Vous avez un workflow ComfyUI ? Obtenez une documentation lisible directement dans votre navigateur.Rendre un workflow lisible.Essayer maintenant ↗

Obtenir l’application

Cryosentinel Terramind v3 | ST Studio

Readme11 juil.huggingface.co/abzal-glw/cryosentinel-terramind-v3

--- license: apache-2.0 library_name: pytorch tags: - earth-observation - remote-sensing - semantic-segmentation - glacial-lakes - GLOF - foundation-model - terramind - multimodal - sentinel-2 - sentinel-1 - copernicus-dem - high-mountain-asia - tien-shan - climate-risk datasets: - abzal-glw/cryosentinel-glof-v3 language: - en metrics: - iou - f1 pipeline_tag: image-segmentation base_model: ibm-esa-geospatial/TerraMind-1.0-large model-index: - name: cryosentinel-terramind-v3 results: - task: type: image-segmentation name: Glacial-lake semantic segmentation dataset: type: abzal-glw/cryosentinel-glof-v3 name: CryoSentinel multimodal chips v3 (Almaty corridor) split: validation metrics: - type: iou value: 0.9557 name: IoU @ thr 0.5, TTA - type: iou value: 0.9596 name: IoU @ best thr 0.70, TTA - task: type: image-segmentation name: Glacial-lake semantic segmentation dataset: type: abzal-glw/cryosentinel-glof-v3 name: CryoSentinel multimodal chips v3 (Almaty corridor) split: test metrics: - type: iou value: 0.8918 name: IoU @ thr 0.5, TTA (raw) - type: iou value: 0.9082 name: IoU @ thr 0.5, TTA (label-corrected, n=658) --- # 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](https://github.com/abzalabdrash/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...