POSTER Session 3

AI-EMPOWERED PHYSICAL INVERSION OF WATER QUALITY AND BENTHIC PARAMETERS FROM MULTI- AND HYPERSPECTRAL IMAGES

This study introduces a novel methodology that seamlessly merges AI with physics-based modeling for retrieving water quality and benthic parameters from multi- and hyperspectral imagery. The developed method, termed WASI-AI, is integrated as a new module within the Water Colour Simulator (WASI) software. WASI-AI tackles the spectral ambiguity issue and expedites the inversion process. Furthermore, despite the existing AI-based models, WASI-AI is sensor-independent and adaptable to a wide range of bio-optical conditions in both optically shallow and deep waters. WASI-AI uses the physics-based WASI-2D module to retrieve the biophysical parameters for a small subset of image pixels. A portion of the inverted samples is then utilized to train neural networks, which predict the parameters of interest for all water pixels. The remaining portion of the samples is then used to assess the agreement between WASI-AI and WASI-2D retrievals through correlation plots. Without ambiguity problems, both methods produce similar results for validation samples due to WASI-AI being trained on the output of WASI-2D. However, in the presence of strong ambiguities, the results become less correlated. The correlation plots of WASI-AI vs. WASI-2D, using the R2 metric, prove effective in identifying the ambiguity issues and fine-tuning inversion parameters. We demonstrate the WASI-AI’s efficacy using various hyperspectral and multispectral imagery in diverse aquatic environments. After refining the inversion process for spectral ambiguities, results demonstrate a strong agreement between WASI-AI and WASI-2D inversions, with WASI-AI producing less noisy maps. Moreover, AI integration significantly speeds up the inversion from hours/days to mere minutes.

Milad Niroumand-Jadidi, University of Florida, USA, [email protected], https://orcid.org/0000-0002-9432-3032

Peter Gege, German Aerospace Center (DLR), Germany, [email protected]

Code of Conduct
Privacy Policy