Multiple endmember spectral-angle-mapper (SAM) analysis improves discrimination of Savanna tree species

Abstract

Differences in within-species phenology and structure driven by factors including topography, edaphic properties, and climatic variables across the landscape present important challenges to species differentiation with remote sensing. The objective of this paper was to evaluate the classification performance of a multiple-endmember spectral angle mapper (SAM) classification approach in discriminating seven common African savanna tree species and to compare the results with the traditional SAM classifier based on a single endmember per species or class. The leaf spectral reflectances of seven common tree species in the Kruger National Park, South Africa, Combretum apiculatum, Combretum hereroense, Combretum zeyheri, Gymnosporia buxifolia, Gymnosporia senegalensis, Lonchocarpus capassa and Terminalia sericea were used in this study. Discriminating species using all training spectra for each species as reference endmembers (the multiple endmember approach or more conventionally termed K-nearest neighbour classifier) yielded a higher classification accuracy of 60% compared to the conventional SAM classifier based on the mean of the training spectra for each species (overall accuracy = 44%). Further analysis using endmembers selected after cluster analysis of all the spectra for each species yielded the highest classification accuracy for the species (overall accuracy = 74%). This study underscores two important phenomena; (i) within-species spectral variability affects the discrimination of savanna tree species with the SAM classifier and (ii) the effect of within-species spectral variability can be minimised by adopting a multiple endmember approach with the SAM classifier. This study further highlights the importance of the quality of the reference endmember or spectral library.