This is the official PyTorch codes for the paper.

Learning High-frequency Feature Enhancement and Alignment for Pan-sharpening
Yingying Wang, Yunlong Lin, Ge Meng, Zhenqi Fu, Yuhang Dong, Linyu Fan, Hedeng Yu, Xinghao Ding^*, Yue Huang（ *indicates corresponding author)

• SOTA performance: The proposed HFEAN outperforms SOTA pan-sharpening methods over multiple satellite datasets.

• Ubuntu >= 18.04
• CUDA >= 11.0
• NumPy
• Matplotlib
• OpenCV
• PyYAML

# git clone this repository
git clone https://github.com/Gracewangyy/HFEAN.git
cd HFEAN

# create new anaconda env
conda create -n HFEAN python=3.8
conda activate HFEAN

pip install torch numpy matplotlib opencv-python pyyaml

Training dataset, testing dataset are available at Data.

The directory structure will be arranged as:

Data
    |- WV3_data
        |- train128
            |- pan
                |- xxx.tif
            |- ms
                |- xxx.tif
        |- test128
            |- pan
            |- ms
    |-  WV2_data
        |- train128
            |- pan
            |- ms
        |- test128
            |- pan
            |- ms
    |-  GF2_data
        |- ...

To test the trained pan-sharpening model, you can run the following command:

python test.py

The configuration options are stored in the option.yaml file and test.py. Here is an explanation of each of the options:

• algorithm: The model for testing

• algorithm: The algorithm to use for testing.
• type: The type of testing, test
• data_dir: The location of the test data.
• source_ms: The source of the multi-spectral data.
• source_pan: The source of the panchromatic data.
• model: The model path to use for testing.
• save_dir: The location to save the test results.
• test_config_path : The configuration file path for models

• upscale: The upscale factor.
• batch_size: The size of each batch.
• patch_size: The size of each patch.
• data_augmentation: Whether to use data augmentation.
• n_colors: The number of color channels.
• rgb_range: The range of the RGB values.
• normalize: Whether to normalize the data.

Our work is based on the following projects:

• Awaresome-pansharpening

If you find DIRFL is useful in your research, please cite our paper:

@inproceedings{wang2023learning,
  title={Learning High-frequency Feature Enhancement and Alignment for Pan-sharpening},
  author={Wang, Yingying and Lin, Yunlong and Meng, Ge and Fu, Zhenqi and Dong, Yuhang and Fan, Linyu and Yu, Hedeng and Ding, Xinghao and Huang, Yue},
  booktitle={Proceedings of the 31st ACM International Conference on Multimedia},
  pages={358--367},
  year={2023}
}