Enhancing hydrological modeling of ungauged watersheds through machine learning and physical similarity-based regionalization of calibration parameters
Arun Bawa, Katie Mendoza, Raghavan Srinivasan, Fearghal O'Donchha, Deron Smith, Kurt Wolfe, Rajbir Parmar, John M. Johnston, Joel Corona
Abstract
This study enhances hydrological modeling in ungauged watersheds by employing physical similarity and machine learning-based clustering for regionalizing the Soil and Water Assessment Tool (SWAT) model parameters at the HUC12 (hydrological unit code) watershed scale within a HUC02 basin. Eleven features, including environmental, topographical, soil, and hydrological properties, were utilized to identify physical similarities for watershed clustering. Machine learning techniques, including random forest and hierarchical clustering, were employed to transfer calibrated parameters from gauged to ungauged watersheds. Validation of parameter transfer over gauged SWAT model projects showed that 88% of the projects achieved calibrated status (KGE ≥0.5; PBIAS ≤25%). Additional validation using MODIS satellite evapotranspiration measurements confirmed the robustness of the approach. Results indicated that the proposed approach successfully captures physical similarities, and effectively captures flow patterns. Overall, the study highlights the potential of physical similarity-based clustering and machine learning techniques for improving hydrological modeling in ungauged watersheds. • Watershed clustering by physical similarity aids in SWAT parameter regionalization. • Parameter regionalization validated across eight HUC02 regions in the US. • Parameter regionalization achieved ∼88% success in hydrological validation. • Hydrological calibration parameters were developed at HUC12 scale for eight HUC02s.