6533b82dfe1ef96bd12912b6

RESEARCH PRODUCT

Extraction of Endmembers from Spectral Mixtures

subject

description

Abstract Linear spectral mixture modeling (LSMM) divides each ground resolution element into its constituent materials using endmembers which represent the spectral characteristics of the cover types. However, it is difficult to identify and estimate the spectral signature of pure components or endmembers which form the scene, since they vary with the scale and purpose of the study. We propose three different methods to estimate the spectra of pure components from a set of unknown mixture spectra. Two of the methods consist in different optimization procedures based on objective functions defined from the coordinate axes of the dominant factors. The third one consists in the design of a neural network whose architecture implements the LSMM principles. The different procedures have been tested for the case of three endmembers. First, were used simulated and real data corresponding to mixtures of vegetation and soil. Factors that limit the accuracy of the results, such as the number of channels and the level of data noise have been analyzed. Results have indicated that the three methods provide accurate estimations of the spectral endmembers, especially the third one. Moreover, the second method, that is based on the exploration of the mixture positions in the factor space, has demonstrated to be the most appropriate when the dimensionality of the data is reduced. Finally, this procedure was applied on a Landsat-5 TM scene.