Figure 6b)
and blocks (Figure 6c), and the imagedata (Figure 6d),
all show similar patterns. Lower canopy water content
was observed in the Spartina and Scirpus zones, and
higher water content in the Salicornia zone, particular-
ly along the channel networks.
The relationship between block kriged canopy
water estimates and image derived CORA data showed
some scatter, but had a strong, linearly increasing trend
(see Figure 4). The Pearson correlation coefficient (r)
between block and image derived water content esti-
mates was 0.66. The regression equation had an R2of
0.44 (p 0.001, n = 157).
This site-based regression relationship was applied
across the image to estimate the canopy water con-
tent distribution for the entire marsh landscape. The
overall distribution was similar to the site level dis-
tribution, having similar mean and standard deviation
(Table 1). The number of observations, however, was
muchlarger( 39,000), and the range was wider, from
0–10.39 kg/m2. Zero values corresponded to areas of
open or nearly open water which were not masked out
by the channel mask. The shape of the distribution
(not shown) was normal, but with a long, positively
skewed tail of high values. We verified the landscape
level results by comparing image derived estimates to
point measurements made at the Pond Site. If the high-
est canopy water content values were omitted (points
4.5 kg/m2), then the regression showed an approxi-
mately linear relationship between pixel and point val-
ues at the Pond Site (R2= 0.60; p 0.003, n=12) with
a slope near one (m = 1.27) (Figure 5).
87
Figure 6. (left panel, a–d). Similarity of spatial patterns of observed and estimated canopy water content at the River Site, Petaluma Marsh. a. Spectral measurement (CORA) of canopy water
content; b. Point kriged estimates of canopy water content; c. Block kriged estimates of canopy water content; d. Image derived measurements of canopy water content.
Figure 7. (right panel). Spatial pattern of canopy water content for Petaluma Marsh landscape.
88
Table 1. Similarity of statistical distributions of canopy water content (kg/m2)forfieldand
image measurements and kriging estimates at River Site and across the Petaluma Marsh.
Canopy water content at river site (kg/m2)
Mean Standard Minimum Maximum Number
deviation
Destructive harvest
measurements 1.93 1.18 0.22 5.23 38
Spectral
measurements (CORA) 2.08 0.82 0.20 3.93 169
Point kriged
estimates 2.08 0.61 0.63 3.72 697
Block kriged
estimatesa2.10b0.57 0.91 3.48 157
Remotely sensed
observationsa2.12b0.80 0.26 3.99 157
Canopy water content across marsh landscape (kg/m2)
Remotely sensed
observations 2.37 1.34 0 10.39 39,282
aThese distributions are statistically indistinguishable by Wilcoxon Rank Sum test (Z = –
0.2872; p = 0.7740)
bThese means are statistically indistinguishable by Student’s t test (t = 0.033; p = 0.8552).
Discussion
Site level pattern
Thepatternof high and low canopywatercontentat the
River Site suggested two casual processes (Figure 6d).
The canopy water content borderingthe Petaluma Riv-
er, in the eastern third of the site, was generally low,
in the 0–2 kg/m2range. In contrast the canopy water
content away from the Petaluma River, in the western
two-thirds of the site, was generally higher, in the 2–4
kg/m2range. The highest canopy water content values
were associated with the tidal channel network in the
center of the site.
Based on the spatial pattern of canopy water con-
tent, we hypothesized that the pattern was governed
primarily by species distributions, and governed sec-
ondarily by the locations of tidal creeks, which may
provide benefits to the neighboring vegetation (Zhang
etal.,1997;Sanderson,in manuscript). Speciesdistrib-
utions at the site were largely monotypic exceptimme-
diately beside the channel networks and in the Scirpus
zone, which was partially undergrown by Salicornia
along its margins. Salicornia was clearly the dominant
byarea,covering83% of theRiverSite (Figure2). Are-
al percent coverage by other species declined sharply,
Scirpus 11%, Frankenia 1.5%, Spartina 1.3% and all
other species less than 1% coverage.
