Identifying land cover variability distinct from land cover change:

Cheatgrass in the Great Basin

Bethany A. Bradley
*
, John F. Mustard
Brown University, Department of Geological Sciences, 324 Brook Street Providence, RI 02912, United States
Received 3 May 2004; received in revised form 24 August 2004; accepted 29 August 2004
Abstract
An understanding of land use/land cover change at local, regional, and global scales is important in an increasingly human-dominated
biosphere. Here, we report on an under-appreciated complexity in the analysis of land cover change important in arid and semi-arid
environments. In these environments, some land cover types show a high degree of inter-annual variability in productivity. In this study, we
show that ecosystems dominated by non-native cheatgrass (Bromus tectorum) show an inter-annual amplified response to rainfall distinct
from native shrub/bunch grass in the Great Basin, US. This response is apparent in time series of Landsat and Advanced Very High
Resolution Radiometer (AVHRR) that encompass enough time to include years with high and low rainfall. Based on areas showing a similar
amplified response elsewhere in the Great Basin, 20,000 km
2
, or 7% of land cover, are currently dominated by cheatgrass. Inter-annual
patterns, like the high variability seen in cheatgrass-dominated areas, should be considered for more accurate land cover classification. Land
cover change science should be aware that high inter-annual variability is inherent in annual dominated ecosystems and does not necessarily
correspond to active land cover change.
D 2004 Elsevier Inc. All rights reserved.
Keywords: Land use/land cover change; Great Basin; Cheatgrass; Bromus tectorum; Invasive species; Time series; NDVI; Inter-annual variability
1. Introduction
A challenge to land cover change science is distinguish-
ing change in land cover from variability within land cover
types. Change implies a shift in land cover type, frequently
associated with land use, while inter-annual variability in
growth is an inherent characteristic of some vegetated
communities. In the semi-arid Great Basin, US (Fig. 1),
inter-annual variability distinct from land cover change is
prominent in land cover types dominated by cheatgrass
(Bromus tectorum). Correctly identifying and interpreting
change in land cover and/or land use is of great interest in
the field of environmental change (Dale, 1997; Lambin et
al., 2001; Turner, 2003; Vitousek et al., 1997), so inter-
annual patterns caused by communities like cheatgrass
should be recognized.
In order to identify change, it is important to have a
baseline of land cover at national and international scales.
Global assessments of land cover have been performed
based on single year phenologies using Normalized Differ-
ence Vegetation Index (NDVI) data from Advanced Very
High Resolution Radiometer (AVHRR) and the Moderate
Resolution Imaging Spectroradiometer (MODIS) (Defries &
Townshend, 1994; DeFries et al., 1995; Friedl et al., 2002;
Loveland et al., 2000, 1999). Multiple phenology-based
mapping methods have been employed, including principle
component analysis, fourier analysis, and decision tree
approaches (Defries & Townshend, 1994; Homer et al.,
1997; Jakubauskas et al., 2002). However, because most
land cover algorithms rely on single year or averaged
phenologies to differentiate between land cover types, they
do not capture inter-annual variability. Land cover with high
inter-annual variability could distort the baseline from
which change is measured depending on growing conditions
during the baseline year (Lambin, 1996).
0034-4257/$ - see front matter D 2004 Elsevier Inc. All rights reserved.
doi:10.1016/j.rse.2004.08.016
* Corresponding author. Tel.: +1 401 863 9845.
E-mail address: Bethany_Bradley@brown.edu (B.A. Bradley).
Remote Sensing of Environment 94 (2005) 204213
www.elsevier.com/locate/rse
Landsat-scale land cover change detection relies on an
accurate interpretation of baseline conditions and change in
surface spectral properties over time. For example, the
conversion of vegetated surfaces to man-made materials
through urbanization can be readily identified and mapped
using remotely sensed tools (Masek et al., 2000). Similarly,
land cover dynamics in tropical regions such as cutting of
forest for pasture and secondary regrowth have well
documented trajectories of spectral properties over time
tied to land cover (Adams et al., 1995; Skole & Tucker,
1993). In the Great Basin, baseline conditions from which to
measure change are difficult to define. Here, changes in
remotely sensed surface properties can be a result of a high
degree of species elasticity in response to rainfall, not land
cover change. A comparison of NDVI for this land cover
type between low and high rainfall periods could suggest a
significant increase in vegetation, or the reverse, a dramatic
loss. Inter-annual variability makes identification of a
baseline increasingly complex.
Regional scale change detection typically involves
interpretation of NDVI time series from sensors such as
AVHRR and MODIS. Time series analyses have been used
to identify changes in the Sahel (Tucker et al., 1991), and to
estimate changes to growing season attributed to climate in
the northern latitudes (Myneni et al., 1997, 1998; Shabanov
et al., 2002; Zhou et al., 2003, 2001) and in China (Young &
Wang, 2001). Potter et al. (2003) use AVHRR derived
fraction absorbed of photosynthetically active radiation
(FPAR) estimates to identify global disturbance events
during an 18-year record. With an increasing interest in
change detection over large spatial scales, it is important to
understand how a range of ecosystems are affected by short-
term climate patterns. In the Amazon, AVHRR NDVI time
series have been shown to be greatly influenced by
precipitation, and thus coupled to the El Nino cycle (Asner
et al., 2000). As we will show, an even more pronounced
effect exists in cheatgrass-dominated annual grasslands in
the Great Basin.
