< Back  ||  Home > Reports > Range Transect Data

Correlation of Range Transect Data with Satellite Remote Sensing Vegetation Indices

Goals and Objectives of Range Monitoring

Range monitoring addresses questions related to range condition and change through time that reflect both utility to livestock and broader environmental quality. Many techniques have been developed to describe and quantify range attributes such as vegetation and soil condition, land cover, and forage quality. Monitoring data may suggest to a range resource manager where and for how long to allow livestock to graze, as well as to inform management decisions such as stocking rates and rest-rotation schedules.

Often, range resource measurements and observations are made at a limited set of locations along line transects or small plots, typically at annual or longer intervals. These data become the bases for estimates of condition at larger spatial scales (e.g., pastures or other management units) and for determinations of trends through time.

Remote sensing data have the potential to extend knowledge of vegetation condition to much larger areas than could be practicably monitored from the ground, to provide this information on time scales relevant to management decision making, and to permit trends through time to be analyzed with continuous time series. Including remote sensing data in a monitoring strategy may allow "normal" patterns of spatial and temporal variability in the resource to be more objectively distinguished from changes related to livestock management, and may even permit prediction of forage availability influenced by certain climatic patterns.

Environmental and analytical context for range monitoring and remote sensing data on the V-bar-V Ranch

As part of RangeView, we have been investigating relationships between point-intercept land cover data from permanent transects on the V-bar-V Ranch and vegetation indices from coincident AVHRR and Landsat imagery. Results have been mixed. Interpretation is complicated both by the very limited ground data and by differences in the timing of monitoring among the three years.

The outcome of this activity presents an opportunity to examine what we know about range condition and trend and to critically evaluate the adequacy of current monitoring approaches. Perhaps the most important product will be a revision to monitoring methodology that will produce data of utility both to managers on the ground and to remote sensing-based approaches to resource assessment.

Historic range monitoring data are analyzed in conjunction with vegetation indices derived for NOAA-AVHRR and Landsat Thematic Mapper imagery for the V-V Ranch. This analysis is part of an ongoing effort to develop a means of estimating or predicting vegetation (i.e. forage) condition from satellite imagery that can be applied in range management decision-making. A measure is being sought that produces significant correlation between these two sources of data that is consistent across space and through time. The measure should be sensitive to interannual variability in land cover, as well.

The three monitoring seasons in 1985, 1992, and 1999, followed winters that were neutral, negative (El Niño), and positive (La Niña), respectively, with respect to the ENSO climatic cycle (Figure 1). Although the winter of 1984-1985 was neutral, winter precipitation was comparable to that of 1991-1992, a wetter-than-normal winter. The winter of 1998-1999 was much drier than normal. Although all three years experienced substantial monsoonal precipitation, the pattern of rainfall was different for the summer months immediately preceding data collection (Figure 2): 1992 experienced a wet mid-summer but a dry early- and late-summer whereas, in 1985 and 1999, the early- and late-summer were wet and the mid-summer was dry.

Figure 1. Average June-November Southern Oscillation Index (SOI) plotted for the ensuing December-March,
with transect data years indicated (symbols: white=neutral ENSO year, red=La Niña, blue=El Niño). Data:
Western Regional Climate Center, Classification of El Niño and La Niña Winters, 1933-1999 based on average
June-November SOI, <http://www.wrcc.dri.edu/enso/ensodef.htm>.

Figure 2. Total monthly precipitation at four weather stations surrounding the V-V Ranch for the year preceding
each transect reading. Brackets with arrows indicate the time frame during which transects were read each year.
Data: NOAA-NCDC Climate Data Online <http://cdo.ncdc.noaa.gov/plclimprod/plsql/poemain.poe>.

Vegetation differences between the two ENSO extremes, in particular, are of interest to range and wildlife management because of both their potential consequences for forage availability and the ability to predict the strength of ENSO cycles months in advance of their manifestations. The behavior of cover type classes and NDVI in each of these two years may provide some indication of the degree to which forage is influenced by interannual variability in precipitation patterns at different locations on the ranch.

Data and Image Processing

Land cover

Point-intercept land cover data were recorded by U.S. Forest Service personnel for 7 sites (all summer range) in September-November 1985, 17 sites (8 summer range, 9 winter range) in September-October 1992, and 22 sites (9 summer range, 13 winter range) in August-September 1999. Coordinates for the start and end of each 100-ft transect in each cluster, and for the monument marking each cluster, were recorded with a Garmin 12XL hand-held GPS receiver in May 2001.

