Abstrak Artikel
Extreme-rainfall detection in small tropical islands is difficult because extreme events are rare and false negatives have substantial disaster-mitigation consequences. This study aims to (1) compare imbalance-aware machine-learning and deep temporal models, (2) determine whether an out-of-fold stacking ensemble can maximize sensitivity, and (3) explain model behavior using SHAP. ERA5 hourly data for a representative Ternate grid point were collected from 1 January 2005 to 6 February 2026 and aggregated into 7,707 daily records; only 12 days met the extreme threshold of 75 mm/day. A 30-day sequence was used to predict the next-day class, and three expanding temporal test folds were evaluated. Weighted loss, focal loss, balanced tree learning, validation-based threshold selection, and leakage-controlled stacking were applied. Across 2,229 pooled test days containing four extreme events, the stacking model detected all events (recall 1.000) but produced low precision (0.0023), F2-score 0.0115, specificity 0.2256, and 1,723 false positives. Balanced Random Forest provided the best overall trade-off, with recall 0.750, F2-score 0.0316, specificity 0.7951, and MCC 0.0570. Therefore, stacking is useful as a high-sensitivity screening layer, but it is not yet suitable as a standalone operational warning model. Local observations and additional event samples are required for calibration.
Pratinjau PDF
Pastikan PyMuPDF terpasang lalu coba muat ulang.