A Rainfall-Based Machine Learning Decision-Support Prototype for Flood Risk Prediction in Malaysia

Abstract

This article presents a rainfall-based machine-learning decision-support prototype for flood and non-flood classification in Malaysia using historical flood data from 2000-2010. The dataset contained 825 records, with 480 non-flood and 345 flood records. Thirteen rainfall predictors were used to train Support Vector Machine (SVM), Random Forest (RF), and a voting-based SVMRF ensemble. The tuned Random Forest model under an 80:20 split produced accuracy = 0.7212, precision = 0.6667, recall = 0.6667, F1-score = 0.6667, and ROC-AUC = 0.7685. The confusion matrix produced 73 true negatives, 23 false positives, 23 false negatives, and 46 true positives. Because the dataset is historical, the model is positioned as a baseline decision-support prototype rather than an operational public warning system. The proposed layer links model outputs with district monitoring, error-to-action interpretation, recent Malaysian flood-impact context, and future validation needs.