DCAP for landslide segmentation
Core checkpoints for Detail–Context Adaptive Structured Pruning for Landslide Segmentation by Deepak Ghimire and Seong-heum Kim. These are the tensor-only checkpoints from the DCAP v1.1.0 release.
| Ensemble | Members | Input |
|---|---|---|
| Dense ResUNet | dense_seed42.pt, dense_seed123.pt, dense_seed2026.pt |
21-channel reflectance and change features |
| DCAP70 | dcap70_seed42.pt, dcap70_seed123.pt, dcap70_seed2026.pt |
Same 21-channel input |
Each ensemble averages the predicted probabilities of its three members. models.json records membership, input mode, counts, and checkpoint hashes. thresholds.json records the thresholds for the released ensembles. Use the project's preprocessing and threshold policy for evaluation.
Download and evaluate
Install the Hugging Face CLI with pip install huggingface_hub. From the root of the v1.1.0 source tree, after installing the project, download and install the six checkpoints.
hf download ghimiredhikura/DCAP --include "*.pt" --include MODEL_NOTICES.txt --local-dir downloads/models
python -m dcap models install --source downloads/models
Follow the reproduction guide to obtain and install the evaluation data, then run the evaluator with --ensemble dense or --ensemble dcap70.
License and citation
These author-trained ResUNet weights are released under the MIT license. See MODEL_NOTICES.txt for provenance and CITATION.cff for citation details.