JL1-ChangeMambaSCD

ChangeMamba SCD weights trained for the JL1_second setup in SCD-CropLand-HZ.

Files

• Base/, Small/, Tiny/ — one folder per backbone scale; each holds best_model.pth when released
• */ckpt/*_model.pth — periodic training saves (optional)
• You still need VMamba backbone weights on disk for inference (--pretrained_weight_path); those are separate from this repo (same files as in the training YAMLs)

Quick installation

1. Clone SCD-CropLand-HZ — it contains src/ChangeMamba, infer_MambaSCD.py, and dataset notes under docs/.

2. Create a Python 3.10 environment with CUDA PyTorch, install ChangeMamba deps, and build the selective-scan kernel:

conda create -n mambascd python=3.10 -y && conda activate mambascd
pip install torch==2.1.1 torchvision==0.16.1 --index-url https://download.pytorch.org/whl/cu118
cd src/ChangeMamba && pip install -r requirements.txt
cd kernels/selective_scan && pip install . && cd ../../..
pip install rasterio

3. Download this checkpoint tree into the project (paths below assume repo root):

pip install -U huggingface_hub
huggingface-cli download BiliSakura/JL1-ChangeMambaSCD --local-dir models/BiliSakura/JL1-ChangeMambaSCD

4. Download the matching VMamba ImageNet checkpoint for your variant (file names must match --pretrained_weight_path). The training configs use MzeroMiko/VMamba — e.g. vssm_small_0229_ckpt_epoch_238.pth for Small.

More detail: installation.md in the same repository.

Inference

Inference runs infer_MambaSCD.py. Test data must be in ChangeMamba SCD layout (bi-temporal images + labels; list file of basenames). See docs/dataset.md in SCD-CropLand-HZ.

From the repository root, after export PYTHONPATH=src:${PYTHONPATH}:

VMamba-Small + this repo’s Small weights

python src/ChangeMamba/changedetection/script/infer_MambaSCD.py \
  --cfg src/ChangeMamba/changedetection/configs/vssm1/vssm_small_224.yaml \
  --pretrained_weight_path models/MzeroMiko/VMamba/vssm_small_0229_ckpt_epoch_238.pth \
  --resume models/BiliSakura/JL1-ChangeMambaSCD/Small/best_model.pth \
  --dataset JL1_second \
  --model_type ChangeMambaSCD \
  --test_dataset_path datasets/JL1_second/val \
  --test_data_list_path datasets/JL1_second/val.txt \
  --infer_size 256 \
  --result_saved_path results

Other variants — keep --resume aligned with the folder you use and swap backbone config + pretrained file:

Example --resume paths: models/BiliSakura/JL1-ChangeMambaSCD/Base/best_model.pth, .../Small/best_model.pth, .../Tiny/best_model.pth.

Outputs: RGB PNGs under results/<dataset>/ChangeMambaSCD/change_map_T1/ and .../change_map_T2/ (semantic maps masked by predicted change).

Train / reproduce

From the repository root:

./scripts/train_changemamba_scd.sh configs/train_changemamba_scd_vmamba_base.yaml

Checkpoints are written to model_save_path in the YAML (this folder by default).

Citation

@article{chen2024changemamba,
  author={Hongruixuan Chen and Jian Song and Chengxi Han and Junshi Xia and Naoto Yokoya},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={ChangeMamba: Remote Sensing Change Detection with Spatiotemporal State Space Model}, 
  year={2024},
  volume={62},
  number={},
  pages={1-20},
  doi={10.1109/TGRS.2024.3417253}
}