Cheatgrass inter-annual variability is distinct in the
Great Basin due to its link to rainfall. Precipitation patterns
in the Great Basin show a high degree of temporal
variability. Twenty six distributed National Climatic Data
Center rain gages with continuous records from 1990 to
2002 averaged 18 cm of rainfall per year with an inter-
annual variance about that mean of 26 cm. Rainfall during
wet years is frequently three to four times higher than during
dry years. This pattern of extreme wet and dry years causes
a dramatic response in cheatgrass. Rangeland studies have
shown that during wet years cheatgrass can be 10 times
taller and denser than during dry years (Hull & Pechanec,
1947; Stewart & Hull, 1949; Stewart & Young, 1939).
High inter-annual variability is unique to annual grass-
lands in the Great Basin. Shrub/perennial bunch grass
ecosystems are adapted to rainfall patterns. While they show
some change in productivity coupled with rainfall, the range
of variance is limited (Prince et al., 1998). Elmore et al.
(2003) showed that for a 20-cm variance in rainfall, native
perennial communities exhibited a 5% variance in live
cover, while invasive annuals had an inter-annual variance
in live cover up to five times higher. Remote sensing is a
good candidate for identifying a similar pattern in the Great
Basin because regionally, cheatgrass monocultures have
become increasingly prominent in formerly shrub/bunch
grass ecosystems (Mack, 1981; Young et al., 1972).
Our goals in this analysis are two-fold. First, we
demonstrate that annual grassland in the Great Basin, US
Fig. 1. Classifications of Great Basin land cover. (A) Native vegetation zones mapped by David Charlet and produced by the Biological Resources Research
Center, University of Nevada Reno. (B) AVHRR-based USGS classification of land cover in the Great Basin.
B.A. Bradley, J.F. Mustard / Remote Sensing of Environment 94 (2005) 204213 205
exhibits inter-annual variability of greenness in response to
rainfall that exceeds the range expected for shrub-dominated
land cover types. Second, we show that this signal of
amplified response to environmental conditions can be used
to map a land cover type not previously captured by
standard land cover mapping methods. This land cover type
is dominated by the invasive annual cheatgrass. It is
important to note that the extreme variability in inter-annual
NDVI resulting from cheatgrass dominance is not caused by
land cover change during the period of record. Rather, the
variability is an effect of cheatgrass presence. These goals
are accomplished through coordinated analysis of coarse
spatial, high temporal resolution AVHRR data and high
spatial resolution Landsat data.
2. Great Basin vegetation
Native vegetation in the Great Basin is adapted to the
deserts highly variable precipitation. The Great Basin
frequently has years of persistent drought followed by
extremely wet years. Native perennial shrubs and grasses
have adapted to these conditions by setting deep roots (in
some cases up to 30 ft), growing photosynthetically active
leaves judiciously, and investing resources in seed produc-
tion only in years when conditions are optimal (Grayson,
1993; Meyer & Monsen, 1992). Perennial shrub and bunch
grass communities typically green up in early to mid-May
and remain photosynthetically active through the hot
summer months.
Cheatgrass, an invasive annual, germinates in mid-April
in the Great Basin (except in rare years when wet conditions
permit fall germination), but remains green for only a few
weeks (Rice et al., 1992). It invests all of its resources in
growing larger to support the production of a maximum
number of seeds (Knapp, 1996; Young & Allen, 1997).
Cheatgrass growth peaks in mid-May, after which point, it
quickly senesces and remains dormant until seeds germinate
the following season. The distinct inter-annual response of
perennial vs. annual communities presented here is caused
by these differences in plant physiology.
3. Datasets
Three types of data were used in this study: field
observations of land cover, a time series of Landsat data
centred on the lower Humboldt River, Nevada, and a time
series of AVHRR weekly and biweekly data. All three were
used to assess temporal vegetation patterns in the Great
Basin.
Field observations were made in May 2003 in areas
surrounding the lower Humboldt River, Nevada (40418N,
117.51198W). Thirteen sites dominated by either cheat-
grass monoculture, or native shadscale (Atriplex spp.) and
bunch grass (Poa secunda, Agropyron spicatum) were
identified and geolocated. Polygons were then created
surrounding the observed areas. Five sites, each spanning
z4 km
2
, contained cheatgrass monoculture. The remaining
eight sites, also spanning z4 km
2
, contained native
shadscale/bunch grass and no cheatgrass. An additional
three sites, each spanning z9 km
2
, were selected on non-
vegetated salt flats. A location map and pictures of
representative land cover types are shown in Fig. 2.
