chengxihan / hanet-cd

HANet: A Hierarchical Attention Network for Change Detection

HANet-Change-Detection :https://chengxihan.github.io/

The Pytorch implementation for::gift::gift::gift: “HANet: A hierarchical attention network for change detection with bi-temporal very-high-resolution remote sensing images,” IEEE J. SEL. TOP. APPL. EARTH OBS. REMOTE SENS., PP. 1–17, 2023, DOI: 10.1109/JSTARS.2023.3264802. C. HAN, C. WU, H. GUO, M. HU, AND H. CHEN, yum::yum::yum:

[14 Aril. 2023] Release the first version of the HANet

Requirement

-Pytorch 1.8.0  
-torchvision 0.9.0  
-python 3.8  
-opencv-python  4.5.3.56  
-tensorboardx 2.4  
-Cuda 11.3.1  
-Cudnn 11.3  

Revised parameters

You can revise related parameters in the metadata.json file.

Training, Test and Visualization Process

python trainHCX.py 
python test.py 
python Output_Results.py

Test our trained model result

You can directly test our model by our provided training weights in tmp/WHU, LEVIR, SYSU, and S2Looking . And make sure the weight name is right. Of course, for different datasets, the Dataset mean and std setting is different.

path = opt.weight_dir+'final_epoch99.pt'

Dataset Download

LEVIR-CD：https://justchenhao.github.io/LEVIR/

WHU-CD：http://gpcv.whu.edu.cn/data/building_dataset.html ,our paper split in Baidu Disk,pwd:6969

SYSU-CD: Our paper split in Baidu Disk,pwd:2023

S2Looking-CD: Our paper split in Baidu Disk,pwd:2023

CDD-CD: Our split in Baidu Disk,pwd:2023

DSIFN-CD: Our split in Baidu Disk,pwd:2023

Note: Please crop the LEVIR dataset to a slice of 256×256 before training with it.

And also we provide all test results of our HANet in the HANetTestResult!!!! Download in HANetTestResult or Baidu Disk,pwd:2023 😋😋😋

Dataset Path Setting

 LEVIR-CD or WHU-CD  
     |—train  
          |   |—A  
          |   |—B  
          |   |—label  
     |—val  
          |   |—A  
          |   |—B  
          |   |—label  
     |—test  
          |   |—A  
          |   |—B  
          |   |—label

Where A contains images of the first temporal image, B contains images of the second temporal images, and the label contains ground truth maps.

Dataset mean and std setting

We calculated mean and std for seven data sets in line 27-38 of utils/datasetHCX , you can use one directly and then annotate the others.

# It is for LEVIR!
# self.mean1, self.std1, self.mean2, self.std2 =[0.45025915, 0.44666713, 0.38134697],[0.21711577, 0.20401315, 0.18665968],[0.3455239, 0.33819652, 0.2888149],[0.157594, 0.15198614, 0.14440961]
# It is for WHU!
self.mean1, self.std1, self.mean2, self.std2 = [0.49069053, 0.44911194, 0.39301977], [0.17230505, 0.16819492,0.17020544],[0.49139765,0.49035382,0.46980983], [0.2150498, 0.20449342, 0.21956162]

PFBS(Progressive Foreground-Balanced Sampling)

you can set Normal Train,Fixed-X,Linear-Y,Fixed-X Linear-Y method in line 113-135 of trainHCX.py .You just need to choose one sampling method, and annotate the others, About 'X' and 'Y', you can set epochs_threshold number in metadata.json.

#Normal Train：正常训练，确保dataloader的方式一样
# train_loader.dataset.curr_num = len(train_loader.dataset)
#Fixed-X:如固定的15个
if epoch < opt.epochs_threshold:
   pass
else:  # 15
   train_loader.dataset.curr_num=len(train_loader.dataset)
#Fixed-X Linear-Y：先固定，后增加，前10个是前景影像，然后线性增加10个，后是正常训练
# if epoch < opt.epochs_threshold:
#     pass
# elif epoch<opt.epochs_threshold+5:
#     train_loader.dataset.curr_num += add_per_epoch
# else:  # 20
#     train_loader.dataset.curr_num=len(train_loader.dataset)
# # Linear-Y：前20个线性增加
# if epoch == 0:
#     pass
# elif epoch < opt.epochs_threshold:  # 20
#     train_loader.dataset.curr_num += add_per_epoch
# else:
#     train_loader.dataset.curr_num = len(train_loader.dataset)

Citation

If you use this code for your research, please cite our papers.

@ARTICLE{10093022,
  author={Han, Chengxi and Wu, Chen and Guo, Haonan and Hu, Meiqi and Chen, Hongruixuan},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing}, 
  title={HANet: A hierarchical attention network for change detection with bi-temporal very-high-resolution remote sensing images}, 
  year={2023},
  volume={},
  number={},
  pages={1-17},
  doi={10.1109/JSTARS.2023.3264802}}

Acknowledgments

Our code is inspired and revised by pytorch-MSPSNet,pytorch-SNUNet, Thanks for their great work!!

Reference

[1] C. HAN, C. WU, H. GUO, M. HU, AND H. CHEN, “HANet: A hierarchical attention network for change detection with bi-temporal very-high-resolution remote sensing images,” IEEE J. SEL. TOP. APPL.EARTH OBS. REMOTE SENS., PP. 1–17, 2023, DOI: 10.1109/JSTARS.2023.3264802.

[2] HCGMNET: A Hierarchical Change Guiding Map Network For Change Detection.

[3]C. Wu et al., "Traffic Density Reduction Caused by City Lockdowns Across the World During the COVID-19 Epidemic: From the View of High-Resolution Remote Sensing Imagery," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 14, pp. 5180-5193, 2021, doi: 10.1109/JSTARS.2021.3078611.

hanet-cd's Issues

无法正确加载您给出的训练结果

Traceback (most recent call last):
File "Output_Results.py", line 39, in
model = torch.load(path)
File "/root/miniconda3/lib/python3.8/site-packages/torch/serialization.py", line 712, in load
return _load(opened_zipfile, map_location, pickle_module, **pickle_load_args)
File "/root/miniconda3/lib/python3.8/site-packages/torch/serialization.py", line 1046, in _load
result = unpickler.load()
File "/root/miniconda3/lib/python3.8/site-packages/torch/serialization.py", line 1039, in find_class
return super().find_class(mod_name, name)
ModuleNotFoundError: No module named 'models.MSPSNetHCX'
我在加载您的训练结果模型时出现了以下情况，似乎在您的训练结果之后还修改过代码结构，不知道您是否方便更新代码，或者给我字典形式的模型参数

No module named 'models.MSPSNetHCX'

Traceback (most recent call last):
File "E:\Futherstudy\change-detection\reproduction\HANet-CD-main\test.py", line 36, in
model = torch.load(path)
File "D:\ANACONDA\envs\d2l\lib\site-packages\torch\serialization.py", line 712, in load
return _load(opened_zipfile, map_location, pickle_module, **pickle_load_args)
File "D:\ANACONDA\envs\d2l\lib\site-packages\torch\serialization.py", line 1049, in _load
result = unpickler.load()
File "D:\ANACONDA\envs\d2l\lib\site-packages\torch\serialization.py", line 1042, in find_class
return super().find_class(mod_name, name)
ModuleNotFoundError: No module named 'models.MSPSNetHCX'

Process finished with exit code 1

i have a question

Author I would like to ask a question, why is the log recorded in the val section of the train code using mean_train_metrics?

Please share your trained models.

I want to check the results, and request you to please share the pre-trained modea l. i do not have good GPU to train.

What