Peter Leth Geo 133 4/27/00
Remote sensing of vegetation abundance in arid valleys of Inyo County, California: What was the effect of the 1987-1992 drought?
Remotely sensed data allow us to study large regional and temporal variations in surface features. Here we track natural (non-anthropogenic) changes in xeric vegetation abundance in arid valleys east of the Sierra Nevada Mountains of California, from 1985 to 1998, with a series of Landsat Thematic Mapper (TM) and Advanced Very High Resolution Radiometer (AVHRR) images. The percentage of live vegetation cover (%LC) for each Landsat TM pixel was calculated by the technique of Spectral Mixture Analysis (SMA). For AVHRR, vegetation strength is characterized by the Normalized Difference Vegetation Index (NDVI). We find that the large Owens Valley, immediately adjacent to the Sierras, consistently has about 5% more %LC than smaller valleys to the east. Additionally, a sustained drought in precipitation, which occurred from 1987-1992, has a strong signature in Owens Valley %LC (i.e. loss of vegetation), but is remarkably absent in the eastern valleys. AVHRR data indicate that these eastern valleys have a short, early growing season compared to the Owens Valley. These findings lead us to conclude that climatic conditions are harsher in the valleys east of the White-Inyo range than in the Owens Valley.
Large-scale, regional measurements of vegetation abundance with time are important for measuring the effect of climatic, biologic and anthropogenic forces on the natural ecosystem. Specific questions that may arise from such studies include, for example: Did a recent drought kill off more vegetation than a normal dry season? How quickly does vegetation recover from such a drought? Did a strong wet season enhance vegetation growth? Does the encroachment of exotic plant species adversely affect the vegetation abundance of native species? How do groundwater pumping, stream diversion and urbanization affect all of this?
For this project we examine vegetation abundance on the floors of several valleys in Inyo County, California (see Figure 1). This area lies in the eastern rain shadow of the Sierra Nevada Mountains, and consequently is very dry. The Owens Valley runs north-south along the eastern edge of the range, and has an average annual rainfall of 14cm. Further east, Death Valley runs NW-SE, and has an average annual rainfall of just 5cm (California Department of Water Resources (CDWR), 2000). In between lie the smaller Eureka and Saline Valleys, where no rainfall data is available. Although the floor of the Owens Valley has seen extensive human impact, including urbanization, agricultural development, groundwater pumping and stream diversion, there are large bajadas and alluvial fans on the sides of the valley that are relatively untouched. Here, as in the drier valleys to the east, vegetation is primarily xeric. It is not largely affected by changes in the groundwater table, as is the case on the valley floor (Elmore et al., in press).
From 1987-1992 a large drought hit much of the western US. In the Owens Valley, average annual rainfall dropped to about 10cm at this time (see Figure 2; Stations BIS and CTT are in the north and south of the valley, respectively, as marked on Figure 1; CDWR, 2000). The primary objective of this study is to address the effect that this drought had on vegetation abundance in these valleys. Specifically, is there a drought signature in %LC? Is this signature prevalent in areas east of the Owens Valley that see very little rainfall to begin with? What can we say about the climatic forcing of vegetation communities based on these results?
Answers to these questions are well suited to methods involving remotely sensed data. In-situ measurements of %LC are cumbersome at best, and limited in spatial coverage, although they can yield highly accurate results (Elmore et al., in press). Given the size of the region in question, large-scale, easily acquired, yet still accurate data are needed.
