Discriminating Semiarid Vegetation Using
Airborne Imaging Spectrometer Data:
A Preliminary Assessment

Randall W. Thomas1 and Susan L. Ustin2
 
1Remote Sensing Research Program, Space Sciences Laboratory
University of California, Berkeley, California 94720
 
2Remote Sensing Research Program, Space Sciences Laboratory
University of California, Berkeley, California 94720
 
Department of Botany
University of California, Davis, California 95616
Received 10 March 1987; revised 10 June 1987.

Abstract

A preliminary assessment was made of Airborne Imaging Spectrometer (AIS) data for discriminating and characterizing vegetation in a semiarid environment. May and October AIS data sets were acquired over a large alluvial fan in eastern California, on which were found Great Basin desert shrub communities. Maximum likelihood classification of a principal components representation of the May AIS data enabled discrimination of subtle spatial detail in images relating to vegetation and soil characteristics. The spatial patterns in the May AIS classification were, however, too detailed for complete interpretation with existing ground data. A similar analysis of the October AIS data yielded poor results. Comparison of AIS results with a similar analysis of May Landsat Thematic Mapper data showed that the May AIS data contained approximately three to four times as much spectrally coherent information. When only two shortwave infrared TM bands were used, results were similar to those from AIS data acquired in October.

Introduction

Ecologists and resource managers wish to use remote sensing techniques to monitor vegetation condition, patterns, and trends in arid regions where rugged terrain, poor access, and extreme climate make field research difficult. Unfortunately, because of low canopy density the synoptic data provided by the relatively broad-band moderate resolution Multispectral Scanner (MSS) and Thematic Mapper (TM) satellite sensors have been used with only limited success for extracting detailed vegetation information (Tueller, 1980; Myers, 1983). In contrast, geologists have been able to take advantage of low vegetation cover in arid regions. Such areas are particularly appropriate to remote-sensing-aided mapping of mineralogy or Ethology because the contribution of vegetation to mixed pixels often can be ignored (Goetz et al., 1983; Siegal and Gillespie, 1980).

A new sensor system, the Airborne Imaging Spectrometer (AIS), has been developed to provide both high spectral and spatial resolution information. This system is potentially more useful for discriminating arid and semiarid shrub communities (Goetz, 1985; Goetz et al., 1983; 1985). The first generation instrument, the AIS-1, has 128 near and shortwave infrared channels averaging approximately 10 nm in width per band. The AVIRIS (Airborne Visible Infrared Imaging Spectrometer), available in 1987-1988, will have 224 channels covering the visible to shortwave infrared regions. These instruments provide the first opportunity to produce high spectral resolution images capable of detecting subtle changes in canopy water or pigment content, to detect changes in abundance or density of dominant species among communities, and to detect narrow-band absorption features potentially useful for the unique identification of surface materials.

In this paper we investigate the spatial and spectral resolution of AIS-1 data and provide a preliminary assessment regarding its potential for discriminating vegetation in semiarid environments. Specifically, AIS results are compared with a corresponding analysis of Landsat TM data to examine the information content of AIS data relative to the TM. The data presented here was obtained during two seasons when shrub vegetation were either at minimum or maximum physiological activity.

Study Site

A map of the larger study area shows the approximate positions of the AIS and TM data included in this report and symbols designating specific areas discussed in text (Fig. 1). The Owens Valley in southeastern California is located in the rainshadow of the Sierra Nevada range and west of the White-Inyo Mountains. An alluvial fan was examined in detail (36°44'-36°47' N latitude and 118°17' W longitude), between 1300 and 1500 m elevation with slopes of 5-15%, west of the town of Independence (Fig. 1). The shrub vegetation of the fan is part of a broad ecotone between Great Basin desert communities on the upper slopes, predominantly sagebrush (Artemisia tridentata) and blackbrush (Coleogyne ramosissima), and Mojave elements (e.g., Franseria dumosa) on the lower slopes. Patches of vegetation on the fans have been removed by past wildfires. These patches exhibit partial recovery toward reestablishment of the original shrub cover and species composition. Two such areas permit comparison with adjacent shrubland, the youngest Symmes Creek burn (SCB): resulted from a wildfire in 1977 and the older Onion Valley burn (OVB), from a wildfire in the 1930s. In addition, two salisic soil types of differing age are present also, each soil derived from granitic parent materials in the Sierra Nevada. The older soil (Tinemaha, TIN) to the north has proportionally more iron and magnesium oxides and is visually darker and more finely structured than the younger soil (Lubkin-Tinemaha, LUT) to the south.

Percent vegetation cover, height, and species composition on the burned and unburned surfaces of the younger alluvial substrate and on the older alluvial substrate were determined by 50 m line transects (n = 18-21) using a stratified random sampling design (Table 1; data from Ustin et al., 1986b). The following perennial species vary in proportion from site to site but comprise 93-98% of the total vegetation cover: Coleogyne ramosissima (Blackbrush), Eriogonum fasiculatum (California Buckwheat), Artemisia tridentata (Great Basin Sagebrush), Chrysothamnus teretifolius (Long Leaf Rabbitbush), Haplopappus Cooperi (Cooper's Goldenbush). Annuals and grasses were an insignificant proportion of the total cover during 1984-1985. Soil mapping was performed by the Bureau of Land Management (Vaughn, 1983) and Soil Conservation Service, Bishop, CA (F. Fischer, personal communication). Both soil units are classified as loamy-skeletal, mixed, thermic Xeralfic Haplargids and differ principally in age and weathering. All soils are gravelly with scattered boulderfields.

