Hyperspectral discrimination of tropical rain forest tree species at leaf to crown scales

Abstract

We investigated the utility of high spectral and spatial resolution imagery for the automated species-level classification of individual tree crowns (ITCs) in a tropical rain forest (TRF). Laboratory spectrometer and airborne reflectance spectra (161 bands, 437–2434 nm) were acquired from seven species of emergent trees. Analyses focused on leaf-, pixel- and crown-scale spectra. We first described the spectral regions and factors that most influence spectral separability among species. Next, spectral-based species classification was performed using linear discriminant analysis (LDA), maximum likelihood (ML) and spectral angle mapper (SAM) classifiers applied to combinations of bands from a stepwise-selection procedure. Optimal regions of the spectrum for species discrimination varied with scale. However, near-infrared (700–1327 nm) bands were consistently important regions across all scales. Bands in the visible region (437–700 nm) and shortwave infrared (1994–2435 nm) were more important at pixel and crown scales. Overall classification accuracy decreased from leaf scales measured in the laboratory to pixel and crown scales measured from the airborne sensor. Leaf-scale classification using LDA and 40 bands had 100% overall accuracy. Pixel-scale spectra from sunlit regions of crowns were classified with 88% overall accuracy using a ML classifier and 60 bands. The highest crown-scale (ITC) accuracy was 92% with LDA and 30 bands. Producer's accuracies ranged from 70% to 100% and User's accuracies ranged from 81% to 100%. The SAM classifier performed poorly at all scales and spectral regions of analysis. ITCs were also classified using an object-based approach in which crown species labels were assigned according to the majority class of classified pixels within a crown. An overall accuracy of 86% was achieved with an object-based LDA classifier applied to 30 bands of data. Object-based and crown-scale ITC classifications were significantly more accurate with 10 narrow-bands relative to accuracies achieved with simulated multispectral, broadband data. We concluded that high spectral and spatial resolution imagery acquired over TRF canopy has substantial potential for automated ITC species discrimination.