Field observations were repeated in 2004 at 659
distributed localities in order to validate the predicted
Fig. 2. Location map of field sites with known land cover types. Green rectangles are cheatgrass monoculture, blue rectangles are native shadscale/bunch grass,
orange rectangles are salt flats. Sites A, B, and C correspond to the photographs shown at right.
B.A. Bradley, J.F. Mustard / Remote Sensing of Environment 94 (2005) 204213 206
distribution of high density cheatgrass (Fig. 3). Validation
localities span the Landsat scene in areas that are accessible
by road. Validation points were selected to include three
types of non-cheatgrass-dominated native vegetation: dry
desert shrubs, sagebrush steppe, and pinyon-juniper wood-
land. Cheatgrass-dominated validation points included
monoculture and dense understories of cheatgrass in dry
desert shrub and sagebrush steppe ecosystems. The incor-
poration of validation points in the study allows us to assess
the accuracy of the distribution maps created with the
remotely sensed data.
A series of 10 Landsat TM and ETM+ scenes (Path 42
Row 32) were acquired for the lower Humboldt River,
Nevada between 5/15/88 and 5/11/2001 (Table 1). The 30 m
resolution of Landsat captures local, landscape-scale varia-
bility over time. Dates of the scenes were as close to May 15
as possible. By this time, residual snow has melted from the
valleys, and Great Basin vegetation, both shrubs and
cheatgrass, is near peak greenness. Landsat data were
converted to reflectance based on the spectral signatures of
five sites whose reflectance spectra were measured with an
ASD FieldSpec spectrometer (Elmore et al., 2000; Hall et al.,
1991; Schott et al., 1988). The use of invariant targets rather
than Landsat calibration coefficients increases the accuracy
of the reflectance conversion by better accounting for
atmospheric aerosols for each observation. The sites con-
tained no vegetation, and were therefore assumed to be
spectrally invariant over time. The five spectrally invariant
targets encompass a range of reflectance values from dark
(Pyramid Lake) to bright (salt flat) (Fig. 4). The mean spectral
signature of 25 spectra acquired at each of the five sites was
used to convert Landsat-measured radiance to reflectance. A
linear regression fit to DN vs. reflectance had an R
2
value
greater than 0.98 in all cases. NDVI values for each scene
were then calculated using Landsat bands 3 and 4.
Weekly and biweekly AVHRR NDVI composites for the
conterminous US were used for the regional analysis
(Eidenshink, 1992; Teillet & Holben, 1994). The 1-km
resolution of AVHRR allows for regional detection of
prominent processes found at the landscape scale. The
dataset spans 1/1/199012/31/2001 and was clipped to
include only the Great Basin (boundaries in Fig. 1). The
visible band for every scene was checked for snow and/or
cloud cover. Parts of scenes containing snow or clouds
(primarily in winter scenes) were discarded, creating a final
NDVI dataset consisting of 394 scenes. Signal variations
caused by orbital drift and sensor degradation have
previously been recognized (Privette et al., 1995) and
removed by subtracting NDVI time series over unvegetated
surfaces from vegetation NDVI time series (Myneni et al.,
1998). In order to account for long-term sensor bias in the
Great Basin, we removed a curve fit to the mean NDVI time
series of the unvegetated Bonneville Salt Flat, UT and Black
Rock Desert, NV. A smooth curve fit rather than the mean
signal is used to prevent any propagation of stochastic error
in the data. A linear continuum was fit to the NOAA-11
NDVI data, and a 3rd order polynomial fit to the NOAA-14
NDVI data after (Kastens et al., 2003). This continuum was
subtracted from every pixel.
Fig. 3. Locations of validation points collected in May, 2004.
Table 1
Dates of Landsat Scenes acquired for p42, r32 in western Nevada
Sensor type Date
Landsat TM 5/15/1988
Landsat TM 5/26/1989
Landsat TM 6/9/1991
Landsat TM 5/10/1992
Landsat TM 6/14/1993
Landsat TM 6/4/1995
Landsat TM 6/6/1996
Landsat TM 5/8/1997
Landsat TM 6/28/1998
Landsat ETM+ 5/11/2001
Fig. 4. Reflectance in Landsat TM bands 3 and 4 of non-vegetated,
spectrally invariant targets used to calibrate Landsat scenes to reflectance.
B.A. Bradley, J.F. Mustard / Remote Sensing of Environment 94 (2005) 204213 207
4. Amplified response of cheatgrass-dominated land
cover
Previous work has shown that Landsat time series can
detect an amplified response to rainfall characteristic of
invasive annuals in Owens Valley, CA (Elmore et al., 2003).
Cheatgrass has also been observed to have a similar
amplified response following years with above average
rainfall (Hull & Pechanec, 1947; Stewart & Hull, 1949).
Therefore, we predict that time-series of NDVI determined
from remotely sensed data will show an inter-annual
amplified response to rainfall in cheatgrass-dominated areas.
We expect to see this pattern in time series of Landsat as
well as AVHRR data.