The present study is made possible by satellite acquisition of electromagnetic radiation that is reflected and emitted by the surface of the Earth. The intensity of this reflected radiation received at the satellite depends on a number of factors, most notably the physical structure and temperature of the surficial material, its distance from the satellite, and wavelength. A perfect emitter ("Blackbody") has a very predictable energy radiance curve with wavelength given temperature. This relationship is summarized by
known as the Stefan-Boltzmann Law, where Fb is the radiant flux (radiant energy integrated over wavelength) of the Blackbody, s is the Stefan-Boltzmann constant and Tk is the kinetic temperature of the blackbody. Since a Blackbody is an idealized conception, real radiant bodies will be governed by
where Fr is the radiant flux of the real body, and e = Fr/ Fb is known as the emissivity. e is wavelength and material dependent, and must be determined in the lab before heading to the field. For typical geologic materials, e ranges from 0.81 (granite) to 0.99 (water) (Sabins, 1997). It is Fr, converted to DN at the desired wavelength, that is used in the analysis techniques described below.
Previous work on quantifying %LC in Inyo County has been mainly limited to the Owens Valley. Studies by Ustin et al. (1986a) in this area addressed various remote sensing techniques for the analysis of semi-arid vegetation health. Like NDVI (see below), some of these utilize the difference in vegetation reflectance between TM band 3 (visible) and band 4 (near IR). For example, the Perpendicular Vegetation Index (PVI) measures the departure of pixel reflectances from a band 3/band 4 linear soil baseline. This difference is presumed to be due to vegetation. One problem found with PVI is that it can have large errors at low leaf area indices (<1), precisely those found in arid environments. At the same time, Ustin et al. (1986b) also addressed multispectral mixing models (e.g. Adams and Adams, 1984). Their endmember analysis for the Owens Valley was later adapted by Elmore et al. (in press) and is used in this work (see below). They found %LC on the bajadas west of Independence (against the Sierras) to be greater than 25%, but much less on bajadas against the White/Inyo ranges, probably because they are drier.
Due to the success of SMA in quantifying %LC in the Owens Valley, little work has been pursued with NDVI. Nonetheless, since AVHRR NDVI calculations have been shown to accurately capture regional variations in vegetation health (e.g. Kalluri and Borak, 2000), we will look at this data to see if we can draw some comparisons with SMA results.
Data & Methods
Landsat TM data has been acquired continuously since 1982 in seven spectral bands: 3 visible, 1 near-IR, 2 mid-IR and 1 thermal IR. Data is acquired passively, with a spatial resolution of 30m (120m for thermal IR). At mid-latitudes such as Inyo County, daytime images are acquired on southbound orbits at 10:30am local time (Sabins, 1997). This time is chosen as the optimum balance between morning shadow loss and afternoon cloud gathering. The images used in this study were selected from relatively cloud-free scenes acquired during late summer or early fall, from 1984 to 1998 (15 images). Most were acquired in September, which ensures that spring ephemeral species are avoided and perennial species have completed their seasonal growth, but not yet senesced (Elmore et al., in press; Sorensen et al., 1991). 1984, 1986, 1989 and 1993 were acquired in late August, while 1985 was acquired in early October, after the first frost. This is important to note since frost kills some vegetation and consequently reduces %LC.
To quantitatively analyze our scene, we need a method whereby the reflectance data acquired by Landsat TM can be accurately reduced to give a precise measure of %LC. This method is the Spectral Mixture Analysis (SMA), whereby the spectral properties of a pixel are modeled as a linear combination of endmember spectra weighted by the percent ground coverage of each endmember. That is, for each pixel
where DNb is the pixel intensity in band b, Fi is the fractional abundance of endmember i (i.e. all Fi sum to 1), DNi,b is the pixel intensity of endmember i in band b, and Eb is the error for band b (Elmore et al., in press). Determination of endmembers is done in a process called classification. This approach exploits the fact that different land covers exhibit relatively uniform spectral properties that differ from other land covers (Mustard, 2000). The obvious limitation here is that endmembers cannot overlap in odd ways (non-linearity), but nonetheless this assumption has been shown to accurately map %LC (Adams, et al., 1993; Smith, et al, 1990a,b). This error is minimized in arid environments, where vegetation is generally a small fraction of the ground space (i.e. does not overlap with other endmembers). In these cases %LC has been shown to be accurate to roughly ± 4% (Elmore et al., in press).
