Machine Learning–Based Reconstruction of Missing Streamflow Records in Data-Scarce River Basins

Abstract

Stable streamflow records are essential in water-resources planning, flood prediction, and effecting climatic change; when plentiful, however, in high-altitude and trans-boundary river basins because of rugged terrain, severe climatic conditions, and insufficient monitoring facilities. The paper fills in missing streamflow records in the Chitral River Basin, a snow- and glacier-fed basin of the Upper Indus Basin, by reconstructing missing discharge at the downstream Arandu station with paradoxical observations of the Chitral station. The statistical and machine-learn models that were used to analyze daily discharge data between 1981-2024 are Linear Regression, Artificial Neural Network (ANN), and Extreme Gradient Boosting (XG Boost). The coefficient of determination (R 2), root mean square error, mean squared error, and mean absolute error were used to assess how well a model works. Although the predictive power of all models was quite robust with R 2 exceeding 0. 95, ANN showed the most balanced and stable performance at both training and testing stages with the lowest prediction errors and best generalization. On these findings, the ANN model was used to recreate missing discharge at the Arandu station to generate a long-run series of streamflow between 1981 and 2024. The restored hydrograph maintains the seasonal snowmelt-based flow regime and fits well with observed records to actually serve as a virtual discharge sensor. The suggested framework is a feasible approach to hydrological reconstruction in mountainous basins with limited data, as well as the enhanced water-resources management in the context of variable climatic conditions