Beyond spectral signals: Geographic features drive bathymetric accuracy in the turbid Sancha Lake using machine learning

Accurate bathymetric mapping in inland water bodies presents significant challenges for conventional optical remote sensing due to complex water quality conditions and variable bottom types. This study introduces a novel Spectral-Geospatial XGBoost Regression (SG-XGBoost) model that revolutionizes depth estimation by integrating comprehensive spectral transformations with explicit geographic coordinates through gradient boosting methodology. Applied to Sancha Lake, a morphologically complex reservoir in China's Upper Yangtze watershed, the model achieved exceptional performance with R 2 =0.91 and RMSE=1.66m, representing 70% improvement over traditional empirical methods (Stumpf, Log-Linear) and 21% advancement beyond Random Forest. The iterative error correction and sophisticated regularization of the gradient boosting methodology not only enable the effective exploitation of spatial-spectral interactions but also ensure better accuracy is maintained across all depth ranges (2-31m). The feature importance analysis revealed an unexpected finding, the geographic coordinates dominated predictive power (85% contribution), while spectral features contributed minimally, challenging fundamental assumptions about optical bathymetry. The iterative error correction and sophisticated regularization of the gradient boosting methodology not only enable the effective exploitation of spatial-spectral interactions but also ensure better accuracy is maintained across all depth ranges (2-31m). Bathymetric maps generated by SG-XGBoost successfully captured fine-scale morphological features invisible to conventional approaches, including channels <30m wide and subtle depth variations of 1-2m. Despite limitations in extreme turbidity and site-specificity requiring readjustment for new water bodies, this research establishes gradient boosting with spatial-spectral integration as a transformative approach for inland water bathymetry, with broader implications for aquatic remote sensing applications including water quality monitoring and habitat mapping. • Geographic coordinates contribute 85% predictive power vs. 15% from spectral bands • SG-XGBoost achieves R 2 =0.91, RMSE=1.66m, outperforming traditional methods by 70% • Gradient boosting captures spatial-spectral interactions in turbid inland waters • Longitude encodes stable valley morphology; latitude reflects dynamic water quality • Model maintains accuracy across 2-31m depths via hybrid optical-spatial approach