Regularization of discriminant analysis for the study of biodiversity in humid tropical forests

The performance of two supervised classifiers, linear and regularized discriminant analysis (LDA and RDA), is compared here for canopy species discrimination in humid tropical forest, based on airborne hyperspectral imagery acquired with the sensor Carnegie Airborne Observatory Alpha System (CAO-Alpha). Classification is performed to identify 13 species at pixel scale, crown scale, and using an object-based approach. The results show that for each scale of study, 70% to 75% overall accuracy is obtained with LDA. RDA allows improved classification for more than half species, and 5% increase of overall accuracy compared to LDA. The extended spectral range of the forthcoming CAO AToMS system (380-2500 nm) will allow for even more accurate classifications of tropical canopy species.