Spatio-temporal Attention driven Rainfall Inference using Transformative Architecture (SARITA)

Short-term rainfall forecasting presents a complex spatiotemporal modeling challenge, critical for adaptation and mitigation of high-impact hydrometeorological events such as flash floods and cloudbursts. This study examines hourly rainfall patterns in the Northwest Himalayan (NWH) region of India, enhancing forecasting accuracy using spatial correlations. The region shows a significant decline in hourly rainfall trends over the period 2019-2024, with approximately 50% of the area exhibiting statistically significant decreasing trends. The analysis reveals that high-elevation regions exhibit stronger initial 1-hour lag correlation (>0.8) and a rapid temporal decay over 9 hours. In contrast, lower-altitude areas (below 600 m) display a more gradual decrease in correlation, suggesting broader spatial influence with less intense propagation. Across all locations, rainfall variability as defined by statistically significant correlations, typically dissipates within 6 to 7 hours. Our analysis further identifies a prominent belt of high rainfall coherence aligned along the central Himalayan zone, highlighting pronounced spatial homogeneity. To improve predictive performance, this study proposes an attention-guided spatial correlation mechanism integrated within a Deformable Convolutional Long Short-Term Memory (DConvLSTM) framework. The proposed model, termed SARITA, processes hourly rainfall and spatial correlations, with the attention mechanism dynamically using spatial dependencies. This integration enhances the model's ability to generalize across spatial-temporal patterns. Model evaluation using statistical metrics demonstrates that SARITA consistently outperforms baseline models such as ConvLSTM and standard DConvLSTM, achieving a 25% improvement in hourly rainfall forecasts. Furthermore, the model also improves anomaly detection by 3.5% over the standard DConvLSTM model.