HydroSym datasets

Built and verified data for HydroSym: A Neuro-Symbolic System for Physically Consistent Flood Mapping. Source code: see the project repository (not yet public). Full provenance, sources and licences for every input are in paper/01-data-sources.md if you have the repo, summarised below.

data/ — Phase 1, flood extent mapping

Sentinel-1 SAR (VV, VH) + Copernicus DEM GLO-30, from Sen1Floods11.

• raw/ — 431 hand-labelled chips (10 events; Bolivia is absent from every official split and is excluded), 512x512, EPSG:4326, plus the DEM reprojected onto each chip's exact grid.
• processed/ — 1,563 tiles, 256x256, 5 channels (vv, vh, vv_minus_vh, dem_norm, slope), split train/val/test = 915/326/322. metadata.csv gives event, date, bounds and water fraction per tile. Each .npz also carries dem_m (raw metres) and the affine transform, for the symbolic hydrology stage.

Sources: Sen1Floods11 v1.1 (Bonafilia et al., 2020, CC BY 4.0), Copernicus DEM GLO-30 (AWS Open Data), JRC Global Surface Water.

data24/ — Phase 2, 24-hour flood forecast

MODIS flood labels + GPM IMERG observed rain + NOAA GEFS / ECMWF forecast rain, gridded to ~1 km over Kerala.

• raw/ — cached source granules (MODIS HDF4, IMERG subsets, GEFS GRIB2).
• gridded/kerala/ — daily layers on the common grid, 2018-2020.
• processed/ — 19 built (X_t, Y_t+24h) samples. Small: monsoon cloud blinds the optical label on most days, which is a property of the source data, not the pipeline.

Licence

Redistributed under CC BY 4.0, matching the most restrictive upstream source (Sen1Floods11). Attribute Sen1Floods11, Copernicus DEM GLO-30, JRC Global Surface Water, MODIS/MCDWD, GPM IMERG and NOAA GEFS when reusing.