Process-Based and Machine Learning Assessment of SPI-Derived Sediment Transport and Deposition Dynamics in Reservoir Systems: Limits of Cross-System Transferability

Understanding sediment transport and deposition in dam reservoirs requires linking bathymetric evidence with physically interpretable terrain and connectivity indicators. This study develops a Stream Power Index (SPI)-based process-oriented machine learning framework using 58 bathymetric datasets from dam reservoirs. Reservoir elevations were normalized into standardized zones to enable cross-system comparison, and sediment-related volumetric loss was evaluated along these normalized reservoir profiles.Within this framework, positive SPI values are interpreted as transport- or stream-power-dominated areas, negative values as deposition-prone or low-transport-capacity zones, and near-zero values as alluvial or buffering domains. The physical consistency of this interpretation was evaluated using curvature-based terrain information as an independent geomorphometric indicator.Zone-wise explanatory modeling showed that the full process-based configuration consistently outperformed simpler models across all reservoir zones, indicating that sediment-related volume loss is best explained when reservoir storage, channel slope, drainage density, channel length, SPI-derived transport–deposition structure, sediment supply, and hydroclimatic context are considered together. Random Forest analysis demonstrated strong internal predictive performance under random train–test validation (R² ≈ 0.84), but performance declined under dam-grouped validation (R² ≈ 0.51), indicating that sediment-process relationships are only partially transferable across independent reservoir systems.Feature-importance results revealed that geomorphic and connectivity-related predictors dominate model behavior, particularly maximum main-channel slope, main-channel length, drainage density, reservoir capacity, catchment scale, SPI-derived transport–deposition metrics, channel dominance, and sediment supply, including contributions from near-zero SPI buffering areas. In contrast, Köppen–Geiger climate class contributed only marginally to model performance.These results demonstrate that SPI-derived transport, deposition, and alluvial-buffering indicators provide a physically interpretable framework for analyzing sediment-related volume loss in normalized bathymetric datasets, while predictive relationships remain conditioned by system-specific geomorphic connectivity, storage configuration, and hydroclimatic setting, limiting their transferability across reservoirs.