Predicting Flood Potential Using Machine Learning with the XGBoost and Logistic Regression Approaches

Abstract

Flooding is one of the most frequent natural disasters in Indonesia, causing significant material losses and casualties. This study aims to develop a flood potential prediction model based on weather data using machine learning approaches, namely XGBoost and Logistic Regression. The dataset consists of 1,513,505 weather records with 1,165 flood events (0.077%). The features include temperature, humidity, wind speed and direction, weather codes, and temporal features generated using a sliding window approach for H-1, H-2, and H-3. Data imbalance was addressed using a combination of stratified undersampling and SMOTE, changing the class ratio from 1:1,298 to 1:3.3. Experimental results show that XGBoost outperforms Logistic Regression, achieving an accuracy of 98.40%, precision of 97.93%, recall of 95.07%, and an ROC-AUC of 99.38%, while Logistic Regression achieved an accuracy of 62.77%. Feature importance analysis indicates that weather codes at H-3 and H-1 are the most influential predictors. With a low false negative rate of 4.9%, the proposed XGBoost model is considered reliable for implementation as a flood early warning system.