Résumé
The rapid loss of biodiversity in tropical rainforests calls for new remote sensing approaches capable of providing rapid estimates of biodiversity over large areas. Imaging spectroscopy has shown potential for the estimation of taxonomic diversity, but the link with spectral diversity has not been investigated extensively with experimental data so far. We explored the relationship between taxonomic diversity and visible to near infrared spectral variance derived from various spectral processing techniques by means of a labeled dataset comprising 2000 individual tree crowns from 200 species from an experimental tropical forest station in French Guiana. We generated a set of artificially assembled communities covering a broad range of taxonomic diversity from this experimental dataset. We analyzed the impact of various processing steps: spectral normalization, spectral transformation through principal component analysis, and feature selection. Correlation between taxonomic diversity and inter-specific spectral variance was strong. Correlation was lower with total spectral variance, with or without normalization and transformation. Dimensionality reduction through feature selection resulted in dramatic improvement of the correlation between Shannon index and spectral variance. While airborne diversity mapping of tropical forest may not be at hand yet, our results confirm that spectral diversity metrics, when computed on properly preprocessed and selected spectral information can predict taxonomic diversity in tropical ecosystems.