Advancing Subseasonal Extreme Rainfall Forecasting in Vietnam Using Machine Learning

抄録

Subseasonal forecasting for extreme precipitation represents a critical yet challenging frontier in weather prediction, particularly in regions like Vietnam, where monsoons, tropical cyclones, and diverse topography complicate the precipitation patterns. This study explores the integration of machine learning techniques—Random Forest (RF) and Extreme Gradient Boosting (XGB)—into model output statistics to enhance subseasonal extreme rainfall forecasts across Vietnam's seven climatic regions. ECMWF S2S hindcast data for the Madden-Julian Oscillation, monsoon indices, and soil moisture are used to predict rainfall extremes. The models are trained over 2001-2014 and evaluated over 2015-2023 against observational data. Evaluation metrics, including probability of detection, false alarm ratio, critical success index, and Brier skill score, highlight the superior performance of RF and XGB over raw ECMWF forecasts, particularly in North West, North East, Red River Plain, Central North and Central South regions. However, challenges remain in the Central Highland and South regions, where both deterministic and probabilistic skills are weaker. Overall, this study underscores the potential of machine learning to address regional and temporal variability in extreme rainfall prediction, offering a transformative tool for disaster preparedness in Vietnam.