Remote tree species identification in a diverse tropical forest using airborne imaging spectroscopy

Plant species identification and mapping based on remotely-sensed spectral signatures is a challenging task with the potential to contribute enormously to ecological studies. This task is especially difficult in highly diverse ecosystems such as tropical forests, and for these ecosystems it may be more strategic to direct efforts to identifying crowns of a focal species. We used imaging spectrometer data collected by the Carnegie Airborne Observatory over Barro Colorado Island, Panama, to develop classification models for the identification of tree crowns belonging to selected focal species. We explored alternative methods for detecting crowns of focal species, which included binary, one-class, and biased support vector machines (SVM). Best performance was given by binary and biased SVM, with poor performance observed for one-class SVM. Binary and biased SVM were able to identify crowns of focal species with classification sensitivity and specificity of 87-91% and 89-94%, respectively. The main tradeoff between binary and biased SVM is that construction of binary SVM requires a far greater amount of training data while biased SVM is more difficult to parameterize. Our results show that with sufficient training data, focal species can be mapped with a high degree of accuracy, in terms of both sensitivity and specificity, in this diverse tropical forest.