GravBathy-Diff dataset: paired SRTM15+ and multibeam bathymetry with SWOT-derived marine gravity fields for seafloor super-resolution

This archive contains the complete dataset used to train and evaluate GravBathy-Diff, a gravity-conditioned diffusion model for 5x seafloor bathymetry super-resolution. The dataset pairs two independently observed bathymetric products with co-registered marine gravity fields derived from SWOT wide-swath altimetry. The high-resolution reference is NCEI multibeam bathymetry at approximately 3 arc seconds (about 90 m); the low-resolution input is SRTM15+ at approximately 15 arc seconds (about 450 m). The low-resolution layer is the released SRTM15+ grid and was not produced by downsampling the reference, which distinguishes this archive from synthetic super-resolution benchmarks. Contents: - 44 large co-registered source tile groups (40 training, 4 testing) covering four ocean regions: the central Pacific, the northwestern Atlantic, the northeastern Atlantic and the western Pacific. - 8,205 model-ready patches (7,078 training, 1,127 testing), each providing six co-registered layers: multibeam bathymetry (320x320), SRTM15+ bathymetry (64x64), the east and north components of the deflection of the vertical, the free-air gravity anomaly, and the vertical gravity gradient (all 64x64). - tile_index.csv and patch_index.csv, which link every patch to its source tile and geographic bounding box and report per-patch depth statistics. All files are single-band float32 GeoTIFFs in EPSG:4326. Depths are in metres, negative below sea level, with no normalisation applied. The training and test sets are split at the level of complete source tiles, so no test patch shares any pixel or neighbouring context with a training patch. See README.md in the archive for the full data descriptor, including grid specification, naming conventions, provenance, processing steps and known caveats.