Antecedent rainfall dominates flood peak prediction in the Ganges-Padma Basin: A machine learning study

In the Ganges-Padma Basin, flood peaks integrate complex signals from monsoon rainfall, antecedent basin wetness, and upstream discharge. Understanding these interactions is critical for flood forecasting in Bangladesh. This study investigates rainfall–flood peak relationships at four stations along the Ganges–Padma corridor using a 29-year daily record (1996–2024). Four ensemble tree-based machine learning algorithms - Random Forest, XGBoost, LightGBM, and CatBoost - were developed using observed (Bangladesh Meteorological Department, BMD) and satellite (Climate Hazards Group InfraRed Precipitation with Station data, CHIRPS) basin-average rainfall. Predictors included total, maximum, minimum, and antecedent rainfall. CatBoost performed best at the Ganges–Brahmaputra confluence (Goalondo: Nash-Sutcliffe Efficiency, NSE = 0.754, R 2 = 0.842), while XGBoost performed best at upstream and near-estuary stations (NSE ≈ 0.38). All models failed at Mawa (NSE < 0) due to limited sample size and routing-dominated dynamics. Feature importance revealed antecedent rainfall over 3–10-day windows consistently outranked storm-window metrics, confirming pre-event basin wetness as the dominant flood control. Satellite-based models showed lower performance due to CHIRPS underestimation of extreme rainfall, yet remained viable in data-scarce settings. These results demonstrate rainfall–flood nonlinearity and establish CatBoost as a robust algorithm for flood peak prediction at confluence stations in transboundary settings.