A depth-dependent machine learning framework for satellite-derived bathymetry of morphologically complex reservoirs

Satellite-derived bathymetry (SDB) in morphologically complex reservoirs is challenged by depth-dependent optical variability, where shallow bottom-controlled and deep attenuation-dominated regimes coexist. To address this issue, we developed a two-regime framework that partitions the depth domain at an empirically optimized threshold and models regime-specific optical–bathymetric relationships. Using Sentinel-2 images acquired under different optical conditions, we identified the optimal regime boundary at 6 m in Uiam Reservoir, South Korea. Spectral controls differed clearly between regimes: blue-green band ratios dominated in shallow water, whereas deeper water was governed by longer-wavelength ratios and attenuation-sensitive bands. Compared with a single-model approach, the two-regime Random Forest (TR-RF) achieved higher accuracy and lower sampling-induced uncertainty, with greater advantages under sparse transect sampling. Even with limited training data, TR-RF better preserved baseline accuracy and key morphological features, demonstrating that depth-aware regime partitioning can improve operational SDB while reducing in-situ sampling demands.