Robert O. Green

Mapping Chaparral in the Santa Monica Mountains Using Multiple Endmember Spectral Mixture Models

Abstract A new technique, called multiple endmember spectral mixture analysis (MESMA), was developed and tested in the Santa Monica Mountains, using Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data acquired in the fall of 1994 to map California chaparral. The technique models remotely measured spectra as linear combinations of pure spectra, called endmembers, while allowing the types and number of endmembers to vary on a per pixel basis. In this manner, vegetation is characterized by a unique set of endmembers as well as by the fractions. Reference endmembers were selected from a library of field and laboratory measured spectra of leaves, canopies, nonphotosynthetic materials (e.g., stems), and soils and used to develop a series of candidate models. Each candidate model was applied to the image, then, on a per pixel basis, assessed in terms of fractions, root mean squared (RMS) error, and residuals. If a model met all criteria, it was listed as a candidate for that pixel. For this study, selection criteria included fractions between −0.01 and 1.01, an RMS less than 0.025 and a residual less than 0.025 in seven or more contiguous bands. A total of 889 two-endmember models were evaluated and used to generate 276 three-endmember models. To facilitate model selection from a large pool of candidates, an optimal set was selected to provide maximal areal coverage. A total of 24 two-endmember and 12 three-endmember models were chosen. These models were used to generate fraction images and vegetation maps showing evergreen and drought deciduous or senesced vegetation. We found that a majority of the image could be modeled as two-endmember models. Three-endmember models provided greater areal coverage, yet provided poorer vegetation discrimination due to an increase in model overlap (two or more model candidates modeling the same pixel). The vegetation maps demonstrate that the technique is capable of discriminating a large number of spectrally distinct types of vegetation while capturing the mosaic-like spatial distribution typical of chaparral. However, additional research is required to fully evaluate the technique and validate the vegetation maps that were produced.

Temporal and spatial patterns in vegetation and atmospheric properties from AVIRIS

Abstract Little research has focused on the use o f intaging spectrometry for change detection. In this paper, we apply Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data to the monitoring of seasonal changes in atinospheric water vapor, liquid water, and surface cover in the vicinity of the Jasper Ridge, CA, for three dates in 1992. Apparent surface reflectance was retrieved and water vapor and liquid water mapped by using a radiativetransfer-based inversion that accounts for spatially variable atmospheres. Spectral mixture analysis (SMA) was used to model reflectance data as mixtures of green vegetation (GV), nonphotosynthetic vegetation (NPV), .soil, and shade. Temporal and spatial patterns in endineinber fractions and liquid water were compared to the normalized difference vegetation index (NDVI). The reflectance retrieval algorithm was tested by using a temporally invariant target. Atmospheric analysis showed a strongly negative linear relation between water vapor and elevation with significant seasonal variation in water vapor. Comparison of AVIRIS estimates of specific humidity to ground-based measures showed good correspondence for all three dates. Analysis of .surface properties showed that GV, NDVI, and liquid water varied in. response to green vegetation and were highly correlated. However, whereas the NDVI peaked between 0.7 and 0.85 in forests, liquid water continued to vary by as much as a factor of two. Seasonal patterns included senescence in herbaceous and nonconiferous vegetation, potential leaf growth in cortiferous forests, and a general increase in shadows. This resulted in seasonal declines in NDVI, GV, and liquid water for non forested vegetation and increases in NPV. Nonconiferous forests showed similar declines in liquid water rend GV and increases in shade and NPV, but they showed an. increase in NDVI. In coniferouns forests, liquid water and NDVI increased seasonally, matching an interpretation of continued growth, but GV decreased owing to increased .shade. The combination of retrieved surface reflectance, atmospheric modeling, and mapping of liquid water demonstrates the utility of imaging spectrometry for change detection. SMA with the use of reference endmembers is an c effective method for monitoring surficial changes. Temporal patterns in NDVI that contradict trends of GV and liquid water in nonconiferous forests raise additional questions about the NDVI. Liquid water may be more appropriate for analysis of high-leaf-area, shadowed forests because it overcomes the problem of saturation with NDVI.

The Effect of Grain Size on Spectral Mixture Analysis of Snow-Covered Area from AVIRIS Data

Abstract We developed a technique to improve spectral mixture analysis of snow-covered area in alpine regions through the use of multiple snow endmembers. Snow reflectance in near-infrared wavelengths is sensitive to snow grain size while in visible wavelengths it is relatively insensitive. Snow-covered alpine regions often exhibit large surface grain size gradients due to changes in aspect and elevation. The sensitivity of snow spectral reflectance to grain size translates these grain size gradients into spectral gradients. To spectrally characterize a snow-covered image domain with mixture analysis, the variable spectral nature of snow must be accounted for by use of multiple snow endmembers of varying grain size. We performed numerical simulations to demonstrate the sensitivity of mixture analysis to grain size for a range of sizes and snow fractions. From Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data collected over Mammoth Mountain, CA on 5 April 1994, a suite of snow image endmembers spanning the imaged region’s grain size range were extracted. Mixture models with fixed vegetation, rock, and shade were applied with each snow endmember. For each pixel, the snow fraction estimated by the model with least mixing error (RMS) was chosen to produce an optimal map of subpixel snow-covered area. Results were verified with a high spatial resolution aerial photograph demonstrating equivalent accuracy. Analysis of fraction under/overflow and residuals confirmed mixture analysis sensitivity to grain size gradients.

Estimating snow cover and grain size from AVIRIS data with spectral mixture analysis and modeled snow spectra

Models of hydrology and climate in alpine and other seasonally snow-covered regions require input of snow- covered area (SCA) and snow surface grain size. The spectral signature of snow depends on the snow grain size. We have shown in earlier work that to map either SCA or grain size with optical data, one must know the distribution of the other variable. Hence, we solve for SCA or grain size simultaneously with spectral mixture analysis. Alpine regions frequently exhibit large snow grain size gradients due to rugged terrain. Because the spectral signature of snow is dependent on grain size, the grain size gradients translate into spectral gradients. Snow must then be represented by a range of endmembers. To provide the range and resolution of grain size, we use modeled snow spectra of varying brain size as snow endmembers. To complete the spectral mixture library, we incorporate reference endmembers of vegetation, rock, and soil. On airborne visible/IR imaging spectrometer data of the Sierra Nevada, CA, we ran multiple mixture models. We established constraints on mixture RMS, residual, and fractions to select a subset of physically realistic models. The optimal mixture was then selected from this subset by means of the least RMS. The snow endmember fraction and grain size of the optimal mixture provide the estimates of sub-pixel SCA and surface grain size, respectively.

Improving alpine region spectral mixture analysis estimates of snow-covered area

A technique has been developed to improve alpine-region spectral mixture analysis estimates of snow-covered area. Snow reflectance in near infrared wavelengths is sensitive to snow grain size while insensitive in visible wavelengths. Alpine regions often exhibit significant snow grain size gradients due to changes in aspect and elevation. A suite of snow image endmembers corresponding to the regions snow grain size range were extracted. Mixture models with fixed vegetation, rock, and shade were applied with each snow endmember to AVIRIS data collected over Mammoth Mountain, Calif., April 5, 1994. For each pixel, the snow-fraction estimated by the model with least mixing error (rms) was chosen to produce an optimal map of snow-covered area. Fraction under/overflow analysis and limited residuals analysis were performed on the test results.