Characterizing rangeland vegetation using Landsat and 1-mm VLSA data in central Wyoming (USA)

Abstract

As an alternative to ground-cover data collection by conventional and expensive sampling techniques, we compared measurements obtained from very large scale aerial (VLSA) imagery for calibrating moderate resolution Landsat data. Using a grid-based sampling scheme, 162 VLSA images were acquired at 100 m above ground level. The percent vegetation cover in each photo was derived using SamplePoint (a manual inventory method) and VegMeasure (a reflectance based, automated method). Approximately two-thirds of the VLSA images were used for calibrating Landsat data while the remainder was used for validation. Regression models with Landsat bands accounted for 55% of the VegMeasure-based measurements of vegetation, whereas models that included both Landsat bands and elevation data accounted for 67%. The relationship between the Landsat bands and the percent vegetation cover measured by SamplePoint was lower (R
2 = 20%), highlighting the differences between the inventory and reflectance based protocols. Results from the model validation indicated that the model’s predictive power was lower when the vegetation cover was either <20% or >55%. Additional work is needed in these ecosystems to improve the calibration techniques for sites with low and high vegetation cover; however, these results demonstrate the VLSA imagery could be used for calibrating Landsat data and deriving rangeland vegetation cover. By adopting such methodologies the US Federal land management agencies can increase the efficiency of the monitoring programs in Wyoming and in other western states of the US.

Full text

Characterizing rangeland vegetation using Landsat
and 1-mm VLSA data in central Wyoming (USA)
Ramesh Sivanpillai ÆD. Terrance Booth
Received: 3 September 2006 / Accepted: 14 February 2008 / Published online: 4 March 2008
ÓSpringer Science+Business Media B.V. 2008
Abstract As an alternative to ground-cover data
collection by conventional and expensive sampling
techniques, we compared measurements obtained
from very large scale aerial (VLSA) imagery for
calibrating moderate resolution Landsat data. Using a
grid-based sampling scheme, 162 VLSA images were
acquired at 100 m above ground level. The percent
vegetation cover in each photo was derived using
SamplePoint (a manual inventory method) and Veg-
Measure (a reflectance based, automated method).
Approximately two-thirds of the VLSA images were
used for calibrating Landsat data while the remainder
was used for validation. Regression models with
Landsat bands accounted for 55% of the VegMea-
sure-based measurements of vegetation, whereas
models that included both Landsat bands and elevation
data accounted for 67%. The relationship between the
Landsat bands and the percent vegetation cover
measured by SamplePoint was lower (R
2
=20%),
highlighting the differences between the inventory and
reflectance based protocols. Results from the model
validation indicated that the model’s predictive power
was lower when the vegetation cover was either\20%
or [55%. Additional work is needed in these ecosys-
tems to improve the calibration techniques for sites
with low and high vegetation cover; however, these
results demonstrate the VLSA imagery could be used
for calibrating Landsat data and deriving rangeland
vegetation cover. By adopting such methodologies the
US Federal land management agencies can increase
the efficiency of the monitoring programs in Wyoming
and in other western states of the US.
Keywords Shrub land Sagebrush 
Remote sensing Satellite data Regression analysis
Introduction
Rangelands are characterized by grass, forb, and
shrub vegetation (e.g., the Sonoran desert, the
sagebrush (Artemesia spp.) steppe, tall- and short-
grass prairies). In the US, rangelands occupy 324 M
ha, providing numerous ecological services of which
grazing, wildlife habitat (including habitat for several
Mention of trade names is for information only and does not
imply endorsement by USDA over comparable products or
services.
R. Sivanpillai (&)
Wyoming Geographic Information Science Center,
University of Wyoming, 1000 East University Avenue,
Laramie, WY 82071, USA
e-mail: sivan@uwyo.edu
D. T. Booth
US Department of Agriculture – Agriculture Research
Service, High Plains Research Center, Cheyenne, WY
82009, USA
e-mail: Terry.Booth@ars.usda.gov
123
Agroforest Syst (2008) 73:55–64
DOI 10.1007/s10457-008-9115-8

species listed as threatened under the Federal Endan-
gered Species Act (NS-GCT 2004; FWS 2005)),
watersheds, recreation, and extractable minerals are
commonly recognized. Less obvious are the range-
land biophysical processes important to a well-
functioning global environment. Natural and
human-induced rangeland disturbances such as wild-
fires, droughts, cattle grazing, and housing
developments all influence rangeland vegetation
and, in turn, affect the carbon and water cycles (Kolb
and Sperry 1999; Gilmanov et al. 2004; Morgan et al.
2004). Given their ecological and economic impor-
tance, and increasing multiple uses, accurate and
continuous assessment of rangeland is fundamental to
ongoing sustainable use. Public US rangeland is
primarily managed by the Department of Interior’s
Bureau of Land Management (BLM) and the Depart-
ment of Agriculture’s Forest Service (USFS). These
agencies operate under several public laws including
the Resources Planning Act (1974), Federal Land
Policy and Management Act (1976), Soil and Water
Resources Conservation Act (1977), Forest and
Rangeland Resources Planning Act (1978), and the
Environmental Monitoring and Assessment Act
(1988). Included in each of these acts is the directive
to monitor and account for the ecological ‘‘health’’ of
the resources for which they are responsible.
