abzal-glw/cryosentinel-terramind-v3

CryoSentinel — TerraMind-v3 Almaty Corridor Soup

A foundation-model semantic-segmentation checkpoint for glacial lakes from joint Sentinel-1 SAR, Sentinel-2 optical, and Copernicus DEM imagery in High Mountain Asia. Released under Apache 2.0 as the supervision component of the GLOFcast operational early-warning system.

This model card documents the Stage 4b v2 production soup — a uniform weight-space average of five SWA snapshots from the Almaty-corridor finetune. It is the recommended checkpoint for inference.

Key results

Metric	Value
Validation IoU @ thr = 0.5, TTA on (n = 666)	0.9557
Validation IoU @ best thr (0.70), TTA on	0.9596
Held-out test IoU @ thr = 0.5, TTA on (n = 665)	0.8918
Held-out test IoU, label-corrected (n = 658)	0.9082

Comparison to literature — same train/val protocol as Adhikari & Regmi (2025), arXiv:2512.24117:

Reference	Modality	Val IoU
Adhikari & Regmi (2025)	S1 only	0.9130
CryoSentinel (this model)	S1 + S2 + DEM	0.9557

Model details

• Architecture: TerraMind 1.0 Large encoder (1.1 B parameters; Jakubik et al., 2025) + UperNet decoder (channel sequence 256 → 128 → 64 → 32; Xiao et al., 2018) with a LearnedInterpolateToPyramidal neck.
• Inputs: 224 × 224 chips at 10 m/pixel:
  • 12 Sentinel-2 L2A bands (B01–B12 minus B10), uint16 reflectance
  • 2 Sentinel-1 GRD polarisations (VV, VH), float32 dB
  • 1 Copernicus DEM 30 m band, int16 m elevation
• Output: per-pixel water sigmoid logit at the input resolution.
• Training: 30 epochs Stage 4a HMA pretrain → 30 epochs Stage 4b Almaty-corridor finetune.
• Soup: uniform mean of five SWA snapshots (epochs 12, 14, 17, 22, 30) per Wortsman et al. (2022). EMA decay 0.999. Flip-only TTA at validation and test time.
• Loss: BCE (pos_weight = 100) + flat Dice + per-image Lovász softmax + asymmetric Tversky (α = 0.25, β = 0.75) + Focal (γ = 2) + Boundary, with OHEM hard-negative mining.
• Train compute: ~ 12 hours on a Modal H100 80 GB, ~ $50 in cloud credits. Robust checkpoint resume via _sanitize_resume_ckpt.

Files

terramind_v3_finetune_almaty_v2/
  checkpoints/
    soup.ckpt                       ← recommended for inference (this card)
    step001605-iou0.952.ckpt        ← single-best fallback
    last.ckpt                       ← resume-from-here
  diagnostics/
    per_chip_val.parquet
    per_chip_test.parquet
    per_region_breakdown.json
    threshold_sweep.json
  logs/
    metrics.csv
    config.yaml

Intended uses

Direct uses. Operational glacial-lake monitoring in High Mountain Asia. Multi-year change detection by running per-year inference and differencing masks. Research on multi-modal foundation-model fine-tuning, label-noise robustness, and snapshot averaging in segmentation.

Downstream uses. Component of GLOF early-warning systems (CryoSentinel produces the lake-area inputs; the breach forecast lives in GLOFcast). Per-lake hazard scoring pipelines that consume per-pixel water masks. Comparative benchmarking against new foundation models on the public test split.

Out of scope. General water mapping (rivers, irrigation reservoirs, coastal water). The model is intentionally trained to suppress non-glacial water signatures; see docs/LIMITATIONS.md § 4. Areas outside HMA — no out-of-domain validation has been performed.

Limitations

The full list is in docs/LIMITATIONS.md; the headline items:

• Not a forecaster. Segmentation only; no breach probability or time-to-failure output.
• HMA only. Twelve sub-regions trained, no Andes / Alps / Caucasus / Patagonia validation.
• Sub-hectare lakes are unreliable. Minimum trustworthy area ~ 0.5 ha (≈ 50 px at 10 m/px).
• Single-snapshot. No temporal modelling.
• No uncertainty quantification. Single sigmoid logit per pixel.
• Single-seed. All numbers from one run with seed = 42. Multi-seed variance (3 seeds) shipping in v1.1, June 2026; budget allocated.
• Not affiliated. CryoSentinel is a research-stage prototype, not affiliated with «Казселезащита», UNESCO GLOFCA, or any other public agency. See docs/LIMITATIONS.md § 9.

Bias, risks, and ethical considerations

The model is trained on a Kumar & Vijay (2026) inventory that integrates Landsat-8, Sentinel-1, and Sentinel-2 imagery with manual quality control. Documented bias modes:

• Label undersizing on a small fraction (~ 1 %) of chips. We have audited seven instances in the test split where the Kumar polygon undersizes a real lake; the model correctly identifies the larger water body in all seven (see docs/LABEL_NOISE_AUDIT.md). This bias is documented but not corrected at training time, because the inverse direction — chips where Kumar oversizes a small puddle — would require an audit we have not performed.
• Cloud climatology bias. The training imagery is selected from late-summer composites with cloud cover < 30 %. Performance on persistently cloudy regions (eastern Himalaya monsoon, Andes Pacific slope) will degrade; we have not quantified by how much.
• Operational misuse. The biggest risk we are aware of is treating CryoSentinel as a forecaster rather than a segmenter. We document this prominently in the README and LIMITATIONS.md. If you find the model deployed in a way that conflates it with a breach forecaster, please flag it.