Cover types in the transect data set include plants (green vegetation), rock, litter (non-green vegetation), and soil. The derived category, non-vegetation, is the sum of rock and soil. The vegetation:non-vegetation ratio is computed as plants divided by non-vegetation. Percent cover measures for individual transects were used with Landsat TM data. Both the sum and the average of the transects in each cluster were used with AVHRR data as aggregate measures for the transect cluster.

Satellite remote sensing data and vegetation indices

Figure 3 illustrates the temporal relationships of the transect reading dates and the satellite imagery used in this analysis. In each case, the biweekly composite AVHRR-NDVI scene encompassing a transect reading date was used as the source of NDVI values for the transect(s) read on that date. Cloud-free Landsat images were selected to correspond as nearly as possible to the transect reading dates, and a single image was used for all of the transects for each monitoring season. Landsat images were first orthorectified using the geometric correction module in ERDAS Imagine (RMSE <10 m for all images) and then radiometrically and atmospherically corrected using the dark object subtraction method defined in the COST model (Chavez, 1996).

Figure 3. Temporal relationship of transect data and satellite remote sensing data sets used in comparisons between
land cover data and vegetation indices for each of three monitoring seasons (no AVHRR data used in 1985).  A
single Landsat scene was used in comparisons with all of the transects in each year; the AVHRR biweekly composite
image corresponding to the reading date for each transect was used.

The Normalized Difference Vegetation Index (NDVI) was computed using the reflectance output for each of the three Landsat images within the Spectral Enhancement module in ERDAS Imagine. The original NDVI values (ranging from -1 to +1) were used in this analysis for the TM and ETM scenes. AVHRR biweekly composite NDVI values used in this analysis were computed and rescaled to 8-bit format (0 to 200) by the USGS (Eidenshink, 1992). The Tasseled Cap transformation for Landsat TM bands (Crist, 1985) was applied to the TM reflectance output using the Tasseled Cap model for Landsat 5 in ERDAS Imagine and a modified version of this, with newly derived coefficients (C.Huang, USGS/EDC, by email), for the 1999 Landsat 7 ETM scene. The original (unrescaled) output from the Tasseled Cap for each scene was used in this analysis.

Image sampling

A database of vegetation index values was created using ArcInfo to sample grids of the corresponding images with point coverages of transect locations. In the Landsat imagery, the 30-m^2 pixel underlying the half-way point of each transect was used to represent that transect. In the AVHRR imagery, the center of the 1-km^2 pixel in which the transect cluster is located was used to represent the entire transect cluster in the image.

Analysis and Discussion

Spectral correlation

Linear regression was used to relate each cover type to the vegetation indices to evaluate which measure or combination of ground-based measures might best relate to satellite data. Simple linear regression was performed using the Data Analysis extension of Microsoft Excel (v.5.0). Correlations were considered to be significant when R^2 exceeded .40 at p=.05. Results are depicted graphically in Figure 4.

Figure 4. Correlation (R^2; p=.05) of transect-derived land cover classes and derived measures with vegetation
indices (Tasseled Cap brightness, greenness, wetness; TM and AVHRR NDVI) for each of the three monitoring
years. The number of individual transects read and transect clusters visited in each year is indicated. The
AVHRR dataset used in this analysis does not include 1985. The R^2=.40 level of significance adopted in this
anaylsis is marked with a dashed line.

There were few instances of significant correlation between Landsat TM vegetation indices and land cover data. No significant correlations obtained for data from 1985. Percent soil cover was significantly correlated with TM brightness in 1992 and 1999, and with TM greenness and NDVI in 1999; the non-vegetation category (rock + soil) was also significantly correlated with TM NDVI in 1999. It is encouraging that the highest correlation of all (R^2=.67) obtained for TM brightness, which is designed to emphasize the contribution of soil to total reflectance, with percent soil cover. Simultaneous moderate correlation between TM greenness and percent soil and no correlation between TM greenness and plant cover is a more difficult pattern to interpret; autocorrelation between percent plant cover and percent soil cover (R^2=.17) in 1999 does not appear to be a contributing factor.