AIS and TM Data

AIS data reported in this study were acquired from the AIS-1 sensor flown on a NASA C-130 aircraft. This instrument uses a 32 x 32 element mercury-cadmium-telluride area detector array (Vane, 1986). Ground resolution is approximately 10x10 m (3.7° FOV) in a 32-pixel swath. AIS data are collected in 128 spectral bands with 9.3 nm sampling intervals in either of two modes, "rock" from 1.21 to 2.40 mm or "tree" from 0.91 to 2.10 mm, depending on choice of detector grating positions. Flight lines were nearly parallel and somewhat offset (roughly separated by 100 pixels) along the length (Fig. 1). Flights occurred between 11.30 and 12.30 h local time on 30 October 1984 (line 402, "rock" mode) and 23 May 1985 (line 507, "tree" mode) at an approximate altitude of 5 km above mean terrain. Initial reviews of the data have indicated that a blocking filter failed to prevent contamination by second-order light at wavelengths longer than 1.6-1.7 mm, i.e., from 0.8 mm and longer wavelengths contributing to the detector signal at the harmonic interval (Vane, 1986). Thus, we have chosen not to include data for bands 1.70 mm and longer in the statistical analyses.

Within the region of wavelength overlap between the two AIS data sets, the comparisons between dates were primarily centered on two wavelength regions: 1.21-1.29 mm and 1.53-1.69 mm (Fig. 2). These correspond roughly to reflectance peaks in the near and middle infrared reflectance range, common to the two dates but excluding wavelengths having atmospheric water absorption (1.30-1.52 mm) (Fig. 3). These spectral regions do, however, include the shoulders of wavelength regions affected by atmospheric water absorption. Murray et al. (1986) found the AIS spectral region between 1.53 and 1.63 mm to best characterize between-class differences in a desert shrub community in central Oregon.

Digital TM data covering the Owens Valley area were acquired from a 16 May 1985 scene (row 41, pass 35) with a 09.45 h overpass time. These data were obtained 1 week prior to the AIS acquisition. All six reflected visible and infrared bands having eight-bit spectral resolution (256 levels) and 30 m on a side ground resolution per pixel were included in the data set. These bands covered the following wavelength regions: 0.45-0.52 mm (blue), 0.52-0.60 mm (green), 0.63-0.69 mm (red), 0.76-0.90 mm (near infrared, NIR), 1.55-1.75 mm and 2.08-2.35 mm (shortwave infrared, SWIR). In addition a thermal band (10.3-12.5 mm) having a 120 m on-a-side ground resolution was also obtained as part of the standard Landsat digital TM data set.
 

AIS data processing

AIS data were displayed and spectral plots analyzed using the SPAM (Spectral Analysis Manager) software developed by the Jet Propulsion Laboratory (JPL), Version 4.2 bsd UNIX, as implemented at U. C. Berkeley. The data were rectified for radiometric differences and corrected for solar irradiance by the JPL prior to our analyses (Vane and Goetz, 1985; 1986). However, due to calibration difficulties, considerable residual vertical striping was observed which was more evident in May than October data. Vertical striping was removed in both data sets through a column normalization procedure, prior to statistical analyses, by calculating the means of each pixel column and the mean of the 32x512 pixel line segments. The difference between each column mean and the scene mean was determined and this difference added or subtracted to the digital reflectance value of each pixel in the column (Ustin et al., 1986a).

The eight destriped AIS bands in the May data covering the 1.21-1.28 mm wavelength region were treated as one set of variables. The 16 bands covering the 1.54-1.69 mm region were separated into two eight-band sets of variables. The October data was treated similarly except that the data selected was displaced one band toward shorter wavelengths as a result of rounding-error uncertainty. Each band set was subjected to statistical analysis using the Earth Resources Laboratory Applications Software (ELAS) (Graham et al., 1984) as implemented at U.C. Berkeley under Version 4.2 bsd UNIX. In this case, the Supervised Encounter (SUPE) module was used to produce a mean vector (by summation over pixels), covariance matrix, and correlation matrix for each band set. Principal components analysis (e.g., Anderson, 1958; Morrison, 1976) was performed with the Statistical Analysis System (SAS, 1982), using the correlation matrix to give eight eigenvalues and eigenvectors for each set. The eigenvectors (containing the direction cosines) were then multiplied by appropriate destriped AIS bands in ELAS (via the Programmable Calculator module, PCAL) to compute the first and second principal component (PC) scores for each AIS pixel. A similar computation of PC scores was performed for 48 contiguous AIS bands, divided into eight-band sets, covering the 0.91-1.38 mm wavelength region in May and the 1.21-1.69 mm region in October.

