Paper results¶

The accepted paper studies whether shared spatial and scalar supervision improves tropical-cyclone intensity estimates. These tables retain its reported values, with the cohort distinctions preserved below. RI means an IBTrACS wind increase of at least 30 kt over the preceding 24 hours; it is an evaluation subset, not a separate training set.

Architecture ablation¶

Maximum-wind MAE in m/s on the original architecture test cohort (paper Table 1a).

Model	ERA5	All	RI
Field U-Net diagnostic	No	7.518	17.859
Post-hoc MLP	No	6.715	12.507
Joint latent MLP	No	5.885	8.934
Field U-Net diagnostic	Yes	6.446	14.475
Post-hoc MLP	Yes	5.529	10.512
Joint latent MLP	Yes	7.302	7.716

Without ERA5, the joint model improves on both alternatives, especially during RI. ERA5 improves the field-only and post-hoc models and the joint model's RI error, but worsens the joint model's all-observation error. Its benefit depends on the architecture and regime.

Latent-supervision ablation¶

Maximum-wind MAE in m/s and RMW MAE in km. RMW is the paper's radius of maximum wind, also called Rmax. Dashes mean no scalar radius head. The four no-ERA5 rows correspond to paper Table 1b; the ERA5 rows retain the companion comparison.

ERA5	SAR supervision	Radius supervision	Wind All	Wind RI	RMW All	RMW RI
No	Yes	No	6.694	6.867	—	—
No	No	Yes	6.297	7.198	19.44	13.14
No	No	No	6.828	6.645	—	—
No	Yes	Yes	6.344	5.469	18.24	11.18
Yes	Yes	No	5.795	8.128	—	—
Yes	Yes	Yes	5.859	7.316	19.51	13.78
Yes	No	No	5.446	6.555	—	—
Yes	No	Yes	5.688	6.904	20.71	16.27

Adding SAR reconstruction to the GEO-only radius-supervised model reduces RI wind MAE from 7.198 to 5.469 m/s and RI RMW MAE from 13.14 to 11.18 km. SAR supervision alone does not improve RI intensity in the wind-only control. This supports the paper's argument that the spatial and scalar objectives are complementary.

Cohort provenance

The supplied paper captions Table 1b as a reduced test set with SAR-valid centers. The retained reports and the Hugging Face release identify these latent-ablation values as validation results. They are kept distinct from the architecture test table. Use the release's effective-cohort and training-membership records for reproduction; there is no common split covering every published artifact.

Complete-storm case studies¶

Humberto 2025, Kiko 2025, and Otis 2023 provide 3,266 valid observations, including 466 RI observations. Each prediction is an independent instantaneous nowcast. The joint model here uses SAR and radius supervision with no ERA5 input. These dense storm diagnostics are separate from the paired test table.

Model / reference	All MAE (m/s)	RI MAE (m/s)
GEO-only field U-Net	9.197	12.203
GEO-only latent MLP, SAR + radii	7.434	5.441
ERA5 10 m wind maximum	22.119	40.550

Curves use hourly means and a centered five-hour rolling mean; all metrics use unsmoothed native predictions. Shading marks RI. ERA5 is an external reference: maximum 10 m wind in the same 5.184° storm-centered crop at the nearest analysis time. Two invalid GEO observations are excluded consistently.

Download figure PDF Download full case-study metrics Open StormSense

Wind-field reconstruction¶

These validation examples use the SAR + ERA5 latent model with wind/radius supervision. The orange footprint marks observed SAR; the red cross is the IBTrACS center. Predictions extend beyond the swath, where they are conditional estimates rather than verified wind observations. The field is smoother than the SAR reference, so scalar skill should be read alongside spatial diagnostics.

Detailed reports¶

The retained exports provide additional diagnostics without repeating every matrix here:

See evaluation for masks, units, and aggregation, and paper reasoning for interpretation and limitations.