Overview
Research introduces MAGPIE-Net, a deep-learning method designed for predicting short-duration heavy-rainfall events within specific station neighborhoods. This system utilizes multitemporal infrared and water-vapor observations from the Fengyun-4A Advanced Geostationary Radiation Imager (FY-4A AGRI) to identify atmospheric signals preceding significant surface rainfall. MAGPIE-Net's approach integrates a geographically adaptive, differentiable grid-to-station mapping process within a framework that combines convection-initiation features, multiscale encoding, and auxiliary gridded precipitation diagnosis. The model's training process directly supervises satellite-to-station learning pathways using station-neighborhood event losses, thereby constraining the satellite representation and its mapping to irregular station locations for 0–3 hour event predictions.
Research Context
Short-duration heavy-rainfall warning systems aim to predict if 1-hour rainfall will exceed a predefined threshold within a target-station neighborhood over the subsequent few hours. Prior to substantial surface rainfall development, multitemporal FY-4A AGRI observations capture indicative atmospheric phenomena such as cloud-top cooling, moisture evolution, and cloud expansion. However, a limitation identified in many existing deep-learning nowcasting methods is their reliance on converting these satellite signals into local warnings through post-processing of gridded precipitation predictions. This indirect conversion prevents direct supervision of the satellite-to-station learning pathway by the actual station-neighborhood event targets.
Approach
The MAGPIE-Net architecture addresses the challenge of directly linking satellite observations to station-neighborhood event predictions. It embeds a geographically adaptive, differentiable grid-to-station mapping component. This mapping is integrated into a pathway that also incorporates convection-initiation features and multiscale encoding of the satellite data. An auxiliary gridded precipitation diagnosis further complements the system. The model's learning objective uses station-neighborhood event losses to directly constrain both the satellite representation and its subsequent mapping to the irregular geographical locations of individual stations. This event-oriented satellite-to-station modeling paradigm was developed for 0–3 hour prediction horizons.
Findings
Independent tests conducted during the 2023 warm season across central and eastern China provided quantitative performance metrics for MAGPIE-Net:
- Critical Success Index (CSI): Under the primary definition of 40 km/20 mm h-1, CSI values were 0.371 for 0–1 h, 0.304 for 1–2 h, and 0.238 for 2–3 h prediction windows.
- Detection Rate: Across all tested episodes, MAGPIE-Net achieved a detection rate of 65.1%.
- Mean Lead Time: The model demonstrated a mean lead time of 64.6 minutes for detected events.
- Baseline Comparison: In contrast, the best gridded-output baseline model achieved a detection rate of 23.6% and a mean lead time of 18.3 minutes.
- Performance Robustness: MAGPIE-Net maintained its superior performance for smaller neighborhoods and when applying a higher threshold of 50 mm h-1.
- Early Warning Stage Performance: During the critical early-warning stage, defined as when antecedent 1-hour rainfall within 40 km remained below 1 mm, MAGPIE-Net detected 51.9% of episodes with a mean lead time of 38.5 minutes.
These results indicate that event-oriented satellite-to-station modeling effectively converts multitemporal geostationary cloud and moisture observations into localized heavy-rainfall warnings, outperforming gridded-precipitation modeling approaches.
Why This Matters
The observed efficacy of MAGPIE-Net in converting geostationary cloud and moisture observations directly into local heavy-rainfall warnings suggests a more effective approach compared to methods reliant on gridded-precipitation modeling. The higher detection rates and significantly longer lead times achieved by MAGPIE-Net, particularly during early warning stages and for localized events, indicate its potential to enhance the timeliness and accuracy of warnings for short-duration heavy-rainfall events.