CoDiT - ocean drifter trajectory datasets

Training, validation, and test data for CoDiT, a conditional denoising diffusion model for multi-day ocean drifter trajectory prediction. Two datasets are released: open_ocean — 91 spatial patches (58 train / 16 val / 17 test) across the global open ocean, one patch-year per file. coastline — 7,236 patch-trajectories (bulk in train, with val and test splits) focused on coastal regions. Each patch bundles the drifter GPS trajectories (Global Drifter Program) with co-registered environmental context: HYCOM ocean currents, Copernicus/GLORYS ocean currents, and ERA5 10 m winds. Ground truth GPS trajectories are provided as NetCDF; context is provided as NetCDF/NetCDF4. Contents (8 archives): open_ocean_train.tar.gz, open_ocean_val.tar.gz, open_ocean_test.tar.gz coastline_train.tar.gz, coastline_val.tar.gz, coastline_test.tar.gz open_ocean_configs.tar.gz, coastline_configs.tar.gz — per-patch YAML configs Directory layout after extraction (via the download script in the repo): datasets/{dataset}/{split}/{patch_name}/ copernicus/copernicus_daily_time*.nc era5/era5_time*.nc hycom/hycom_time*_d0.nc4 trajectories/trajectories_time*.nc configs/{dataset}/{split}/{patch_name}.yml Bathymetry (GEBCO) is hosted separately and fetched by the same download script. Reproducing the paper: clone https://github.com/SwissDataScienceCenter/CoDiT and run python scripts/download_data.py --dataset all --bathymetry. Training and evaluation entry points are documented in the repository README. License: CC-BY 4.0. Citation: please cite the CoDiT paper (see repository) alongside this dataset DOI.