Three-stage hybrid modeling for real-time streamflow prediction in data-scarce regions

Study Region: The Upper Blue Nile Basin, Ethiopia Study focus: This study addresses the challenge of utilizing satellite-based precipitation data in rainfall-runoff models for regions with limited ground observations. We propose a three-stage methodology incorporating Variational Mode Decomposition (VMD) into a conceptual data-driven framework (CHM-VMD-ML). The method was tested on four PERSIANN family precipitation products (2005–2019) using two conceptual hydrological models (CHM: HBV and GR6J) and three machine learning models (ML: Random Forest Regression, Boosted Regression Forest, and CatBoost Regression), with VMD applied to improve model inputs. New hydrological insights: Our results highlight that integrating VMD significantly enhances the reliability of hydrological simulations driven by satellite precipitation data, particularly during low-flow periods. This approach reduces biases in PERSIANN products and improves overall model performance, as evidenced by an increase in Nash–Sutcliffe Efficiency values from 0.22–0.87 in the initial stage (CHM) to 0.74–0.92 in the final stage (CHM-VMD-ML). These findings underscore the importance of signal decomposition for refining data-driven models, facilitating better hydrological prediction and decision-making in data-scarce regions.