Estimating taxonomic diversity using spectral variance of imaging spectroscopy data collected over a tropical rainforest

Tropical forests are the largest reservoir of terrestrial biodiversity. Today, this biodiversity is rapidly eroding due to climate change, land use change and human pressure. Management and monitoring of tropical forests are difficult and costly in terms of financial and human resources. Forest inventories are generally conducted at a limited spatial scale and our knowledge of floristic composition and species distribution in tropical forests is still incomplete. Remote sensing data are a promising tool for the development of biodiversity monitoring systems and the development of hyperspectral sensors has greatly contributed to advances in remote sensing vegetation analysis. Spectral diversity, here considered as the variation in space of the spectral information, can be calculated as the total variance of the reflectance table. The structure of spectral variance itself and its relationship to the compositional structure of tropical forest communities has not been thoroughly studied yet, mainly due to lack of sufficient field data. However, in an operational framework of biodiversity estimation without prior identification of species, it is essential to address this issue in order to understand precisely what the spectral signal is able to measure of taxonomic diversity, but also to establish the potential and limitations of current space hyperspectral imaging sensors (EnMAP, DESIS, PRISMA), and the instrumental needs in terms of spectral information and spatial resolution for future satellite missions (SBG, CHIME, HYSP).
Thus, our objective is to explore the relationship between taxonomic and spectral diversity derived from airborne hyperspectral imaging acquisitions. Specifically, we want to assess the strength and sensitivity of the relationship between taxonomic and alpha and beta spectral diversity at the plot scale of our study site: is it possible, on a highly diverse landscape, to measure subtle compositional differences using hyperspectral imagery?
We used georeferenced forest inventory data Abstract 165 collected at the Paracou experimental rainforest station in French Guiana, complemented by airborne acquisitions including very high spatial resolution hyperspectral imagery from visible to near infrared, high spatial resolution orthophotos and LiDAR data. We analysed the influence of different data preprocessing and evaluated the limitations of the spectral variance approach.
We finally applied spectral variance partitioning to Paracou plots to evaluate the ability of the method to estimate alpha and beta diversity of canopy species in an operational context. This thesis work confirms the potential of hyperspectral remote sensing for vegetation analysis, but also highlights the fact that the ability of these data to estimate biodiversity directly at a global scale should not be overestimated. While remote sensing data are a powerful tool for monitoring vegetation, over less contrasting and biologically diverse landscapes, spectral diversity is not a reliable indicator of local biodiversity. Large-scale mapping of biodiversity on different ecosystem types using spectral data is not yet within reach: the methods used have proven to be not robust enough and, above all, not generalizable enough to succeed without abundant, high-quality floristic surveys to calibrate the estimation models.

Composition du jury :
Maria J. SANTOS, Professor, University of Zurich (rapportrice)
Fabian FASSNACHT, Professor, Frei Universität Berlin (rapporteur)
Anna K. SCHWEIGER, Researcher, University of Zurich (examinatrice)
Pierre COUTERON, Directeur de recherche, UMR AMAP (examinateur)