These species differed in their canopy water con-
tents (Table 2). The destructive harvests showed Sal-
icornia had the highest canopy water content, consis-
tent with its succulent leaves, followed by Scirpus,
then Spartina; however the differences between Scir-
pus and Spartina were largely a function of phenology
at sampling time. Neither were at peak biomass: Scir-
pus typically has more new biomass than Spartina by
late May (Cameron, 1972). Frankenia had an interme-
diate canopy water content.
The second aspect of the pattern, higher water con-
tent associated with the channel network, can also be
explained by thedistribution and relative water content
of plant species (Figure 2; Table 2). Several species
(Frankenia is representative) are associated with the
levees along the channel network, most which had
healthy, green biomass at the time of sampling. These
levee species are hypothesized to be associated with
tidal channels because of tidal subsidies and reduced
anoxia in the elevated sediments along the channel
(Hinde, 1954). Salicornia near the channels may also
benefitfromtidalsubsidies,causingthemtogrowmore
robustly than plants farther away, although we had too
few points near the channels to confirm this pattern in
our data.
89
Table 2. Distributions of canopy water content by species at
River Site depend on observational method, either destructive
harvests or spectral measurements (CORA).
Species Mean water N Mean water N
content (kg/m2), content (kg/m2),
destructive harvest spectral
interpretation
(CORA)
All 1.93 38 2.05 169
Salicornia 2.56 21 2.32 124
Spartina 0.78 10 1.05 19
Scirpus 2.11 3 1.04 11
Frankenia 1.28 3 2.02 10
Other 1.71 1 1.56 4
Species estimated to have at least 50% of cover at that point,
or in cases where total cover was less than 50%, species with
highest cover by ocular estimate.
Landscape level pattern
For most parts of the marsh, the pattern of species
distribution and canopy water content suggested at the
River Site are probably adequate to explain the spa-
tial pattern of canopy water content for the marsh as
a whole (Figure 7). Canopy water content seemed to
be highest along small channels, and lower immediate-
ly along the Petaluma River and large sloughs, where
Spartina and Scirpus grow. Although it is difficult to
discernlow-orderchannel networksinthe image, elon-
gate structures of higher water content are suggestive
of channel network effects and correspond roughly to
the locations of channels near the Pond Site.
For some localities, however, further explanations
were needed to explain the pattern. For example, much
higher water content was observed where San Antonio
Creek, a fresh water stream, empties into the marsh in
the northwest corner. Field reconnaissance after image
analysis revealed large areas of upland, glycophytic
species including Lolium multiflorum and Poa annua,
and tall, lush broadleaf species lining the slough/creek
in this area, including Raphanus sativus and one or
more Conium spp. The biomass of these communities
was probably higher on an area basis than the Salicor-
nia dominated community characteristic of the main
salt marsh, resulting in an apparently higher canopy
water content, though we did not measure biomass or
water content in this vegetation type. A series of aban-
doned fence posts demarcated the area, suggestingthat
at one time this portion of the marsh might have been
reclaimed for pasture. The influx of freshwater from
SanAntonio Creek may bemaintainingnon-halophytic
species in this area.
In contrast low canopy water content was observed
in a well-defined area neighboring the sewage treat-
ment plant northeast of the river (Figure 7). Field
reconnaissance of this area showed that levees define
an area which receives fresh-water (treated) effluent
from the treatment plant. As a result of the lower salin-
ities and periodic flooding, several species typical of
brackish marshes grow in this area including Scirpus
acutus and Typha spp., in addition to large areas of the
meso-halophytic Scirpus robustus. These plants grow
in stands alternating with flooded areas, resulting in a
patchwork of low to zero canopy water contents in this
area.
