PaperlayerView exact revision on arXiv

Source-linked AI summary

MAGPIE-Net: Predicting short-duration heavy-rainfall events in station neighborhoods from multitemporal FY-4A AGRI observations

Xiang Lin, Yunying Li, Chengzhi Ye, Zitong Chen, Jing Sun

arXiv:2608.17753v1cs.LG

TL;DR

Short-duration heavy-rainfall warnings must translate satellite cloud signals into station-neighborhood events despite the indirect radiance–rainfall relationship. MAGPIE-Net directly learns this satellite-to-station pathway with event supervision and outperformed gridded-output baselines, achieving a 65.1% episode detection rate and 64.6 min mean warning lead time versus 23.6% and 18.3 min.

  • Problem

    Satellite cloud signals have an indirect relationship with rainfall, while conventional methods derive station-neighborhood warnings only after gridded precipitation prediction.

  • Method

    MAGPIE-Net combines multitemporal FY-4A AGRI encoding, convection-initiation features, auxiliary precipitation diagnosis, and geographically adaptive grid-to-station event decoding.

  • Results

    65.1% episode detection and 64.6 min mean warning lead time versus 23.6% and 18.3 min for the best gridded-output baseline.

  • Takeaways & Limitations

    Event-oriented satellite-to-station modeling more effectively translates multitemporal AGRI observations into local short-duration heavy-rainfall warnings than gridded-output modeling.

  • Takeaways & Limitations

    Performance is lowest for localized extreme-rainfall events, and AGRI observations do not directly resolve several near-surface rainfall processes.

Abstract

from arXiv · show

Short-duration heavy-rainfall warning determines whether 1 h rainfall will exceed a threshold within a target-station neighborhood over the next few hours. Multitemporal infrared and water-vapor observations from the Fengyun-4A Advanced Geostationary Radiation Imager (FY-4A AGRI) capture cloud-top cooling, moisture evolution, and cloud expansion before substantial surface rainfall develops. However, most deep-learning nowcasting methods convert these signals into local warnings by post-processing gridded precipitation predictions, preventing station-neighborhood event targets from directly supervising the satellite-to-station learning pathway. We propose MAGPIE-Net, which embeds a geographically adaptive, differentiable grid-to-station mapping in a pathway combining convection-initiation features, multiscale encoding, and auxiliary gridded precipitation diagnosis. Station-neighborhood event losses thereby constrain the satellite representation and its mapping to irregular station locations for 0-3 h event prediction. In independent 2023 warm-season tests over central and eastern China, critical success index (CSI) values under the primary 40 km/20 mm h-1 definition were 0.371, 0.304, and 0.238 at 0-1, 1-2, and 2-3 h. Across episodes, MAGPIE-Net achieved a detection rate of 65.1% and a mean lead time of 64.6 min, compared with 23.6% and 18.3 min for the best gridded-output baseline, and remained superior for smaller neighborhoods and the 50 mm h-1 threshold. During the critical early-warning stage, when antecedent 1 h rainfall within 40 km remained below 1 mm, MAGPIE-Net detected 51.9% of episodes with a mean lead time of 38.5 min. These results show that event-oriented satellite-to-station modeling converts multitemporal geostationary cloud and moisture observations into local heavy-rainfall warnings more effectively than gridded-precipitation modeling.

1 Introduction

The introduction motivates station-neighborhood heavy-rainfall warnings by contrasting early, broad satellite observations with radar and gridded precipitation approaches. It presents MAGPIE-Net as an event-oriented FY-4A AGRI model that directly learns station-neighborhood events through a learnable grid-to-station pathway.

