Deep learning nowcast using 3D atmospheric forecasts beats baselines for radar echoes and rain

Why it matters

Radar-only nowcasts miss storms that have not yet appeared on radar, and they let strong echoes fade. This system targets both problems. It runs a 12 h forecast in about 10 s on one GPU, and it has been deployed at the Jiangsu Meteorological Observatory.

What they did

The authors built FuXi-Nowcast. It combines radar, station and land-analysis data with 70 atmospheric variables from the FuXi-2.0 forecast model. It was trained on April to September 2019 to 2023 data over East China on a 1 km grid. It was tested on April to July 2024 against CMA-MESO, persistence and PySTEPS. Two case studies and ablation runs probed why it works. Special modules and loss terms keep strong signals from fading.

Key findings

• FuXi-Nowcast had the best overall critical success index for composite reflectivity, with the gap growing at longer lead times, and it stayed best at 50 dBZ.
• For total precipitation it beat all baselines at 1, 20 and 50 mm, though scores were low for every method.
• For wind gusts it led at 10.8 m/s after the first few hours, but persistence and PySTEPS did better at 13.9 and 17.2 m/s.
• In a dryline case on 16 June 2025, the full model predicted the new storm line. The version without 3D atmospheric fields did not.
• Removing specific humidity hurt reflectivity skill most among input groups. Removing the pooling module or Balanced L1 loss weakened strong echoes or heavy rain.

Limitations

• Storm initiation positions are off by tens of kilometers in case studies, and the forecasts have less small-scale detail than the 1 km grid suggests.
• Testing covers only April to July 2024 over East China, so cold-season skill is not assessed. The case evidence for initiation rests on single events.
• The model is deterministic and gives no forecast uncertainty.

Glossary

• Nowcasting: Very short-range weather prediction, usually the next 0 to 6 hours.
• Convective initiation: The first appearance of new storm echoes, defined as radar reflectivity above 35 to 40 dBZ.
• Critical success index (CSI): A score from 0 to 1 for event forecasts. Higher means more hits with fewer misses and false alarms.
• Composite reflectivity: A radar-derived map of the strongest echo in each column of air, used to show storm intensity.