GeoBASP
The dense and selected pruned ConvNeXt-V2-Tiny U-Net checkpoints from GeoBASP: Geography-Aware Structured Pruning for Landslide-Scar Segmentation in the Nepal Himalaya by Deepak Ghimire. See the published article and v1.0.0 source and reproduction guide.
| Checkpoint | Reported MAC reduction | Test Dice in the operating-point comparison |
|---|---|---|
| Dense reference | 0% | 0.5159 |
| GeoBASP, 25% target | 24.22% | 0.5054 |
The models take four channels in this order: Sentinel-2 B11, B08, B04 and the DEM-derived slope channel. Follow the original preprocessing pipeline and training normalization statistics. The reported results apply to the paper's spatial splits in the Koshi Basin.
Download and evaluate
Install the Hugging Face CLI with pip install huggingface_hub. From the GeoBASP v1.0.0 checkout, set up the dependencies and prepare the data using the reproduction guide. Then download the pruned checkpoint and select its probability threshold on the validation split before running the test split.
hf download ghimiredhikura/GeoBASP models/convnextv2_tiny/geobasp_mac25_seed2028.pt --local-dir .
python main.py threshold --checkpoint models/convnextv2_tiny/geobasp_mac25_seed2028.pt --step 0.005 --output generated/threshold.json
python main.py evaluate --checkpoint models/convnextv2_tiny/geobasp_mac25_seed2028.pt --threshold-report generated/threshold.json --split test --output generated/test.json
Attribution and license
These checkpoints derive from ConvNeXt V2 by Woo et al., via the timm ConvNeXt-V2-Tiny weights. Changes include four-channel input, a segmentation decoder, landslide fine-tuning, and structured pruning for the pruned variant. They retain the upstream CC BY-NC 4.0 terms. See NOTICES.md and THIRD_PARTY_LICENSES.txt. Cite the GeoBASP article, the original encoder paper, and the ICIMOD landslide inventory when applicable.
Model tree for ghimiredhikura/GeoBASP
Base model
timm/convnextv2_tiny.fcmae_ft_in1k