Mapping forest biodiversity from optical imagery: a plant trait-based method

Earth observation is a key component for the establishment of biodiversity monitoring systems. The increasing number of instruments acquiring information on Earth surface provides opportunities to assess and monitor various properties of vegetated ecosystems, including vegetation biochemical and biophysical traits and associated processes such as photosynthesis, growth and adaptation to environmental stress. Multiple approaches have been developed during the past decades to link forest taxonomic diversity with remotely sensed information, with varying degrees of success. Statistical metrics directly derived from the vegetation reflectance such as spectral variance have shown limitations for the estimation of taxonomic diversity. One reason is that factors intrinsic and extrinsic to vegetation influence reflectance and contribute to this variance. On the other hand, this reflectance can be converted into optically effective plant properties (optical traits) using statistical methods (e.g. spectral transformation, machine learning or spectral indices) or physical methods (e.g. physical model inversion) applied to optical imagery, with an objective to reduce the influence of extrinsic factors. The spatial heterogeneity of a set of optical traits may then be used as a relevant proxy for vegetation diversity. Statistical methods are computationally efficient, but lack generalization ability, while physical approaches show better potential for generalization ability, but show limitations when applied on complex systems. Moreover, the set of optical traits accessible from optical data varies with sensor characteristics: new imaging spectroscopy missions expand the range of variables for which quantitative assessment is possible compared to multispectral imagery.

We introduce a framework taking advantage of physical modelling to assess a set of vegetation traits then used to feed remotely sensed diversity mapping techniques in the context of forest ecosystems. This approach intends to convert the optical information on a physical basis, in terms of vegetation traits related to structural, compositional and functional properties prior to computing diversity metrics. Physical modeling contributes to minimizing the influence of factors extrinsic to vegetation on optical traits, as a way to improve the generalization ability of existing frameworks taking advantage of Earth observation through space and time. To illustrate it, we used the model PROSAIL to assess Leaf Area Index, leaf chlorophyll content, equivalent water thickness and leaf mass per area from imaging spectroscopy acquired over forested areas. The method implemented in the R package biodivMapR was then applied to compute various diversity metrics from these vegetation biophysical properties, including α- and β-diversity metrics usually obtained from species inventories in ecological applications. We illustrate this framework with data acquired over different sites and with various optical sensors, including airborne and spaceborne imaging spectroscopy, and discuss current limitations.