Improving regional low-flow modeling in ungauged French catchments using the SMASH platform: A comparative assessment of grid-based parsimonious GR hydrological models and ANN-based regionalization strategies

Accurate streamflow prediction in ungauged basins remains a major challenge in hydrology. SMASH is a regionalizable and differentiable modeling framework that simulates hydrological responses and states at any spatial location within a basin, at relatively high spatial and temporal resolutions. The model includes gridded operators based on a parsimonious GR-like model, with spatially distributed conceptual parameters being regionalized using physical descriptors through regionalization mappings, such as artificial neural networks (ANN). Analyses across 248 French catchments with three models indicate that parsimonious models, when coupled with regionalization ANN, exhibit enhanced robustness, particularly for low-flow indices, highlighting their potential for regional hydrological prediction.