FSMamba

Code release for the paper:

FSMamba: A Frequency-Enhanced Spatial-Temporal Mamba Network for Local-Global Remote Sensing Image Change Detection

Paper link: IEEE Xplore

Environment

Ubuntu with a CUDA GPU is recommended. The conda environment name is rscd.

conda create -n rscd python=3.8 -y
conda activate rscd

# Example for CUDA 11.8. Change this command if your CUDA version is different.
pip install torch==2.0.1 torchvision==0.15.2 --index-url https://download.pytorch.org/whl/cu118

pip install -r requirements.txt
pip install causal-conv1d==1.2.0.post1
pip install mamba-ssm==1.2.0.post1

If causal-conv1d or mamba-ssm cannot be installed from pip, install wheel files matching your Ubuntu, CUDA, PyTorch, and Python versions.

Data

Put datasets under data/ or edit data_root in configs/_base_/.

Default CLCD path:

data/CLCD/

Training

python train.py -c configs/FSMamba.py

Training logs and checkpoints will be saved under:

work_dirs/clcd_epoch100/FSMamba/

Testing

python test.py \
  -c configs/FSMamba.py \
  --ckpt work_dirs/clcd_epoch100/FSMamba/version_x/ckpts/test/test_change_f1/best.ckpt \
  --output_dir work_dirs/eval_fsmamba_clcd

Evaluation metrics are printed in the terminal. RGB masks are saved to:

work_dirs/eval_fsmamba_clcd/mask_rgb/

Citation

@article{guo2026fsmamba,
  title={FSMamba: A Frequency-Enhanced Spatial-Temporal Mamba Network for Local-Global Remote Sensing Image Change Detection},
  author={Yingting Guo and Yan Xu},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  year={2026}
}