In order to identify amplified response in the Landsat
time series, we determine the DNDVI, or the difference in
NDVI values between a high rainfall year and a low rainfall
year for each pixel. By subtracting the low rainfall NDVI
values, we normalize the high rainfall scene across
vegetation types to target areas showing an amplified
response. Native perennial communities should have low
DNDVI values because they do not respond strongly to
inter-annual rainfall variability (Elmore et al., 2003).
However, cheatgrass-dominated areas should have high
DNDVI values as a result of their amplified response to
rainfall. We selected 1995 as the high rainfall scene and
1991 as the low rainfall scene. The 1995 scene was chosen
for DNDVI calculation because its timing, June 4th, was
closer to cheatgrass peak greenness than the June 28, 1998
scene (Table 1). The low rainfall scene was selected due to
its seasonal proximity, June 9, 1991, to the high rainfall
1995 scene.
In the case of AVHRR, 2 years are not adequate for
identification of regional amplified response because rainfall
patterns are heterogeneous throughout the Great Basin. We
identified regionally high and low rainfall years based on 26
rain gages with continuous records from 1990 to 2001
distributed through Utah, Nevada, and Oregon (Fig. 5A). To
account for differences in rainfall patterns between stations,
rainfall is normalized by the decadal mean at individual
stations, making mean rainfall equal to 1. The resulting
normalized OctoberMay rainfall patterns for all 26 gages
were then averaged to identify regionally high rainfall years
(Fig. 5B). The 3 years of regionally high rainfall were 1993,
1995, and 1998. Annual rainfall was above the station
decadal average at 65%, 96%, and 88% of the rain gages
during 1993, 1995, and 1998, respectively. This indicates
that these years were regionally wet, rather than affecting
only localized areas. The remaining years were below
average at between 62% and 92% of the stations, indicating
that these years were regionally dry. To identify amplified
response, we found the maximum NDVI between April 15
and June 15 for each year. Spring timing was chosen to
coincide with peak cheatgrass productivity. DNDVI was
calculated by subtracting the median of the spring maxima
during the 8 low rainfall years from the median of the spring
maxima during the 3 high rainfall years. By using medians
rather than means, the result is not skewed by anomalously
low or high values caused by local rainfall variability,
agriculture, or sensor error. The method used to identify
regional amplified response with AVHRR assumes that
cheatgrass was present during all 3 high rainfall years in the
1990s. Changes in the distribution of cheatgrass during this
time period via expansion or decline add uncertainty to the
regional estimate.
5. Results
Fig. 6 shows mean Landsat and AVHRR time series for
the three land cover type localities identified in Fig. 2:
cheatgrass monoculture, native shadscale/bunchgrass, and
non-vegetated salt flats. In the Landsat time series,
cheatgrass shows a highly amplified response to rainfall
compared to the other land cover types following the wet
years of 1988, 1995, and 1998. Likewise, in the AVHRR
Fig. 5. Rainfall patterns in the Great Basin. (A) Distribution of 26 Great Basin rain gages with continuous 19902001 records. (B) Mean-normalized rainfall
patterns for those gages. 1993, 1995, and 1998 were the wettest years, while the remaining years were average or below average.
B.A. Bradley, J.F. Mustard / Remote Sensing of Environment 94 (2005) 204213 208
time series, cheatgrass shows an amplified response
following 1995 and 1998. The amplified response of
cheatgrass is also apparent when DNDVI values are
compared (Fig. 7). With Landsat, cheatgrass-dominated
areas have a mean DNDVI of 0.202F0.091 (1r), native
shadscale/bunch grass ecosystems have a mean DNDVI of
0.055F0.032 (1r), and salt flats have a mean DNDVI of
0F0.015 (1r). Histograms of DNDVI values for the three
land cover types show that salt flats and native ecosystems
are normally distributed. Cheatgrass, however, shows a
bimodal distribution of DNDVI values with the lower peak
covering a similar range of values as native vegetation (Fig.
7C). This suggests that the cheatgrass-dominated field sites
contained patches of native vegetation in 1995. These areas
may have since been converted to cheatgrass monoculture,
accounting for the discrepancy between the remotely sensed
data and the field observations.
With AVHRR data, we see a similar separation between
land cover types. Cheatgrass-dominated areas have a mean
DNDVI of 0.12F0.045 (1r), native shadscale/bunch grass
ecosystems have a mean DNDVI of 0.049F0.021 (1r),
and salt flats have a mean DNDVI of 0.004F0.005 (1r).
Fig. 6. Temporal NDVI responses for three land cover types. The mean time
series for the land cover type areas from Fig. 2 are shown. Due to its smaller
pixel size, many more samples from the type localities are available for
Landsat than for AVHRR. Cheatgrass shows an amplified NDVI response
following high rainfall years in both (A) Landsat and (B) AVHRRtime series.
Fig. 7. Box plots and histograms indicate the distribution of DNDVI values
at three land cover types. Boxes encompass the central 50% of values, lines
encompass 95%. Cheatgrass-dominated areas show higher DNDVI values,
a result of their amplified response to rainfall. (A) Landsat box plots. (B)
AVHRR box plots. (C) Landsat histograms.
