Streamflow forecasting using machine learning and remote sensing data in the Himalayan region

ABSTRACT This study aimed to predict daily runoff using three machine learning models: artificial neural networks (ANNs), random forests (RFs), and extreme gradient boosting (XGBoost). Thirty-five years of discharge records and remote-sensing climate data were used to develop and validate the models, demonstrating the potential of satellite data for machine-learning-based runoff prediction, especially in data-poor regions like Nepal. The study identified 1-day lagged discharge (lag_1), average temperature, and maximum temperature as the three most important input variables. The developed ANN model demonstrates high accuracy in streamflow prediction, achieving an R2 of 0.928, a root-mean-square error (RMSE) of 435.43 m3/s, a Nash–Sutcliffe efficiency (NSE) of 0.93, and a Kling–Gupta efficiency (KGE) of 0.944 for the Narayani River Basin. When applied to the Trishuli River, a major tributary of the Narayani, the model demonstrated strong performance, with an R2 of 0.895, an RMSE of 425.92 m3/s, an NSE of 0.89, and a KGE of 0.908. Nonetheless, RF outperformed ANN in capturing low-flow conditions in the Narayani River. Overall, the models provide valuable tools for retrospective streamflow prediction, gap-filling in hydrological records, and supporting effective and sustainable water management.