Optimizing tree-species classification in hyperspectal images

Abstract

The authors investigate the classification of eight prominent savanna tree species, based on hyperspectral reflectance data. Although two principal components account for 95% of the variance of the data, up to 20 components are found to be useful for classification. Scaling of these components so that all features have equal variance is found to be useful, and their best performance (88.9% accurate classification) is achieved with 15 scaled features and a support vector machine as classifier. A graphical analysis suggests that several exemplars (“endmembers”) are required for each class, and this observation is confirmed by the large number of support vectors employed by the best classifier.