First and second PC images were produced for the two AIS data sets. The first PC scores in the 1.21-1.28 m m) 1.54-1.69 mm wavelength regions were then subjected to unsupervised clustering in ELAS using the Search (SRCH) module. Initial clustering was performed with a minimum required intercluster distance (sdis) of 1.5. Final clusters were defined using various minimum required intercluster distances between 0.375 and 1.5, according to the dynamic range and number of distinct spectral classes available in each data set. Resulting cluster statistics (mean vector and covariance matrix) were used to classify each pixel in each of the 32 x 512 pixel data sets to one unique class using the ELAS Maximum Likelihood Classifier (MAXL).
 

TM data processing

Landsat TM data were also processed using ELAS unsupervised clustering and maximum likelihood classification algorithms. In this case, only original TM bands were used because of the limited numbers of bands for PC analysis. For purposes of comparison, a TM classification set was produced using data for the six reflected visible and infrared bands from the 16 May 1985 TM scene selected from a 151 x 151 pixel area centered on the AIS flight lines (Fig. 1). Inspection of the thermal channel (TM Band 6) data had suggested that inclusion would degrade the classification so that it was omitted. Three classification maps were produced using intercluster distances of 0.75, 1.00, and 1.5 for the final clusters. A second TM classification set consisting of six classification maps (sdis 0.375-1.5) of May data using TM Bands 5 and 7 was produced to evaluate TM classification performance using the wavelength region that has the closest correspondence with the AIS data sets.

All AIS and TM classification maps were displayed using ELAS and examined for relative level of classification detail and approximate spatial correspondence with ground cover and soil types. In addition to examination of the classification statistics, the location and relative distribution of individual AIS classes were evaluated by displaying each class individually to determine the spatial coherence and fidelity with known ground features. Because this analysis of the classification maps was based on comparisons with preexisting ground data, no attempts were made to locate specific pixels on the ground. The surface characteristics considered were the mean and variance of the vegetation cover, height, and species composition (based on transect data) from each of the four surface types. Additionally, the positions of terrain features such as streams and washes were examined on the AIS data using 35-mm black/white and color infrared ground-track photos obtained from the AIS flights and from black/white aerial photos (1973 and 1983) obtained from the U.S.D.A. Forest Service.

Results and Discussion

Data quality

The AIS-1 was the prototype test bed instrument of the high spectral resolution imaging spectrometry program. As such, the AIS suffered from a number of sources of instrument noise, which reduced the signal-to-noise ratio to 10-40% of the theoretical efficiency (Vane, 1986). These noise sources included electronic noise (signal chain noise and instrument-aircraft ground loops), mechanical noise (horizontal stripping, from instrument vibration and grating position wobble), and detector noise which resulted in vertical striping, reportedly from detector response AIS spectra changes between calibrations and from second-order energy overlap (Cocks and Green, 1986; Conel et al., 1986; Tucker and Vane, 1986; Vane, 1986). These problems apparent in the data we analyzed, have been corrected or significantly reduced in the second generation AIS-2 and AVIRIS instruments. Despite the difficulty with data quality, this system has demonstrated the feasibility and potential usefulness of high spectral resolution data for a variety of applications (Goetz et al., 1985; Vane, 1986). However, results presented at the AIS Data Analysis Workshops in 1985 and 1986 have suggested that interpreting vegetation remains more difficult and ambiguous than analyses for many geological purposes (Vane and Goetz, 1985; 1986).
 

AIS spectral band images

AIS flight lines showing the 24 destriped channels from October and May are illustrated in Fig. 2. Most vertical striping was removed by the destriping procedure (described in the AIS Data Processing section above), although horizontal striping remains. Despite the fact that information in adjacent AIS bands tends to be correlated (i.e., varies in the same direction), some differences, principally in brightness, are observed. Sharp contrasts in brightness are observed at grating position changes. Nonetheless, some spatial patterns are apparent. For example, Independence Creek and the recently burned shrub community (SCB) are clearly seen on the two images as are some of the smaller anastomosing washes and channels on the fan. However, the older burned surface (OVB) is less distinct, and the two differentially weathered soil terraces cannot be distinguished.
 

AIS spectra

AIS spectra for the four surface types on the alluvial fan are presented in Fig. 3. Note that digital number (DN) scales are not equal on the two plots. Plots are 5 x 5 pixel averages typical of reflectance characteristics from each surface type. These AIS data are from October and May acquisition dates, and correspond to periods when physiological condition of the plant communities were near minimum and maximum levels of growth activity, respectively. The October measurements occurred during a seasonal drought that had lasted 10 months. Thus, both vegetation canopies and soil surfaces were quite dry, and the vegetation was in a near-dormant state. In contrast, the May measurements were obtained after a wet spring and after mild but overcast sky conditions on the preceding day. Therefore, the vegetation was considered to have been in a relatively moist, vigorous state. Both assessments are supported by growth measurements on the five dominant shrub species from a site within the study locality (Ustin et al., 1986b; unpublished data).