The difficulties associated with statistically-ade-
quate ecological assessments make conventional field
survey and sampling impractical if not impossible to
apply to the vast US rangelands (West 1999; Schino
et al. 2003), and do not provide timely information
for those regions that undergo frequent changes.
Remote sensing provides a viable method for mon-
itoring and characterizing rangeland vegetation.
Moderate spatial resolution data collected by Landsat
and other satellites provide a temporal resolution and
economy not matched by aerial or ground methods.
Several studies have demonstrated the utility of
Landsat data for mapping rangeland vegetation (Todd
et al. 1998; Ikeda et al. 1999; Hostert et al. 2003;
Schino et al. 2003; Cingolani et al. 2004). Todd et al.
(1998) found that Landsat bands were able to account
for 67% of the variance in the measured biomass in
grazed rangelands of Colorado (USA). However,
most of the studies used correlation to analyze the
relationship between field-measured vegetation data
and remotely sensed reflectance values, and fewer
studies have tested the relationship with regression
models. Ancillary data such as elevation, slope,
aspect, and soil type are often used together with
the reflectance values recorded by Landsat to assist
classification (Kozar et al. 2002; Zambon et al.
2006). For example, the association between eleva-
tion and species distribution was used to separate
mountain fir (elevation range 2652–2774 m) from
spruce-fir (elevation [2774 m), though their spectral
reflectance values were identical (Homer et al. 1997).
Precise ground-based measurements are required
to calibrate the Landsat pixels prior to characterizing
the rangeland vegetation. However, conventional
ground-based methods of collecting calibration data
are expensive. Seefeldt and Booth (2006) have
demonstrated the utility of aerial images for moni-
toring and measuring rangeland ground cover using
Very Large Scale Aerial (VLSA) photography.
VLSA photography refers to low-altitude, high-
resolution (1–20 mm GSD) imagery that is acquired
intermittently during an aerial rangeland survey
(Booth and Cox 2006). Booth et al. (2005a) reported
that a sub-millimeter contact area had the highest
correlation of measured cover to known values.
VLSA surveys are excellent for acquiring large
sample numbers, but they are a sampling, not a
mapping method.
The goal of this study was to assess the potential of
VLSA data (spatial resolution 1 mm) for calibrating
Landsat Thematic Mapper 5 (TM5) data (spatial
resolution 30 m) so as to extend the reach of highly
detailed imaging with the field-of-view and period-
icity of space imagery. Specific objectives of this
study were: (1) Test the relationship between ground
cover measurements derived from VLSA photos and
Landsat reflectance values through regression analy-
ses; (2) Compare the advantages of using transformed
bands instead of raw Landsat reflectance values; and
(3) Assess the value of incorporating the physio-
graphic variables (elevation, slope and aspect) in the
regression models. We used approximately two thirds
of the VLSA data for testing the relationship and
model development between the percent vegetation
derived from VLSA and Landsat reflectance values.
The rest of the data was used to test regression model
validity. If a useful relationship exists, then a multi-
scale approach that combines VLSA images and early
growing season medium resolution satellite data
could be used for some routine monitoring and for
mapping rangeland plant communities.
56 Agroforest Syst (2008) 73:55–64
123

Materials and methods
Study area
The 9,000-ha Hay Press Creek Pasture is in the
northeast part of the Green Mountain Common
Allotment, southwest of Jeffrey City, Wyoming
(42°270N, 107°550W) between Green Mountain
and the flood plain of the Sweetwater River (Fig. 1).
It is managed by the US Department of the Interior,
Bureau of Land Management, Lander Field Office.
The pasture contains 85% sagebrush/grassland (Arte-
misia tridentata Nutt. ssp. wyomingensis Beetle and
Young) (7,526 ha), 12% riparian area (1,096 ha), 2%
Playa (161 ha) and 0.5% road (30 ha) (Beetle 1960;
Booth et al. unpublished data).
Very large scale aerial photography (VLSA)
We acquired 162 color digital aerial images on 8th
and 9th June 2004, using a light airplane (225-kg
empty weight, fixed wing, three-axis), a navigation
and camera-triggering system, a digital camera, and a
laser rangefinder (Booth and Cox 2006). The aircraft
speed ranged between 68 and 95 km/h. The naviga-
tion system was powered by Tracker software
(Track’Air B.V., Oldenzaal, The Netherlands) on a
laptop computer interfaced with (1) a central navi-
gation box, (2) a differentially-corrected geographic
positioning system and (3) a 15-cm in-cockpit pilot
display. The navigation system was programmed
using a laptop PC to automatically trigger the camera
at 800-m intervals along 12 flight lines. We used a
Canon EOS 1Ds 11.1-megapixel single lens reflex,
color (RGB) digital camera with a Canon 600 mm
f/4.0 EF lens plus a 1.49teleconverter to yield the
equivalent of a 840 mm, f/5.6 lens. (Canon USA,
Lake Success, NY, USA). Shutter speed was manu-
ally set for 1/4,000th second with safety shift enabled
to allow the shutter speed to slow in inadequate light.
