SCAFNet

SCAFNet: Semantic Compensated Adaptive Fusion Network for Remote Sensing Change Detection

This is the official implementation of our paper "SCAFNet: A Semantic Compensated Adaptive Fusion Network for Remote Sensing Images Change Detection".

Abstract Current CNN-Transformer hybrid methods for remote sensing change detection suffer from semantic misalignment and non-adaptive fusion between dual branches, resulting in persistent sensitivity to pseudo-changes. SCAFNet addresses these issues through three innovative components:

Semantic Compensation Module (SCM): Aligns local-global features via cross-attention

CNN-Transformer Feature Adaptive Fusion (CTFAF): Dynamically balances local details and global context

Change Feature Identification Module (CFIM): Enhances true changes while suppressing seasonal variations

Experiments on CDD and WHU-CD datasets demonstrate SCAFNet's superior robustness and accuracy.

Model Architecture SCAFNet employs an encoder-decoder architecture with three core modules:

text Input → Siamese Encoder (CNN + Transformer) → SCM → CTFAF → CFIM → Output Change Map Key Components SCM (Semantic Compensation Module): Bridges semantic gap through cross-attention with self-correlation masking

CTFAF (CNN-Transformer Feature Adaptive Fusion): Adaptive fusion with learnable parameters for dynamic feature recalibration

CFIM (Change Feature Identification Module): Multi-stage feature interaction for robust change identification

Installation Requirements Python 3.8+

PyTorch 1.9+

CUDA 11.0+

Create conda environment

conda create -n scafnet python=3.8 conda activate scafnet

Install PyTorch

conda install pytorch==1.9.0 torchvision==0.10.0 torchaudio==0.9.0 cudatoolkit=11.3 -c pytorch -c conda-forge

Install requirements

pip install -r requirements.txt

requirements.txt

einops==0.8.1 matplotlib==3.8.3 misc==0.0.0 mmengine==0.10.6 numpy==1.23.0 Pillow==12.0.0 thop==0.1.1.post2209072238 timm==1.0.22 torch==2.3.0+cu118 torchvision==0.18.0+cu118 tqdm==4.67.1 ultralytics_yolo==0.0.1

Dataset

The dataset we use is public. Here we provide a link to facilitate readers' access.

CDD: https://drive.google.com/file/d/1GX656JqqOyBi_Ef0w65kDGVto-nHrNs9/edit.

WHU-CD: https://gpcv.whu.edu.cn/data/building_dataset.html.

Directory Structure

Organize your data as follows: text data/

• CDD/
  • train/
    • A/ # Time1 images (10,000)
    • B/ # Time2 images (10,000)
    • label/ # Change masks (10,000)
  • val/ # 3,000 images each
  • test/ # 3,000 images each
• WHU-CD/
  • train/ # 5,947 images
  • val/ # 743 images
  • test/ # 744 images

Train

We suggest using the provided scripts for training or testing.

Here is a command-line startup example provided：

python main_cd.py --img_size ${img_size} --checkpoint_root ${checkpoint_root} --lr_policy ${lr_policy} --split ${split} --split_val ${split_val} --net_G ${net_G} --gpu_ids ${gpus} --max_epochs ${max_epochs} --project_name ${project_name} --batch_size ${batch_size} --data_name ${data_name} --lr ${lr} --optimizer ${optimizer}

Evaluation Metrics

The code automatically computes:

F1-Score

IoU (Intersection over Union)

Precision

Recall

Overall Accuracy

Result in FLOPs、Time and Parameters

Due to the limitation of the number of text pages, we have placed the performance results of the model on size, FLOPs, and both training and inference times in the open-source repository. The compare_in_FLOPs_time_Parameters.JP shows the specific experimental results

We respectfully acknowledge that SCAFNet's computational time, parameters and FLOPs is not the lowest among compared methods. However, this moderate computational investment reflects our deliberate emphasis on achieving breakthrough accuracy rather than optimizing solely for speed. The complex interactions between our novel SCM, CTFAF, and CFIM modules—while computationally demanding—were essential to overcome the fundamental challenges of semantic misalignment and pseudo-change suppression that limit current methods.

Repository Structure

SCAFNet/

• data_preprocess/
  • compute_mean_std.py
• datasets/
  • CD_dataset.py
  • data_utils.py
• models/
  • init.py
  • basic_model.py
  • BitemporalFusion.py
  • cc_attention.py
  • DualFusion.py
  • evaluator.py
  • FCTransformer.py
  • help_funcs.py
  • HybridModel.py
  • losses.py
  • networks.py
  • resnet.py
  • trainer.py
• scripts/
  • eval.sh
  • eval_HY_CDD.sh
  • eval_HY_WHU.sh
  • run_baseline_SCM_CTFF_CDD.sh
  • run_baseline_SCM_CTFF_WHU.sh
  • run_hy_CDD.sh
  • run_hy_WHU.sh
• README.md
• requirements.txt
• compare_in_FLOPs_time_Parameters.jpg

Citation

If you use this code in your research, please cite our paper: bibtex @article{zhang2025scafnet, title={SCAFNet: A Semantic Compensated Adaptive Fusion Network for Remote Sensing Images Change Detection}, author={Zhang, Yunzuo and Zhen, Jiawen and Sun, Shibo and Liu, Ting and Huo, Lei and Wang, Tong}, journal={IEEE Geoscience and Remote Sensing Letters}, year={2025}, publisher={IEEE} }

License

This project is licensed under the MIT License - see the LICENSE file for details.