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 six cryosphere segmentation datasets spanning five components: supraglacial debris, glacial lakes (two sensing configurations), sea ice, calving fronts, and Antarctic ice-shelf extent. 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 Github/ directory, which contains dataset configs, encoder definitions, and decoder heads built on top of PANGAEA.

📋 Tasks & Datasets

Cryo-Bench includes six benchmark tasks covering five 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
GLB	Glacial Lakes (multi-source)	High Mountain Asia	Sentinel-2, Sentinel-1, terrain (11 bands)	Binary	Slope, Elevation	in review	HuggingFace
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
Shelf-Bench	Ice-Shelf Extent	Antarctica	SAR	Binary	—	TBD	HuggingFace

GLB and Shelf-Bench replace the earlier GLD dataset in this benchmark (GLD is no longer part of Cryo-Bench). GLB's own paper is still in review; a citation will be added once it's public.

---

The dataset contains the exact training, validation, and test splits used in Cryo-Bench, covering all six datasets above.

📥 Download Data

1. Install dependency

pip install huggingface_hub

2. Download all datasets

python download_data.py

3. Download specific datasets only

python download_data.py --datasets GLID GSDD SICD GLB Shelf-Bench

---

🏆 Benchmark Results

Table below reports mIoU (↑) for all models evaluated with frozen encoders and 100% training data using the UPerNet decoder, across all six datasets. Rank (↓) is by six-dataset average mIoU. Baseline models (U-Net, ViT) are trained from scratch. Every value below is parsed directly from run logs (tables_v2.json).

Bold = best performance · Italic = second best

Model	GSDD	GLID	GLB	SICD	CaFFe	Shelf-Bench	Avg. mIoU ↑	Rank ↓
U-Net	73.89	91.58	80.46	20.61	59.82	82.97	68.22	1
TerraMind	74.63	88.26	79.91	33.27	46.64	84.46	67.86	2
DOFA	72.96	92.61	79.39	20.41	50.71	87.08	67.19	3
RemoteCLIP	73.42	90.89	78.09	14.51	56.64	83.59	66.19	4
GFM-Swin	73.00	89.69	78.42	11.55	58.12	85.63	66.07	5
Scale-MAE	73.47	90.91	79.16	6.02	58.19	83.28	65.17	6
CROMA	74.15	78.52	77.68	21.81	42.03	81.15	62.56	7
ViT	74.41	78.14	79.81	15.98	40.25	81.86	61.74	8
S12-MoCo	73.03	75.51	75.45	20.66	36.21	74.97	59.31	9
S12-DINO	71.19	75.69	75.17	20.59	35.58	75.28	58.92	10
S12-Data2Vec	73.68	75.19	74.59	17.25	35.96	75.79	58.74	11
SatlasNet	73.70	77.03	74.74	19.93	33.96	73.00	58.73	12
S12-MAE	73.51	75.71	75.43	11.21	36.99	75.93	58.13	13
Prithvi	70.52	71.11	69.08	15.37	32.01	74.41	55.42	14
SpectralGPT	73.22	70.87	72.82	12.57	32.70	66.75	54.82	15

Encoders are kept frozen for all GFMs; U-Net and ViT are trained from scratch. GLB and Shelf-Bench are new in this update, replacing GLD.

---

🧠 Pretrained Checkpoints

Frozen-encoder checkpoints (UPerNet decoder trained, backbone frozen) are included for all 15 models across all 6 datasets — 90 checkpoints in total, corresponding exactly to the cells in the results table above. Each is a PyTorch state_dict (checkpoint__best.pth) saved at the best-validation-mIoU epoch under the standard PANGAEA training protocol.

checkpoints/
├── <DATASET>/
│   └── frozen/
│       └── <encoder>/
│           └── checkpoint__best.pth

where <DATASET> is one of GSDD, GLID, GLB, SICD, CaFFe, Shelf-Bench and <encoder> is the PANGAEA encoder config name (e.g. croma, dofa, terramind_large, unet_encoder, vit_scratch, ...).

Download a single checkpoint

from huggingface_hub import hf_hub_download

ckpt = hf_hub_download(
    repo_id="Sk-21/Cryo-Bench",
    filename="checkpoints/GLID/frozen/dofa/checkpoint__best.pth",
)

Download all checkpoints

from huggingface_hub import snapshot_download

snapshot_download(repo_id="Sk-21/Cryo-Bench", allow_patterns="checkpoints/*")

U-Net and ViT are trained from scratch (no pretrained backbone), so their "frozen" checkpoint is the same end-to-end trained model reported elsewhere as their baseline result.

📜 License

This project is licensed under the MIT License.

---

🙏 Acknowledgements

Cryo-Bench builds on the PANGAEA benchmark. 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, GLB, SICD, CaFFe, and Shelf-Bench for making their data publicly available.