B.A. Bradley, J.F. Mustard / Remote Sensing of Environment 94 (2005) 204213 209
All three land cover types appear normally distributed,
although there are fewer samples because of the 1-km
2
AVHRR pixel size. The mean DNDVI value for cheatgrass
land cover calculated from AVHRR is lower than the mean
DNDVI calculated from Landsat. Mixtures of shrub/
bunchgrass and cheatgrass are more likely within the 1-
km
2
spatial resolution of AVHRR, resulting in a lower
mean. The AVHRR calculation also combines multiple
years to identify DNDVI, which may account for the lower
values.
5.1. Regional amplified response
Using the distributions of DNDVI values at known land
cover types, we estimate regional occurrence of amplified
response in the Great Basin. We chose to increase the
sensitivity, or the likelihood that cheatgrass will be present
given a positive detection, by setting a detection level above
the 95% confidence interval (C.I.) for the native DNDVI
distribution. Localities with DNDVI values above this
detection level are predicted to contain cheatgrass. Predicted
cheatgrass-dominated areas are well separated from the
native population while still capturing most of the cheatgrass
population (Fig. 7). To demonstrate that populations showing
an amplified response are distinct despite changes to the
threshold level, the distributions of areas with DNDVI values
above both the 95%and 99%C.I.s are presented in Fig. 8. For
Landsat, the 95% C.I. for native shadscale/bunch grass is
N0.103 DNDVI and the 99% C.I. is N0.124 DNDVI. For
AVHRR, the 95% C.I. is N0.082 DNDVI and the 99% C.I. is
N0.102 DNDVI.
In the Great Basin, the AVHRR 95% C.I. amplified
response map predicts 20,000 km
2
, or 7% of the total land
cover to be dominated by cheatgrass. The Landsat 95% C.I.
amplified response map predicts 5000 km
2
, or 17% of the
total land cover within the Landsat scene to be dominated by
cheatgrass. In the Landsat scene, AVHRR predicts 4500 km
2
to be dominated by cheatgrass. Areas of overlap between the
two sensors are shown in Fig. 8C. Both sensors identify
similar spatial patterns of cheatgrass presence in valleys of
the eastern half of the Landsat scene. The inter-annual
variability of cheatgrass is distinct enough that it can be
identified over spatial resolutions ranging from 30 mto 1 km.
Any discrepancy between Landsat and AVHRR distributions
of DNDVI values may be due to differences in timing of
scene acquisition, differences in band width at the red and
infra-red wavelengths, and the difference in sensor resolution
mentioned previously.
Fig. 8. Predicted distribution of cheatgrass-dominated areas based on an
amplified response to rainfall. (A) Landsat distribution of amplified response
higher than the 95th (blue) and 99th (red) percentile for native vegetation.
Cultivated areas and peaks above 1800 m are masked. (B) AVHRR
distribution of amplified response higher than the 95th (blue) and 99th
(red) percentile for native vegetation. Landsat extents are outlined in black.
(C) Overlap of Landsat and AVHRR distributions of amplified response.
B.A. Bradley, J.F. Mustard / Remote Sensing of Environment 94 (2005) 204213 210
5.2. Validation
The accuracy of both Landsat and AVHRR cheatgrass
distributions was validated through field observations at
659 localities in May 2004 (Fig. 3). These localities are
distributed along accessible roadways within the Landsat
boundaries. Sites containing no cheatgrass or trace
cheatgrass (only along roadsides or under shrub canopies)
were considered to have no cheatgrass and account for 250
of the 659 observations. Sites containing cheatrass mono-
culture or a dense understory of cheatgrass in shrub
interspaces were considered to have cheatgrass and
account for 409 of the 659 observations. Localities with
a sparse understory of cheatgrass in shrub interspaces were
not considered.
Receiver operator calibration (ROC) curves for both
the Landsat and AVHRR predictions are shown in Fig.
9. Based on the validation results, a threshold of N0.103
DNDVI for Landsat accurately identifies 72% of
cheatgrass cover while misidentifying 17% of other land
cover types. A threshold of N0.082 DNDVI for AVHRR
accurately identifies 64% of cheatgrass cover while
misidentifying 15% of other land cover types. By
combining the two sensors, we identify a higher
percentage of cheatgrass occurrence. Sites identified by
either Landsat or AVHRR using the previous threshold
levels accurately identify 81% of cheatgrass cover while
misidentifying 25% of other land cover types. The
validation results show a decreased sensitivity compared
to the original 13 field sites. The higher estimate derived
from the original field sites is likely a result of spatial
autocorrelation. By using validation points distributed
across the Landsat scene, we can more precisely define
the accuracy of the maps.
6. Discussion
This study shows that surfaces with high inter-annual
variability, a response linked to cheatgrass presence, are
widespread throughout the Great Basin. This pattern is
distinct from other land cover types and can be used to
identify areas where cheatgrass dominates. Cheatgrass,
native shadscale/bunch grass, and salt flats are distinct in
both time series (Fig. 6) and DNDVI values (Fig. 7).
