Year: 2026
Type: article
Abstract: Extreme rainfall is a primary cause of flood disasters, particularly in tropical regions such as Indonesia. Kota Ternate is highly vulnerable to such events due to its geographical and climatic conditions. However, predicting extreme rainfall remains challenging because of the highly imbalanced nature of rainfall data, where extreme events occur very rarely. This study proposes a recall-oriented stacking ensemble model that integrates Long Short-Term Memory (LSTM) and Transformer as base learners, with XGBoost as a meta-learner. The model is trained using ERA5 reanalysis data from 2005 to 2026 and incorporates time series feature engineering, including lag and rolling statistics. The experimental results show that traditional machine learning models fail to detect extreme rainfall events, achieving zero recall. In contrast, the proposed model achieves perfect recall (1.0), ensuring that all extreme events are successfully detected. However, this performance is accompanied by low precision, indicating a high number of false positives. This trade-off is acceptable in early warning systems, where missing extreme events is more critical than generating false alarms. Furthermore, SHAP analysis reveals that lag-based rainfall features, rolling rainfall accumulation, and atmospheric variables such as dewpoint and wind significantly influence model predictions. The results demonstrate that the proposed model is effective for extreme rainfall detection and has strong potential for application in flood early warning systems.). Keywords: Keywords—Extreme Rainfall, Early Warning System, LSTM, Transformer, XGBoost, SHAP, Time Series
Language: English
Subfield: Environmental Engineering
Field: Environmental Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: diamond
OpenAlex ID: W7208718693