Analyzing the spectral variability of tropical tree species using hyperspectral feature selection and leaf optical modeling

Abstract Hyperspectral remote sensing can provide information about species richness over large areas and may be useful for species discrimination in tropical environments. Here, we analyze the main sources of variability in leaf spectral signatures of tropical trees and examine the potential of spectroscopic reflectance measurements (450 to 2450 nm) for tree species discrimination. We assess within- and among-species spectral variability and perform a feature selection procedure to identify wavebands in which the species most differ from each other. We assess the discriminative power of these wavebands by calculating a separability index and then classifying the species. Finally, leaf chemical and structural parameters of each species are retrieved by inversion of the leaf optical model PROSPECT-5. Among-species spectral variability is almost five times greater than within-species spectral variability. The feature selection procedure reveals that wavebands, where species most differ, are located at the visible, red edge, and shortwave infrared regions. Classification of the species using these wavebands reaches 96% overall accuracy. Leaf chemical and structural properties retrieve by model inversion show that differences in leaf pigment concentrations and leaf internal structure are the most important parameters controlling the spectral variability of species. These parameters also contribute to the variation in red edge position among species.