Mapping vegetation, soils, and geology in semiarid shrublands using spectral matching and mixture modeling of SWIR AVIRIS imagery

Spectral matching and linear mixture modeling techniques have been applied to synthetic imagery and AVIRIS SWIR imagery of a semiarid rangeland in ord er to determine their effectiveness as mapping tools, the synergism between the two methods, and their advantages, and limitations for rangeland resou rce exploitation and management. Spectral matching of pure library spectra was found to be an effective method of locating and identifying endmembers for mixture modeling although some problems were found with the false ident ification of gypsum. Mixture modeling could accurately estimate proportions for a large number of materials in synthetic imagery; however, it produced high variance estimates and high error estimates when presented with all n ine AVIRIS endmembers because of high noise levels in the imagery. The prob lem of which endmembers to select was addressed by implementing a mixture m odel that allowed estimation of the errors on the proportions estimates, di scarding the endmembers with the highest errors, recomputing the errors, an d the proportions estimates, and iterating this process until the mixture m aps were relatively free from noise. This methodology ensured that the lowe st contrast materials were discarded. The inevitable confusion that followe d was monitored the using the maps produced by spectral matching. Spectral matching was more effective than mixture modeling for geological mapping be cause it allowed identification and mapping of the relatively pure regions of all the surficial materials that exert an influence on the spectral resp onse. The maps of the different clay minerals were of considerable value fo r mineral exploration purposes. Conversely, spectral matching was less usef ul than mixture modeling for rangeland vegetation studies because a classif ication of all pixels is needed and abundance estimates are required for ma ny applications. Mixture modeling allowed identification of both nonphotosy nthetic and green vegetation cover and thus total cover. Though the green v egetation mixture map appears to be very precise, the nonphotosynthetic veg etation estimates were poor. (C)Elsevier Science Inc., 1999.