Abstract: Shallow coastal waters are critical zones for navigation, engineering, resource development, and environmental management. Acquiring accurate depth data in these areas economically and efficiently remains a pressing challenge. To improve the accuracy of multispectral remote sensing bathymetry and reduce human intervention in parameter tuning, this study proposes a CatBoost model optimized via Bayesian hyperparameter tuning (BO-CatBoost). Bathymetric inversion experiments were conducted in two representative nearshore environments: Nanshan Port in Sanya and Dazhou Island in Wanning. The method establishes a hyperparameter search space including learning rate, maximum tree depth, number of iterations, and L2 regularization coefficient, and applies a multi-objective Bayesian optimization strategy to optimize model fit and error metrics simultaneously. Using four performance metrics—the coefficient of determination (R²), mean absolute error (MAE), root mean square error (RMSE), and mean relative error (MRE) —a dynamically weighted objective function based on inverse weights was constructed to achieve an adaptive trade-off among multiple indicators. Results show that the proposed approach efficiently determines optimal hyperparameters under limited computational resources, significantly improving bathymetric accuracy and robustness in turbid nearshore waters. In Nanshan Port, the BO-CatBoost model achieved R², MAE, RMSE, and MRE values of 0.98, 0.32 m, 0.68 m, and 4.51%, respectively. In Dazhou Island, the corresponding values were 0.97, 0.46 m, 0.85 m, and 5.55%. The results outperform those of commonly used models such as Stumpf, random forest, and unoptimized CatBoost, maintaining stable performance across different depth ranges. This work provides a high-precision, low-intervention technical solution for shallow-water bathymetry, with practical value for large-scale coastal mapping and engineering applications.