For our dataset, vegetation, organic-rich soil, non-organic-rich soil and shade were chosen as the endmembers within the Owens Valley, following Ustin et al. (1986b) and Elmore et al. (in press). %LC is simply a measure of how much the vegetation endmember dominates the pixel in question. Since SMA work done by researchers at Brown has focused on the Owens Valley, their endmembers need to be extrapolated to the entire scene. This introduces a source of error to our data. It is likely that more robust endmembers do exist, since the Owens Valley occupies just a fraction of the entire scene. Consequently, there are areas which measure less than 0 %LC. All TM SMA data preparation for 1984-1998 was carried out by Bill Fripp, and is available at /research2/owens/owens_veg_reg.pix. 1984 was thrown out because it has large, inexplicable splotches of high vegetation that cover much of the scene.
AVHRR was launched as a NOAA weather satellite in 1978 and acquires data in five spectrral bands: 1 visible, 1 near-IR, and 3 thermal IR. Data is acquired passively, with a spatial resolution of about 1 km. There is a 12 hour repeat time, one in the morning and one in the evening (Sabins, 1997). We have 280 images, roughly 23 per year (15 days apart) from 1989 to 1999.
Because of the low pixel resolution and the variety of terrain covered in each scene (the entire southwestern US), AVHRR data is not well suited to SMA (it would be highly nonlinear). Instead, we calculate NDVI, given by
where NIR is the measured reflectance of the near-IR band and VIS is the measured reflectance of the visible band. NDVI takes advantage of the jump in vegetation reflectance at the VIS/NIR boundary, the famed "red edge." AVHRR is well suited to this index because it has both of these bands.
NDVI is inherently a measure of vegetation health (i.e. dead vegetation has a weaker red edge than fresh, live vegetation). While it is expected to be somewhat correlated to %LC, by no means should it be a simple relationship. Other factors, notably soils, are present that will contribute to the spectrum. A pure soil spectrum, for example, would have a small positive NDVI. Studies have also shown that NDVI saturates at high vegetation abundance, making it difficult to distinguish between significant differences in high %LC.
Despite the drawbacks of NDVI, AVHRR could provide insight into short-term variability that is lost in the TM data. The NDVI scene for the southwestern US is readily available on the department sever at /research3/fripp/swus/sw_us.pix.
All work for this project was done on PCI ImageWorks. For TM scenes, each pixel is %LC+75 (to increase brightness). The Histogram feature shows the distribution of %LC for a selected area. Time series were taken using the Spectral Plot feature. This allows us to look at changes in %LC with time (i.e. from scene to scene, loaded into channels in chronological order). Individual spectra are plotted as an average of an 11x11 box (330mx330m), to reduce the effect of local error. Each plot shows the spectra for 8 of these boxes to give a sense for the variability within each valley. Data were only extracted from the bajadas, in order to stay away from non-xeric vegetation which may be present on valley floors. The locations of these bajadas are marked by the red lines on Figure 1. For AVHRR NDVI data, individual spectra are plotted as an average of a 3x3 box (3kmx3km), again to reduce the effect of local error.
For all SMA spectral graphs, the numbers 2-15 on the horizontal axis correspond to the years 1985-1998, and, again, the vertical axis corresponds to %LC+75 (i.e. 75 is 0 %LC with the chosen endmembers).
Figure 3 shows %LC over this time period for a bajada to the west of the town of Independence, in the Owens Valley. To a first order, %LC follows rainfall quite well (see BIS and CTT on Figure 2). There are large peaks in 1986 and 1998 (big rainfall years), minor peaks in 1993 and 1995 (also big rainfall years) and lows in intervening dry years. Note that the 1987-1992 drought shows up as a broad low in %LC. Overall, vegetation abundance ranges from 4-18 %LC. Recall from above that this is lower than the 25% reported by Ustin et al. (1986b) for the same bajadas, using 1982 TM images.
