Nowcasting of Tropical Cyclone Wind Fields and Intensity from Geostationary Imagery

AI summary

Overview

Research area: Machine learning for Earth observation, specifically nowcasting tropical cyclone (TC) surface wind fields and intensity from geostationary satellite imagery.

Technical level: Intermediate. The paper uses standard, well-known architectures (U-Net, MLP) and is readable with basic familiarity with convolutional neural networks and remote sensing; no exotic methods are introduced.

Scope (one sentence): A set of ablation studies showing that simple neural networks can recover 2D SAR-quality surface wind fields and storm-scale intensity metrics from geostationary imagery alone, with joint latent-space supervision giving the largest gains during rapid intensification.

What This Paper Is About

Tropical cyclone destructive potential depends on the full surface wind field, not just a single peak wind number, but the only high-resolution surface wind measurements come from synthetic aperture radar (SAR), which is spatially and temporally sparse (roughly twice a day across a storm's lifetime) and not available in near-real-time. Geostationary satellites, by contrast, image cloud tops continuously, but never see the surface. The paper asks whether simple neural networks can learn the mapping from geostationary infrared imagery to SAR-derived surface wind fields and, at the same time, to storm-scale intensity metrics such as maximum sustained wind and radius of maximum wind, including during rapid intensification (RI, defined as at least a 30 kt increase in 24 hours).

Key Contributions

1. An observation-only nowcasting pipeline. To the authors' knowledge, no prior work has used only satellite-based observations as input to predict both high-resolution SAR wind fields and storm-scale intensity metrics. The model is trained against SAR fields but never requires SAR at inference time.
2. A systematic comparison of four supervision strategies: a pooled-feature CNN baseline, a U-Net predicting only the wind field, a two-stage post-hoc MLP on the frozen U-Net output, and a joint "latent MLP" that shares the encoder features between the field decoder and the scalar predictor.
3. Evidence that joint field-and-scalar learning helps most for RI. Jointly supervising the latent representation with both the 2D wind field and scalar metrics improves retrieval of maximum wind and radius of maximum wind, particularly during rapid intensification.
4. An ablation showing reanalysis is optional. Adding coarse ERA5 reanalysis gives mixed results and is not required, which keeps the workflow usable in near-real-time settings where reanalysis products are unavailable.

Main Findings

• Latent joint modelling is the best observation-only configuration. Without ERA5, the latent MLP achieves a V_max MAE of 5.885 m s⁻¹ across all test observations, versus 6.715 m s⁻¹ for the post-hoc MLP and 7.518 m s⁻¹ for the field-only U-Net.
• The gain is largest during rapid intensification. In the same no-ERA5 setting, RI V_max MAE is 8.934 m s⁻¹ for the latent MLP, compared with 12.507 m s⁻¹ (post-hoc MLP) and 17.859 m s⁻¹ (U-Net).
• ERA5 helps some models and hurts others. Adding ERA5 improves the U-Net (6.446 m s⁻¹) and the post-hoc MLP (5.529 m s⁻¹, the lowest all-observation value in the table), and reduces RI error for the latent MLP to 7.716 m s⁻¹, but does not reduce the latent MLP's all-storm error (7.302 m s⁻¹).
• Field and scalar objectives are complementary. For the GEO-only latent MLP, combining SAR wind field reconstruction with radius supervision lowers RI V_max MAE from 7.198 to 5.469 m s⁻¹ and RI R_max MAE from 13.14 to 11.18 km. Without radius supervision, SAR reconstruction alone does not improve RI V_max (6.867 m s⁻¹ versus 6.645 m s⁻¹ with neither SAR nor radius supervision). These numbers come from a reduced test set where the storm centre falls on a valid SAR pixel, reducing the storm count from 38 to 33, so they are not directly comparable to the table above.
• Improvements hold across complete storm life cycles. For Hurricane Humberto (North Atlantic, 2019), Kiko (East Pacific, 2025) and Otis (East Pacific, 2023), spanning 3,266 observations, the latent MLP lowers V_max MAE from 9.197 to 7.434 m s⁻¹ relative to the simple U-Net, and from 12.203 to 5.441 m s⁻¹ during RI.
• Reconstructions are smooth but structurally useful. In a category 5 example, the predicted field is smoother than the SAR reference yet correctly captures the extent of winds at or above 30 m s⁻¹ and fills in regions beyond the edge of the SAR swath.
• Bigger models did not help. Initial exploration of more complex, higher-parameter architectures did not outperform the simple approach, which the authors attribute largely to the limited amount, quality and balance of training data.

