Mineral mapping and landsat thematic mapper image classification using spectral unmixing

Abstract

In this paper, spectral image unmixing applied to Landsat Thematic mapper data from southern Spain is described to obtain a classified image based on abundance estimates of a number of spectral endmembers assuming linear mixing systematics. Spectral angle mapping (e.g. a technique by which a pixel spectrum is compared with a reference spectrum using the spectral angle between the two in a vector space) is used to distil the five most important endmembers out of a total of 12: (A) carbonate, (B) green vegetation, (C) dry vegetation, (D) hematite, and (E) kaolinite. The spectral unmixing final product is compared with classified images obtained using parallelepiped, maximumlikelihood, and k‐nearest neighbour classification. This comparison demonstrates that high precision results can be obtained from spectral unmixing. Furthermore spectral unmixing overcomes some drawbacks of conventional classification methods: the root‐mean square error and the total abundance image provide a means of assessing the accuracy of the analysis and spectral unmixing yields abundance estimates at a pixel support for all endmembers and thus allows a fuzzy‐type of classification in which more then one class may be present at a pixel.