Figures 4-7 show %LC over this time period for bajadas in the Eureka, Saline and Death Valleys. Although similarities do exist, these spectra generally are quite different from those in the Owens Valley. 1985, 1993 and 1998 are peaks, whereas 1986, 1991 and 1997 are lows. Although rainfall data is not available for most of these valleys, the observed trends in %LC do not follow the data available for Death Valley (see DTV on Figure 2). Perhaps the most striking difference between these valleys and the Owens Valley is the complete opposite sense of %LC in 1986. This is a very strong peak in %LC in the Owens Valley, but is a very strong low in %LC in the other valleys. Consequently, the entire drought period from 1987-1992 does not appear to be abnormally low in %LC in any of these valleys. Note also that vegetation abundance is rarely greater than 8 %LC (DN=83), and is often close to zero. The Saline and Eureka Valleys have many negative %LC results (DN<75), again showing the difficulties associated with choosing perfect endmembers.
To further investigate the effects of the 1987 to 1992 drought on %LC in these valleys, we constructed difference images using the Modeling Tool. 1986 was a wet year throughout most of California, including the Owens Valley, and 1989 was in the middle of the drought, so we expect that drought-affected areas will show up with positive %LC if we subtract 1989 from 1986. Values for this difference vary greatly from pixel to pixel, so we use construct histograms to look at the overall distribution and statistics of the data.
Figure 8 is the histogram of pixel values for a 1986-1989 difference image on the aforementioned bajadas west of Independence in the Owens Valley. Actual "grey level values" are (1986-1989)*10+100, so that a grey level value of 100 would be no change in %LC, 200 would be a 10% decrease in %LC and 80 would be a 2% increase in %LC, from 1986 to 1989. The mean for this distribution is about 150, meaning the bajadas of Owens Valley lost roughly 5 %LC from 1986 to 1989, presumably due to the drought.
Figures 9-12 are the histograms of change in %LC from 1986 to 1989 for the other valleys, under the same stretch. The means of these distributions range from 62 to 91, meaning that these valleys gained roughly 1-4 %LC from 1986 to 1989. These may be insignificant given the 4 %LC error (at 65% confidence) in SMA mentioned above. This supports our previous spectra and rainfall observations that the drought signature does not exist east of the Owens Valley.
Figures 13-17 are a similar series of histograms to Figures 8-12, except these are for 1998 minus 1989. 1998 was a wet year in both Owens and Death Valleys (see Figure 2), so this difference image would characterize the recovery of %LC from the drought. The Owens Valley bajadas show an increase of 8 %LC (Figure 13), while the others show an increase of 2-5 %LC (Figures 14-17). This indicates that recovery from drought in the valleys east of Owens Valley was not as significant as in Owens Valley itself. This apparent "recovery" is likely due to the fact that 1998 was a big rain year more than 1989 being a drought year.
Figure 18 shows NDVI for this region in late spring, 1989. The gray scale is adjusted to (NDVI+1)*100, so that a DN of 100 is 0 NDVI. This data has much lower resolution than TM data, so it is difficult to pick out features clearly, but our respective data collection sites are penned in. Note the white cross in the Owens Valley probes DN=109, indicating NDVI=0.09. The western slope of the Sierras is very bright, indicating richer vegetation growth, while the high peaks are very dark, indicating snow cover (negative NDVI).
Figure 19 is a spectral plot of NDVI with time for the bajadas west of Independence in the Owens Valley and the Eureka Valley. Other valleys are not plotted since they are similar to the Eureka Valley spectrum. The numbers 1-250 on the horizontal axis correspond to roughly January 1989 to December 1999. A gentle sloping upward of the plots in the late 1990s is due to sensor drift, not increases in NDVI (Elmore, pers. comm.). Consequently, there is no apparent affect on NDVI from the 1987-1992 drought. The seasonal growth and senescence of vegetation is evident as the dominant one-year wavelength in both the Owens and Eureka spectra. Owens Valley NDVI is about 0.05-0.10 greater than Eureka Valley, suggestive of greater vegetation health, or greater %LC, or both. This supports TM SMA results discussed above.
