Regression Models for Dense Bathymetry in Turbid Coastal Areas

We trained and analyzed the behavior and robustness of two regression models, Random Forest and Support Vector Machine, with aerial hyperspectral images and echosounder measurements in an area of the Bahia Blanca estuary (Buenos Aires, Argentina) to generate a dense bathymetric map. This region of the estuary is characterized by high sediment transport, which makes its waters turbid, which makes bathymetric-optical estimates difficult. Images of 24 NIR and visible spectral bands acquired using a hyperspectral camera on board a UAV were used, together with 100 bathymetric data points surveyed with a sonar sensor on board a USV in an area of approximately 800 m2. The best model was Random Forest with a coefficient determination of 0.815 (for the test data), an RSME = 0.160 m, and an absolute mean error less than 1.3%. We performed ablation tests to evaluate the robustness of the models and using SHAP values we determined the bands with the highest incidence in the model. The results allow for dense and accurate reconstructions of the underwater profile in shallow and muddy regions of the Bahia Blanca estuary, showing the feasibility of merging hyperspectral images with sonar data in turbid shallow waters.