Explainable Spatiotemporal Deep Learning for Extreme Rainfall Prediction Using Multi-Station Climate Data in Indonesia
Year: 2026
Type: article
Abstract: Extreme rainfall prediction remains an important challenge for hydrometeorological disaster mitigation in Indonesia, where climate conditions and rainfall distributions vary considerably. This study evaluates four machine-learning models—Logistic Regression, Random Forest, HistGradientBoosting, and XGBoost—for classifying extreme rainfall events using multi-station daily climate data. The Climate Data Daily IDN dataset contains 589,265 observations from 173 stations over 2010–2020. Extreme events are defined using station-specific 95th-percentile (P95) thresholds calculated from the training data. A chronological split is used: 2010–2017 for training, 2018–2019 for validation, and 2020 for testing. Performance is assessed using accuracy, precision, recall, F1-score, ROC-AUC, and PR-AUC. On the test set, XGBoost achieves 86.24% accuracy, an F1-score of 0.2304, ROC-AUC of 0.7346, and PR-AUC of 0.1689, whereas Logistic Regression obtains the highest recall (39.19%). SHAP analysis of XGBoost identifies previous-day relative humidity as the feature with the highest mean absolute SHAP value (0.4097), followed by previous-day rainfall (0.1661). These results constitute an explainable machine-learning baseline; no deep-learning architecture or explicit inter-station spatial relationship is evaluated in this study. The limited precision and recall indicate that the evaluated models are not yet ready for operational early-warning use.
Institution: Universitas Pancasakti Makassar
Language: English
Subfield: Environmental Engineering
Field: Environmental Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: hybrid
OpenAlex ID: W7234476400