It’s apparent from this plot that the growing season (width of the NDVI peak) in the Eureka Valley is, on average, earlier and shorter than in the Owens Valley. Since a monthly breakdown of rainfall over this time period (see Figure 20; CDWR, 2000) indicates that peak rainfall months are generally the same in Owens and Death Valleys, we infer that vegetation in the eastern valleys completes growth and begins senescence much quicker than in the Owens Valley. This is indicative of a harsher climate.
These results point to some very simple, yet intriguing conclusions. The 1987-1992 drought, very significant in much of the western US, caused a loss of 5 %LC in the Owens Valley xeric shrub community, but was insignificant in similar plant communities in valleys to the east. This supports rainfall data (what little there is), which indicates that these years were not, on average, drier than normal for these valleys. The rainshadow effect of the Sierras, yielding just 14cm of precipitation yearly in the Owens Valley, is probably amplified to the east by the White/Inyo range, creating an even harsher desert environment. Here rainfall likely averages a meager 5-10cm per year, and plant communities respond by thinning out (less %LC) and taking on shorter growing seasons and longer dormant seasons.
Further work on this topic might address the responses of different vegetation communities within the eastern valleys to the drought. Although xeric shrubs are likely to be dominant where the groundwater table is deep, there is enough variability in the %LC data to suggest that other communities may be present, especially around small intermittent streams, or salt playas. To ensure a predominance of xeric vegetation, we had to gather data from alluvial fans and bajadas only. In-situ field work, involving species identification and mapping and emplacement of more rainfall gauges, would greatly aid this analysis. Meanwhile your researcher is left with the impression that so much of this work is one great guessing game!
Adams, J.B., et al., ‘Imaging Spectroscopy: Interpretations Based on Spectral Mixture Analysis,’ in Remote Geochemical Analysis: Elemental and Mineralogical Composition, 145-166, 1993.
Adams, J.B. and Adams, J.D., ‘Geologic Mapping Using Landsat MSS and TM Images: Remving Vegetation by Modeling Spectral Mixtures,’ in Third Thematic Conference on Remote Sensing for Exploration Geology, 615-622, 1984.
California Department of Water Resources, California Data Exchange Center, http://cdec.water.ca.gov, 2000.
Elmore, A.J., et al., ‘Quantifying Vegetation Change in Semi-Arid Environments: Precision and Accuracy of Spectral Mixture Analysis and the Normalized Difference Vegetation Index,’ submitted to Remote Sens. Environ., in press.
Kalluri , S.N.V. and Borak, J.S., ‘AVHRR Images Capture Natural Calamities in Mid-Atlantic Region,’ Eos, 81,16, 169-173, 2000.
Mustard, J., ‘Lab 8: Classification: Remote Sensing for Mapping Land Cover,’ for GE133 Lab, Spring 2000.
Sabins, F.F., Remote Sensing: Principles and Interpretation, 1997.
Smith, M.O., et al., ‘Vegetation in Deserts: I. A regional measure of abundance from multispectral images,’ Remote Sens. Environ., 31,1-26, 1990a.
Smith, M.O., et al., ‘Vegetation in Deserts: II. Environmental influences on regional abundance,’ Remote Sens. Environ., 31,27-52, 1990b.
Sorenson, S.K., et al., Soil and Vegetation Responses to Precipitation and Changes in Depth to Ground Water in Owens Valley, USGS Paper 2370, Ch. G, 1991.
Ustin, S.L., et al., ‘Use of Remote Sensing Techniques in the Analysis of Semi-Arid Shrub Communities,’ in Proceedings of the White Mountain Research Station High Altitude Symposium , 84-98, 1986a.
Ustin, S.L., et al., ‘Thematic Mapper Studies of Semiarid Shrub Communities,’ BioScience, 36,7,446-452, 1986b.