Improvement of the river discharge reanalysis dataset for event-based flood estimation using HEC-HMS

Global-scale reanalysis products have created new opportunities for streamflow estimation in ungauged basins, yet hydrological model outputs still suffer from significant uncertainties due to systematic biases. In this study, an XGBoost machine learning model under three input scenarios has been used to remove bias in the GloFAS discharge data and to explore the suitability of the same as a benchmark dataset for the calibration and validation of event-based HEC-HMS runoff simulations in the Shilabati River basin in eastern India. Results indicate that neither GloFAS discharge (Scenario 1) nor engineered rainfall features (Scenario 2) as input features can capture the complex nonlinearity in observed discharge. Incorporating time series of lag and rolling-window statistics for rainfall and GloFAS discharge as model inputs (Scenario 3) substantially improves XGBoost’s predictive performance by capturing richer hydrological memory and flow dynamics. The accuracy statistical metrics RMSE, R2, KGE and PBIAS have improved from 75.60 m3/s, 0.49, 0.56 and 17.10% for Scenario 1 to 11.38 m3/s, 0.99, 0.98 and 0.44% for Scenario 3. The well-improved GloFAS discharge from Scenario 3 also performs well as a calibration and validation benchmark in the HEC-HMS model (KGE ~0.65–0.95), indicating its reliability for rainfall–runoff modelling.