Note that the atmospheric water bands (centered at 1.4 and 1.9 mm) showed greater absorptance in May relative to October. This suggests that the AIS is sensitive to variations in atmospheric water vapor concentration. The greater atmospheric water absorption in May also affects the shoulders of the water bands, making the region between the water bands appear narrower. In addition to these major absorption features, secondary water absorption bands are seen at 0.94, 1.13, 1.25, 1.54, and 1.67 mm) Two smaller CO2 absorptance bands occur at 2.00 and 2.05 mm and another on the shoulder of the water band, at 1.46 mm. Secondary water and CO2 absorptance bands also show deeper absorption features in May relative to October.

In general, the two data sets show similar patterns for the overlapping spectral regions (Fig. 3). The spectra on both days basically follow the same signature patterns, with differences between surface types being expressed principally through albedo. In fact, the primary difference between the dates is in curve amplitude (digital numbers, DN), with maximum reflectance values in May data about 82% DN reflectance values in October (the proportion varied across the spectrum between 60 and 90%). This pattern is consistent with field measured spectra which showed increased reflectance during winter compared to summer (Ustin et al., 1985; 1986b; unpublished data). Observations on the field-collected spectral data suggest that the seasonal trend is linked with changes in the solar angle, possibly through altered canopy shading.

In both October and May, the relative brightness shows that the unburned shrub community on the older soil surface (TIN) had the lowest DN values, while the recent burned surface (SCB) on the younger soil terrace had the highest DN values (Fig. 3, Table 1). Both the unburned shrub community surface (LUT) and the older burned surface (OVB) on the younger soil terrace had similar spectral signatures and intermediate DN values. One of the important differences between dates is the ability to separate the soils of different ages (TIN and LUT) in October but not in May. In contrast, the May AIS image allowed better separation of the older burn surface (OVB) from the unburned shrub surface (LUT) which were only marginally separable in October. Although statistical tests were not performed, these results are consistent with numerous spectral plots examined on these data. It appears that the ability to distinguish spectrally similar surface types is somewhat dependent on the timing of data collection and possibly related to the phenological activity of the species. In our AIS data we have not attempted to separate sun angle effects from seasonal phenological differences in vegetation.

The primary change observed between the dates was in the region between the 1.4 and 1.9 mm water absorption bands (Fig. 3). This difference is principally seen in the negative slope between 1.51 and 1.70 mm in October and the relatively flat response across this region in May. If these differences were due to atmospheric water absorption on the shoulder of the 1.9 mm water band, the slope was expected to have been greater in May rather than October. However, other physical bases for the slope change cannot be readily determined since, in this case, all surface types responded similarly. Nonetheless, slope changes in reflectance in the NIR to SWIR region have been related to ecological differences by other authors, and have been consistently observed in comparisons among various vegetation surfaces and community types, as reported for marshes (Gross and Klemas, 1985, 1986), agriculture (Wood and Wrigley, 1985), urban (Vanderbilt, 1985), forest (Milton et al., 1986), and desert marsh and meadow communities (Ustin et al., 1986a).

In general, areas with dense vegetation tend to have positive slopes rather than negative slopes in the 1.5-1.7 mm region. Thus, in these data the general spectral pattern appears to be responding predominantly to soil features rather than vegetation per se. However, because these communities have similar species composition and structure they primarily differ only in total vegetation cover (Table 1). This is spectrally evident by the decrease in brightness as canopy cover increases. Ustin and Rock (1985) and Ustin et al. (1986a) suggested that the predominant effect of semiarid vegetation on these fans is to lower spectral brightness nearly equally in all AIS bands. This effect is at least partially due to the canopy geometry of semiarid shrubs, having low leaf area and relatively sparse, open branching (deliquescent structures) that cast shadows and shade the soil surface but have relatively low biomass. Based on an analysis of multidimensional TM spectral space in linear mixing models of this fan region, Smith and Adams (Univ. of Washington, personal communication) have found that the mixing space in their vegetation models occurs along a similar trajectory with the shadow-shade spectral mixing space. As a consequence, separation of vegetation and shadowed surfaces is difficult in semiarid communities with a physiogonomic structure such as these.
 

Principal Components Analysis

Results of the Principal Components Analysis (PCA) confirmed the pattern shown in the spectral plots. Table 2(a) and 2(b) present the PCA results for the first two PCs in the three wavelength regions of primary interest in this study. Inspection of these tables shows that a majority of the variance (61-96%) was accounted for by the first PC. Spatial pattern consistency between images of the first PC suggests that it represents brightness. When less than 90% of the variance was accounted for by the first PC, the eight-band groups included some spectral channels in an atmospheric water band, or included spectral bands at grating position changes which are reported to be noisy (Vane, 1986). For example, notice Bands 3 and 4 on Fig. 2(b).

Table 3 gives the percentage of the total variance over all bands in each eight-band data set accounted for by the first two principal components. Differences in percentage of variance accounted for by each PC may be explained by the inclusion of some AIS channels affected by atmospheric water in the analyses [see Fig. 3(a)-(b) for position and width of water bands]. This assumption is consistent with examination of the PC statistics and inspection of the PC images. For instance, the steep shoulders on the 1.4 mm water absorption band affect few spectral channels in the May image and, thus, have little effect on the PCs. This is confirmed by the relatively small proportion of variance accounted for by the second PC in May. On the other hand, the gentler slope of the water absorption regions in the October data trig. 3(a)] affected more of the spectral: bands, therefore, creating second and subsequent PCs accounting for a significant proportion of the variance in the NIR and second SWIR region.

