Identification of spectral features in the longwave infrared (LWIR) spectra of leaves for the discrimination of tropical dry forest tree species

With the emergence of longwave hyperspectral imaging systems, studies are revealing the potential of these data for discriminating tree species. However, few studies have applied statistical methods of band selection to select and characterize features at the species level that can then be used for improved classification. A dataset of leaf spectra was recently collected in-situ from twenty-six tree species in a Costa Rican tropical dry forest. The spectra of the species present overall low contrast and a range in spectral shapes, with some species displaying spectral similarity. This motivates our study to explore the performance of band selection tools to help identify key spectral features for the classification of these species. The bands selected using an ensemble of multiple methods improved the Logistic Regression classification performance by 3% in comparison to a result without band selection. The multiple methods encompassed the random forest, minimum redundancy maximum relevance and n-dimensional spectral solid angle methods. Bands selected by the ensemble methods agree well with the features previously identified based on expert knowledge and can be understood in the context of leaf constitutional compounds and related spectral features. The longwave hyperspectral bands or features identified in this study can potentially assist the future image mapping of tree species at large scales. The ensemble strategy is recommended for the band analysis of vegetation for its highest accuracy and stability.