Code for this paper.

Blind Super-Resolution of Single Remotely Sensed Hyperspectral Image, IEEE TGSR (2023)

Zhiyuan Liang, Shuai Wang, Tao Zhang, and Ying Fu.

In this paper, we introduce a two-step framework for blind remotely sensed HSI super-resolution, where the degradation is unknown.

• Python >= 3.6, PyTorch >= 1.7.1

conda create -n tsbsr python=3.6
conda activate tsbsr

conda install -c conda-forge python-lmdb
conda install caffe
pip install --upgrade git+https://github.com/pytorch/tnt.git@master

pip install -r requirements.txt

• Pavia Centre and University
• Washington DC Mall
• Salinas
• Kennedy Space Center
• Botswana
• Houston2013
• Houston2018

Prepare AID training set for training SwinIR.

python utils/gen_data/gen_train.py

• Train SwinIR

  python main.py -a swinir -p aid_blind --bandwise --lr 1e-4 -mlr 1e-5 -n 40 --ri 1 --dir /media/exthdd/datasets/hsi/lzy_data/AID/AID_64_Y.db -b 32  # Please change --dir to your AID.db path

Take Pavia University dataset as example.

• Generate the pseudo HR HSI

  # 1. Generate testing noises
    python utils/transforms/noise/gen_noise.py -ht 340 -w 340 -b 103
  # 2. Generate LR HSI
    python utils/gen_data/gen_test.py # generate norm test
    python utils/gen_data/gen_lr.py --sigma 10 -k k1 --sf 4 -dn paviau
  # 3. Apply SwinIR to the LR HSI, generating pseudo HR HSI
    python test.py -a swinir -p noise10_k1_aid --noise 10 -ds k1 -dn paviau -rp logs/checkpoint/swinir/aid_blind/model_best.pth --bandwise

• Unsupervised training

  python main_transfer.py -a cnn_103_128_64_blind_sparse -p noise10_paviau -ds k1 --noise 10 --dir logs/result/swinirY/paviau/noise10_k1_aid/PaviaU.mat -n 400 --lr 1e-3 -mlr 5e-5 --ri 50 -b 1

python test_transfer.py -a cnn_103_128_64_blind_sparse -p noise10_k1 --noise 10 -ds k1 -dn paviau --dir logs/result/swinirY/paviau/noise10_k1_aid -fn PaviaU.mat -rp logs/checkpoint/cnn_103_128_64_blind_sparse/noise10_paviau/model_best.pth

If you find this work useful for your research, please cite:

@ARTICLE{liang2023blind,
  author={Liang, Zhiyuan and Wang, Shuai and Zhang, Tao and Fu, Ying},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Blind Super-Resolution of Single Remotely Sensed Hyperspectral Image}, 
  year={2023},
  volume={61},
  number={},
  pages={1-14},
  doi={10.1109/TGRS.2023.3302128}
}