  • Motivation: Short-duration heavy rainfall can rapidly trigger damaging flash and urban floods, especially in small catchments and densely populated areas.Extreme hourly rainfall over eastern China is highly localized and often associated with warm-season mesoscale convective systems.
  • Observational basis: Radar may show weak echoes during early convective growth, whereas geostationary infrared and water-vapor sequences capture cloud development before precipitation echoes become pronounced.These observations include cloud-top cooling, vertical growth, horizontal expansion, glaciation-related spectral changes, and tropospheric moistening.
  • Problem: Satellite observations provide broad day-and-night coverage, but the indirect radiance–rainfall relationship makes surface-rainfall warnings physically difficult.Rainfall also depends on updraft strength, cloud microphysics, moisture supply, boundary-layer convergence, and vertical structure.
  • Target formulation: The warning target is a station-neighborhood event defined by 1 h rainfall reaching or exceeding a prescribed threshold within a specified radius of a target station.This target accommodates gauge sampling limitations, shifting rain cores, and uncertainty in future rainfall-core location and morphology.
  • Proposed approach: MAGPIE-Net predicts 0–3 h station-neighborhood events from multitemporal FY-4A AGRI observations using direct event supervision, learnable grid-to-station mapping, multiscale cloud-moisture encoding, and auxiliary gridded precipitation diagnosis.It is evaluated against gridded-output baselines on an independent 2023 warm-season test set over central and eastern China.

2 Data

The study evaluates station-neighborhood heavy-rainfall warnings across central and eastern China using multitemporal FY-4A AGRI observations and station precipitation records. Samples combine four antecedent satellite scans with six station-based event definitions and forecast targets spanning 0–3 h.

  • Study domain: The domain spans central and eastern China, covering inland and coastal regions, plains, and mountainous terrain with diverse heavy-rainfall regimes.The domain extends from 20.40°–38.28°N and 107.04°–124.92°E.
  • Satellite inputs: FY-4A AGRI channels 7–14 jointly capture cloud-top temperature, cloud phase, moisture, cloud growth, expansion, and cooling across day and night.Observations are remapped to a common 448 × 448 latitude–longitude grid with approximately 0.04° spacing.
  • Precipitation observations: Ground observations provide 1 h accumulated precipitation at 15 min intervals, station coordinates, elevation, and valid national or regional automatic-station records.Neighborhood labels use every station with a valid rainfall record at the relevant time.
  • Event labels: Six event definitions pair radii R ∈{40, 20, 10} km with thresholds q ∈{20, 50} mm h−1; 40 km/20 mm h−1 is primary.The other definitions support auxiliary training and evaluation under stricter spatial or intensity criteria.
  • Forecast samples: Each sample uses four AGRI scans from t0 −60, −45, −30, and −15 min to predict station-neighborhood event-label maps for the 0–1, 1–2, and 2–3 h windows.Training uses 2018–2021, validation uses 2022, and independent testing uses 2023 warm-season data.

3 MAGPIE-Net: An Event-Oriented Satellite-to-Station Architecture

MAGPIE-Net is an event-oriented architecture that combines convection-initiation features, multiscale satellite encoding, auxiliary gridded precipitation diagnosis, and a differentiable grid-to-station decoder. Its geographically adaptive mapping lets station-neighborhood event losses constrain the satellite representation and predictions at irregular stations.

  • Overall architecture: MAGPIE-Net combines a CI feature module, multiscale satellite encoder, auxiliary precipitation-diagnosis head, GA-SetConv decoder, and multi-head event predictor.The encoder preserves regional cloud and rainfall structures before mapping them to irregular target stations.
  • Convection-initiation features: CI masks use motion-compensated AGRI observations to identify localized rapid cloud development along estimated cloud trajectories.Masks are constructed at t0 −30 min and t0 −15 min using three preceding AGRI scans and optical-flow fields from 10.7 µm imagery.
  • Multiscale satellite encoding: The encoder uses UNet++ nested skip connections across four resolution levels to combine localized cloud-development features with broader cloud and moisture patterns.The four levels use 64, 128, 256, and 512 feature channels.
  • Auxiliary precipitation diagnosis: An auxiliary 1 × 1 regression head projects the regional representation into three forecast-window precipitation diagnosis fields that supply gridded inputs to the station decoder.The diagnosis target is generated from station precipitation observations using Barnes objective analysis on the satellite grid.
  • Grid-to-station decoding: GA-SetConv uses an axis-aligned learnable kernel whose east–west and north–south length scales depend on target-station attributes and forecast window.This geographically adaptive aggregation connects regular-grid features to irregular station locations and accommodates differing spatial support.

