Answers are grounded in this paper and cite their sources.
Accurate rainfall forecasting is crucial for disaster preparedness, agricultural planning, and water resource management, particularly in countries like Bangladesh that are highly vulnerable to the impacts of climate change. Traditional forecasting methods often struggle to capture complex spatiotemporal relationships among meteorological stations, handle missing timestamps, and provide reliable uncertainty estimates for multi-horizon and extreme rainfall events. To address these challenges, this study proposes BanglaRain, a missingness-aware spatiotemporal graph transformer for probabilistic multi-horizon rainfall and extreme-event forecasting. The proposed framework integrates a temporal transformer with graph attention networks to capture spatiotemporal relationships among meteorological stations, while incorporating a missingness-aware mask and seasonal regime embeddings. Experiments using 3-hourly rainfall data from 34 weather stations spanning 2003–2023 show that BanglaRain outperforms baselines including LightGBM, CatBoost, Quantile LSTM, DeepAR, Temporal Transformer, and Graph Attention Networks. The model achieves the lowest CRPSlog=0.1068 and CRPSₘₘ=0.5726 , and the highest performance for extreme-event prediction with short-duration extreme rainfall (SDER) risk ROC-AUC = 0.9224 and PR-AUC = 0.4182, evaluated across seven quantile levels. Ablation studies reveal genuine trade-offs across metric types, with the missingness mask contributing primarily to PR-AUC-based extreme-event discrimination, including a 2.3% improvement in accumulation-based PR-AUC (0.2673 vs. 0.2614) and improved CRPSₘₘ (0.5726 vs. 0.5781) relative to forward-fill imputation. After isotonic calibration fitted on the validation set, the expected calibration error reduces to 0.0209 for SDER risk, 0.0139 for peak exceedance, and 0.0345 for accumulation (24h). The proposed framework provides a promising probabilistic rainfall and extreme-event forecasting approach that may support disaster preparedness, flood risk management, and climate-resilient agricultural planning, while requiring further operational-scale validation.
No takes yet. Share an insight, caveat, or question.
Khanom et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: