Mapping Vegetation, Soils, and Geology in Semiarid Shrublands Using Spectral Matching and Mixture Modeling of SWIR AVIRIS Imagery

Spectral matching and linear mixture modeling techgeological mapping because it allowed identification and niques have been applied to synthetic imagery and mapping of the relatively pure regions of all the surficial AVIRIS SWIR imagery of a semiarid rangeland in order materials that exert an influence on the spectral response. to determine their effectiveness as mapping tools, the The maps of the different clay minerals were of considersynergism between the two methods, and their advanable value for mineral exploration purposes. Conversely, tages, and limitations for rangeland resource exploitation spectral matching was less useful than mixture modeling and management. Spectral matching of pure library specfor rangeland vegetation studies because a classification of tra was found to be an effective method of locating and all pixels is needed and abundance estimates are required identifying endmembers for mixture modeling although for many applications. Mixture modeling allowed identifisome problems were found with the false identification cation of both nonphotosynthetic and green vegetation of gypsum. Mixture modeling could accurately estimate cover and thus total cover. Though the green vegetation proportions for a large number of materials in synthetic mixture map appears to be very precise, the nonphotoimagery; however, it produced high variance estimates synthetic vegetation estimates were poor. Elsevier Sciand high error estimates when presented with all nine ence Inc., 1999 AVIRIS endmembers because of high noise levels in the imagery. The problem of which endmembers to select was addressed by implementing a mixture model that allowed INTRODUCTION estimation of the errors on the proportions estimates, disThis study investigates the capabilities, and limitations of carding the endmembers with the highest errors, recomusing imaging spectroscopy data in the short wavelength puting the errors, and the proportions estimates, and itinfrared (SWIR 2–2.5 lm) to map the vegetation, geolerating this process until the mixture maps were ogy, and soils of a semiarid rangeland. This wavelength relatively free from noise. This methodology ensured that range has been shown to be a promising one for mineral the lowest contrast materials were discarded. The ineviidentification and mapping (Mackin et al., 1990; Hook et table confusion that followed was monitored the using al., 1991), but also has the potential to identify, and map the maps produced by spectral matching. Spectral some of the different constituents of the vegetation canmatching was more effective than mixture modeling for opy (i.e., green leaves and woody material) because green plant materials exhibit a spectrum dominated by *Department of Geography, King’s College, London water absorption while spectra of nonphotosynthetic †Departmento Quimica Agricola, Geologia y Geoquimica, Uniplant materials exhibit absorption features due to lignin, versidad Autonoma, Madrid, Spain ‡ESSC, University of Reading, Whiteknights, Reading, Berks, and hollocellulose (Elvidge, 1990). United Kingdom. To accomplish this aim, we compared and combined Address correspondence to N. A. Drake, Dept. of Geography, two methods of mapping surface materials using imaging King’s College, Strand, London, WC2R 2LS, UK. E-mail: nick.drake@ spectrometry data. First, we adopted a spectral matching kcl.ac.uk Received 4 August 1997; revised 16 September 1998. approach to match library spectra to those in the image, in