<p>Summary of datasets used.</p>

The VegET–mizuRoute model reproduced seasonal flow dynamics but showed station-dependent biases, with NSE from −2.28 to 0.23, RMB from −0.59 to 0.77, and CC from 0.46 to 0.71. Following bias correction, Random Forest yielded substantial improvements for both precipitation and discharge, underscoring the capacity of machine learning to address complex, nonlinear biases. It raised precipitation NSE to a consistently positive 0.85–0.95 across all eight stations while halving RMSE at most sites (9.05–46.23 mm).

For discharge, Random Forest improved NSE at most stations to 0.80–0.88, with CC rising to 0.92–0.95, except at one station. Our findings confirm that systematic bias remains a dominant error source in global hydrological products in this region, and that machine learning-based bias correction is essential for the operational reliability and accuracy of these datasets for water management and flood applications.</p></div>