Classification of multispectral images based on fractions of endmembers: Application to land-cover change in the Brazilian Amazon

Abstract

Four time-sequential Landsat Thematic Mapper (TM) images of an area of Amazon forest, pasture, and second growth near Manaus, Brazil were classified according to dominant ground cover, using a new technique based on fractions of spectral endmembers. A simple four-endmember model consisting of reflectance spectra of green vegetation, nonphotosynthetic vegetation, soil, and shade was applied to all four images. Fractions of endmembers were used to define seven categories, each of which consisted of one or more classes of ground cover, where class names were based on field observations. Endmember fractions varied over time for many pixels, reflecting processes operating on the ground such as felling of forest, or regrowth of vegetation in previously cleared areas. Changes in classes over time were used to establish superclasses which grouped pixels having common histories. Sources of classification error were evaluated, including system noise, endmember variability, and low spectral contrast. Field work during each of the four years showed consistently high accuracy in per-image classification. Classification accuracy in any one year was improved by considering the multiyear context. Although the method was tested in the Amazon basin, the results suggest that endmember classification may be generally useful for comparing multispectral images in space and time.