PLUME: Staged Fusion of FY-4A AGRI and Rain-Gauge Observations for Bidirectional Refinement of Short-Duration Heavy-Rainfall Warnings

Local short-duration heavy-rainfall warning requires early assessment of hazardous rainfall within a specified neighborhood. Geostationary meteorological satellites provide frequent, spatially continuous observations of cloud systems and storm evolution, but only indirectly reflect surface rainfall. Rain gauges measure surface accumulation directly, but only at irregular locations. Deep learning fusion methods commonly convert gauge observations to gridded representations before combining them with satellite observations. This treatment merges gauge measurements and local gauge support into a single representation, limiting how local surface conditions inform warning decisions and increasing the likelihood that local heavy-rainfall risk is underestimated or overestimated. We present PLUME (Precipitation Nowcasting with Late-Stage Residual Updates from Rain-Gauge Measurements for Short-Duration Heavy-Rainfall Events), a staged fusion model with gridded and point-origin gauge pathways. PLUME first combines a gauge-derived gridded rainfall background with FY-4A AGRI observations to form a base rainfall representation. The point-origin pathway uses local gauge features and observation support to update this representation through signed residuals, which the station decoder then maps to neighborhood-event probabilities. This staged design preserves the complementary roles of continuous spatial context, local rainfall measurements, and gauge availability. We evaluated PLUME for 0–3 h local short-duration heavy-rainfall warning over central and eastern China using an independent test set from May to September 2023. Across Barnes analysis, inverse distance weighting, and kriging, PLUME consistently improved the critical success index (CSI) over matched satellite–grid baselines. Relative CSI gains were 2.6–3.3% under complete gauge input and 3.6–5.9% under simulated gauge missingness. PLUME also reduced the associated CSI loss by 45.8–67.4%. Analysis of the variant with frozen batch normalization (BN) statistics showed that, under all three backgrounds and both input conditions, the largest positive probability updates were concentrated among baseline misses and converted some to hits, whereas the largest negative probability updates were concentrated among baseline false alarms and converted some to correct negatives. The dominant conditional CSI contribution shifted from false-alarm suppression at 0–1 h to missed-event recovery at 1–3 h. By using gridded and point-origin gauge information at different fusion stages, PLUME improves local heavy-rainfall warning through lead-dependent bidirectional refinement.