Accurate prediction of extreme rainfall events in the Indian subcontinent remains a critical challenge due to the inherent rarity of such phenomena and the resulting data imbalance in meteorological observations. We propose a rarity-aware spatiotemporal attention mechanism that dynamically prioritizes precursors of extreme rainfall without relying on explicit data resampling or loss reweighting. The method introduces a novel rarity score computed from historical extreme event clusters, which modulates attention weights to amplify the influence of rare but physically significant patterns in multi-modal meteorological data. A dual-path architecture processes spatial and temporal dependencies separately, then fuses them through a gated cross-attention layer to capture both regional precursors and evolving atmospheric dynamics. The proposed mechanism integrates seamlessly with standard Transformer backbones while maintaining computational efficiency through sparse attention patterns and parallel rarity score computation. Furthermore, a joint optimization objective combines prediction accuracy with physical consistency, ensuring that learned attention aligns with expert-annotated importance maps of known extreme rainfall drivers. Evaluated on satellite, reanalysis, and ground station data, our approach demonstrates significant improvements in predicting both rainfall intensity and extreme event occurrence. The rarity-aware attention not only addresses the class imbalance problem but also provides interpretable insights into the spatiotemporal processes governing extreme rainfall, offering practical value for disaster preparedness in vulnerable regions.
Sinha et al. (Wed,) studied this question.