MAGPIE-Net Improves Heavy Rainfall Warnings with Satellite Data.

Predicting short-duration heavy rainfall events, crucial for timely warnings, is challenging, especially when relying on satellite data. Traditional deep-learning nowcasting methods often convert satellite signals into gridded precipitation predictions, which then require post-processing to generate local warnings. This indirect approach prevents direct supervision of the satellite-to-station learning pathway for specific station-neighborhood event targets. Introducing MAGPIE-Net, a novel model that directly addresses this by embedding a geographically adaptive, differentiable grid-to-station mapping within its architecture. This allows station-neighborhood event losses to directly constrain the satellite representation and its mapping to irregular station locations, enabling 0-3 hour event predictions. Tested independently during the 2023 warm season over central and eastern China, MAGPIE-Net demonstrated superior performance. For the primary 40 km/20 mm h-1 definition, it achieved critical success index (CSI) values of 0.371, 0.304, and 0.238 at 0-1, 1-2, and 2-3 hours respectively. Critically, MAGPIE-Net achieved a detection rate of 65.1% and a mean lead time of 64.6 minutes, significantly outperforming the best gridded-output baseline which managed only 23.6% detection and 18.3 minutes lead time. Its effectiveness was particularly notable during the early-warning stage, detecting 51.9% of episodes with a 38.5-minute lead time when antecedent rainfall was minimal.