SINGLE-TREE SPECIES MAPPING USING ONE-CLASS CLASSIFICATION METHODS AND UAV HYPERSPECTRAL IMAGES

Progress in tree species mapping with hyperspectral data usually is limited by the multi-class classification framework, which imposes the requirement of exhaustively defining all species encountered in a landscape. As the research objective may be to map only one or a few species of interest, it is necessary to explore alternative classification methods that may be used to more efficiently detect a single species. In this study, we used UAV hyperspectral data to detect one endangered tree species, Araucaria angustifolia, in a subtropical forest area comparing the performance of two one-class classifiers (OCC): OCSVM and OCRF. Besides the 25 spectral bands (SB), we also tested two other datasets: one comprising the first five MNF components, and the other one comprising the first five PCA. Both algorithms and all the datasets reached good results, with F-score varying from 0.81 for OCRF and SB dataset, to 1 for OCSVM associated with the PCA dataset.