README.md

metadata

license: mit
task_categories:
  - image-segmentation

Cryo-Bench 🧊

A Benchmark for Evaluating Geospatial Foundation Models on Cryosphere Applications

Cryo-Bench is a community benchmark that evaluates geospatial foundation models (GFMs) on five cryosphere remote sensing tasks spanning glacial lakes, supraglacial debris, sea ice, and calving fronts. It is built on top of the PANGAEA evaluation protocol using multi-sensor satellite imagery from Sentinel-1/2, Landsat-8, WorldView-2, and historical SAR missions.

---

🧰 Code

The full benchmarking code is available in the Cryo-Bench/ directory, which contains dataset configs, encoder definitions, and decoder heads built on top of PANGAEA.

📋 Tasks & Datasets

Cryo-Bench includes five benchmark tasks covering key components of the cryosphere:

Dataset	Component	Location	Sensors	Classes	Ancillary Data	Paper	Download
GSDD	Supraglacial Debris	Global	Sentinel-2	Binary	Slope, Elevation, Velocity	Article	Zenodo
GLID	Glacial Lakes	Himalayas	WorldView-2, Sentinel-2, Landsat-8, Gaofen-2	Binary	—	Article	Zenodo
GLD	Glacial Lakes	Himalayas	Sentinel-2	Binary	SAR Coherence, Slope, Elevation	Article	Zenodo
SICD	Sea Ice	Canadian & Greenlandic Arctic	Sentinel-1	Multiclass	Incidence Angle	Article	HuggingFace
CaFFe	Calving Fronts	Greenland, Alaska, Antarctic Peninsula	ERS-1/2, Envisat, RADARSAT-1, ALOS PALSAR, TSX, TDX, Sentinel-1	Multiclass	—	Article	PANGAEA

---

📥 Sample Usage (Download Data)

The dataset contains the exact training, validation, and test splits used in Cryo-Bench, covering the SICD, GLID, GLD, GSDD, and CaFFe datasets.

• Install the dependency:

pip install huggingface_hub

• Download all datasets at once using the script provided in the GitHub repository:

python download_data.py

• Download specific datasets only:

python download_data.py --datasets GLID GLD SICD

🏆 Benchmark Results

Table below reports mIoU (↑) for all models evaluated with frozen encoders and 100% training data using the UPerNet decoder. Rank (↓) is averaged across all five tasks. Baseline models (U-Net, ViT) are trained from scratch.

Bold = best performance · Italic = second best

Model	GLID	GLD	SICD	CaFFe	GSDD	Avg. mIoU ↑	Avg. Rank ↓
CROMA	78.52	76.84	24.84	42.03	74.15	59.28	6.60
DOFA	92.61	80.44	19.20	50.71	72.96	63.18	6.20
GFM-Swin	89.68	72.42	18.98	58.13	73.00	62.44	9.40
Prithvi	71.11	75.84	20.59	32.01	70.52	54.01	13.60
RemoteCLIP	90.88	69.52	22.71	56.64	73.42	62.63	8.00
SatlasNet	77.02	77.11	24.04	33.96	73.70	57.17	8.40
Scale-MAE	90.13	72.65	12.90	58.19	73.47	61.47	8.80
SpectralGPT	70.87	78.90	15.98	32.70	73.22	54.33	11.80
S12-MoCo	75.51	77.38	26.09	36.21	73.03	57.64	8.80
S12-DINO	75.69	75.91	27.28	35.58	71.19	57.13	10.20
S12-MAE	75.71	77.39	20.63	36.99	73.51	56.85	8.20
S12-Data2Vec	75.19	77.10	24.15	35.96	73.68	57.22	9.00
TerraMind	88.26	79.10	31.48	46.64	74.63	64.02	3.40
RAMEN	82.17	73.67	16.52	25.10	70.56	57.17	12.80
U-Net (baseline)	91.58	77.51	29.11	59.82	73.89	66.38	2.80
ViT (baseline)	71.58	80.18	16.17	39.90	74.41	56.45	8.00

Encoders are kept frozen for all GFMs. U-Net and ViT are trained from scratch.

📜 Citation

If you use this benchmark in your research, please cite:

@article{kaushik2026cryobench,
  title={Cryo-Bench: Benchmarking Foundation Models for Cryosphere Applications},
  author={Kaushik, Saurabh and Maurya, Lalit and Tellman, Beth},
  journal={arXiv preprint arXiv:2603.01576},
  year={2026}
}

📜 License

This project is licensed under the MIT License.

---

🙏 Acknowledgements

Cryo-Bench builds on the PANGAEA benchmark and the RAMEN framework. We thank the developers of DOFA, TerraMind, Prithvi, SatlasNet, and all other foundation models included in this benchmark. We also thank the dataset authors of GSDD, GLID, GLD, SICD, and CaFFe for making their data publicly available.