SemiBCD：Semi-Supervised Building Change Detection From Bitemporal Remote Sensing Images Leveraging Visual–Language Models and Consistency Learning

Code for SemiBCD paper: Semi-Supervised Building Change Detection From Bitemporal Remote Sensing Images Leveraging Visual–Language Models and Consistency Learning.

---

🔧 Environment Setup

We recommend using conda:

conda create -n semibcd python=3.9 -y
conda activate semibcd
pip install -r requirements.txt

📥 Download Backbone Pretrained Weights

SemiBCD uses a ResNet-50 backbone. Download the pretrained checkpoint: 👉 ResNet-50

📥 Download Dataset

1. LEVIR-CD-256

👉 LEVIR-CD-256 Dataset
Extract the downloaded file to the data/LEVIR-CD-256/ folder.

2. WHU-CD-256

👉 WHU-CD-256 Dataset
Extract the downloaded file to the data/WHU-CD-256/ folder.

🚀 Run Testing

1. Download pretrained experiment weights

👉 SemiBCD Experiment Weights

2. Run testing

Run WHU test:

python eval.py --config configs/eval_whu_config.yaml --checkpoint ./best.pth

Run LEVIR test:

python eval.py --config configs/eval_levir_config.yaml --checkpoint ./best.pth

🚀 Run Training

Train on WHU-CD:

python experiments.py --exp 48 --run RUN_ID

# e.g. RUN_ID=0 for SemiBCD on WHU-CD with 5% labels
# RUN_ID controls the labeled data ratio:
# 0 → 5% labels
# 1 → 10% labels
# 2 → 20% labels
# 3 → 40% labels

Train on LEVIR-CD:

python experiments.py --exp 47 --run RUN_ID

# e.g. RUN_ID=0 for SemiBCD on LEVIR-CD with 5% labels
# RUN_ID controls the labeled data ratio:
# 0 → 5% labels
# 1 → 10% labels
# 2 → 20% labels
# 3 → 40% labels

✒️ Citation

If you find our work helpful for your research, please consider citing.

@ARTICLE{11414160,
  author={Liu, Wei and He, Jianglin and Zhong, Yuhang and Yu, Yongtao and Luo, Zhiming and Guan, Haiyan and Li, Jonathan},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Semi-Supervised Building Change Detection From Bitemporal Remote Sensing Images Leveraging Visual–Language Models and Consistency Learning}, 
  year={2026},
  volume={64},
  number={},
  pages={1-15},
  keywords={Buildings;Semantics;Remote sensing;Transformers;Uncertainty;Training;Feature extraction;Convolutional neural networks;Annotations;Adaptation models;Building change detection (CD);consistency learning;semi-supervised learning;visual–language model (VLM)},
  doi={10.1109/TGRS.2026.3668369}}

Acknowledgements

SemiBCD is based on SemiCD-VL, SemiVL, UniMatch, APE, and MMSegmentation. We thank their authors for making the source code publicly available.