Referring again to Table 2, note that when the first PC dominates (i.e.:, accounts for a large percentage of the variance) that the PC score is formed by an approximately even weighting (using the direction cosines) of the digital response from each original AIS band. When the first PC is less dominant, the weighting is more uneven. Less dominance in the first PC reflects inclusion of the edge of an atmospheric absorption window or choppiness in the spectral response curve due to plant or soil absorption, or other phenomena such as shading. However, direction cosines for the second PC nearly always represented a contrast (i.e., a differencing) between the first three to four AIS channels and the remaining bands in the set suggesting atmospheric water absorption. Even when the second PC accounted for greater than 5-20% of the variance, the dynamic range of DN counts was primarily restricted to fewer than five. Thus, in each case, only one PC was significant. A similar pattern was evident in consecutive eight-band groups referred to in Table 3. Other studies have reported that a significant proportion of the variance was attributable to brightness. For instance, Murray (1983) found that the first two PCs accounted for 91% of scene variance in Landsat multispectral data.

When a larger number of AIS channels are included in the PCA, other authors (Smith and Adams, 1985; Murray et al., 1986) found additional PCs to be significant. Murray et al. (1986) evaluated 32 channels in the 1.5-1.8 mm region obtained over desert sagebrush and soil targets in central Oregon. They found the largest proportion of variance was accounted for by the first two PCs which were successfully used to separate the major cover classes. The next four PCs, however, were found to contain a mix of real factor and noise components. Contamination of these subsequent PCs with noise supports our decision to destripe data and eliminate wavelength regions affected by atmospheric water absorption prior to statistical analyses. Unfortunately, because of analytical restrictions, we could not perform a PCA on our entire data set simultaneously, so we cannot eliminate the possibility that a larger number of significant PCs could be identified. In contrast, Smith and Adams (1985) analyzed 32-band AIS data in the 2.1-2.4 mm region (which does not include atmospheric water bands) from a site having greater geologic diversity at Cuperite, NV. They found additional significant PCs that were useful in separating six surface mineral and vegetation types. Subsequently, instead of using their PCs to define an image classification map as done in this study, they utilized the spectral dimensions associated with the PCs to decompose the matrix into spectral signatures. They demonstrated that it was possible to interpret these signatures by comparisons to spectra of various ground categories. This approach is similar to that proposed by Huete (1986) using ground-based radiometry.
 

Clustering and Classification

The number of classes resulting from clustering both the AIS and TM data sets are shown in Table 4. Note that the number of classes increased as the minimum intercluster distance (sdis) was reduced. In order to determine the number of meaningful classes available from each data set, the sdis was varied between 0.375 and 1.5 units and the resulting images were examined for spatial pattern consistency with known ground conditions. As seen in Table 4, considerable differences exist between the number of spectral classes identified at different intercluster distances among AIS and TM data sets. Table 4 also gives a subjective ranking of the adequacy of the resulting classifications in terms of three surface characteristics. These characteristics were 1) spatial coherence with known ground class pattern, 2) fidelity with ground labels for individual class map mosaic elements, and 3) class resolution, or the adequacy of the spectral partitioning of the data set to represent the number of ground categories.

May vs. October AIS data. Greater resolution of classification detail is apparent in the May map when compared with the October maps [Plate VII (a)-(d)]. However, not all clusters could be associated with ground surface features. The effect of the clustering algorithm, in this case, may have been almost equivalent to density slicing the PC spectral data belonging to these highly correlated ground density and cover classes. Inspection of the maximum likelihood results [Plate VII (a)-(d)] relative to known ground cover within each remote sensing data set showed the following: Despite apparent similarity of spectral signatures, May AIS data provided much better discrimination between surfaces and considerably more subtle structure. The number of distinct, spectrally meaningful (nonnoise) classes in the May AIS data appeared to be two to three times that of the October AIS data at all intercluster distances (sdis). For example, 10-15 such spectral classes were found on the more recent burn (SCB) in May in contrast to approximately half that number in the October data. The same pattern held on the more recent soil unit (LUT) and the older soil unit (TIN). Both the Symmes Creek and Independence Creeks were evident on the May image, whereas only the latter creek was detectable in October. Also the OVB area was more difficult to resolve in October. These observations regarding the October AIS classification results are reflected in the low rankings shown in Table 4 for spatial data coherence, correspondence with labels, and adequacy of spectral partitioning to represent ground cover categories. Of the October classifications, the eight-class resolution presented the best representation of ground categories. The 51-class product had many classes which lacked spatial continuity and which appeared to represent noise.