Finally, an overall gradient in canopy water con-
tent is observableat the landscape level.It appears that
canopy water content rises from the margins of the
marsh from the Petaluma River to the interior, with the
highest canopy water contents observed in the region
where Zhanget al. (1997) reporteda lowgrowing, very
dense form of Salicornia. This pattern is not observed
in strip marshes along the Petaluma River or its trib-
utaries, suggesting that some additional dynamic may
be at work in the large, interior marsh which is not
expressed in the smaller strip marshes.
Landscape ecology and remote sensing: assessing
errors
It is important when evaluating remotely sensed land-
scape patterns to remember what remote sensing data
is: the measurement of reflected light from the land-
scape. In general, ecologists are not interested in
reflected light per se, but in landscape properties (e.g.
vegetation type) and/or processes (e.g. productivity)
which may influence reflected light. Thus for remote
sensing data to be useful, the radiance data must be
transformed, through use of an interpretative tech-
nique, into a measure of the phenomenon of interest
(Verstraete et al., 1996). For example in this paper, we
use the continuum removalmethod to relate the area of
a spectral feature to canopy water content. This semi-
empirical interpretation had a measurable error at the
point scale based on the fit of the regression line (R2=
0.63, from Zhang et al. (1997)) (Table 3). We desired
to make a similar estimate of error at the pixel scale to
verify our interpretation of the image data.
However evaluating the interpretation at the pixel
scale is more difficult, because often we can not sep-
arate error in the interpretative methods from errors
90
Table 3. Estimation of error at different steps in the scaling procedure.
Procedural step Unexplained variance Suggested sources of error
Point scale spectral interpretation 37% Saturation; species specific canopy architec-
ture and phenology
Grain to grain scale translation 9–63% (mean 14%) Estimation of autocorrelation; modeling of
autocorrelation
Block scale spectral interpretation 56% Georeferencing; pixel sampling; atmospheric
correction; errors from earlier steps.
Extrapolation of extent (verification atsecond
site) 40% Out of range; errors from earlier steps
From Zhang et al. (1997)
due to scale differences between the remotely sensed
and field data (Verstraete et al., 1996). In fact, were
we to compare the point measurements directly to
image measurements, the observed error would sub-
sume both interpretative and scaling errors. Combin-
ing these errors obscures both evaluation of the remote
sensing method and the scaling procedure.
Our approach was to explicitly address and eval-
uate the scaling methodology separate from errors in
remote sensing interpretation (Table 3). For the scal-
ing methodology, kriging is advantageous because it
allowsestimation ofthe kriging variance at everypoint.
In this example the kriging variances varied between
9–63%, with a mean variance of 14%. These variances
measure the errors associated with the modeling of the
autocorrelation structure (the variogram), the extent to
whichthedistribution of canopy watercontentmatched
the assumptions of the kriging algorithms (stationarity
and a normal distribution of the data), and the averag-
ing of point estimates to pixel size areas. The criteria
for evaluating the seriousness of these errors depends
on the goals of the application, their magnitude and
spatial distribution: here, the variances appeared to be
randomly distributed spatially (not shown) and their
magnitudesseemed reasonable,givenknown variances
in the field measurements (37% unexplainedvariance)
and errors in their locations.
The relative success of the scaling methodology
provides additional information. Since kriging provid-
ed a satisfactory scaling method, we can infer that
scaling of canopy water content met the assumptions
of kriging as well, that is, canopy water content scales
linearly over scales with grain from less than a square
meter to over 300 m2and within an extent of approx-
imately 67,000 m2(the River Site), despite spatial
discontinuities due to species zonation. Moreover the
canopy water content appears to be normally distrib-
uted over this range of scales. Canopy water content is
autocorrelated at distances less than 100 meters, and
the autocorrelation can be described using a exponen-
tial variogram model (Figure 3).