The camera was interfaced with a laptop PC
(3.2-GHz, 40-GB-hard drive) running Canon Remote
Capture software and images were stored directly on
the hard drive. Images were initially saved as RAW
(10MB compressed) files and later converted to
24-bit, 31 MB, 4064 92704-pixel TIFF files for
analysis. A Riegl 3100VHS laser rangefinder (Riegl,
Orlando, FL, USA) was used as an altimeter in
conjunction with LaserLOG software (Booth et al.
2006a) to continuously read and record the airplane’s
altitude above-ground-level (AGL) below 300 m.
Altitude was displayed for the pilot on the screen of
the laptop storing the images, while stored data were
saved for later correlation with images. Planned flight
altitude for the upland survey was 100 m AGL with
an expected image resolution of 1 mm ground surface
distance (GSD) and a 3 94 m field of view. The
flight plan of 12 E-W flight lines totaling 121 km was
created by extracting coordinates of user-defined
points drawn on a digital raster graphic in ArcView
GIS 3.3 (ESRI, Redlands, CA), then using Track’Air
SnapXYZ flight planning software to enter the
coordinates into a flight plan utilized in flight by
Track’Air SnapShot software. Photo targets were
planned on a 0.8-km grid covering the entire pasture.
A DGPS Max differentially-corrected global posi-
tioning system (DGPS) unit (CSI Wireless, Calgary,
Alberta) with sub-meter accuracy was used to guide
the pilot to the photo targets. Booth and Cox (2006)
estimated that the cost of acquiring VLSA images
was approximately $0.08 per hectare, based on the
images acquired for a 70,800-ha rangeland watershed
in Wyoming.
SamplePoint analysis
SamplePoint is a digital ‘pointframe’ designed for
point sampling digital images. With 1 mm GSD
ground-acquired images it is has comparable accu-
racy to conventional field-methods for ground-cover
Fig. 1 Location of the study area in Wyoming, USA
Agroforest Syst (2008) 73:55–64 57
123

measurements (Booth et al. 2006b). The program
loads the images from a specified data base then
systematically locates a user-defined number of
sample points in the image—in this study we used
100 and each image had about a 3 94 m field-
of-view. The software takes the user from one sample
point to the next so that errors from double counting
or missing a point are avoided. SamplePoint identifies
each sample point by 4 red, 1-pixel-thick lines in a
crosshair pattern that surround but do not cover, the
sample-point pixel. Thirty buttons under the image
are labeled for users to identify designated ground-
cover characteristics. When a user classifies a point
by clicking a button, the user’s classification is saved
to the database and the next point automatically
shows up in the image window at the user-defined
zoom level. Labels can be user-defined and we
defined 16 categories: bare ground, litter, rock,
biological crust, perennial grass, annual grass, peren-
nial forb, annual forb, little sagebrush (Artemisia
arbuscula Nutt. ssp. arbuscula), silver sagebrush
(Artemisia cana Pursh), big sagebrush (Artemisia
tridentata Nutt. ssp. wyomingensis Beetle and
Young), Greene’s rabbitbrush (Chrysothamnus gree-
nei (Gray) Greene), greasewood (Sarcobatus
vermiculatus (Hook.) Torr.), spineless horsebrush
(Tetradymia canescens DC.), plains pricklypear
(Opuntia polyacantha Haw.), and unknown. The
software allows a user to zoom in or out as needed to
understand the context or detail of an image pixel.
Measuring vegetation cover (%) using
VegMeasure
We used VegMeasure v1.6.0, a software program
developed at Oregon State University to measure plant
cover on rangeland (Louhaichi et al. 2001; Johnson
et al. 2003). VegMeasure quantifies areas of specific
color, and does so for large batches of digital images
through rapid binary classification. The green leaf
algorithm (Louhaichi et al. 2001) was used to measure
green cover and the blue band and brightness algo-
rithms (Johnson et al. 2003) were used to measure bare
ground. The blue band and brightness algorithms were
used for bare ground because, in our experience, they
more accurately separated bare ground from other
parameters of ground cover. The detection threshold in
VegMeasure is a pre-process user-adjustable method
in which users are presented with side by side views of
the original image and its simplified, black and white
binary-classified depiction. The threshold for the
characteristic under consideration is manipulated
looking at the original color image. To objectively
calibrate the threshold we used SamplePoint software
(Booth et al. 2006b) to classify a 10% subset of the
images by setting the threshold to reflect the Sample-
Point measure of vegetative cover (Booth et al. 2005b,
2006b). SamplePoint has a 98% potential accuracy and
a 92% practical accuracy, the difference being due to
the pixel mixing inherent in imaging methods (Booth
et al. 2006b).
