Hybrid WRF‑hydro and random forest framework for improved flood reconstruction in data‑sparse river basins: A case study of the transboundary Nu‑Salween Basin

Study region The Nu-Salween Basin, a transboundary, monsoon-driven river system in Asia characterized by complex topography and sparse hydrological observations. Study focus This study aims to assess the efficacy of a hybrid framework that integrates the physically based, distributed WRF-Hydro model with a Random Forest (RF) machine learning post-processor to dynamically correct systematic simulation biases for flood reconstruction in data-sparse regions. The research methodology employs a multi-source forcing approach (ERA5-Land, CHIRPS, GPCP) and evaluates model performance against a global runoff reanalysis (G‑RUN ENSEMBLE) as a consistent reference benchmark, given the absence of long-term in-situ discharge observations. Spatial flood patterns were assessed using a hydrogeomorphologically constrained probabilistic benchmark during core monsoon seasons (June–August) from 1993 to 2023. Model validation was rigorously conducted through relative comparison against the G‑RUN ENSEMBLE benchmark. The results demonstrate the hybrid model’s robust performance, characterized by a 32% reduction in RMSE relative to the standalone WRF‑Hydro baseline and an increase in Nash-Sutcliffe Efficiency from 0.61 to 0.78. Comparative assessments showed strong improvements in spatial agreement of flood extent (from 76.2% to 92.8%) and decadal flood frequency bias (reduced to less than +0.02). Diagnostic analysis revealed that lagged simulated discharge and root-zone soil moisture were the most important predictors of residual error, stressing the role of hydrological memory in bias correction. New hydrological insight This study addresses a key gap in flood reconstruction for poorly monitored, topographically complex transboundary basins through a benchmark application in the Nu-Salween system. Results indicated significant sensitivity to hydrological memory effects, with lagged discharge and soil moisture dominating the bias correction process. Notably, the hybrid physical-statistical framework consistently outperformed the standalone physics-based model across all metrics, achieving a 92.8% spatial agreement in flood extent. Ultimately, this study stresses the critical role of integrating machine learning post-processors with physically based hydrological models, offering a scalable and cost-effective pathway for enhancing flood risk assessment and climate adaptation strategies in data-sparse regions where traditional methods fail.