Although not perfect, these rescaled field measure-
ments are at least known quantities, with known errors,
which we can compare to the image. Like Zhang et
al. (1997), we used linear regression to compare the
remote sensing data to the field data, but here we com-
pare them at the pixel scale (Figure 4). Scatter in this
relationship (R2= 0.44) derives from problems in geo-
referencingandregistering,atmospherically correcting
and then interpreting the image, as well as variance in
the kriged field data. Despite all these potential errors,
thestatistical distributionsof the rescaled field data and
the image data are remarkably similar (Table 1). Thus
the scatter probably results mostly from slight pixel to
pixel discrepancies in the spatial distributions of the
blocked data (Figure 6 c,d), related to georeferencing
and registering of the image, and pixel sampling by
the sensor. Recall that AVIRIS pixels overlap approxi-
mately 2.5 m on each edge.
The final step is to evaluate the remote sensingpro-
cedure when it is extrapolated beyond the primary site
where it was developed, typically by making measure-
ments at a second site. Here the regressionat the Pond
Site showed that the interpretation is satisfactory for
canopy water content values within the same range as
was observed at the River Site, but may be inadequate
for predicting values outside that range ( 4 kg/m2),
whichisconsistent with deficiencies in boththescaling
method(contractionof the distributionsdue to smooth-
ing) and in the spectral interpretation (saturation of
the water absorption feature) (Figure 5). Verifying the
extrapolation informs us of the bounds of our interpre-
tation at the landscape level. For example, predictions
ofcanopywater content greaterthan4 kg/m2are known
91
to be unreliable, based on the verification process, so
they are uniformly color coded on Figure 7. Despite
this kind of limitation on our analysis, a synthesis of
field and remotely sensed data is the only tool we cur-
rently have to non-intrusively estimate canopy water
content, and many other ecological properties, over
landscape sized areas.
Conclusion
This paper has shown a practical scaling strategy to
relate ecological field data to a remote sensing image
in order to verify and interpret landscapepatterns. We
showed that field measurements of canopy water con-
tent in a tidal salt marsh could be rescaled to remote
sensing pixel estimates of the same parameter using
geostatistics, and that the relationship obtained at the
site level could beextended across the landscape. Pre-
liminary analysis of the landscape pattern of canopy
water content indicates processes at multiple scales are
important for structuring the observed pattern. Errors
due to the scaling procedure and the remote sensing
interpretation were separately evaluated and used to
inform the interpretation of pattern.
Acknowledgments
The authors would like to acknowledge the following
individualsfor help in conducting field measurements:
Claudia Castaneda, Larry Costick, Martha Diaz, John
Gabriel, Quinn Hart, Robert Haxo, Helen Hansen,
Han-YuHung, Stephane Jacquemoud,KentJorgensen,
Alicia Palacios, Jorge Pinzon, Rebecca Post, George
Scheer, Harry Spanglet, Lai-Han Szeto, Gail Wheel-
er, Linette Young, and QingFu Xiao. We additionally
thank Harry Spanglet for comments on an earlier draft
of this paper and for assistance with plant identifica-
tions, and Jorge Pinzon for many useful discussions.
We wish to recognize the support of the US Environ-
mentalProtectionAgency grant R82-1695-010andthe
NASA EOS program under grants NAS5-31359 and
NAS5-31714, and support for EWS under a NASA
Global Change Fellowship, reference number 1995-
GlobalCh00404. We also acknowledge the support of
the Digital Equipment Corporation under a Sequoia
2000 grant for computer hardware.
Disclaimer
This research was supportedin part by grant R82-1695
from the US Environmental Protection Agency (EPA)
– National Center for Environmental Research and
Quality Insurance and in part by the US EPA Cen-
ter for Environmental Health Research (R81-9658) at
UC Davis. Although the information in this document
has been funded wholly or in part by the US EPA, it
may not necessarily reflect the views of the Agency
and no official endorsementshould be inferred.
References
Allen, T.F.H., A.W. King, B.T. Milne, A. Johnson, and S. Turner.
