OSFNet

Optical Cloud Removal with Microwave Image Structural Information Fusion and Spatial-Frequency Alignment

This repository provides the official PyTorch implementation of OSFNet, the method proposed in our paper 融合微波图像结构信息的空频对齐光学去云方法.

Highlights

• Structure-aware SAR guidance. A dedicated SAR branch based on ReAFConv extracts cloud-independent geometric and texture cues while reducing modality interference.
• Decoupled spatial-frequency modeling. SSAB and FSAB model optical spatial context and frequency degradation separately, avoiding premature feature mixing.
• Adaptive spatial-frequency alignment. SF-AlignNet performs correlation matching and adaptive fusion to recover reliable structures under stripe clouds, fragmented clouds, and high cloud coverage.

Quantitative Results

Dataset	PSNR (dB)	SSIM	SAM	MAE	ColorLoss
SMILE-CR	30.8528	0.8451	-	-	-
SEN12MS-CR	29.0844	0.8940	7.6636	0.02725	0.1508

Datasets

The datasets are not redistributed in this repository. Download them from their official pages:

• SEN12MS-CR dataset
• SEN12MS-CR-TS official split
• SMILE-CR dataset

Organize SEN12MS-CR as follows:

SEN12MS-CR/
├── train/
│   ├── s1/
│   ├── s2_cloudfree/
│   ├── s2_cloudy/
│   └── data.csv
├── val/
│   ├── s1/
│   ├── s2_cloudfree/
│   ├── s2_cloudy/
│   └── data.csv
└── test/
    ├── s1/
    ├── s2_cloudfree/
    ├── s2_cloudy/
    └── data.csv

Organize SMILE-CR as follows:

SMILE-CR/
├── train/
│   ├── s1/
│   ├── L8_cloudfree/
│   ├── L8_cloudy/
│   └── train_data.csv
├── val/
│   ├── s1/
│   ├── L8_cloudfree/
│   ├── L8_cloudy/
│   └── val_data.csv
└── test/
    ├── s1/
    ├── L8_cloudfree/
    ├── L8_cloudy/
    └── test_data.csv

If the SMILE-CR CSV files are unavailable, generate them with:

python create_smile_csv.py

Installation

conda create -n osfnet python=3.8
conda activate osfnet
pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

Training and Evaluation

Train and evaluate on SEN12MS-CR:

python train_test.py

Train and evaluate on SMILE-CR:

python train_test_smile.py

Evaluate a trained SMILE-CR model:

python test.py --weight_path /path/to/weight.pth

Dataset locations, CSV paths, checkpoint directories, and training parameters can be changed through the command-line arguments defined in each script.

Citation

The manuscript is currently under review. Citation information will be updated after publication.

Contact

For questions or suggestions, please open a GitHub issue.