AVHRR NDVI was significantly correlated with both the sum and the average of percent soil cover in both 1992 (R^2=.43 and .42) and 1999 (R^2=.618 and .481). Moderate correlations also obtained with the sum of non-vegetation counts in both years and with the average of non-vegetation counts in 1992, with average litter cover in 1992, and with the ratio of summed plants to summed non-vegetation in 1999. Values for TM NDVI and AVHRR NDVI were relatively highly correlated with each other in both 1992 (R^2=.69, N=17) and 1999 (R^2=.65, N=22) for pixels corresponding to transect locations. This relationship is reflected most consistently by percent soil cover.

Spatial Variability

In that transect locations are arrayed along an elevation gradient that spans the length of the V-V ranch, elevation has been used as a proxy for spatial distribution in this analysis (Figure 5). Bivariate correlations between each land cover class and elevation were sought to evaluate the spatial pattern of percent cover for each of the transect years. Results are depicted graphically in Figure 6. Significant correlations with elevation obtained for both the sum and average of percent soil cover for both 1992 (R^2=.67 and .72) and 1999 (R^2=.75 and .62); significant correlations with the non-vegetation category follow from these. The reversal in the relative strength of these correlations between years gives some indication that the way in which data are aggregated (i.e. sum v. average) to represent a site is important to the result and, thus, to comparisons between sites and dates.  The coincidence between significant correlations between NDVI and percent soil cover, as discussed above, and between elevation and percent soil cover is conditioned by a strong positive correlation between NDVI and elevation in both sets of imagery, both at the transect locations and across the study area. This relationship is discussed further, below.

Figure 5. Transect cluster locations with elevation. Elevations are derived from a 30-m USGS DEM for the
center of the 1-km AVHRR pixel in which each transect cluster occurs.

Figure 6. Correlation (R^2; p=.05) of transect-derived land cover classes with elevation. For each transect cluster,
both the sum and the average of the individual transects are used in the regressions. The number of transect
clusters visited in each year is indicated. The R^2=.40 level of significance adopted in this anaylsis is marked with
a dashed line.

The sharp elevation gradient across the V-bar-V Ranch is consistently reflected in the NDVI values in both Landsat TM and AVHRR imagery (Figure 7). Correlations between elevation and NDVI range from R^2=.71 to R^2=.86 for both sensors and across years (an outlying point was excluded in the 1985 series in obtaining these values). One-way ANOVA indicate that the distribution of NDVI values between years is significantly different for the three Landsat TM scenes (F=41.6, d.f.2,63, p<.001) but is statistically identical between years for the AVHRR scenes (F=1.00, d.f.1,44, p=.32). This difference is probably due to differences in the timing of the TM scenes (refer to Figure 3), reflecting phenological change, rather than differences in vegetation conditions between years. The AVHRR scenes used in this comparison are from the same biweekly compositing period where the two monitoring seasons corresponded between years (the period ending about September 23), and thus are reflecting vegetation condition at the same stage of the phenological cycle. Temporal variability within and between years, and the potential consequences for interpretation of comparing non-coincident imagery, are described in the following section.

Figure 7. Relationship between NDVI and elevation by transect location and monitoring date for Landsat TM and AVHRR
imagery. Values for all transect locations were used for each year whether or not the transect was read. An outlying point
(W-15) in the TM 1985 series is labeled. The AVHRR dataset used in this analysis does not include 1985. Correlations
(R^2; p=.05) are shown for the set of transects in each monitoring year. NDVI values used for Landsat TM are the average
of values for the 30-m pixels corresponding to individual transects in the transect cluster. NDVI values for AVHRR are
those of the 1-km pixel in which the transect cluster is located. AVHRR scenes from the same date each year (biweek
ending about September 23) are used in the comparison. The scales on the graphs are comparable horizontally (i.e.,
TM NDVI = .50 is equivalent to rescaled AVHRR NDVI = 150).

Temporal variability

Range monitoring data on the V-bar-V Ranch have historically been gathered in late summer to late fall (refer to Figure 3). Figure 8 illustrates the temporal dynamics of the NDVI at transect locations for this time frame. Across the ranch, long-term (11-year) averages for these sites reveal a pattern of gradual greening following the summer monsoon season, peaking in early October, and declining thereafter. This pattern of change may relate to various phenological stimuli, including changes in photoperiod and, importantly for herbaceous plants and deciduous-leaved perennials, the timing of the first killing frost of the season which, in northern Arizona, typically occurs around the end of September (Barbour et al., 1999; dashed white line in Figure 8).