Physiographic data
Each VLSA image is geocoded by the Track’Air
system using the system’s GPS (Booth and Cox
2006). Image centerpoints were added as a layer to a
GIS (ArcView 3.3, ESRI, Redlands, CA, USA)
containing a digital elevation model (DEM) acquired
from the Wyoming Geographic Information Science
Center (WYGISC, University of Wyoming, Laramie,
WY, USA). The DEM was used to generate aspect
and slope raster layers. Using the StatMod extension
(Garrard 2002), all VLSA image centerpoints were
queried for elevation, aspect and slope.
Landsat TM5 data
A cloud-free, TM5 scene (Path 36 - Row 30) acquired
by Landsat on June 2nd, 2004 was obtained from the
Upper Midwest Aerospace Consortium (UMAC).
The scene was corrected for geometric and terrain
distortions at the US Geological Survey—Earth
Resource Observation Satellite Data Center in Sioux
Falls, SD (USA). Raw digital numbers associated
with the TM5 pixels were converted to at-satellite
radiance using gain and offset values provided by the
USGS (Markham and Barker 1986; Chander and
Markham 2003). Using ERDAS Imagine
Ò
software
(Atlanta, GA, USA) geographic location of each
VLSA image (center point) was located on the
Landsat image. Mean reflectance values in each band
were obtained for the 162 VLSA images. Several
transformed bands and vegetation indices were
derived from the six multi-spectral bands of Landsat
image (Table 1). Image transformation methods were
used to reduce the dimensionality of the six Landsat
bands to fewer bands or indices that could be related
58 Agroforest Syst (2008) 73:55–64
123

to certain vegetation phenomenon (Jenson 2000;
Campbell 2002).
The 162 VLSA data points were randomly divided
into two groups. Approximately two thirds (or 105) of
the data points were used for assessing the relation-
ship between the percent vegetation estimated from
the photos and Landsat image. The remaining one
third (or 57) of the data points were set aside for
validation.
Regression model development and validation
Stepwise multivariate regression analysis was used to
select a subset from the six multi-spectral Landsat
bands that accounted for most of the variability in the
dependent variable (Montgomery and Peck 1992;
Sivanpillai et al. 2006). We used an alpha value of
0.05 as the criteria for retaining a variable that was
selected in each step. The vegetation-cover measure-
ments (dependent variable) derived from VLSA
imagery using VegMeasure software were regressed
against the following four sets of independent
variables: (1) six Landsat bands, (2) six Landsat
bands and physiographic data, (3) seven transformed
bands described in Table 1, and (4) seven trans-
formed bands and physiographic data. Independent
variables of the final model were selected based on a
combination of both their individual contribution to
the model and the overall adjusted R
2
value. A similar
procedure was repeated with vegetation-cover mea-
surements derived from the SamplePoint software.
Regression models developed for predicting veg-
etation cover were validated using the 57 data points
that were set aside. VegMeasure cover measurements
and estimates derived from the four regression
models were compared for assessing the adequacy
of the regression models. This was followed by a
comparison of the SamplePoint measurements and
estimates derived from the four regression models.
We also measured the correlation between Sample-
Point and VegMeasure data. To determine if the
SamplePoint and VegMeasure data sets were statis-
tically independent of each other and that an
assumption of homoscedasticity was valid even
though the VegMeasure detection threshold was
calibrated using SamplePoint, we tested the null
hypothesis that the mean of regression residuals (the
correlation coefficient is identical to the slope of the
regression line) was not equal to zero.
Results
Model calibration
The regression model containing Landsat bands 1
(blue), 3 (red), 5 (near-infrared 2) and 7 (mid-
infrared) was the best subset among the 6 raw
Landsat bands to predict the percent vegetation
measured by VegMeasure from 1 GSD (100-m
AGL) VLSA imagery (Table 2). All the independent
variables were significant in the regression model
(F=29.16; P\0.001) and the adjusted R
2
value
was 52%. The root mean square error associated with
this model was 8. However, the regression model
containing the Landsat bands 1, 3 and 7 and elevation
was the best predictor of measured percent vegeta-
tion. All four independent variables in the above
Table 1 Transformed
bands and indices derived
from Landsat data that were
used in this study
Name
Greenness Condition Index (GI): Band 4/Band 3
Normalized Difference Vegetation Index (NDVI): Band 4 -Band 3/Band 4 +Band 3
Vegetation Condition Index (VCI): Band 7/Band 4
Mid-IR/Red Reflectance Index (MIRI): Band 7 -Band 3/Band 7 +Band 3
Tasseled Cap Brightness: 0.29 Band 1 + 0.25 Band 2 + 0.48 Band 3 + 0.56 Band 4 + 0.44
Band 5 + 0.17 Band 7
Tasseled Cap Greenness: -0.27 Band 1 -0.22 Band 2 -0.55 Band 3 + 0.72 Band 4 + 0.07
Band 5 -0.16 Band 7
Tasseled Cap Wetness: 0.14 Band 1 + 0.18 Band 2 + 0.33 Band 3 + 0.34 Band 4 -0.62
Band 5 -0.42 Band 7
Agroforest Syst (2008) 73:55–64 59
123

regression model were significant (F= 49.95;
P\0.001) and the adjusted R
2
value increased to
65%. Reflectance values recorded by Landsat bands 3
and 7 decreased with increases in vegetation cover as
measured by VegMeasure (inversely proportional),
whereas values in band 1 increased with vegetation
cover. Elevation was the only physiographic variable
included in the model. The root mean square error
associated with this model was 7.
