Endmember selection for multiple endmember spectral mixture analysis using endmember average RMSE

Abstract

Multiple endmember spectral mixture analysis (MESMA) models mixed spectra as a linear combination of endmembers that are allowed to vary in number and type on a per pixel basis. For modeling an image using MESMA, a parsimonious set of endmembers is desirable for computational efficiency and operational simplicity. This paper presents a method of selecting endmembers from a spectral library for use in MESMA. Endmember average root mean square error (EAR) uses MESMA to determine the average error of an endmember modeling spectra within its land cover class. The minimum EAR endmember is the most representative endmember for a land cover class within the spectral library and can be used to model the larger image. These techniques were used to map land cover, including four dominant vegetation species, soil, and senesced grass, in the Santa Ynez Mountains above Santa Barbara, CA, USA. Image spectra were extracted from a 20-m resolution airborne visible infrared imaging spectrometer (AVIRIS) reflectance image using reference polygons and combined into a library of 915 spectra. Possible confusion between land cover classes was determined using the class average RMSE (CAR). EAR was used to select the single most representative endmember within each land cover class. The six minimum EAR endmembers were used to map the AVIRIS image. Land cover class accuracy was assessed at 88.6%. Using a fractional accuracy assessment, undermodeling of dominant land cover classes and overmodeling of absent land cover classes was found at the pixel scale. Land cover mapped using the minimum EAR endmembers represents a substantial improvement in accuracy over previous efforts.