Improving annual streamflow estimates using Budyko parameterization driven by catchment attributes

The Budyko framework is widely used to infer long-term water-energy partitioning. Its original uniform parameters limited accuracy, while locally calibrated formulations are difficult to apply in ungauged basins. This study develops a machine-learning (ML)-based parameterization linking Budyko parameters to catchment attributes, enabling annual estimates of evapotranspiration (ET) and streamflow in gauged and ungauged basins. Using 671 CAMELS catchments, we optimized parameters for six Budyko-type equations with the Shuffled Complex Evolution (SCE) algorithm, then adjusted them to catchment attributes using a ML model. The resulting attribute-based parameterization improved evaporative-index estimates relative to both SCE-based calibration and a calibrated SAC-SMA model. Applied to test catchments, it achieved a mean KGE of 0.67 for annual streamflow, outperforming calibrated SAC-SMA (0.48) and site-specific Budyko predictions (0.61). Results highlight a scalable approach for annual water-budget estimation in data-limited regions.