First-hand information for everyone
Scientists have shown that a neural network can take only the infrared pictures that a geostationary satellite continuously records and turn them into high‑resolution maps of surface winds around tropical cyclones. The procedure uses a U‑Net encoder–decoder architecture that reads ten infrared channels and four spatial‑context channels (local solar‑time sine and cosine, solar‑zenith angle, distance to the storm centre). An optional multi‑layer perceptron (MLP) is attached to the latent space, or run after the 2‑D field output, to predict key storm metrics such as the maximum wind speed (Vmax), the radius of maximum wind (Rmax), and the 34, 50 and 64 kt wind radii.
Training data come from pairs of satellite images and wind fields derived from synthetic‑aperture‑radar (SAR) observations. The nearest SAR image is used in most cases, giving satellite inputs that are within ± 15 minutes of the wind measurement. Each image is cropped to 196 × 196 pixels and centred on the storm so that no single spatial location dominates the learning.
When the U‑Net is trained to reconstruct the SAR wind field and the MLP is supervised by Vmax and Rmax, the model achieves the lowest errors on a held‑out test set. The mean absolute error (MAE) for Vmax drops to 5.88 m s‑1, and for Rmax to 8.93 km, compared with 7.52 m s‑1/17.86 km for a plain U‑Net and 6.71 m s‑1/12.50 km for a post‑hoc MLP. When the model is further trained only for rapid‑intensification cases (RI), the errors reduce from 7.20 m s‑1 / 13.14 km to 5.47 m s‑1 / 11.18 km.
| Model | ERA5 | All | RI |
|---|---|---|---|
| U-Net | ✗ | 7.518 | 17.859 |
| Post-hoc MLP | ✗ | 6.715 | 12.507 |
| Latent MLP | ✗ | 5.885 | 8.934 |
| U-Net | ✓ | 6.446 | 14.475 |
| Post-hoc MLP | ✓ | 5.529 | 10.512 |
| Latent MLP | ✓ | 7.302 | 7.716 |
| Vmax | Rmax | ||||
|---|---|---|---|---|---|
| SAR | Radius | All | RI | All | RI |
| ✓ | ✗ | 6.694 | 6.867 | – | – |
| ✗ | ✓ | 6.297 | 7.198 | 19.44 | 13.14 |
| ✗ | ✗ | 6.828 | 6.645 | – | – |
| ✓ | ✓ | 6.344 | 5.469 | 18.24 | 11.18 |
Including nine additional ERA5 reanalysis channels can improve performance slightly, but it also introduces non‑real‑time data and is not part of the core approach. The model does not predict wind fields at aircraft flight level and does not refine existing SAR products; it focuses solely on surface winds.
Limitations remain. The need for paired SAR wind fields limits where the method can be trained, and the available sample size (841 training, 232 validation, 212 test images from 115, 34 and 38 storms) is modest, especially for the 134 rare rapid‑intensification events. The study did not exhaustively explore alternative architectures or fine‑tune hyper‑parameters, leaving further gains possible.
Nevertheless, the results demonstrate that continuous, high‑resolution wind field estimates are obtainable from satellite imagery alone, offering a powerful tool for studying tropical cyclone structure and intensity.