4 Results

The experiments test whether pre-initialization AGRI sequences contain information about subsequent station-neighborhood heavy rainfall and whether MAGPIE-Net converts that information into local event probabilities more effectively than gridded-output baselines.

  • Results: The experiments test whether pre-initialization AGRI sequences contain information relevant to subsequent station-neighborhood heavy rainfall.
  • Results: A second question assesses whether MAGPIE-Net converts pre-initialization AGRI information into local event probabilities.
  • Results: The experiments compare MAGPIE-Net’s conversion of AGRI information with gridded-output baselines.

4.1 Experimental Setup and Evaluation Protocol

The evaluation uses identical four-scan FY-4A AGRI inputs and year-based splits across models, comparing direct station-neighborhood events with gridded-output baselines under six event definitions. Performance is assessed at station–time and episode levels, including antecedent-rainfall stratification.

  • Model comparison: All models receive the same four-scan FY-4A AGRI sequence and year-based data split, while baselines predict gridded 1 h precipitation across three forecast windows.The baselines are adapted Earthformer, PhyDNet, NPM, and NowcastNet; MAGPIE-Net directly predicts the station-neighborhood target pathway.
  • Event definitions: Six event definitions combine neighborhood radii R ∈{40, 20, 10} km with rainfall thresholds q ∈{20, 50} mm h−1.Gridded baselines use radius-based neighborhood maxima in physical rainfall units and the full verification radius.
  • Verification metrics: Station–time performance is evaluated using critical success index (CSI), probability of detection (POD), and false alarm ratio (FAR).Counts are accumulated over valid station–time instances for each event definition and forecast window.
  • Episode verification: Episode-level verification uses the primary 40 km/20 mm h−1 definition and groups consecutive positive labels at a station during sustained rainfall into one episode.An episode is detected when the model issues a positive forecast no later than the end of its first positive window.
  • Antecedent-rainfall stratification: 9,843, 10,838, and 23,121 episodes belong to antecedent-rainfall groups Pante < 1 mm, 1 ≤Pante < 5 mm, and 5 ≤Pante < 20 mm, respectively.Another 8,480 episodes have Pante ≥20 mm and are excluded from stratified analysis because they do not begin below threshold.

4.2 Primary Station-Neighborhood Warning Performance

MAGPIE-Net outperforms NowcastNet for primary station-neighborhood warnings across forecast windows and episode-level metrics. Its advantage persists across antecedent-rainfall conditions, including an early-warning stage before substantial surface rainfall develops.

  • Forecast-window performance: CSI values for MAGPIE-Net are 0.371, 0.304, and 0.238 at 0–1, 1–2, and 2–3 h, versus 0.167, 0.107, and 0.069 for NowcastNet.All models lose skill as forecast lead time increases.
  • Forecast-window performance: During the 1–2 h and 2–3 h windows, MAGPIE-Net has substantially higher POD and lower FAR than NowcastNet.In the first forecast hour, higher POD is accompanied by a moderate FAR increase.
  • Episode-level performance: 65.1% of observed heavy-rainfall episodes are detected by MAGPIE-Net, compared with 23.6% for NowcastNet; mean warning lead times are 64.6 and 18.3 min, respectively.The results show that AGRI sequences can anticipate impending heavy-rainfall episodes within target-station neighborhoods.
  • Antecedent-rainfall dependence: Episode-level performance for both models improves as Pante increases, consistent with stronger persistence and spatial organization after surface rainfall begins developing.Antecedent-rainfall analysis evaluates performance before the first threshold exceedance.
  • Antecedent-rainfall dependence: For Pante < 1 mm, MAGPIE-Net detects 51.9% of 9,843 episodes with a 38.5 min mean lead time, versus 18.6% and 11.5 min for NowcastNet.Across all three antecedent-rainfall classes, MAGPIE-Net’s detection rate is approximately 2.8 times NowcastNet’s, while its lead-time advantage rises from 27.0 to 51.0 min.

