Export GEO–SAR pairs¶

Use the published dataset when its processed inputs fit your task. Rebuilding an export requires an observation manifest and the original source products; upstream raw training granules are not part of the Hub release.

Export ten GEO bands with ERA5¶

uv run geo2wf-export geo-sar \
  --config configs/config.yaml \
  --data-root /path/to/source-archive \
  --manifest-file /path/to/observation_manifest.csv \
  --geo-channel-set common10 --include-era5 \
  --output-root data/geotiff/geo_sar_10bands_era5 \
  --limit 2

--limit 2 bounds successful pairs per split for an initial check; omit it for an export of the full supplied manifest. Explicit flags override the export section of the config. Supply local source paths: the checked-in export config contains historical machine defaults.

The exporter matches each SAR acquisition to the nearest eligible GEO observation within each storm/split, regrids them onto a shared grid, and writes raw GeoTIFFs, internal masks, manifests, and training-only statistics. Missing channels, unreadable files, or invalid geometry are recorded and skipped.

Settings that define the dataset¶

Setting	Current default / behavior
`geo_channel_set`	`common4`; explicitly select `common10` for the documented models
`grid_size`, `grid_resolution`	256 pixels, 0.027° spacing in EPSG:4326
`closest_match_hours`	0.5 h; the paper describes a wider ±90-minute pairing window
`include_era5`	Explicit opt-in; source fields plus derived wind speed/vorticity
`include_pmw`, `include_ibtracs`	Optional companion and nearest best-track-row export
`center`, `shift_center`	Image-centered crop, shifted when needed to include the storm center

Export defaults therefore do not by themselves recreate the paper's cohort. Retain the source manifest, matching settings, and normalization statistics for each version. center_lat/center_lon describe the raster crop; ibtracs_center_lat/ibtracs_center_lon describe the storm and drive runtime geometry. Scalar training interpolates its own USA_WIND targets rather than using the exporter's nearest-row convenience intensity.

See uv run python scripts/export_geo_sar_geotiffs.py --help for flags and the local data contract for loading the result.