Although a third of the May classification may represent noise classes (e.g., classes having speckled, vertical or horizontal distributions, or classes of especially small size), an equivalent proportion of the class structure seen in the May AIS data did not appear to be noise and may in fact represent real variation in surface components. While the ground data was not sufficiently detailed to validate this conclusion, the spatial pattern inherent in the assignment of the pixels (when examined on a class-by-class basis) suggested that at least 25-35 classes were potentially nonnoise classes. Their spatial patterns were consistent with features (rock, shadow, riparian and shrub vegetation, and combinations of these) known to exist within the plant communities surveyed. An analysis of subpixel site variation (5 x 5 m plots) revealed that cover could vary greatly between adjacent plots (e.g., 15-30%). Additionally, sample sizes of 20 50-m transects were insufficient to produce accurate estimates of the mean cover of the dominant species individually. In fact, the variance of the total shrub cover averaged nearly 25%. Therefore, the considerable spatial heterogeneity of the fans could account for this level of class distinction. Thus, the May AIS 99-class classification, produced with an intercluster distance of 1.50, was ranked in Table 4 as "generally sufficient," in terms of class spatial coherence and correspondence to ground class labels. The exceeding degree of detail represented by the large number of classes considered to be nonnoise may be a product of slight variations in the proportion of various ground class components, such as vegetation density, composition, or surface texture.

TM data. As expected, the clustering and classification of the May TM data produced successively more detailed partitions of the landscape as the minimum intercluster distance was reduced (Table 4). At an sdis of 0.75, the TIN soil units and the SCB burn were accurately distinguished from other surface types. This classification map [Plate VII (e)] is similar to the less detailed July 1984 TM scene reported by Ustin et al. (1986c). Considerably more classes were produced from TM data when all visible and reflected infrared bands were used compared to just the two band shortwave infrared data sets. Use of six bands produced good class resolution (broad category) and the best spatial coherence of any data set examined. However, at sdis below 0.75, the classification categories were partitioned into units having much less spatial organization. A similar pattern was observed in comparison of six-, three-, and two-band July TM data of this region (unpublished data). Plate VII (f) illustrates the loss of resolution when only the two SWIR TM bands are used. When visible bands were included in the analysis, significantly better resolution of broad category classes was developed at all sdis. This suggests that the AVIRIS system, which will have visible and IR bands combined with an intermediate ground resolution between the AIS-1 and TM (i.e., 20 x 20 m pixels) will produce better spatial pattern categories than either AIS1 or TM data. Plate VII (g) is an aerial color infrared photo obtained from the October AIS flight. Comparison of this figure with the AIS and TM maps facilitates image comparison with ground features.

AIS vs. TM data. AIS maps show greater detail than TM maps, especially when the comparison is restricted to the TM two-band SWIR wavelength region. The May AIS data contained at least three times the spatially coherent classification detail as the corresponding May TM data. For example, six-channel TM classifications had three to four classes for the SCB burn compared to at least 10 or more meaningful classes in the AIS data. A similar ratio held on the older soil type mentioned previously and on the LUT soil type. Small washes and creeks running through the study site were not detectable with TM but could be resolved with both the October and May AIS data. Classes covering small areas but having coherent spatial patterns within the burns and in other vegetated areas could be distinguished with the May AIS data but not with the TM data. The significance of this AIS detail could not be determined, however, since it exceeded that of the corresponding ground sample data.

The situation was generally reversed with the October AIS data. May six-channel TM data provided both a more detailed and a more spatially coherent representation of ground cover phenomena. The two band, 11-class TM data set also produced classification results [Plate VII (f)] which were somewhat more spatially coherent (for broad scale patterns) and consistent with ground data than those obtained with the October AIS set. However, the October AIS data did enable better detection of fine scale spatial pattern, particularly the smaller streams and washes, than either of the two TM data sets.

The increased information content of the May AIS vs. May TM data sets may be due to a combination of better spatial resolution (ca. 10 m vs. 30 m), spectral resolution, and/or the location of the wavelength regions utilized. The relative importance of these factors will require additional research, which should include determination of the pixel-level mixtures of vegetation, rock, soil, and shade, contributing to the reflectance. Linear mixing models of reflectance data have been utilized by Adams and colleagues (Adams et al., 1986) to examine Viking Lander and TM data and by Huete (1986) using field radiometry measurements to identify spectral surface types in spatially averaged data. Mixing models have been examined by Adams and colleagues for TM data from the fan region studied here (Ustin et al., 1986c). Such an approach may prove useful in establishing a physical basis for the detailed information apparent in this initial analysis of AIS data for mapping arid land surfaces and vegetation cover.

Summary and Concluding Remarks

The degree of spectral discrimination obtained here can be considered to represent the minimum resolution possible with an AIS-like system. The AIS is an experimental instrument with attendant problems, and is limited to a portion of the reflective infrared wavelength region. As was shown by results from the six-channel TM data, inclusion of the NIR and visible wavelengths can produce valuable improvement in discriminating capability. This is because significant information regarding vegetation condition resides in the visible and NIR spectral regions. The AIS follow-on instrument, AVIRIS, will be capable of imaging in these wavelengths as well.

Nonetheless, it was shown that despite the apparent similarity of spectral signatures between vegetation and soil surfaces, a sufficient degree of fine-scale spectral structure existed in AIS wavelength regions to define major and minor surface features. The degree to which differentiation was possible varied with time of year, being greater in spring. The level of potentially useful ground class information in the May AIS data was found to significantly exceed that in approximately date-matched TM six-channel data. One objective of further work should be determining what these small spatially coherent class patterns represent.

