The rapid and accurate acquisition of nearshore bathymetry is crucial for coastal economic development, safe navigation, and marine ecological protection. This study focuses on the shallow coastal waters of Beihai and Fangchenggang in the Guangxi Beibu Gulf, China. Utilizing Landsat-9 multispectral imagery and Electronic Nautical Chart (ENC) data, three empirical Satellite-Derived Bathymetry (SDB) algorithms—single-band linear regression, dual-band ratio, and multi-band combined regression—were formulated and comparatively analyzed. Furthermore, a depth-stratified inversion approach was introduced to evaluate its impact on retrieval accuracy across different depth zones. Experimental results demonstrated that the multi-band combined regression model yielded the highest inversion accuracy in both study areas. For the unzoned global inversion, the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) were 1.38 m and 1.76 m in Beihai, and 1.86 m and 2.46 m in Fangchenggang, respectively. Notably, following the implementation of the depth-stratified inversion strategy, the weighted average errors were substantially reduced. In the Beihai region, the MAE and RMSE decreased by 0.64 m and 0.80 m, respectively, while in Fangchenggang, they witnessed significant reductions of 1.68 m and 1.92 m. The findings confirm that the depth-stratified multi-band combined regression model provides superior performance and robustness for nearshore bathymetric mapping.