Hyperspectral remote sensing of coral reefs by semi-analytical model inversion – Comparison of different inversion setups

Abstract Due to its high spectral and spatial resolutions, airborne hyperspectral imaging has great potential for becoming a powerful large-scale monitoring tool for coral reef communities. In recent years, methods based on radiative transfer model inversion have shown promising results for extracting information about seabed type, bottom depth and water constituents from hyperspectral imagery. However, low signal-to-noise ratios (SNR) due to low water-leaving radiance combined with environmental variability make it very difficult to design an optimal processing algorithm. Here, we selected a state-of-the-art, forward semi-analytical model in which we included a mixing model of four seabed albedo, namely sand, corals, algae and seagrass. The purpose of this paper was then to compare different setups of the inversion scheme, each one having its own theoretical strengths and weaknesses regarding the different confounding factors. Six inversion setups were implemented, corresponding to the combinations between ( i ) three cost functions: least square (LS), spectral angle mapper (SAM) and least square on spectral derivative (LSD), and ( ii ) two physical constraints imposed on the seabed type retrieval: abundance sum-to-one constraint (ASC) and a relaxed version (RASC). Performances of bathymetry and seabed type retrieval were evaluated on hyperspectral data acquired in a coral reef environment in Reunion Island. Our results showed that the accuracy and robustness of the bathymetric estimation were greatly influenced by the choice of the inversion setup. RASC-LSD produced the overall best performances even if SAM-based inversion setups showed particularly low error dispersion with respect to Lidar derived bathymetry. RASC-LSD also produced the most accurate results in terms of spatial coverage of benthic components on a very shallow area (inner part of a fringing reef). The results on our study areas clearly highlighted the interest of relaxing the ASC when the bottom depth is shallow. In deeper areas, ASC versions of LS- and LSD-based inversion setups produced the best seabed type mapping results. Only broad seabed types could be retrieved in areas deeper than 10 m.