In the AIS data analyzed here, only the first PCs were significant and these were related to scene brightness. Even though they were developed from small eight-band sets covering a limited portion of the SWIR spectrum, these PCs produced scene classifications at, or exceeding, knowledge of the surface conditions on the fan. A PCA over a larger portion of the data set could reveal additional PCs related to physical condition. Further improvements in extractable information might result from other partitionings and/or transformations of the data. For example, selection of individual or very small groups of AIS bands to take advantage of the microphenomena seen in the AIS spectral plots might enable a more accurate identification of subtle vegetation condition differences. Or construction of band ratios, and generation of greenness or other spectral indices related to biomass or vegetation condition, may be possible. In addition, inspection of our data suggests that information recovery from AIS for many purposes may be improved by further normalization of the raw data.

Our results represent a preliminary assessment of the usefulness of high resolution radiometry systems for arid vegetation studies. At present, facilities for processing and analysis of the large data arrays present the most severe limitation for use of such data. The SPAM software package, developed by the JPL, provides the most widely used method of analysis. While the capabilities within this package have grown, it remains limited because of the difficulty in interfacing the data easily with readily available statistical packages or other software useful for specific research purposes. Even when such packages are available, because of the large number of AIS bands, software programs commonly used for other remotely sensed data are inadequate. Unlike analysis of mineralogical features which depend on identification of distinctly different absorption features across the spectrum, analysis of vegetation features depends upon more subtle changes in spectral patterns. This becomes extremely critical in arid regions where low vegetation density makes mere identification difficult. Better analytical methods, particularly for curve shape, are required before the full potential of this data is realized.

Acknowledgments

We wish to thank Mr. Paul Ritter (U.C. Berkeley) for assistance with data analysis, the White Mountain Research Station and staff for use of research facilities, and Dr. Roy Woodward (U.C. Davis) for assistance with the field research. This research was partially supported by NASA Grant NAS7-918 JPL Subcontract 9568899.

References

Adams, J. B., Smith, M. O., and Johnson, P. E. (1986), Spectral mixture modeling: a new analysis of rock and soil types at the Viking Lander I site, J. Geophys. Res. forthcoming.

Anderson, T. W. (1958), An Introduction to Multivariate Statistical Analysis, Wiley, New York.

Cocks, T. D., and Green, A. A. (1986), Airborne spectroradiometry: the application of AIS data to detecting subtle mineral absorption features, In Proc. Second Airborne Imaging Spectrometer Data Analysis Workshop, 6-8 May, (G. Vane and A. F. H. Goetz, Eds.), JPL Publ. No. 86-35, pp. 52-62.

Conel, J. E., Adams, S., Alley, R. E., Hoover, G., and Schultz, S. (1986). Analysis of AIS radiometry with emphasis on determination of atmospheric properties and surface reflectance, In Proc. Second Airborne Imaging Spectrometer Data Analysis Workshop, 6-8 May, (G. Vane and A. F. H. Goetz, Eds.), JPL Publ. No. 86-35, pp. 31-51.

Goetz, A. F. H. (1985), High spectral resolution remote sensing of the land, Proc. Soc. Photo-Opt. Inst. Eng. 475:56-68.

Goetz, A. F. H., Rock, B. N., and Rowan, L. C. (1983), Remote sensing for exploration: an overview, Econ. Geol. 78:573-590.

Goetz, A. F. H., Vane, G., Solomon, J. E., and Rock, B. N. (1985), Imaging spectrometry for earth remote sensing, Science 228:1147-1153.

Graham, M. H., Junkin, B. G., Kalcic, M. T., Pearson, R. W., and Seyfarth, B. R. (1984), ELAS (Earth Resources Laboratory Applications Software), Vol. 11: ELAS Users Guide, NASA and National Space Technology Laboratories, Earth Resources Laboratory, Rept. #183, Bay St. Louis, MI.

Gross, M. F., and Klemas, V. (1985), Discrimination of coastal vegetation and biomass using AIS data. In Proc. Airborne Imaging Spectrometer Data Analysis Workshop, 8-10 April (G. Vane and A. F. H. Goetz, Eds.), JPL Publ. No. 8541, pp. 129-133.

Gross, M. F., and Klemans, V. (1986), Use of Airborne Imaging Spectrometer (AIS) data to differentiate marsh vegetation, Remote Sens. Environ. 19:97-103.

Huete, A. R. (1986). Separation of soil-plant spectral mixtures by factor analysis, Remote Sens. Environ., forthcoming.

Milton, N. M., Walsh, P. A., and Purdy, T. L. (1986), Geobotanical studies at Pilot Mountain, North Carolina, using the Airborne Imaging Spectrometer, In Proc Second Airborne Imaging Spectrometer Data Analysis Workshop, 6-8 May (G. Vane and A. F. H. Goetz, Eds.), JPL Publ. No. 86-35, pp. 162-170.

