A station-scale modeling framework for heavy rainfall classification in tropical weather using representative machine learning approaches

Abstract

Extreme daily rainfall in rapidly urbanizing tropical cities frequently overwhelms drainage and disrupts critical services, yet station-scale forecasting remains limited by convective variability and sparse observations. This motivates lightweight, interpretable machine-learning tools that operate on routine station data. We propose and evaluate a station-scale framework to classify heavy-rainfall days (50 mm) in a humid tropical setting. Using 1,796 daily observations from the Soekarno-Hatta Meteorological Station (20182022), we engineered lag-informed predictors (e.g., previous-day rainfall, 3-day sums/means) and compared three representative classifiers, Logistic Regression (LR), Support Vector Machine (SVM), and Random Forest (RF). Class imbalance was addressed with class-weighted training, and models were assessed on a held-out test set using precision, recall, F1, and Receiver Operating Characteristic Area Under the Curve (ROC-AUC). LR achieved the highest recall (0.429), indicating moderate sensitivity to rare heavy-rainfall events, whereas RF yielded the best probabilistic discrimination (AUC = 0.619) but failed to flag positives at the default threshold; SVM displayed near-random behavior. Feature analyses highlighted humidity, temperature, and recent rainfall accumulation as the most influential predictors, consistent with tropical convective processes. Despite severe class imbalance, simple, station-based classifiers can extract actionable signals for rare-event screening in data-limited tropical regions. Operational value is likely to improve through probability calibration and threshold tuning, ensemble integration, and spatial generalization to multi-station settings.