Résumé
Biodiversity conservation is challenged by multiple factors related to global change. Remote sensing is a crucial source of information towards operational biodiversity monitoring systems. Optical sensors are widely used tostudy vegetation, as canopy reflectance results from interactions between incoming light and key vegetation properties. High spatial resolution imaging spectroscopy has shawn strong potential to estimate plant diversity in various ecosystem types. Spectral diversity metrics are often used as an indicator of various dimensions of plant diversity, including taxonomic, structural and chemical diversity. This hypothesis is scale dependent and avariety of spectral diversity metrics can be derived from remote sensing data, with potentially very different capacity to relate to biodiversityOur objective was to identify relevant spectral information to be used for the computation of spectral variance, and for estimating taxonomic diversity from tropical forested ecosystems using high spatial resolution imaging spectroscopy. The link between spectral diversity and taxonomic diversity is usually limited by ground information, which is mainly obtained from a limited number of inventoried field plots. Here, we took advantageof an experimental dataset including spectral information extracted from visible to near infrared imaging spectroscopy acquired over an experimental tropical forest station in French Guiana, encompassing about two thousand individual tree crowns from two hundred species. Each individual tree was carefully delineated based on a combination of very high spatial resolution imagery, airborne LiDAR, plot inventories, and ground validation.We explored the relationship between spectral variance and taxonomic diversity expressed as Shannon index by generating a set of artificially assembled communities covering a broad range of taxonomic diversity. Eachindividual community included one thousand pixels extracted from one hundred tree crowns, selected from two to one hundred species. We analyzed the correlation between Shannon index and the variance computed fromspectral information following various preprocessing steps, including spectral normalization, spectral transformation through principal component analysis (PCA), and feature selection. The feature selection wasapplied on reflectance, normalized reflectance, and PCA-transformed normalized reflectance. We analyzed total spectral variance, inter/intra specific and inter/intra crown components.The correlation between total variance of reflectance and Shannon index was weak, while reflectance normalization resulted in substantial increase in correlation. This evidenced the influence of multiple factors extrinsic of species and species traits on spectral variance, such as illumination effects, which were partly removed from the signal after normalization. The application of feature selection resulted in dramatic improvement of the correlation between Shannon index and spectral variance for all types of reflectance. The spectral variance computed from normalized and PCA-transformed normalized reflectance showed strong correlation with Shannon index, while the correlation obtained from raw reflectance showed high variability, ranging from poor to moderate.Our results evidence the relationship between spectral variance and taxonomic diversity but highlight i) lack of robustness of the spectral diversity metrics computed from unprocessed reflectance, and ii) strong potential ofproperly preprocessed and selected spectral information acquired at metric spatial resolution to predict taxonomic diversity in tropical ecosystems. The influence of spatial resolution and spectral sampling is now investigated in order to assess the applicability of these results to decametric resolution multispectral satellites (Sentinel-2) and future satellite missions including spaceborne imaging spectroscopy sensors (CHIME, SBG).