blutjens/meltwaterbench

This is the official dataset for the paper: MeltwaterBench: Deep learning for spatiotemporal downscaling of surface meltwater

This dataset sets up a benchmark for spatiotemporal downscaling. Downscaling is similar to superresolution across space and time. The task is to create high-resolution surface meltwater maps from low-resolution (LR) data streams. The LR data contains passive microwave (PMW), a regional climate model (MARv3.14), a temporal interpolation of the targets from dates excluding the target date (time_interpolate_sar), and a static digital elevation map (DEM).

The generated 100m daily surface meltwater product can be found at interim/runs/unet_smp/data_v1_4/deploy/. You can download individual files by clicking the 'download' button.

For downloading the full dataset and training models use the following or see full instructions at https://github.com/blutjens/meltwaterbench:

git clone https://huggingface.co/datasets/blutjens/meltwaterbench --branch main --single-branch

Reference

If this data is useful for your analysis please consider citing:

@misc{lutjens26meltwaterbench,
      title={MeltwaterBench: Deep learning for spatiotemporal downscaling of surface meltwater}, 
      author={Björn Lütjens and Patrick Alexander and Raf Antwerpen and Til Widmann and Guido Cervone and Marco Tedesco},
      year={2026},
      journal = {arXiv},
      url={https://arxiv.org/abs/2512.12142}, 
}