Interpretable monthly streamflow estimation for seasonal water-resource management in highland watershed

Monthly streamflow information supports seasonal water-resource planning in tropical highland basins, where dry-season scarcity and wet-season runoff require different management responses. This study evaluated an integrated hydrological machine-learning and socio-economic water-pressure screening framework (Hydro-ML–SEWP) for the Da Tam watershed, Lam Dong Province, Vietnam. The hydrological models estimate monthly discharge conditionally on target-month rainfall and temperature together with antecedent streamflow, lagged rainfall, rolling hydroclimatic indicators, and seasonal cyclic terms; they are therefore within-month estimators rather than strict lead-time forecasts. A seasonal-naive benchmark, Ridge regression, support vector regression with a radial basis function kernel (SVR-RBF), Random Forest, and Gradient Boosting were assessed using a chronological 2019–2024 hold-out period and rolling-origin evaluation for 2011–2024. Ridge regression gave the highest R 2 and lowest RMSE in both evaluations (hold-out R 2 = 0.863, RMSE = 0.904 m 3 s −1 ; rolling-origin R 2 = 0.831, RMSE = 1.047 m 3 s −1 ), whereas Random Forest gave the lowest MAE and the highest normalized composite score. Target-month rainfall, seasonal cyclicity, 1-month lagged discharge, and 1-month lagged rainfall represented 77.1% of the Random Forest impurity-based importance, although correlated predictors and the use of concurrent rainfall limit causal interpretation. Screening of the 1995–2024 drainage-area-transferred streamflow series identified February–March and September–October as the principal low-flow and high-flow periods. Because the source dataset does not identify the reference gauge and the available socio-economic records are not temporally aligned across all four proposed SEWP components, absolute discharge magnitudes and annual SEWP values are not treated as fully verified basin-specific quantities. The framework is therefore presented as a transparent conditional-estimation and scenario-screening approach whose operational use requires documented reference-basin metadata, target-month meteorological forecasts or lag-only predictors, and watershed-scale socio-economic data.