Among the transformed Landsat bands, the regres-
sion model containing Vegetation Condition Index
(Band 7/Band 4), mid-IR/Red reflectance index
(MIRI), and tasseled cap brightness, greenness and
wetness bands was the best subset (F=26;
P\0.001) to predict the measured percent vegeta-
tion (Table 2). The adjusted R
2
value was 55% and
this was only 3% more than the model containing raw
Landsat bands. VCI, MIRI, tasseled cap brightness,
greenness and wetness bands, and elevation (m) were
the best subset (F=36; P\0.001) to predict the
percent vegetation (Table 2). This model had the
highest adjusted R
2
value of 67% (n=105) and all
six transformed bands were significant (F=36.7;
P\0.001). MIRI values were directly proportional
to percent vegetation content whereas all other
independent variables included in this model were
inversely proportional to percent vegetation cover.
Elevation was the only physiographic variable
included in the model. The root mean square error
associated with this model was 6.8.
The regression model containing Landsat bands 5
(mid-infrared1) and 7 (mid-infrared2) was the best
subset (F=12.87; P\0.001) among the 6 raw
Landsat bands to predict the percent vegetation
measured by SamplePoint, and the adjusted R
2
value
was 19% (Table 3). Inclusion of elevation did not
improve the overall significance (F=9.8;
P\0.001), nor adjusted R
2
value (19%). Reflectance
Table 2 Models selected
by the stepwise regression
procedure when percent
vegetation estimate, derived
from VLSA imagery using
VegMeasure software, were
regressed against
combinations of Landsat
(original and transformed)
bands and physiographic
data as independent
variables
All regression models
significant at 95% level
(a=0.05)
Landsat TM5 bands
% cover =2.3 +0.68 Band 5 -1.91 Band 7 -2.12 Band 3 +2.45 Band 1
Adjusted R
2
=52%; RMSE =8.2
Landsat TM5 bands and physiographic data
% cover =-308 +0.16 Elevation -0.98 Band 7 -1.64 Band 3 +2.27 Band 1
Adjusted R
2
=65%; RMSE =7.0
Transformed Landsat TM5 bands
% cover =870 -965 Band 7/Band 4 +1159 MIRI -1.9 TC_B -12 TC_G -3.7 TC_W
Adjusted R
2
=55%; RMSE =8.0
Transformed Landsat TM5 bands and physiographic data
% cover =347 -744 Band 7/Band 4 +1043 MIRI -1.4 TC_B -10 TC_G -2 TC_W
+0.15 Elevation
Adjusted R
2
=67%; RMSE =6.8
Table 3 Models selected
by the stepwise regression
procedure when percent
vegetation estimate, derived
from VLSA imagery using
SamplePoint software, were
regressed against
combinations of Landsat
(original and transformed)
bands and physiographic
data as independent
variables
All regression models
significant at 95% level
(a=0.05)
Landsat TM5 bands
% cover =86.7 -1.38 Band 7 +0.56 Band 5
Adjusted R
2
=19%; RMSE =9.0
Landsat TM5 bands and physiographic data
% cover =5.6 -1.11 Band 7 +0.4 Band 5 +0.04 Elevation
Adjusted R
2
=19%; RMSE =9.0
Transformed Landsat TM5 bands
% cover =-95.4 +1150 NDVI -89.9 MIRI -0.45 TC_B -6.4 TC_G
Adjusted R
2
=20%; RMSE =8.9
Transformed Landsat TM5 bands and physiographic data
% cover =-95.4 +1150 NDVI -89.9 MIRI -0.45 TC_B -6.4 TC_G
Adjusted R
2
=20%; RMSE =8.9
60 Agroforest Syst (2008) 73:55–64
123

values recorded by Landsat band 7 decreased with
increases in vegetation cover as measured by Sam-
plePoint (inversely proportional), whereas band 5
values increased with vegetation cover. The root
mean square error was 9 in both models.
Among the transformed Landsat bands, the regres-
sion model containing Normalized Vegetation
Difference Index (NDVI), MIRI, and tasseled cap
brightness and greenness bands was the best subset
(F=7.5; P\0.001) to predict the measured percent
vegetation (Table 3). Inclusion of elevation values
did not improve the overall significance (F=7.5;
P\0.001) or the adjusted R
2
value (20%). The root
mean square error associated with this model was 8.9.
