CoastalBench-preview

[Paper] [GitHub]

This repository contains a 90-day subset of the full CoastalBench dataset, provided as a lightweight preview.

Dataset Overview

• Time span: 2008-01-02 to 2008-04-01
• Duration: 90 days
• Temporal resolution: Half-hourly (48 time steps per day)
• Spatial resolution: 898 × 598
• Vertical levels: 12 (for 3D ocean features)
• Data type: float16 (This preview uses lower precision compared to the original NetCDF files for reduced storage)

Variables

Ocean and Meteorological Features

Each .zarr file contains 48 half-hourly time steps across 82 features, stacked along the first dimension.

The 82 features include:

• 3D variables (each with 12 vertical levels, treated as 12 separate 2D channels):
  • u: horizontal velocity (east-west)
  • v: horizontal velocity (north-south)
  • w: vertical velocity
  • temp : temperature
  • salt : salinity
  • AKs : vertical mixing coefficient
• 2D surface variables:
  • zeta: sea surface height
  • Pair: surface air pressure
  • Qair: specific humidity
  • Tair: air temperature
  • rain: rainfall rate
  • swrad: shortwave radiation
  • lwrad: longwave radiation
  • Uwind: u-wind component
  • Vwind: v-wind component
• Land-sea mask:
  • mask: binary mask (1 for ocean, 0 for land), included as the 82nd feature

All features are stacked into a single 4D tensor: (features=82, height=898, width=598, time=48)

River Inputs

Each river Zarr file contains 48 half-hourly time steps of 25 features across 4 river locations. These features include:

• river_transport: river transport (1D)
• river_temp: temperature at 12 depth levels
• river_salt: salinity at 12 depth levels

All features are stored in a 3D array with shape (features=25, rivers=4, time=48) and are temporally aligned with the ocean and meteorological data.

File Structure

CoastalBench-preview/
├── ocean_met/
│   ├── 20080102.zarr/
│   ├── 20080103.zarr/
│   ├── ...
│   └── 20080401.zarr/
│
├── river/
│   ├── 20080102.zarr/
│   ├── 20080103.zarr/
│   ├── ...
│   └── 20080401.zarr/

Citation

Please cite our paper if you find this dataset useful for your work:

@inproceedings{coastalbench,
  title={CoastalBench: A Decade-Long High-Resolution Dataset to Emulate Complex Coastal Processes},
  author={Xu, Zelin and Zhang, Yupu and Xiao, Tingsong and Lizaso, Maitane Olabarrieta and Gonzalez-Ondina, Jose M and Liu, Zibo and Chen, Shigang and Jiang, Zhe},
  booktitle={Forty-second International Conference on Machine Learning},
  year={2025},
}