StormSense¶

StormSense follows Humberto, Kiko, and Otis through their full tracks, comparing model wind diagnostics with sparse SAR retrievals and retrospective IBTrACS records. It is a case-study viewer, not an operational forecast product.

Open StormSense Browse observations Scientific data on Hugging Face

Nowcasts¶

The selected ViT, U-Net, or U-Net+MLP series is evaluated at observation time. SAR points are sparse acquisitions; connecting segments only interpolate visually. IBTrACS provides the intensity reference. Optional numerical-model and ERA5 curves are precomputed comparison series.

U-Net+MLP here means the post-hoc intensity correction applied to a frozen field. Spatial diagnostics remain those of that field. The paper's jointly trained latent MLP is a separate model, shown in the paper results. ViT outputs are imported artifacts; its training implementation is outside this package.

Forecasts¶

Forecast mode shows retrospective results at a fixed +12 h lead. Map time is the issue time; the highlighted point is valid 12 hours later.

• MLP: the retained six-hour scalar model applied twice, starting from current and −6/−12 h IBTrACS anchors.
• ConvLSTM: an external artifact using a 12-frame GEO/PMW context and a 12-hour lead; its model implementation is outside this package.

Best-track context can be unavailable in real time, so these layers do not establish operational forecast skill.

Data and delivery¶

The frontend ships with the docs. Generated JSON and image overlays can be served as immutable R2 data releases. Those browser assets are separate from the full scientific dataset on Hugging Face. The case-study manifest provides a data-only CSV and source JSON; model outputs remain in the JSON/dashboard.