Skip to content

Tropical-cyclone wind reconstruction

A baseline first. Diffusion second.

geo2wf reconstructs SAR-like surface wind fields from geostationary satellite imagery and ERA5 context. The main system is deliberately two-stage: a deterministic model commits to a broad physical field, then diffusion models the signed correction that remains.

Read the two-stage workflow Understand the data Open the dashboard

The central workflow

Stage 1 Deterministic baseline

GEO, ERA5, geometry, solar context, and validity masks produce one dense wind field anchored to ERA5.

baseline = ERA5 + learned correction
→
Stage 2 Residual diffusion

The Stage 1 checkpoint is frozen. Diffusion samples a signed SAR-minus-baseline residual and adds it back in physical m/s.

wind sample = baseline + sampled residual

The split is useful because the two models have different jobs. Stage 1 learns the stable, broad reconstruction. Stage 2 spends its capacity on plausible eye, eyewall, gradient, and asymmetric structure without having to regenerate the whole field from noise. See the equations, channel counts, and training sequence.

What the models receive

The checked-in two-stage setup uses a 23-channel data condition:

  • 10 GEO infrared and water-vapor bands;
  • 9 ERA5 fields: seven exported variables plus derived 10 m wind speed and relative vorticity;
  • 1 normalized distance-to-IBTrACS-center raster; and
  • 3 solar-time fields.

Validity masks and an explicit ERA5 wind anchor are appended by the model. Stage 2 also receives the frozen Stage 1 field and its validity mask. SAR wind is the supervised target during training; it is not an inference-time input.

Real GEO, ERA5, SAR, and mask example

Real exported sample WP232024_sar_geo_20241030095303_bb2c52ca, rendered from the repository GeoTIFFs with Matplotlib. The data page shows every input family.

Read by task

Understand the modelThe main two-stage baseline + diffusion article. Understand the inputsReal examples, channel lists, masks, and tensor assembly. Run an experimentExport a small batch and launch a smoke run. Choose a presetCompare the stacked workflow with standalone controls. Evaluate structurePhysical error, eye, inner core, radial profile, and RMW. Find a file or commandProject map, configuration keys, and troubleshooting.

Scope

geo2wf is paired image-to-image reconstruction on a shared 256 × 256 geospatial grid. It does not currently model arbitrary observation sets, full storm tracks, or multi-temporal windows. Those boundaries keep the data contract, physical units, masks, and model comparisons inspectable.

Start with the two-stage workflow, then follow the data inputs into training and evaluation.