1993. The problem of scaling in ecology. Evol. Trends in Plants
7(1): 3–8.
Atkinson, P.M., R. Dunn and A.R. Harrison. 1996. Measurement
error in reflectance data and its implications for regularizing the
variogram. Int. J. Rem. Sens. 17(8): 3735–3750.
Atkinson, P.M., R. Webster and P.J. Curran. 1994. Cokriging with
airborne MSS imagery. Rem. Sens. Env. 48: 1–25.
Atkinson, P.M., R. Webster and P.J. Curran. 1992. Cokriging with
ground-based radiometery. Rem. Sens. Env. 41:45–60.
Barkhadle, A.M.I., L. Ongaro and S. Pignatti. 1994. Pastoralism
and plant cover in the lower Shabelle region, southern Somalia.
Landscape Ecol. 9(2): 79–88.
Cameron, G.N. 1972. Analysis of insect trophic diversity in two salt
marsh communities. Ecology 53: 58–73.
Clark, R.N. and T.L. Roushe. 1984. Reflectance spectroscopy: Quan-
titative analysis techniques for remote sensing applications. J.
Geophys. Res. 89: 6329–6340.
Curran, P.J. 1988. The semivariogram in remote sensing: an intro-
duction. Rem. Sens. Env. 24: 493–507.
Englund, E. and A.Sparks. 1988, GEO-EAS (GeostatisticalEnviron-
mental Assessment Software) User’s Guide. EPA600/4-88/033.
Environmental Monitoring Systems Laboratory, Las Vegas, NV,
USA.
Frew, J.E. 1990 The Image Processing Workbench. Ph.D. disserta-
tion, University of California, Santa Barbara, CA, USA.
Fuhlendorf, S.D. and F.E. Smeins. 1996. Spatial scale influence on
long-term temporal patterns of a semi-arid grassland. Landscape
Ecol. 11(2): 107–113.
Gao, B., K.B. Heidebrecht and A.F.H. Goetz. 1993. Derivation of
scaled surface reflectances from AVIRIS data. Rem. Sens. Env.
44: 165–178.
Gardner, R.H., W.G. Cale and R.V. O’Neill. 1982. Robust analysis
of aggregation error. Ecology 63(6): 1771–1779.
Haines-Young, R.H. 1992. The use of remotely-sensed satellite
imagery for landscape classification in Wales (U.K.) Landscape
Ecol. 7(4): 253–274.
Hickman, J.C. 1993. The Jepson manual: higher plants of California.
University of California Press, Berkeley, CA, USA.
Hinde, H.P. 1954. Vertical distribution of salt marsh phanerograms
in relation to tidal levels. Ecol. Mon. 24: 209–225.
Hyppanen, H. 1996. Spatial autocorrelation and optimal spatial reso-
lution of optical remote sensing data in boreal fores t environment.
Int. J. Rem. Sens. 17(17): 3441–3452.
92
Issaks, E.H. and R.M. Srivastava. 1989. An Introduction to Applied
Geostatistics, Oxford University Press, New York, NY, USA.
Iverson, L.R., E.A. Cook and R.L. Graham. 1994. Regional for-
est cover estimation via remote sensing: the calibration center
concept. Landscape Ecol. 9(3): 158–174.
Jackson, R.B. and M.M. Caldwell. 1993. The scale of nutrient het-
erogeneity around individual plants and its quantification with
geostatistics. Ecology 74(2): 612–614.
Jelinski, D.E. and J. Wu. 1996. The modifiable areal unit problem
and implications for landscape ecology. Landscape Ecol. 11(3):
129–140.
Josselyn, M. 1983. The Ecology of San Francisco Bay Tidal Marsh-
es: a community profile. US Fish and Wildlife Service, Washing-
ton, D.C. USA.
Knight, C.L., J.M. Briggs and M.D. Nellis. 1994. Expansion of
gallery forest on Konza Prairie Research Natural Area, Kansas,
USA. Landscape Ecol. 9(2): 117–125.
