WASI-AI: Synergistic Integration of AI and Physics for Retrieving Water Quality and Benthic Parameters from Multi- and Hyperspectral Images

This is the authors' abstract. We don't add key points for this paper.

Artificial intelligence (AI) has spurred significant progress in the remote sensing of water constituents through physical models that rely on a predefined database of simulations, e.g., Case-2 Regional/Coast Colour (C2RCC) processor. However, these models are sensor-specific, not applicable in optically shallow waters, and incapable of adapting to different bio-optical conditions. This study introduces a novel approach that synergistically integrates AI with physics-based modeling. The developed method, termed WASI-AI, is implemented as a new module into the Water Colour Simulator (WASI) software. WASI-AI uses the physics-based WASI-2D module of WASI to retrieve the unknown biophysical parameters for a small subset of image pixels selected at random. A portion of the inverted samples is used to train neural networks (NNs). The trained NNs are then applied to predict the unknown biophysical parameters for all water pixels of the image. The remaining portion of the samples is used to assess the agreement between WASI-AI and WASI-2D retrievals. WASI-AI maintains the advantages of WASI-2D regarding sensor independence and flexibility in bio-optical adaptation. The correlation plots of WASI-AI vs. WASI-2D allow recognizing spectral ambiguity and optimizing inverse modeling parametrization. The integration of AI significantly speeds up the inversion, reducing the processing time of a single image from hours/days to mere minutes. We applied WASI-AI to hyperspectral (EnMAP, DESIS) and multispectral (Sentinel-2, Landsat-8/9, Planet SuperDove) imagery in optically deep and shallow waters. After handeling spectral ambiguities, the results indicate a strong correspondence between WASI-AI and WASI-2D inversions, with WASI-AI exhibiting lower noises on the maps.