When the spatial pattern of amplified response is
mapped (Fig. 8), the areas identified by AVHRR are
similar to those identified by Landsat. Changing the
detection confidence interval does not affect the con-
tiguous areas of amplified response seen in the eastern
half of the Landsat scene. The presence of areas
identified as cheatgrass by only one of the two sensors
is consistent with the identification rates of 64% and
71% by AVHRR and Landsat, respectively. Additionally,
areas identified in AVHRR by the DNDVI method
almost all show the targeted amplified response to
rainfall over time. The only exception being scattered
pixels in southwestern Utah. There, time series show an
amplitude increase during the late 1990s which may be
attributable to slight misregistration along steep mountain
gradients of the Fish Lake and Dixie National Forests.
Everywhere else, positively identified pixels show an
amplified response in 1993, 1995, 1998, or a combina-
tion of the three. Because cheatgrass is the primary
invasive annual into native perennial communities,
amplified response is likely caused by cheatgrass
dominance.
The inter-annual pattern caused by cheatgrass amplified
response should not be confused with land cover change.
There is some risk of this type of pattern being erroneously
labeled change, particularly because several of the cheat-
grass-dominated time series show a secondary amplified
response the year following a wet year. This is particularly
notable in 1996 and 1999, despite the fact that rainfall
during these 2 years was below average. This effect may
be a result of an increased seed bank the year following
wet years (Prince et al., 1998), or, in the case of 1999, an
exceptionally wet preceding September/October which
allowed for fall germination of cheatgrass. A time series
spanning the 1990s could be misinterpreted as land cover
change due to this secondary response (Fig. 10). However,
the pattern is entirely attributable to cheatgrass variability.
Land cover change research, particularly studies at regional
scales, should be aware that this type of rainfall-driven
pattern is a temporal signature of an unchanging land
cover type.
Based on the regional map (Fig. 8B), cheatgrass-
dominated ecosystems are widespread in northern
Nevada and western Utah. The counties most affected
by invasion include Pershing, Humboldt, and northern
Lander and Eureka in Nevada. Tooele and Box Elder
counties are most affected in Utah, and Harney County
Fig. 9. Receiver operator characteristic (ROC) curves for predictive
capacity of the Landsat and AVHRR cheatgrass maps. Landsat accurately
identifies 72% of cheatgrass cover with an error of 17%. AVHRR
accurately identifies 64% of cheatgrass cover with an error of 15%.
B.A. Bradley, J.F. Mustard / Remote Sensing of Environment 94 (2005) 204213 211
in Oregon also has a concentrated invasion. In all,
20,000 km
2
of detected amplified response is apparent in
several concentrated localities in the Great Basin. Despite
the large extent of cheatgrass dominance, the land cover
type is not apparent in single year phenology-based
classifications of US land cover, like the North American
Land Cover Characteristics Map (Fig. 1B) (Loveland et
al., 1999). Inter-annual variability in time series encom-
passing wet and dry years should be considered for more
accurate classification.
The broad spatial extent of cheatgrass invasion allows for
easy scaling between field observations, Landsat, and
AVHRR resolutions. However, the amplified response of
cheatgrass is not apparent without a time series of data
encompassing both dry and wet years. Additionally,
cheatgrass phenology is short, lasting only 12 spring
months. As a result, Landsat scenes must be chosen with
care to capture peak productivity. Analysis of a single scene
is not enough to differentiate cheatgrass from native land
cover types. While cheatgrass detection based on its early
green-up has been performed using Landsat scenes from
2001 (Peterson, 2003), identification of cheatgrass is
optimal during wet years. Time series analysis should be
considered for distinguishing between land cover types,
particularly when they have different responses to climatic
patterns on an inter-annual basis.
Variability over time does not necessarily imply active
land cover change. Cheatgrass-dominated areas show
changing productivity over time, but the land cover itself
is not changing. For example, a cheatgrass time series from
northern Nevada may at first suggest a dramatic shift in land
cover between 1994 and 1995, particularly in areas
exhibiting a second year of rainfall-related amplified
response (Fig. 10). Likewise, a comparison of NDVI based
productivity for 1998 and 2000 would indicate a marked
loss of biomass that could be erroneously attributed to land
cover change. In fact, the same land cover has been in place
at these localities for at least the past decade. Land Use/
Land Cover Change science, particularly regional and
global models, must be aware that some responses that
look like change are caused by inherent variability within
ecosystems. Strong correlation to precipitation may be of
particular concern in semi-arid ecosystems; however, differ-
ent environmental variables may affect inter-annual vegeta-
tion patterns in other ecosystems. Variability over time,
particularly when it can be correlated to climatic patterns, is
likely a response of a particular land cover type rather than
an indication of land cover change. Land cover change
scientists should use caution when interpreting signals like
the one presented here.
Acknowledgement
The work was funded by the NASA Land Use Land
Cover Change Program and the American Society for
Engineering Education. We are grateful for the insight and
expertise of Jeff Albert, Jeremy Fisher, Erica Fleishman,
Steven Hamburg, and Robin Tausch.