Model validation
Two models that included the elevation as an
independent variable were selected and validated
using the 57 observations that were set aside earlier
(Table 2). Predicted values from the regression
models were compared with the percent vegetation
derived from VLSA imagery using VegMeasure
(Fig. 2a, b) and checked for model adequacy. Veg-
etation cover, as measured by VegMeasure from
VLSA imagery, was significantly correlated to the
predicted values using both raw (r=0.83,
P\0.001) and transformed (r=0.86, P\0.001)
Landsat bands. Points were spread above and below
the 1:1 line in the 20% to 50% vegetation cover
range. In the lower ranges (\20%) both models tend
to over-predict the percent vegetation cover and in
the higher ranges ([55%) the models under-predict
the values. Since the models developed using Sam-
plePoint measurements had lower R
2
values
(Table 3), the validation process did not yield any
insights regarding the relationship between the mea-
surements and Landsat reflectance values.
Correlation between VegMeasure and
SamplePoint measurements
The percent cover values obtained from VegMeasure
and SamplePoint were well correlated (r=0.64,
P\0.001) and the assumption of data-set homosce-
dasticity was justified (residual mean =0.001,
P=0.99); however, for most of the points the
VegMeasure estimates were lower than the Sample-
Point estimates (Fig. 3). All types of vegetation
including their non-chlorophyll parts (stems) were
included in SamplePoint whereas VegMeasure used
reflectance values. This could explain the lower
values obtained from the VegMeasure protocol.
Fig. 2 Regression of vegetation cover derived from VLSA
imagery with models involving raw (a) and transformed (b)
Landsat bands (Table 1)
Fig. 3 Scatter plot of the percent vegetation cover estimates
obtained from the aerial photos through the SamplePoint and
VegMeasure sampling protocols (r=0.64)
Agroforest Syst (2008) 73:55–64 61
123

Discussion
VegMeasure and Landsat
The regression model with a subset of transformed
Landsat bands and elevation values was able to
account for 67% of the variance in the percent
vegetation derived from VegMeasure, whereas a
similar model with raw Landsat bands accounted
for 65% (Table 2). Models with Landsat bands were
able to account for only 52% and 55% of the
variability in the vegetation estimates derived from
VegMeasure. Several factors on the ground could
have contributed to the differences observed between
Landsat reflectance values and VegMeasure values.
First, the reflectance of wet soil in the mid infrared
region is lower in comparison to the reflectance
values of dry soil of the same type. Second, numerous
studies have demonstrated the near infrared (band 4)
is a better predictor of vegetation biomass than are
bands 1 (blue) and 3 (red). However, the reflectance
values of band 4 are also influenced by soil moisture.
The presence of wet soil patches following snow melt
could have resulted in differences in the near infrared
reflectance values. In addition, background soil color
could have also contributed to the differences though
the vegetation cover could have been similar. Todd
et al. (1998) found the red band (band 3) to be a
useful discriminator of vegetation in semi-arid
regions, since the reflectance values are a combina-
tion of bright soil background and some dry
vegetation. The regression model using the trans-
formed Landsat bands increase the adjusted R
2
by
only 3%. Inclusion of tasseled cap wetness band in
both models (Table 2) indicates that soil moisture
could have influenced the reflectance values.
Incorporating elevation values improved the pre-
dictive power of the regression model. Increasing
elevation in this pasture was correlated with increas-
ing bare ground (Booth et al. unpublished data) and
thus with greater soil reflectance. The observed
increase in bare ground at the higher elevations of
the pasture could be due to a decrease in plant
cover—particularly rhizomatous grasses—resulting
from lower water infiltration and storage. The soils
in this area often appear to be of coarser materials
than soils at lower elevations. The higher-elevations
within this pasture do not appear to receive the
precipitation nor have the abundance of rock that
might account for less bare ground at elevations on
Green Mountain above the Hay Press Creek Pasture.
That elevation was a significant variable in the
regression models (raw and transformed Landsat
bands) is indicative of the ability of TM data to
capture important ecological differences that can be
correlated with finer data. Exclusion of slope and
aspect as a significant independent variable indicates
either these variables were not influencing the
vegetation reflectance values or Landsat instruments
were not sensitive enough to capture the differences.
SamplePoint, Landsat, and VegMeasure
Regression models using SamplePoint measurements
as the dependent variable with combinations of
Landsat data accounted for only 20% of the variance
(Table 3). Mid infrared bands (5 and 7) accounted for
the variability in SamplePoint measurements. Eleva-
tion was not a major contributor to the models using
SamplePoint measurement and the absence of a
significant relationship with the blue and red bands (1
and 3) indicates a lack of influence from the higher-
elevation bare ground—thus highlighting the differ-
ence between SamplePoint and VegMeasurement.
The degree of association between SamplePoint
measurements and Landsat reflectance values were
lower, though significant. Landsat and VegMeasure-
ment values were based on reflectance, whereas
SamplePoint measurements included spectral, tex-
ture, and context information derived through human
interpretation. SamplePoint measurements were anal-
ogous to an inventory of various plant life forms
irrespective of their reflectance characteristics. For
example, when a SamplePoint user found pixels that
fell on non-green parts of vegetation (live stems), it
was included in the vegetation class.
