CryoFM
A foundation model for the cryosphere: a ViT-B/8 encoder pretrained with a masked autoencoder on CryoMesh (877,881 Sentinel-2, Sentinel-1 and Copernicus DEM chips over glacierised terrain). It accepts any subset of ten bands; missing bands are treated as absent, not zero-filled.
Code, notebooks and the full description: anonymous code repository.
Files
| file | content |
|---|---|
config.json, cryofm_encoder.safetensors |
pretrained encoder (122.5M parameters) |
tasks/<DATASET>/config.json |
bands, classes, normalisation and paper scores of a Cryo-Bench task |
tasks/<DATASET>/frozen.safetensors |
UPerNet head trained on the frozen encoder |
tasks/<DATASET>/finetuned.safetensors |
encoder and UPerNet head fine-tuned together |
Tasks: GSDD, GLID, GLB, SICD, CaFFe, Shelf-Bench. All task models are the paper's runs (seed 234).
Bands
B2_blue, B3_green, B4_red, B8_nir, B11_swir1, B12_swir2 (Sentinel-2 L2A, DN), slope, elevation
(Copernicus DEM, degrees and m), VV, VH (Sentinel-1 GRD, dB). Normalise each image per band
(zero mean, unit standard deviation); SICD uses fixed dataset statistics (in its config.json).
Usage
from cryofm import load_encoder, load_task_model, predict_geotiff
encoder = load_encoder("CryoFM-Ano/CryoFM") # features for a new task
model, cfg = load_task_model("GLID", "finetuned", "CryoFM-Ano/CryoFM",
bands=["B4_red", "B3_green", "B2_blue"])
pred = predict_geotiff(model, "scene.tif", "scene_glacial_lakes.tif")
Cryo-Bench test mIoU (seed 234)
| GSDD | GLID | GLB | SICD | CaFFe | Shelf-Bench | Avg | |
|---|---|---|---|---|---|---|---|
| frozen | 79.28 | 88.10 | 82.89 | 27.37 | 62.65 | 84.03 | 70.72 |
| fine-tuned | 79.11 | 91.43 | 84.63 | 31.89 | 67.07 | 87.76 | 73.65 |
Limitations
Band independence covers this ten-band vocabulary only; other bands have no embedding. A task model used with fewer bands than it was trained on still runs, but accuracy can drop.
Pretraining data
CryoFM-Ano/CryoMesh: 2,373 ten-band scenes (2021) over glaciated and ice-covered regions, with georeferencing and RGI 7.0 region labels per scene.
License
Apache-2.0. Pretraining data contain modified Copernicus Sentinel and Copernicus DEM data.
- Downloads last month
- 63