References
Adams, J. B., Sabol, D. E., Kapos, V., Almeida, R., Roberts, D. A., Smith,
M. O., et al. (1995). Classification of multispectral images based on
fractions of endmembersapplication to land-cover change in the
Brazilian Amazon. Remote Sensing of Environment, 52, 137154.
Asner, G. P., Townsend, A. R., & Braswell, B. H. (2000). Satellite
observation of El Nino effects on Amazon forest phenology and
productivity. Geophysical Research Letters, 27, 981984.
Dale, V. H. (1997). The relationship between land-use change and climate
change. Ecological Applications, 7, 753769.
DeFries, R., Hansen, M., & Townshend, J. (1995). Global discrimination of
land cover types from metrics derived from AVHRR pathfinder data.
Remote Sensing of Environment, 54, 209222.
Defries, R. S., & Townshend, J. R. G. (1994). NDVI-derived land-cover
classifications at a global-scale. International Journal of Remote
Sensing, 15, 35673586.
Eidenshink, J. C. (1992). The 1990 conterminous United-States Avhrr data
set. Photogrammetric Engineering and Remote Sensing, 58, 809813.
Elmore, A. J., Mustard, J. F., & Manning, S. J. (2003). Regional patterns of
plant community response to changes in water: Owens Valley,
California. Ecological Applications, 13, 443460.
Elmore, A. J., Mustard, J. F., Manning, S. J., & Lobell, D. B. (2000).
Quantifying vegetation change in semiarid environments: Precision and
accuracy of spectral mixture analysis and the Normalized Difference
Vegetation Index. Remote Sensing of Environment, 73, 87102.
Friedl, M. A., McIver, D. K., Hodges, J. C. F., Zhang, X. Y., Muchoney,
D., Strahler, A. H., et al. (2002). Global land cover mapping from
MODIS: Algorithms and early results. Remote Sensing of Environ-
ment, 83, 287302.
Grayson, D. K. (1993). The deserts past: A natural prehistory of the Great
Basin. Washington7 Smithsonian Institution Press.
Hall, F. G., Strebel, D. E., Nickeson, J. E., & Goetz, S. J. (1991).
Radiometric rectificationtoward a common radiometric response
among multidate, multisensor images. Remote Sensing of Environment,
35, 1127.
Fig. 10. Mean cheatgrass time series of 6 km
2
in Buena Vista Valley
(4085VN, 1188W) for 19902001. While suggestive of land cover change
between 19941995, this pattern is entirely attributable to cheatgrass
variability.
B.A. Bradley, J.F. Mustard / Remote Sensing of Environment 94 (2005) 204213 212
Homer, C. G., Ramsey, R. D., Edwards, T. C., & Falconer, A. (1997).
Landscape cover-type modeling using a multi-scene thematic mapper
mosaic. Photogrammetric Engineering and Remote Sensing, 63, 5967.
Hull, A. C. J., & Pechanec, J. F. (1947). Cheatgrassa challenge to range
research. Journal of Forestry, 45, 555564.
Jakubauskas, M. E., Peterson, D. L., Kastens, J. H., & Legates, D. R.
(2002). Time series remote sensing of landscapevegetation interactions
in the southern Great Plains. Photogrammetric Engineering and Remote
Sensing, 68, 10211030.
Kastens, J. H., Jakubauskas, M. E., & Lerner, D. E. (2003). Using temporal
averaging to decouple annual and nonannual information in AVHRR
NDVI time series. IEEE Transactions on Geoscience and Remote
Sensing, 41, 25902594.
Knapp, P. A. (1996). Cheatgrass (Bromus tectorum L) dominance in the
Great Basin Desert-History, persistence, and influences to human
activities. Global Environmental Change-Human and Policy Dimen-
sions, 6, 3752.
Lambin, E. F. (1996). Change detection at multiple temporal scales:
seasonal and annual variations in landscape variables. Photogrammetric
Engineering and Remote Sensing, 62, 931938.
Lambin, E. F., Turner, B. L., Geist, H. J., Agbola, S. B., Angelsen, A.,
Bruce, J. W., et al. (2001). The causes of land-use and land-cover
change: Moving beyond the myths. Global Environmental Change-
Human and Policy Dimensions, 11, 261269.
Loveland, T. R., Reed, B. C., Brown, J. F., Ohlen, D. O., Zhu, Z., Yang, L.,
et al. (2000). Development of a global land cover characteristics
database and IGBP DISCover from 1 km AVHRR data. International
Journal of Remote Sensing, 21, 13031330.
Loveland, T. R., Zhu, Z. L., Ohlen, D. O., Brown, J. F., Reed, B. C., &
Yang, L. M. (1999). An analysis of the IGBP global land-cover
characterization process. Photogrammetric Engineering and Remote
Sensing, 65, 10211032.
Mack, R. N. (1981). Invasions of Bromus tectorum L. into Western North
America: An ecological chronicle. Agro-Ecosystems, 7, 145165.