Methodology in Plain English

The authors gather three kinds of data and line them up in time and space. SAR retrievals from the CyclObs database give the target surface wind fields at 3 km resolution. Imagery from GOES-16 and -18 and Himawari-8 and -9 provides ten shared thermal infrared channels at 2 km nadir resolution and 10 minute cadence. IBTrACS best-track records supply the scalar metrics: maximum sustained wind, radius of maximum wind, and the radii of 34, 50 and 64 kt winds in each quadrant. Each SAR acquisition is matched to the closest geostationary image within ±90 minutes (in most cases within ±15 minutes), and imagery is cropped to 196×196 pixels around the storm with varying centre positions to reduce spatial bias. The dataset is split 60/20/20 into 841/232/212 samples from 115/34/38 storms, keeping all observations of a given storm in one split, and includes 134 RI samples from 62 storms.

The input is the ten geostationary channels plus four hand-made spatial context channels (local solar time as sine and cosine, solar zenith angle, and distance to the storm centre), optionally nine ERA5 channels, and a validity mask. Four models are compared: a baseline that pools encoder features into an MLP predicting only the scalars; a U-Net that reconstructs the wind field and derives V_max and R_max from it; a two-stage model with a frozen U-Net feeding an MLP; and the joint latent model that splits one shared encoder into a field decoder and a scalar MLP. The U-Net loss upweights winds between 25 and 50 m s⁻¹ by factors from 1× to 8× and adds a 0.05-weighted term on the top 0.5 percentile of values; the scalar loss combines V_max and R_max terms with a 0.25 weighting on the radius. All objectives use Huber losses in physical units. Architectures are deliberately small: four-level residual U-Nets with channel widths (32, 64, 128, 256), GroupNorm, SiLU activations and skip connections on 192×192 centre crops, and MLPs with hidden widths 128 and 64 and dropout 0.1.

Why This Matters

Impact on research. The paper reframes TC intensity estimation as a joint field-and-scalar problem rather than a scalar regression, and shows that the auxiliary dense target does real work during the hardest cases. It also provides a clean, controlled ablation over supervision strategies and input sources on a modest dataset, which is useful as a baseline for future work.

Real-world applications:

• Near-real-time intensity and structure monitoring in ocean basins where reconnaissance aircraft do not fly.
• Earlier situational awareness during rapid intensification, when evacuation and disaster-response decisions have the least lead time.
• Estimating integrated kinetic energy and other structure-aware measures of destructive potential, which a single peak wind number cannot capture, for example for offshore infrastructure and shipping exposure.
• Gap-filling between infrequent SAR overpasses and beyond the edges of a SAR swath, as demonstrated in the category 5 reconstruction example.

Industry relevance. Weather forecasting agencies and satellite operators can use the observation-only design because it needs no reanalysis at inference. Catastrophe-risk insurers and reinsurers, offshore energy operators, and marine logistics providers all depend on storm structure and intensity estimates in regions with sparse in-situ coverage.

Future Directions

• Replacing the U-Net and MLP with alternative architectures, notably vision transformers, which the authors flag as promising for geospatial analysis.
• More extensive ablation of the ERA5 inputs to determine whether alternative channel combinations improve performance, plus broader hyperparameter and loss-function tuning, which the authors state they have not yet explored.
• Incorporating additional geophysical variables, such as upper-ocean conditions along the forecast track, to refine intensity estimates.
• Extending satellite coverage by adding Meteosat Second Generation imagery to complete coverage over the Indian Ocean, and addressing the underlying constraints of training data quantity, quality and balance.

Target Audience

Machine learning researchers working on geospatial and climate applications; meteorologists and remote sensing scientists interested in TC structure and intensity estimation; and operational teams at forecasting agencies or insurers evaluating whether satellite-only, near-real-time wind field products are viable. Readers with basic neural network literacy will get the most from it, while the ablations and data-split details make it a practical reference for anyone building similar observational pipelines.

Authors’ abstract

High-resolution tropical cyclone surface wind speeds derived from synthetic aperture radar (SAR) observations can support storm structure and intensity modelling, but remain spatially and temporally sparse and unavailable near-real-time. Geostationary satellites capture cloud tops continuously in space and time, but do not directly resolve surface winds. Here, we present simple neural networks to predict 2D wind fields and storm-scale intensity metrics from geostationary imagery. Through a series of ablation studies, we find that joint prediction of 2D fields and scalar metrics from the model latent space improves the retrieval of the (radius of) maximum wind, especially for rapid intensification events.