Streamflow data is essential for water resource management, especially in transboundary river basins where data sharing between countries is often limited. Simulating and forecasting streamflow in such basins, particularly those with large upstream reservoir systems, presents significant challenges. This study introduces a novel machine learning (ML) approach to reconstruct streamflow data at intermittent gauging stations in transboundary rivers, using streamflow and water level data from neighboring stations to enhance model performance. This approach contrasts with traditional methods that mainly rely on forcing data. We applied six ML models to the Da River basin in Northern Vietnam, where all models achieved high accuracy, with Nash-Sutcliffe Efficiency and Kling-Gupta Efficiency exceeding 0.9. The LGBM (light gradient boosting machine regressor) performed best overall. We found that combining multiple ML models improved simulation accuracy, and some models performed reliably without precipitation data, highlighting the importance of nearby stream gauge data. Furthermore, the ML models outperformed a process-based distributed model (Variable Infiltration Capacity) in general metrics and hydrological signature evaluations, especially in simulating baseflow, low flow, and high flow conditions. ML also demonstrated faster computational efficiency and required less data for configuration. This research emphasizes the need for tailored approaches and data selection in complex transboundary river systems, offering a promising solution for effective water resource management in regions with limited cross-border data sharing and contributing to more accurate, adaptable hydrological forecasting.

https://doi.org/10.1007/s11269-025-04109-6