Kruse, F.A., A.B. Lefkoff, J.W. Boardman, K.B. Heidebrecht, A.T.
Shapiro, P.J. Barloon and A.F.H. Goetz. 1993. The spectral image
processing system (SIPS)-interactive visualization and analysis
of imaging spectrometer data. Rem. Sens. Env. 44: 145–163.
Legendre, P. and M.-J. Fortin. 1989. Spatial pattern and ecological
analysis. Vegetatio 80: 107–138.
Levin, S.A. 1992. The problem of pattern and scale in ecology.
Ecology 73(6): 1943–1967.
Lillesand, T.M. and R.W. Kiefer. 1987. Remote Sensing and Image
Interpretation. John Wiley and Sons, New York.
Mahall, B.E. and R.B. Park. 1976. The ecotone between Spartina
foliosaTrin. andSalicornia virginicaL. in saltmarshes of northern
San Francisco Bay. I. Biomass and production. J. of Ecol. 64:
421–433.
Mankin, J.B., R.V. O’Neill, H.H. Shugart and B.W. Rust. 1975.
The importance of validation in ecosystem analysis. In: New
Directions in the Analysis of Ecological Systems, Part 1, George
S. Innis, Ed., Simulation Councils Proceedings Series, Volume 5,
Number 1, Simulation Councils, Inc., LaJolla, California
McNaughton, K.G. and P.G. Jarvis. 1991. Effects of spatial scale on
stomatal control of transpiration. Agr. For. Met. 54: 279–301.
Moody, A. and C.E. Woodcock. 1995. The influence of scale and
the spatial characteristics of landscapes on land-cover mapping
using remote sensing. Landscape Ecol. 10(6): 363–379.
Musick, H.B. and H.D. Grover. 1991. Image textural measures as
indices of landscape pattern. In ed. Turner, M.G. and Gard-
ner, R.H. Quantitative Methods in Landscape Ecology. Springer-
Verlag, New York, NY, USA.
O’Neill, R.V., D.L. DeAngelis, J.B. Waide and T.F.H. Allen. 1986. A
Hierarchical Concept of Ecosystems. Princeton University Press:
Princeton, NJ, USA.
O’Neill, R.V., C.T. Hunsaker, S.P. Timmins, B.L. Jackson, K.B.
Jones, K.H. Ritters and J.D. Wickham. 1996. Scale problems in
reporting landscape pattern at the regional scale. Landscape Ecol.
11(3): 169–180.
Pinzon, J.E., S.L. Ustin, Q.J. Hart, S. Jacquemoud and M.O. Smith.
1995. Using foreground-background analysis to determine leaf
and canopy chemistry. In Proc. 5th. Ann. JPL Airborne Earth Sci.
Work. Edited by R.O. Green. Jan. 23–27, 1995. Pasadena, CA,
USA.
Price, J.C. 1994. How unique are spectral signatures? Rem. Sens.
Env. 49(3): 181–186.
Qi, Y., and J. Wu. 1996. Effects of changing spatial resolution on the
results of landscape pattern analysis using spatial autocorrelation
indices. Landscape Ecol. 11(1): 39–49.
Quattrochi, D.A. and R.E . Pelletier. 1991. Remote sensing for analy-
sis of landscapes: an introduction. In ed. Turner, M.G. and Gard-
ner, R.H. Quantitative Methods in Landscape Ecology. Springer-
Verlag, New York, NY, USA .
Raffey, M. 1994b. Heterogeneity and change of scale in models of
remote sensing. Spatialization of multi-spectral methods. Int. J.
Rem. Sens. 15(12): 2359–2380.
Raffey, M. 1994a. Change of scale theory: a capital challenge for
space observation of earth. Int. J. Rem. Sens. 15(12): 2353–2357.
