Comparative machine-learning retrieval of high-precision bathymetry from multispectral remote sensing: a case study of Manzala Lagoon, Egypt.

Manzala Lagoon, the largest coastal lagoon in the Nile Delta, has undergone extensive restoration and dredging in recent years (2017-2022), creating an urgent need for accurate and cost-effective bathymetric mapping. Conventional echo-sounder surveys remain reliable but are difficult to implement over such a large, shallow, and vegetated water body. This study developed an interpretable machine-learning framework for satellite-derived bathymetry using Landsat 8 multispectral imagery calibrated with extensive echo-sounder measurements. Seven machine-learning algorithms, comprising tree-based ensembles and boosting-based regressors, were evaluated using regression metrics, complementary depth-class performance metrics, uncertainty estimation, and SHAP-based model interpretation. Tree-based ensembles consistently outperformed boosting-based regressors, with Decision Tree and Extra Trees achieving the highest test-set accuracy under the adopted validation design (r = 0.996, R2 = 0.992, RMSE = 0.073 m). Although both models produced similar accuracy, Extra Trees showed substantially lower prediction uncertainty (≈0.09 m), making it the most reliable model for operational mapping. SHAP analysis identified the near-infrared band as the most influential predictor, likely reflecting surface and turbidity-related effects rather than direct depth sensing, while logarithmic band ratios improved model robustness by capturing nonlinear spectral relationships. The resulting bathymetric maps resolved depth gradients from ≈ -0.6 to -4.0 m and provide an efficient alternative to extensive field surveys. The proposed framework offers a transparent, scalable approach for bathymetric mapping that supports hydrodynamic modeling, sediment transport studies, and sustainable management of shallow coastal lagoons.