The difference between protocols and high and
low values might also be due to the fact that VLSA
data contain some motion blur at the pixel level. Blur
increases the amount of judgment a SamplePoint user
must exercise and therefore increases the suscepti-
bility of users to a bias for green color (Booth et al.
2005a,b). It is possible that in addition to assigning
brown stem vegetation to total plant cover, there were
also a percentage of drab categories (bare ground,
rock, litter) that were classified as green due to the
combination of motion blur and human bias. The
differences in the methods suggests that accurate
62 Agroforest Syst (2008) 73:55–64
123

Landsat analysis may be enhanced by developing
correlations among SamplePoint and VegMeasure
analyses, and between VegMeasure and Landsat data.
Understanding how these separate analyses relate to
each other may open a door that will allow Landsat
data to be used to signal subtle, but ecologically
important change and a need for a VLSA-type survey
to identify and quantify the change.
Application to rangeland monitoring
A careful assessment of bare-ground measurement
accuracy implies that at 1-mm GSD, SamplePoint
and VegMeasure are 90% and 80% respectively; but,
the latter is true only where there is a clear spectral
separation between ground-cover characteristics
(Booth et al. 2006b). VegMeasure will batch process
5-megapixel images at about 1 image/s (Booth et al.
2005b) and 11 megapixel images at\10 sec/image. It
requires about 15 min for a person to read 100 points
on an image using SamplePoint. Therefore, we
recommend the use of VLSA and VegMeasure with
Landsat for modeling extensive land areas with good
contrast between ground-cover characteristics where
*80% is an acceptable level of accuracy. Where
greater accuracy is required or where the spectral
separation of ground-cover characteristics is not
good, we recommend the use of VLSA and Sample-
Point with Landsat but emphasize that more work is
required to develop a better relationship between
SamplePoint cover data and Landsat bands or
indexes.
Conclusions
Regression models incorporating elevation values
and Landsat bands as independent variables were
better predictors of range vegetation cover (R
2
=65–
67%; P\0.001) than those models containing only
Landsat bands (R
2
=52–55%; P\0.001). However
the ability of Landsat bands to predict vegetation
cover at the low (\20%) and high ([55%) ranges is
somewhat diminished. Landsat reflectance values
accounted for more of the variability in the percent
vegetation cover measurements derived from Veg-
Measure protocol (R
2
=52–67%; P\0.001) than
from those derived from the SamplePoint protocol
(R
2
=19–20%; P\0.001). VLSA imagery can be
used to calibrate Landsat data for estimating percent
vegetation in semi-arid rangelands, thereby reducing
the need to conduct expensive field and plot surveys.
Calibrated Landsat data can be used to model
rangeland vegetation conditions in Wyoming and
similar semi-arid environments but more work is
needed to improve the models relating VLSA and
Landsat data.
Acknowledgements Authors thank Mr. Samuel Cox, Remote
Sensing Technician, USDA-ARS, for technical support for the
aerial survey and for data management and analysis; and Mr. Joe
Nance, pilot and owner, CloudStreet LLC, for acquiring the
VLSA imagery. Technical assistance for the aerial survey was
funded in part by a grant from the Wyoming State Office of the
Bureau of Land Management to D.T. Booth. Cost of the survey
was funded by a grant from the USDA Small Business
Innovative Research program to T. W. Meikle, Vice-President
for Research and Development, Bitterroot Restoration, Inc.
(BRI) through a USDA-ARS Cooperative Research and
Development Agreement with D.T. Booth. The Landsat image
was made available through the Upper Midwest Aerospace
Consortium (UMAC) at the University of North Dakota
(NDUNV48602SVNPL). Portions of this research were
conducted as part of the WyomingView—Applied Research
Program and financial support for the first author was
provided through the USGS/AmericaView grant to UW
(AMVIEW46014). Authors thank the two internal reviewers
and the external reviewers for their valuable suggestions.
References
Beetle AA (1960) A study of sagebrush. The section Triden-
tatae of Artemisia. University of Wyoming, Agricultural
Experiment Station Bulletin #368, Laramie, Wyoming
Booth DT, Cox SE (2006) Very large scale aerial photography
for rangeland monitoring. Geocarto Int 21:27–34
Booth DT, Cox SE, Berryman RD (2006a) Precision mea-
surements from very large-scale aerial imagery. Environ
Monit Assess 112:293–307
Booth DT, Cox SE, Berryman RD (2006b) Point Sampling
Digital Imagery with ‘SamplePoint’. Environ Monit
Assess 112:293–711
Booth DT, Cox SE, Meikle TW, Fitzgerald C (2005a) The
accuracy of ground cover measurements. Rangeland Ecol
Manage 59:179–188
Booth DT, Cox SE, Johnson DE (2005b). Detection-threshold
calibration and other factors influencing digital measure-
ments of bare ground. Rangeland Ecol Manage 58:
598–604
Campbell JB (2002) Introduction to remote sensing. The
Guilford Press, New York
Chander G, Markham B (2003) Revised Landsat-5 TM radio-
metric calibration procedures and postcalibration dynamic
ranges. IEEE Trans Geosci Remote Sens 41:2674–2677
Cingolani AM, Renison D, Zak MR, Cabido MR (2004)
Mapping vegetation in a heterogeneous mountain
Agroforest Syst (2008) 73:55–64 63
123

rangeland using landsat data: an alternative method to
define and classify land-cover units. Remote Sens Environ
92:84–97
FWS (Fish and Wildlife Service) (2005) Status review com-
pleted: greater sage-grouse not warranted for listing as
endangered or threatened. http://news.fws.gov/NewsReleases/
R9/4D0B98E6-DE22-18A8-5FBC25A1E928B298.html.