Rastetter, E.B., A.W. King, B.J. Cosby, G.M. Hornberger, R.V.
O’Neill and J.E. Hobbie. 1992. Aggregating fine-scale ecological
knowledge to model coarser-scale attributes ofecosystems. Ecol.
Appl. 2(1): 55–70.
Rossi, R., D. Mulla, A. Journel and E. Franz. 1992. Geostatistical
toolsfor modelingand interpreting ecological spatial dependence.
Ecology 62: 277–314.
Rossi, R.E., J.L. Dungan and L.R. Beck. 1994. Kriging in the shad-
ows: geostatistical interpolation for remote sensing. Rem. Sens.
Env. 49: 32–40.
Schlesinger, W.H., J.A. Raikes, A.E. Hartley and A.F. Cross. On
the spatial pattern of soil nutrients in desert ecosystems. Ecology
77(2): 364–374.
Tucker, C.J. 1979. Red and photographic infrared linear combina-
tions for monitoring vegetation. Rem. Sens. Env. 8: 127–150.
Turner, S.J., R.V. O’Neill, W. Conley, M.R. Conley and R.E.
Humphries. 1991. Pattern and scale: statistics for landscape ecol-
ogy. In ed. Turner, M.G. and Gardner, R.H. Quantitative Methods
in Landscape Ecology. Springer-Verlag, New York,NY, USA.
Ustin, S.L, R.W. Pearcy and D.E. Bayer. 1981. Plant water relations
in a San Francisco Bay salt marsh. Bot. Gaz. 143(3): 368–373.
Ustin, S.L., M.O. Smith and J.B. Adams. 1993. Remote sensing of
ecological processes: a strategy fordeveloping ecological models
using spectral mixture analysis. In Scaling Physiological Process-
es: Leaf to Globe. pp. 339–357. Edited by J. Ehlringer and C.
Field. Academic Press, New York, New York, USA.
Van Der Meer, F. 1994. Extraction of mineral absorption features
from high-spectral resolution data using non-parametric geosta-
tistical techniques. Int. J. Rem. Sens. 15(11): 2193–2214.
Vane, G., R.O. Green, T.G. Chrien, H.T. Enmark, E.G. Hansen
and W.M. Porter. 1993. The airborne visible/infrared imaging
spectrometer (AVIRIS). Rem. Sens. Env. 44: 127–143.
Verstraete, M.M., B. Plinty and R.B. Myneni. 1996. Potential and
limitations of information extraction on the terrestrial biosphere
from satellite remote sensing. Rem. Sens. Env. 58: 201–204.
Vitousek, P.M. 1991. Global dynamics and ecosystem processes:
scaling up or down? In Ehlringer, J. and Field, C.B. Scaling
Physiological Processes: Leaf to Globe. Academic Press, New
York, NY, USA.
Waring, R.H. 1991. Howecophysiologists can help scale from leaves
to landscapes? in Ehlringer, J. and Field, C.B. Scaling Physiolog-
ical Processes: Leaf to Globe. Academic Press, New York, NY,
USA.
Weins, J.A. 1989. Spatial scaling in ecology. Func. Ecol. 3: 385–397.
Wessman, C.A. 1992. Spatial scales and global change: bridging
the gap from plots to GCM grid cells. Ann. Rev. Ecol. Sys. 23:
175–200.
Woodcock, C.E. and A.H. Strahler. 1987. The factor of scale in
remote sensing. Rem. Sens. Env. 21: 311–312.
Yates, S.R., and M.V. Yates. 1989, Geostatistics for waste manage-
ment: a user’s manual for the GEOPACK Geostatistical Software
System. USDA Salinity Lab, Riverside, CA, USA.
Zhang, M., S.L. Ustin, E. Rejmankova and E.W. Sanderson. 1997.
Monitoring Pacific Coast Salt Marshes Using Remote Sensing.
Ecol. Appl. 7(3): 1039–1053.
