Satellite derived bathymetric retrieval through tree based learning within the shallow hypersaline environment of Lake Urmia

Satellite-derived bathymetric (SDB) retrieval in shallow hypersaline waters is challenging because complex hydro-optical conditions and bottom heterogeneity can produce nonlinear spectral–depth relationships that limit conventional linear approaches. This study develops and systematically evaluates a comparative framework for SDB retrieval in Lake Urmia by integrating Landsat-8 Operational Land Imager (OLI) spectral predictors, the Modified Normalized Difference Water Index (MNDWI), Singular Value Decomposition (SVD)-based predictor transformation, and complementary regression architectures. Multiple Linear Regression (MLR), Decision Tree (DT), Random Forest (RF), and CatBoost regression were evaluated using in situ bathymetric observations under consistent predictor and holdout conditions. In the primary five-repetition assessment, CatBoost achieved the highest held-out performance, with \(\:{R}^{2}=0.928\) , RMSE = 9.11 cm, and CC = 0.959, followed by RF ( \(\:{R}^{2}=0.898\) , RMSE = 13.31 cm, CC = 0.947), DT ( \(\:{R}^{2}=0.852\) , RMSE = 17.84 cm, CC = 0.926), and MLR ( \(\:{R}^{2}=0.799\) , RMSE = 27.97 cm, CC = 0.870). An extended 20-repetition analysis produced closely consistent results, with CatBoost achieving \(\:{R}^{2}=0.927\pm\:0.012\) , RMSE = \(\:9.2\pm\:0.7\) cm, and CC = \(\:0.958\pm\:0.007\) . SVD-based predictor transformation improved testing performance across the evaluated models, although the magnitude of improvement varied with model architecture. The four-model spatial assessment further revealed differences in spatial continuity and depth gradients, with the CatBoost and RF retrievals exhibiting comparatively more coherent spatial patterns than MLR and DT. These findings demonstrate that predictor representation and learning architecture jointly influence SDB retrieval under the investigated hypersaline conditions. The framework provides a controlled and reproducible basis for evaluating complementary SDB modeling strategies while maintaining the interpretation of quantitative performance within the environmental and observational conditions of the investigated Lake Urmia scene.