4.3 Contribution of the Event-Oriented Pathway and GA-SetConv

Controlled comparisons show that MAGPIE-Net’s complete event-oriented pathway provides the main skill gain over gridded-output training, while GA-SetConv adds consistent improvements across all six event definitions. Direct event supervision and trainable spatial conversion optimize satellite representations against station-neighborhood classifications.

  • Controlled configurations: MAGPIE-Net is the complete model, whereas the grid-trained backbone uses the same UNet++ satellite encoder but lacks station-event prediction and gridded precipitation supervision only.Their comparison isolates the overall change from gridded-output training to the complete event-oriented satellite-to-station pathway.
  • Event-oriented pathway: 0.0154 to 0.1433: MAGPIE-Net increases mean CSI across six event definitions relative to the grid-trained backbone, with improvement for every radius–threshold combination.Localized high-intensity rain cores make station-neighborhood classifications sensitive to small positional or intensity errors in thresholded gridded forecasts.
  • Event-oriented pathway: Direct event supervision and trainable spatial conversion optimize the satellite representation against the final station-neighborhood event-classification target.This pathway avoids relying solely on post-processed gridded precipitation outputs for local warnings.
  • GA-SetConv: 0.1331 to 0.1433: GA-SetConv raises overall mean CSI by 7.7% relative to bilinear interpolation, improving all six event definitions, especially smaller neighborhoods and higher thresholds.The stricter definitions are more sensitive to spatial displacement between cloud and rainfall structures and to local geographic conditions.

4.4 Performance Across Event Definitions and Spatiotemporal Subsets

MAGPIE-Net maintains the highest CSI across all tested neighborhood radii, rainfall thresholds, and forecast windows, while performance varies with daytime, season, and region. It retains station-neighborhood warning performance after sunset and performs best during July and August.

  • Event definitions: MAGPIE-Net achieves the highest CSI across all 18 combinations of neighborhood radius, rainfall threshold, and forecast window.Its advantage extends from the 40 km/20 mm h−1 definition to the localized 10 km/50 mm h−1 definition.
  • Event definitions: CSI decreases for all models as neighborhood radius shrinks and rainfall threshold rises, reflecting greater difficulty from displacement and localization errors.A 40-km neighborhood accommodates some mismatch between satellite-observed cloud structures, surface rainfall, and the target station, whereas 10 km is more sensitive to location errors.
  • Day–night differences: Daytime samples have slightly higher CSI and POD than nighttime samples and contain approximately twice as many positive events under the primary event definition.Daytime is defined as 06:00–18:00 Beijing Time (BJT).
  • Day–night differences: MAGPIE-Net retains station-neighborhood warning performance across all three forecast windows after sunset.Infrared and water-vapor channels continue observing cloud-top and moisture evolution in nocturnal convective systems and rainbands.
  • Seasonal variation: 66.4% and 70.5% are the July and August episode detection rates, respectively, when performance is strongest and positive events are most frequent.Across all five months, episode detection ranges from 48.0% in September to 70.5% in August, while mean warning lead time ranges from 47.0 to 74.5 min.
  • Regional variation: The Southeast Coastal Belt has the highest CSI and lowest FAR across all three forecast windows, whereas the Mountain Belt has lower POD and higher FAR at longer lead times.Regional contrasts are associated with differences in organized coastal and tropical-cyclone rainbands, terrain-influenced precipitation, and warm-sector rainfall.

4.5 CI-Feature Ablation and Spatial Attribution

CI-feature ablation and masking experiments showed that CI features improved episode-level warning performance and exerted a localized, lead-time-dependent influence on event probabilities. The model used CI candidates selectively, with surrounding AGRI context helping distinguish regions linked to later heavy rainfall.