Masek, J. G., Lindsay, F. E., & Goward, S. N. (2000). Dynamics of urban
growth in the Washington DC metropolitan area, 19731996, from
Landsat observations. International Journal of Remote Sensing, 21,
34733486.
Meyer, S. E., & Monsen, S. B. (1992). Big sagebrush germination
patternssubspecies and population differences. Journal of Range
Management, 45, 8793.
Myneni, R. B., Keeling, C. D., Tucker, C. J., Asrar, G., & Nemani, R. R.
(1997). Increased plant growth in the northern high latitudes from 1981
to 1991. Nature, 386, 698702.
Myneni, R. B., Tucker, C. J., Asrar, G., & Keeling, C. D. (1998). Interannual
variations in satellite-sensed vegetation index data from 1981 to 1991.
Journal of Geophysical Research-Atmospheres, 103, 61456160.
Peterson, E. B. (2003). Mapping percent-cover of the invasive species
Bromus tectorum (Cheatgrass) over a large portion of Nevada from
satellite imagery. Reno7 Nevada Natural Heritage Program.
Potter, C., Tan, P. N., Steinbach, M., Klooster, S., Kumar, V., Myneni, R., et
al. (2003). Major disturbance events in terrestrial ecosystems detected
using global satellite data sets. Global Change Biology, 9, 10051021.
Prince, S. D., De Colstoun, E. B., & Kravitz, L. L. (1998). Evidence from
rain-use efficiencies does not indicate extensive Sahelian desertifica-
tion. Global Change Biology, 4, 359374.
Privette, J. L., Fowler, C., Wick, G. A., Baldwin, D., & Emery, W. J.
(1995). Effects of orbital drift on advanced very high-resolution
radiometer productsnormalized difference vegetation index and sea-
surface temperature. Remote Sensing of Environment, 53, 164171.
Rice, K. J., Black, R. A., Radamaker, G., & Evans, R. D. (1992).
Photosynthesis, growth, and biomass allocation in habitat ecotypes of
cheatgrass (Bromus tectorum). Functional Ecology, 6, 3240.
Schott, J. R., Salvaggio, C., & Volchok, W. J. (1988). Radiometric scene
normalization using pseudoinvariant features. Remote Sensing of
Environment, 26, 1.
Shabanov, N. V., Zhou, L. M., Knyazikhin, Y., Myneni, R. B., & Tucker,
C. J. (2002). Analysis of interannual changes in northern vegetation
activity observed in AVHRR data from 1981 to 1994. IEEE Trans-
actions on Geoscience and Remote Sensing, 40, 115130.
Skole, D., & Tucker, C. (1993). Tropical deforestation and habitat
fragmentation in the Amazon-satellite data from 1978 to 1988. Science,
260, 19051910.
Stewart, G., & Hull, A. C. J. (1949). Cheatgrass (Bromus tectorum L.)an
ecological intruder in Southern Idaho. Ecology, 30, 5874.
Stewart, G., & Young, A. E. (1939). The hazard of basing permanent
grazing capacity on Bromus tectorum. Journal of the American Society
of Agronomy, 31, 10021015.
Teillet, P. M., & Holben, B. N. (1994). Towards operational radiometric
calibration of NOAA AVHRR imagery in the visible and near-infrared
channels. Canadian Journal of Remote Sensing, 20, 110.
Tucker, C. J., Dregne, H. E., & Newcomb, W. W. (1991). Expansion and
contraction of the Sahara desert from 1980 to 1990. Science, 253,
299301.
Turner, B. L. (1972). Toward integrated land-change science: Advances in
1.5 decades of sustained international research on land-use and land-
cover change. In W. Steffen, J. Jager, D. J. Carson, & C. Bradshaw
(Eds.), Challenges of a changing earth: Global Change Open Science
Conference, Amsterdam.
Vitousek, P. M., Mooney, H. A., Lubchenco, J., & Melillo, J. M. (1997).
Human domination of Earths ecosystems. Science, 277, 494499.
Young, J. A., & Allen, F. L. (1997). Cheatgrass and range science:
19301950. Journal of Range Management, 50, 530535.
Young, J. A., Evans, R. A., & Major, J. (1972). Alien plants in the Great
Basin. Journal of Range Management, 25, 194201.
Young, S. S., & Wang, C. Y. (2001). Land-cover change analysis of China
using global-scale Pathfinder AVHRR Landcover (PAL) data, 198292.
International Journal of Remote Sensing, 22, 14571477.
Zhou, L., Kaufmann, R. K., Tian, Y., Myneni, R. B., & Tucker, C. J. (2003).
Relation between interannual variations in satellite measures of northern
forest greenness and climate between 1982 and 1999. Journal of
Geophysical, 108.
Zhou, L. M., Tucker, C. J., Kaufmann, R. K., Slayback, D., Shabanov, N.
V., & Myneni, R. B. (2001). Variations in northern vegetation activity
inferred from satellite data of vegetation index during 1981 to 1999.
Journal of Geophysical Research-Atmospheres, 106, 2006920083.
B.A. Bradley, J.F. Mustard / Remote Sensing of Environment 94 (2005) 204213 213