Single Remote Sensing Image Super-Resolution via a Generative Adversarial Network With Stratified Dense Sampling and Chain Training

• Fanen Meng, Sensen Wu, Yadong Li, Zhe Zhang, Tian Feng, Renyi Liu, Zhenhong Du
• IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1-22
• https://ieeexplore.ieee.org/document/10375518
• checkpoint: https://pan.baidu.com/s/1eUANo6besXp5HctEqqRVOA?pwd=0115 or https://drive.google.com/drive/folders/1_ncUya30a-AwaDXkys56AMj2m8eSJjfH?usp=sharing

Our folder structure is as follows:

├── dataset (dataset used by SRADSGAN)
│   ├── sradsgan
│   │   ├── AID
│   │   ├── DOTA
│   │   ├── LoveDA
│   │   ├── RSSCN7_2800
│   │   ├── SECOND
│   │   ├── UCMerced_LandUse
├── SRADSGAN (code)
│   ├── data
│   ├── GDP_x0
│   ├── img
│   │   ├── GF2_HR.tif
│   │   ├── GF2_LR.tif
│   │   ├── Sentinel2.tif
│   ├── model
│   │   ├── amssrn.py   (amssrn model)
│   │   ├── drcan.py    (drcan model)
│   │   ├── dssr.py     (dssr model)
│   │   ├── edsr.py     (edsr model)
│   │   ├── hat.py      (hat model)
│   │   ├── ndsrgan.py  (ndsrgan model)
│   │   ├── sradsgan.py (sradsgan model)
│   │   ├── sragan.py   (sragan model)
│   │   ├── srgan.py    (srgan model)
│   ├── utils
│   ├── main_amssrn.py
│   ├── main_drcan.py
│   ├── main_dssr.py
│   ├── main_edsr.py
│   ├── main_hat.py
│   ├── main_ndsrgan.py
│   ├── main_sradsgan.py
│   ├── main_sragan.py
│   ├── main_srgan.py
│   ├── Scene_classification_mfe.py (scene classification)

• SRADSGAN (traditional generative model architecture)

  • Contains nine super-resolution models: ['EDSR', 'SRGAN', 'DRCAN', 'DSSR', 'SRAGAN', 'NDSRGAN', 'AMSSRN', 'HAT', 'SRADSGAN']
  • GDP_x0 (diffusion model architecture): This project is based on [sr3] and [GDP]
  • Scene_classification_mfe.py: Scene classification experiments for super-resolution results

The HAT model uses python 3.8, pytorch 1.9, tensorflow-gpu 2.1.0, and the environment of other models is in requirements.txt

pip install -r requirements.txt

We used six datasets to train our model. After secondary processing, we obtained a total of about 35,000 images of 216*216 size.

• Train

  • ["AID", "DOTA", "LoveDA", "RSSCN7_2800", "SECOND"]

• Test

• ["UCMerced_LandUse"]

• Link: https://drive.google.com/drive/folders/1e4VRUgFL4bDRfrb0CS6-Y9f7bTZhBqiV?usp=sharing or

  ​ https://pan.baidu.com/s/1cXkKu-CI6Q8EF_7Kbfxodg?pwd=w5t8 提取码：w5t8

1. Prepare environment, datasets and code.
2. Run training / evaluation code. The code is for training on 1 GPU.

# sradsgan
cd SRADSGAN
python main_sradsgan.py
---------------------------------------------------------------
net.train()                       # train
net.mfeNew_validate()             # test
net.mfeNew_validateByClass()      # classes test
net.mfe_test_single()             # infer
---------------------------------------------------------------
# GDP_x0
cd GDP_x0
python sr_mfe.py -p train -c config/gdp_train_27_216.json   # train
python sr_mfe.py -p val -c config/gdp_test_27_216.json      # test
python sr_mfe.py -p val -c config/gdp_GF2_x3.json           # infer

Fig. 5. Visualization of different methods on UC Merced dataset. From (a) to (e) are x2, x3, x4, x8 and x9 SR results, respectively.

Fig. 8. 3-time super-resolution results of different methods on GaoFen-2 remote sensing image.

Fig. 9. 9-time super-resolution results on Sentinel-2 remote sensing image without downsampling.

If our code helps your research or work, please consider citing our paper.

F. Meng et al., "Single Remote Sensing Image Super-Resolution via a Generative Adversarial Network With Stratified Dense
Sampling and Chain Training," in IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1-22, 2024, Art no. 5400822,
doi: 10.1109/TGRS.2023.3344112.