  • CI-feature ablation: 34,053 episodes, 65.1% detection rate, and 64.6 min mean warning lead time were achieved with CI features, versus 29,838 detected episodes without them.The ablation retrained MAGPIE-Net without the five CI channels while retaining the same AGRI sequence, event heads, and GA-SetConv decoder.
  • Spatial attribution: Event-probability differences were largest near the CI core, decreased with distance, and were strongest in the 0–1 h forecast window.The inference-time masking experiment kept model parameters and all other inputs unchanged.
  • Spatial attribution: The localized radial response and stronger positive-event response indicate selective use of CI features rather than uniform probability increases around every candidate.Permissive CI masks identify possible rapid-cloud-development regions, while surrounding AGRI context helps distinguish candidates associated with subsequent heavy rainfall.

4.6 Representative Heavy-Rainfall Cases

Representative cases show MAGPIE-Net detecting localized heavy-rainfall initiation before substantial surface rainfall and more continuously capturing organized coastal rainband expansion than gridded-output baselines.

  • Localized convective development: 0.5 mm was the maximum antecedent 1 h rainfall across the three 40-km station neighborhoods before the first threshold-exceeding window.The latest AGRI image showed a localized cold-cloud feature near the Beibu Gulf coast.
  • Localized convective development: MAGPIE-Net detected all three newly positive stations, while PhyDNet, NPM, EarthFormer, and NowcastNet missed all three.Within the enlarged local domain, MAGPIE-Net produced three hits, no misses, and four false alarms, corresponding to a local CSI of 0.429.
  • Typhoon Doksuri landfall: The gridded-output models produced fragmented warnings, whereas MAGPIE-Net formed a more continuous band and achieved the highest case-level CSI in all three forecast windows.The difference was most pronounced during 1–2 h and 2–3 h as the affected area expanded along the Fujian–Zhejiang coast.
  • Complementary case evolution: The two cases span complementary cloud–rain stages: localized development from near-zero antecedent rainfall and inland expansion of a mature tropical-cyclone rainband.MAGPIE-Net identified the newly positive southern Guangxi component missed by all gridded-output baselines and more continuously captured Doksuri rainband expansion.

5 Conclusions

MAGPIE-Net uses an event-oriented satellite-to-station pathway to translate multitemporal FY-4A AGRI observations into 0–3 h station-neighborhood heavy-rainfall warnings. Independent 2023 warm-season evaluation showed stronger performance than gridded-output baselines, while experiments identified configuration gains and remaining limitations.

  • Model design: MAGPIE-Net combines convection-initiation features, multiscale satellite encoding, auxiliary gridded precipitation diagnosis, and geographically adaptive grid-to-station decoding.The design preserves regional cloud- and rainfall-related feature organization while directly supporting station-neighborhood event prediction.
  • Evaluation: 65.1% episode detection rate and 64.6 min mean warning lead time were achieved during independent 2023 warm-season evaluation over central and eastern China.MAGPIE-Net outperformed gridded-output baselines across all three forecast windows, with the advantage persisting for smaller neighborhoods and the 50 mm h−1 threshold.
  • Configuration analysis: The complete event-oriented configuration produced the largest improvement over the gridded-output pathway, while GA-SetConv added gains through station- and lead-dependent spatial aggregation.Including CI features increased episode detection rate and mean warning lead time relative to the retrained no-CI model.
  • Limitations: Performance was lowest for spatially localized extreme-rainfall events under smaller-neighborhood and higher-intensity definitions.AGRI observations do not directly resolve low-level moisture convergence, warm-rain microphysics, terrain-induced near-surface processes, or cloud-top displacement from surface rainfall.
  • Conclusion: The event-oriented satellite-to-station pathway more effectively translates multitemporal AGRI observations into local short-duration heavy-rainfall warnings than conventional gridded-output modeling.This conclusion summarizes the reported comparative evaluation.
Loading 2608.17753v1…