How to use

Hugging Face Hub

from huggingface_hub import hf_hub_download
import torch

ckpt_path = hf_hub_download(
    repo_id="abzal-glw/cryosentinel-terramind-v3",
    filename="terramind_v3_finetune_almaty_v2/checkpoints/soup.ckpt",
)
state = torch.load(ckpt_path, map_location="cpu", weights_only=False)
# Pass to a CryoSentinel Lightning module configured with
#   configs/terramind_v3_finetune_almaty_v2.yaml

One-line evaluation

git clone https://github.com/abzalabdrash/Cryosentinel.git
cd Cryosentinel && pip install -e .
bash scripts/reproduce_benchmarks.sh

Reproduces the benchmark table to four decimal places. ~ 25 min, ~ $3 on Modal H100. See docs/REPRODUCING.md for local-GPU notes.

Inference on released chips

modal run scripts/predict_modal.py::predict \
    --hf-data-repo abzal-glw/cryosentinel-glof-v3 \
    --hf-ckpt-repo abzal-glw/cryosentinel-terramind-v3 \
    --ckpt-uri hf://abzal-glw/cryosentinel-terramind-v3/terramind_v3_finetune_almaty_v2/checkpoints/soup.ckpt \
    --config configs/terramind_v3_finetune_almaty_v2.yaml \
    --region ile_alatau --year 2023 --n-chips 30 \
    --gpu L4

Input chip layout: 12 S2 L2A bands → 2 S1 GRD bands (VV, VH dB) → 1 Copernicus DEM 30 m band, all at 10 m/pixel in a 224 × 224 GeoTIFF.

Training data

abzal-glw/cryosentinel-glof-v3: 42,237 chips across 12 HMA sub-regions × 4 years (2017, 2021, 2022, 2023) at 224 × 224 / 10 m / pixel. 30.4 GiB on disk. Built from Kumar & Vijay (2026) labels + ESA Copernicus Sentinel-1/2 + Copernicus DEM 30 m. Publicly downloadable on Hugging Face under ODC-By 1.0; see docs/DATA.md for download instructions and full preprocessing details.

Stage 4b finetune split (after spatial-block-split with 0.15° blocks and 0.02° / 2.2 km buffer):

Split	Chips	Share
Train	4,283	76 %
Val	666	12 %
Test	665	12 %

The split is hash-based, year-invariant, and rules out spatial leakage by construction. See docs/METHOD.md § 4.

Evaluation protocol

• Production checkpoint: soup.ckpt (uniform mean of five SWA snapshots).
• TTA: flip-only (horizontal + vertical, four passes averaged in logit space).
• EMA: decay 0.999, applied at validation and test time.
• Threshold: 0.5 (default headline) and 0.70 (best-on-val).
• Metric: global IoU computed as (sum of intersection) / (sum of union) across all chips in the split. Per-chip mean IoU also reported as a robustness check.

Citation

@software{abdrash2026cryosentinel,
    title   = {CryoSentinel: A Foundation-Model Glacial Lake Segmenter
               for High Mountain Asia},
    author  = {Abdrash, Abzal},
    year    = {2026},
    url     = {https://github.com/abzalabdrash/cryosentinel},
    version = {v1.0.0},
    license = {Apache-2.0},
    doi     = {10.5281/zenodo.20239229}
}

@dataset{abdrash2026cryosentinelglofv3,
    title     = {CryoSentinel-GLOF v3: Multimodal Glacial Lake Chips
                 for High Mountain Asia},
    author    = {Abdrash, Abzal},
    year      = {2026},
    publisher = {Hugging Face},
    url       = {https://huggingface.co/datasets/abzal-glw/cryosentinel-glof-v3},
    doi       = {10.57967/hf/8823}
}

If you use CryoSentinel in operational early-warning, please also cite the underlying TerraMind backbone:

@article{jakubik2025terramind,
    title   = {TerraMind: Large-Scale Generative Multimodality for Earth Observation},
    author  = {Jakubik, Johannes and others},
    journal = {arXiv:2504.11171},
    year    = {2025}
}

And the Kumar & Vijay glacial-lake inventory:

@dataset{kumar2026hma,
    title     = {Inventory of Glacial Lakes in High Mountain Asia for the Years 2016 and 2022},
    author    = {Kumar, Ravindra and Vijay, Saurabh},
    year      = {2026},
    publisher = {PANGAEA},
    doi       = {10.1594/PANGAEA.983845}
}

Acknowledgments

• TerraMind by IBM Research, the European Space Agency Φ-lab, and the FAST-EO project (Apache 2.0). Released April 2025.
• TerraTorch by IBM Research (Apache 2.0). The training framework that made the multi-modal patch-embedder dispatch and LLRD profile manageable.
• Modal Labs for the H100 80 GB compute.
• Hugging Face for hosting the model checkpoints and the training dataset.
• The «Казселезащита» institutional GLOF early-warning service in Almaty, Kazakhstan, and the UNESCO GLOFCA programme — the institutional context this work is positioned to complement, not replace. CryoSentinel is not affiliated with either organisation; this acknowledgment recognises their decades of operational work on the same problem.

Contact

Open an issue at https://github.com/abzalabdrash/Cryosentinel/issues for bug reports, reproducibility checks, or out-of-domain validation results.