Benchmarking Machine-Learning Classifiers for Heavy-Rainfall Prediction from Monthly Station Data in Bangladesh: A First Step Toward Flood-Risk Assessment for Dhaka

Abstract

Flooding and extreme rainfall pose recurrent risks to Bangladesh, and Dhaka in particular, where unplanned urbanization, drainage channel encroachment, and intensifying monsoon rainfall compound exposure. Reproducible, data-driven tools are needed to complement resource-intensive hydrodynamic models and qualitative surveys. This study benchmarks four machine-learning classifiers (logistic regression, Random Forest, XGBoost, and a sequence-based LSTM) for predicting heavy-rainfall months from monthly meteorological records of 31 stations across Bangladesh (1950–2022), as a first stage toward flood-risk assessment for Dhaka. A month was labelled heavy if the following month's rainfall exceeded the station's 75th percentile, and zero-filled rainfall records (21% of station-years) were removed beforehand. Models were trained on 1950–1999, tuned on 2000–2009, and tested on 2010–2022 (4,805 samples, 24% positive). All four classifiers reached ROC AUC values of 0.906–0.911 (LSTM 0.911, XGBoost 0.910, Random Forest 0.908, logistic regression 0.906), compared with 0.50 for a majority-class baseline. However, a seasonal climatology baseline that uses only station and calendar month reached 0.912, and no model exceeded it on AUC, PR AUC or F1. These results indicate that, at monthly resolution, three months of station-level meteorological history adds little beyond the monsoon cycle for predicting heavy-rainfall months, and that observed flood labels and terrain, drainage, and land-cover data would be needed for reliable flood-risk mapping.