Satellite-based precipitation retrieval using cross-temporal attention and multiscale spatial representation

Geostationary meteorological satellites provide valuable observations for hourly precipitation monitoring. However, single-time snapshot observations struggle to capture the dynamic evolution of cloud systems. Furthermore, the scale mismatch between point-based rain gauges and grid-based satellite pixels, combined with complex terrain, limits the spatial generalization of current retrieval models. To address these challenges, we propose DTPNet, a deep learning framework designed for station-level hourly precipitation retrieval. Instead of relying on a single snapshot, the model utilizes consecutive multi-channel observations from geostationary satellites (e.g., FY-4A/AGRI) to explicitly track cloud movement and development over time. A cross-temporal attention mechanism is introduced to extract these dynamic cloud evolution patterns, alongside multiscale spatial representations to bridge the spatial gap between satellite pixels and ground stations. To ensure robust performance across diverse landscapes, the model dynamically adapts to local terrain elevation and spatiotemporal context. It simultaneously detects precipitation occurrence and classifies precipitation intensity. Evaluated against nationwide hourly rain gauge data in China for 2021, DTPNet significantly outperforms operational products like GPM IMERG and FY-4 QPE. Specifically, the Critical Success Index (CSI) for precipitation detection improves by approximately 53% and 108% compared to these products, respectively, and the Macro-F1 score for intensity classification reaches 0.4370. These results demonstrate the method’s superior accuracy and robust generalization for satellite-based precipitation monitoring.