Cited 3 Feb 2008
Garrard C (2002) StatMod. http://bioweb.usu.edu/gistools/
statmod. Cited 3 Feb 2008
Gilmanov TG, Johnson DA, Saliendra NZ, Svejcar TJ, Angell
RF, Clawson KL (2004) Winter CO2 fluxes above sage-
brush-steppe ecosystems in Idaho and Oregon. Agric For
Meteorol 126:73–88
Homer CG, Ramsey RD, Edwards TC, Falconer A (1997)
Landscape cover-type modeling using a multi-scene TM
mosaic. Photogramm Eng Remote Sens 63:59-67
Hostert P, Ro
¨der A, Hill J (2003) Coupling spectral unmixing
and trend analysis for monitoring of long-term vegetation
dynamics in Mediterranean rangelands. Remote Sens
Environ 87:183–197
Ikeda H, Okamoto K, Fukuhara M (1999) Estimation of
aboveground grassland phytomass with a growth model
using Landsat TM and climate data. Int J Remote Sens
20:2383–2294
Jensen JR (2000) Introductory digital image processing: a
remote sensing perspective. Prentice Hall, New Jersey
Johnson DE, Vulfson M, Louhaichi M, Harris NR (2003)
VegMeasure v.1.6 user’s manual. Department of Range-
land Resources, Oregon State University, Corvallis,
Oregon
Kolb KJ, Sperry JS (1999) Differences in drought adaptation
between subspecies of sagebrush (Artemisia Tridentata).
Ecology 80:2373–2384
Kozar B, Lawrence R, Long D (2002) Soil phosphorous and
potassium mapping using a spatial correlation model
incorporating terrain slope gradient. Precis Agric 3:
407–417
Louhaichi M, Borman MM, Johnson DE (2001) Spatially
located platform and aerial photography for documenta-
tion of grazing impacts on wheat. Geocarta Int 16:63–68
Markham BL, Barker JL (1986) Landsat MMS and TM post-
calibration dynamic ranges, exoatmospheric reflectances
and at-satellite temperatures, Landsat Technical Note 1.
Earth Observation Satellite Co., Lanham, MD
Montgomery DC, Peck EA (1992) Introduction to linear
regression analysis, 2nd edn. John Wiley & Sons, New
York
Morgan JA, Mosier AR, Milchunas DG, LeCain DR, Nelson
JA, Parton WJ (2004) CO2 Enhances productivity, alters
species composition, and reduces digestibility of short-
grass steppe vegetation. Ecol Appl 14:208–219
NS-GCT (Nevada Sage-Grouse Conservation Team) (2004)
Sage-grouse conservation plan for Nevada and eastern
California. Nevada Department of Wildlife, Reno,
Nevada, 108 pp. Available at http://www.ndow.org/wild/
conservation/sg/plan/SGPlan063004.pdf cited 3 Feb 2008
Schino G, Borfecchia F, Cecco LD, Dibari C, Iannetta M,
Martini S, Pedrotti F (2003) Satellite estimate of grass
biomass in a mountainous range in central Italy. Agrofor
Syst 59:157–162
Seefeldt SS, Booth DT (2006) Measuring plant cover in
sagebrush steppe rangelands: a comparison of methods. J
Environ Manage 37:703–711
Sivanpillai R, Smith CT, Srinivasan R, Messina MG, Wu XB
(2006) Estimation of managed loblolly pine stand age and
density with Landsat ETM +data. For Ecol Manage
223:247–254
Todd SW, Hoffer RM, Milchunas DG (1998) Biomass esti-
mation on grazed and ungrazed rangelands using spectral
indices. Int J Remote Sens 19:427–438
West NE (1999) Accounting for rangeland resources over
entire landscapes. In: Eldridge D, Freudenberger D (eds)
People and rangelands: building the future. Proceedings of
the VI International Rangeland Congress, Queensland
4814, Australia
Zambon M, Lawrence RL, Bunn A, Powell S (2006) Effect of
alternative splitting rules on image processing using
classification tree analysis. Photogramm Eng Remote
Sens 72:25–30
64 Agroforest Syst (2008) 73:55–64
123