Figure 8. AVHRR-NDVI time series for the late summer-early fall historic monitoring time frame. Long-term
averages for each transect location are shown (lines color-coded by major TES vegetation unit: pink=desertscrub,
yellow=pinyon-juniper, green=pine-oak). Transects are ordered with elevation in the legend. The average
for the set of transects read in each year under corresponding biweekly NDVI composite scenes is indicated
with a colored symbol above the composite end date. The long-term average for all transect locations is also
shown (black line). Dashed white line indicates the approximate average date of the first killing frost in
northern Arizona (see text).

The timing of transect monitoring in 1992 and 1999 differs slightly; the former occurred from mid-September to mid-October, while the latter began in late August and continued through late September. The corresponding AVHRR average NDVI values (symbols in Figure 8) differ in the direction and magnitude in which they vary around the long-term mean for the cluster locations (black line) in this time frame. Figure 9 illustrates the familiar trend in the NDVI increasing with elevation on the V-bar-V Ranch for monitoring dates in each of these two years. The change in NDVI in each pixel across the six-week monitoring season differs markedly between years. Of importance to the dual range monitoring goals of assessing current condition and analyzing changes through time are the differences across space in the particular time (and the corresponding AVHRR composite) at which each transect was read. The combination of the temporal variability in NDVI and the spatial contiguity of transects (or other factors influencing the order in which transect clusters were or are visited) could introduce bias to comparisons both across space and through time of NDVI data for range monitoring sites.

Figure 9. AVHRR NDVI values for the three biweekly composite periods under which transects were read
each year (1992, left; 1999, right). The scene under which each particular transect cluster was visited is
indicated by a colored symbol.

These same data are plotted in Figure 10, along with the complete series of values for the one biweekly compositing period (the biweek ending about September 23) in which the 1992 and 1999 monitoring seasons overlapped - ideally, the one that should be compared between the two years for this data set. The NDVI values corresponding to the time frame in which each transect was actually read are indicated by symbols, and the long-term average for the overlapping date is also shown. The potential sources of error in interpreting NDVI values from different weeks between years are illustrated graphically in Figure 11.

Figure 10. AVHRR NDVI values corresponding to transect reading dates (symbols), and values for the same biweekly
composite data in each of the two monitoring years (colored lines). The long-term (11-year) average and one standard
deviation for this same composite date at the transect locations is also shown (black line).  Symbol shapes correspond
to the legends in Figure 9, and indicate the date of the particular AVHRR scene under which the transect was read
which supplied the values used here.

Correspondence between NDVI values for reading dates and the same-date scenes is best at lower elevations, where there is also the least interannual variability in the NDVI.  The values at transect cluster W-15, however, in 1999 are illustrative of error introduced by transect reading dates differing between years.  At this transect, the differences between the the two years and between 1999 and the long-term average are not significant, however the differences between the value at the time of the transect reading and both the 1992 and the long-term average at that location are significantly different (see Figure 11).

At mid-elevations, the NDVI values for the reading dates at transects W-3, W-2, S-8, W-1, and S-9 in 1992 are lower than those of the overlap date, although in 1999 these same transects were read during the overlap period.  Qualitatively, the pattern of difference between the two years is the same for these transects whether the NDVI values for the reading dates or for the same-date scene are compared.  The magnitude of the difference, however, becomes significant when the reading date NDVI values are used where the difference is within a single standard deviation of the long-term average for the values from the same-date image (refer to Figure 11b).

The converse is the case for transects at higher elevations (transects including and to the right of S-7 in the Figures).  There, the transects were read during the overlap period in 1992, but were read under another scene in 1999.  The NDVI values for the reading date scene in 1999 are much higher than, and in several cases significantly different from, those of the overlap date (Figure 11c).  Again, the direction of difference is the same for each of the two sets of values, but the size and significance of the differences appear much greater than it would if the values from the same-date scene were used.

Figure 11. Differences from the long-term average (LTA) for September 23 for each AVHRR pixel corresponding to
transect locations on the V-bar-V Ranch. Differences are in the NDVI values for the September 23 ("same date")
scene in each of two years and for the scenes corresponding to transect reading dates:  (a) compares the two
monitoring years, (b) compares values for 1992, (c) compares values for 1999.  The LTA and LTSD for each pixel
is based on an 11-year data set (1989-1999).