Morrison, D. F. (1976), Multivariate Statistical Methods, 2nd ed., McGraw-Hill, New York.

Murray, R. J. (1983), Factor Analysis of multispectral data (abstract), In Proc. Int. Conf. Renewable Resource Inventories for Monitoring Changes and Trends, Corvallis, OR, 15-19 Aug., College of Forestry, OSU, pp. 648-651.

Murray, R., Isaacson, D. L., Schrumpf, B. J., Ripple, W. J., and Lewis, A. J. (1986). AIS spectra of desert shrub canopies, In Proc. Second Airborne Imaging Spectrometer Data Analysis Workshop, 6-8 May, (G. Vane and A. F. H. Goetz, Eds.), JPL Publ. No. 86-35, pp. 187-194.

Myers, V. Ed. (1983), remote sensing applications in agriculture. In Manual of Remote Sensing, 2nd ed. (R. N. Colwell, Ed.), Am. Soc. Photogramm., The Sheridan Press, Falls Church, VA.

SAS (Statistical Analysis System) Institute (1982), SAS Statistics Users Manual, SAS Institute, Cary, NC.

Siegal, B. S., and Gillespie, A. R. (1980), Remote Sensing in Geology, Wiley, New York.

Smith, M. O., and Adams, J. B. (1985). Interpretation of AIS images of Cuperite, Nevada using constraints of spectral mixtures. In Proc. Airborne Imaging Spectrometer Data Analysis Workshop, 8-10 April (G. Vane and A. F. H. Goetz, Eds.), JPL Publ. No. 85-41, pp. 62-67.

Tucker, D., and Vane, G. (1986), Radiometric calibration of the Airborne Imaging Spectrometer, In Proc. Second Airborne Imaging Spectrometer Data Analysis Workshop, 6-8 May (G. Vane and A. F. H. Goetz, Eds.), JPL Publ. No. 86-35, pp. 17-20.

Tueller, P. T. (1980), Remote sensing for range management, presented at the Remote Sensing for Resource Management Conf. 23-30 October, Kansas City, MO, U.S. Soil Conservation Service and the National Aeronautics and Space Administration.

Ustin, S. L., Rock, B. N., and Woodward, R. A. (1985), Analysis of substrate and plant spectral features of semi-arid shrub communities in the Owens Valley, California, In Proc. Int. Symp. Remote Sens. Environ. 4th Thematic Conference: Remote Sensing for Exploration Geology, ERIM, Ann Arbor, MI, pp. 347-359.

Ustin, S. L., Woodward, R. A., and Rock, B. N. (1986a), Patterns of vegetation in the Owens Valley, California, In Proc. Second Airborne Imaging Spectrometer Data Analysis Workshop, 6-8 May (G. Vane and A. F. H. Goetz, Eds.), JPL Publ. No. 86-35, pp. 180-186.

Ustin, S. L., Rock, B. N., and Woodward, R. A. (1986b), Use of remote sensing techniques in the analysis of semi-arid shrub communities. In Proc. White Mountain High Altitude Research Symp., Natural History of the White Inyo Range, Eastern California and Western Nevada and High Altitude Physiology, 23-25 Aug. (A. A. Hall, Jr. and D. J. Young, Eds.), University of California, pp. 84-98.

Ustin, S. L., Adams, J. B., Elvidge, C. D., Rejmanek, M. Rock, B. N., Smith, M. O., Thomas, R. W., and Woodward, R. A. (1986c), Thematic Mapper studies of semiarid shrub communities, BioSci. 36:446-452.

Vanderbilt, V. C. (1985), Urban, forest, and agricultural AIS data: fine spectral structure, In Proc. Airborne Imaging Spectrometer Data Analysis Workshop, 8-10 April (G. Vane and A. F. H. Goetz, Eds.), JPL Publ. No. 85-41, pp. 158-165.

Vane, G. (1986), Introduction, In Proc. Second Airborne Imaging Spectrometer Data Analysis Workshop, 6-8 May (G. Vane and A. F. H. Goetz, Eds.), NASA/Jet Propulsion Laboratory Publ. 86-35, pp. 1-16.

Vane, G., and Goetz, A. F. H., (Eds.) (1985), Proceedings of the Airborne Imaging Spectrometer Data Analysis Workshop, 8-10 April, NASA/Jet Propulsion Laboratory Publ. 85-41.

Vane, G., and Goetz, A. F. H., (Eds.) (1986), Proceedings of the Second Airborne Imaging Spectrometer Data Analysis Workshop, 6-8 May, NASA/Jet Propulsion Laboratory Publ. 86-35.

Vaughn, D. E. (1983), Soil Inventory of the Benton-Owens Valley Area: Parts of Inyo and Mono Counties, California, Bureau of Land Management BLM CA 83 009 7111.

Wood, B. L., and Wrigley, R. C. (1985), AIS investigation of agricultural monocultures, In Proc. Airborne Imaging Spectrometer Data Analysis Workshop, 8-10 April (G. Vane and A. F. H. Goetz, Eds.), JPL Publ. No. 8541, pp. 134-140.

1998, Center for Spatial Technologies and Remote Sensing (CSTARS)
University of California, Davis