Enhanced early warning of extreme river discharge events in the Yangtze River Basin using atmospheric circulation signals

Abstract

Early warning of extreme river discharge events (ERDEs) is crucial for water resources management and disaster reduction in the Yangtze River Basin (YRB), yet substantial uncertainty remains in their prediction. This study proposes a novel strategy to enhance ERDE early warning by detecting its associated atmospheric circulation signals. We develop DetRF, a machine learning model that integrates anomaly detection with a balanced random forest ensemble. DetRF was trained and validated using ERA5 reanalysis data from 2000 to 2019 and independently applied to 2020–2024, with its early warning performance evaluated against satellite-based precipitation and reanalysis river discharge data. Across eight subregions of the YRB, DetRF effectively detects ERDE-related atmospheric circulation signals, achieving mean recall, precision, and F1 score values of 0.756, 0.506, and 0.576, respectively, and consistently outperforming benchmark machine learning models. The detected signals accurately capture key synoptic features associated with hydrological extremes, including the positions and intensities of troughs, vortices, and cyclones, as well as moisture transport. More than 99% of the detected signals are followed by precipitation within five days, while persistent precipitation occurs in over 67% of cases across both local and upstream regions. More importantly, an average of 93.8% of DetRF-detected circulation signals are followed by at least one ERDE within the subsequent five days. These findings establish a direct link between specific atmospheric circulation signals and subsequent hydrological extremes, offering a new pathway for extending ERDE early warning beyond conventional hydrological prediction. The proposed framework is particularly promising for regions where hydrological observations are sparse.