In each of the three years included in this analysis, transect monitoring has occurred over a span of several weeks, complicating interpretation of differences across space (e.g., between pastures or vegetation types) within the same monitoring season. Additionally, these monitoring seasons only partly overlap with one another, hindering comparison of land cover - and corresponding vegetation index - values between years. Differences in the timing of each monitoring season relative to the cessation of the monsoon and the dieback of herbaceous growth have particular implications for discerning, unambiguously, what "change" is due to phenology versus to management practices or to climatic events. Three seasons of range monitoring data are probably insufficient to adequately characterize the temporal variability possible in range conditions due solely to "background" (i.e. not related to livestock) factors, particularly where interannual variability in timing and amount of precipitation around the V-bar-V Ranch approaches 30% (see Figure 2; coefficients of variation computed for four stations using monthly data from 1994-1999) and interannual variability in the NDVI may be between 5% and 10% at different times of the year.

Summary and Conclusions

Linear regression was used to evaluate relationships between transect-derived land cover classes and satellite-derived NDVI values. Significant correlations with vegetation indices occur consistently only for percent soil cover. There may be some importance to this outcome, particularly as it is evident in the indices produced by both Landsat ETM and AVHRR imagery. Although it is not possible at this time to clearly interpret this result, soil cover should be recognized as a potentially useful monitoring attribute relatable to remote sensing imagery. Based on the analyses described here, for this study area, elevation emerges as the strongest and most consistent predictor of greenness both at the transect locations and across the entire ranch.

Differences exist in the values of and strength of correlations between vegetation indices and transect data across space and through time. Determining the source of these differences is fundamental to applying this information to range management decision-making. In relating satellite and ground data, achieving a close temporal match is important to interpreting imagery in terms of ecologically-based management objectives (e.g., range condition or forage availability). In seeking information about trends or changes through time, it is likewise essential to be comparing data that were acquired at nearly the same point in the phenological cycle each year. At the same time, it is important to analyze these sources of data in the context of ecological variability due to larger-scale processes (e.g., the ENSO climatic cycle). The data used in this analysis do not, in all cases, satisfy these criteria.

Two lessons may be drawn from the methodoligical complications encountered in relating field-based land cover measures to satellite-derived vegetation indices in this analysis. First, there is potential for remote sensing data to inform range monitoring strategies. In looking at the differences between NDVI values for transect reading dates compared both to the same date from each of two years and to the long-term average for each corresponding pixel, it is possible to evaluate where, in what direction, and of what magnitude error may occur in comparing ground-based land cover data between years.  Remote sensing imagery might also help inform siting of monitoring locations and, with long-term knowledge of vegetation dynamics, could indicate what times of year monitoring could produce ecologically informative data.  Second, even in the absence of a ground-based measure of range condition strongly correlated with image-based vegetation indices, satellite data still are able to provide reliably comparable information about relative vegetation condition both simultaneously across space and between points in time. This capability attenuates the problems associated with reading transects over the course of several weeks during a monitoring season which can result in readings being made at different times of the growing season in different years for the same transect.  Utilizing time series also allows measurements from a single date or season to be related to a longer-term pattern of variability, where monitoring data collected at irregular and/or long intervals cannot provide this context on their own.

There are other sources of error in interpretation, either or both of satellite-derived indices or of transect data, that have not been considered here, including spatially and seasonally variable effects of clouds and snow on surface reflectance, and the grazing-rotation cycle relative to phenology and both range and satellite data acquisition. In order to develop a means for quantitatively - beyond merely relatively - estimating forage condition using satellite imagery, ground-based data collection must be underatken more systematically, with greater spatial and temporal representation, to adequately characterize inter- and intra-annual variability in vegetation as well as to discover a measure that is strongly and consistently related to image-based vegetation values.  Both ground-based and image-based assessments of range condition and trend could benefit from establishing control (i.e. livestock exclosure) sites. Each of these components, incrementally, is an important consideration in developing a comprehensive model of rangeland vegetation dynamics that can be informed by, and in turn enhance the interpretability of, satellite remote sensing imagery.  

References Cited

Barbour, M. G., J. H. Burk, W. D. Pitts, F. S. Gilliam, & M. W. Schwartz (1999) Terrestrial Plant Ecology. Benjamin-Cummings, Menlo Park.

Chavez, P.S., jr. (1996) Image-based atmospheric corrections - Revisited and improved. Photogrammetric Engineering and Remote Sensing 62 (9): 1025-1036.

Crist, E. P. (1985) A TM Tasseled Cap Equivalent Transformation for Reflectance Factor Data. Remote Sensing of Environment 17: 301-306.

Eidenshink, J. C. (1992) The 1990 Conterminous U. S. AVHRR Data Set. Photogrammetric Engineering and Remote Sensing 58 (6): 809-813.