SDCluster: A clustering based self-supervised pre-training method for semantic segmentation of remote sensing images

This repository is the official implementation of SDCluster: A clustering based self-supervised pre-training method for semantic segmentation of remote sensing images.

Pre-trained Model Weights

Backbone

Model Name	Training Dataset	#params	Download(Google)	Download(Baidu)
Swin Transformer-Small (Swin-S)	Million-AID	49M	Swin-S.pt	Swin-S.pt

Clustering Network

Model Name	Training Dataset	#params	Download(Google)	Download(Baidu)
ClusterEval	Million-AID	55M	ClusterEval.pt	ClusterEval.pt

Pre-training

The main_pretrain.py script demonstrates how to reimplement SDCluster for custom datasets.

Visualization of Clustering Results

The following are some clustering results, which can be output using the vis_cluster.py by loading our prepared models ClusterEval.pt.

Dataset

The Shaanxi building extraction dataset can be available at https://zenodo.org/records/12531251.

Citation

If this work is helpful for your research, please consider citing us.

@article{xu2025sdcluster,
  title={SDCluster: A clustering based self-supervised pre-training method for semantic segmentation of remote sensing images},
  author={Xu, Hanwen and Zhang, Chenxiao and Yue, Peng and Wang, Kaixuan},
  journal={ISPRS Journal of Photogrammetry and Remote Sensing},
  volume={223},
  pages={1--14},
  year={2025},
  publisher={Elsevier}
}

License

This project is for research purpose only.

Acknowledgement

We would like to acknowledge the contributions of public projects, such as Swin-Transformer, leopart, and SlotCon , whose code has been utilized in this repository.