Awesome

Contributors:

Weikang Yu, Xiaokang Zhang, Richard Gloaguen, Xiao Xiang Zhu, Pedram Ghamisi

News

11.11.2024 UCD is open to everyone! Be a contributor by sending a pull request!

11.11.2024 Codes for UCD have been released, if you find any problems or bugs, please leave us a message.

09.11.2024 Our paper of MineNetCD has been published on IEEE TGRS 2024, the repo for MineNetCD is available here.

09.07.2024 The UCD project is announced on IEEE IGARSS 2024, we are organizing the codes.

Environment Preparation:

Create a conda environment for UCD

conda create -n ucd python=3.10
conda activate ucd
conda install pytorch torchvision pytorch-cuda=12.1 -c pytorch -c nvidia
pip install -r requirements.txt

Configurate the accelerate package:

accelerate config

How to use:

To train a model:

accelerate launch train.py --config $CONFIG$

An example of CONFIG can be configs/DTCDSCN_MNCD256.yml

Or if you want to run the framework in Singularity & Slurm, you can use this command:

srun singularity exec --env PREPEND_PATH=/home/yu34/.local/bin --nv /home/yu34/ucd.sif accelerate launch train.py --config $CONFIG$

To test a model:

accelerate launch test.py --model $PRETRAINED_MODEL_PATH$ 

An example of PRETRAINED_MODEL_PATH can be a local path like checkpoints/MNCD256/ResUnet/BestF1/ or a Huggingface hub id like HZDR-FWGEL/UCD-MNCD256-ResUnet

To push a pretrained model to Huggingface hub in the UCD format:

accelerate launch test.py --model $LOCAL_PATH$ --batch-size 10 --push-to-hub $HUB_NAME$

The example of LOCAL_PATH can be checkpoints/MNCD256/ResUnet/BestF1/, the example of HUB_NAME can be HZDR-FWGEL/UCD-MNCD256-ResUnet

To transfer a model from torch format (e.g., pytorch_model.bin) to UCD compatible format:

You should prepare a config file in your directory, an example can be found via configs/MineNetCD_MNCD256.yml

accelerate launch test.py --external-config configs/MineNetCD_MNCD256.yml  --batch-size 10 --push-to-hub HZDR-FWGEL/UCD-MNCD256-ChangeFFT

The use_external_checkpint will be set to false before uploading the model to the hub.

To calculate model parameters & FLOPs:

python params_sum.py --config $CONFIG$

Available Models:

Available Datasets:

Our Results:

Here are results derived from the UCD

CLCD256:

Dataset for these implementations: ericyu/CLCD_Cropped_256

GVLM256:

Dataset for these implementations: ericyu/GVLM_Cropped256

EGYBCD:

Dataset for these implementations: ericyu/EGY_BCD

LEVIRCD256:

Dataset for these implementations: ericyu/LEVIRCD_Cropped256

SYSUCD:

Dataset for these implementations: ericyu/SYSU_CD

MineNetCD256:

Dataset for these implementations: HZDR-FWGEL/MineNetCD256

Tutorial Avaiable!

We just added a very simple example as a tutorial for those who are interested in change detection, check here for more details.

Future Development Schedule:

We will implement more models and datasets. If you are interested in this project and want to make any contributions, please send a pull request and we will add your names under the contributors!

If you have any questions or meeting any difficulties when using this framework, please leave us with an issue or you can contact us with email address: w.yu@hzdr.de

Acknowledgement:

We would like to thank Huggingface for providing a wonderful open-source platform. We would also like to thank all the authors and contributors who open-sourced the datasets and models that we incorporated into the UCD platform.

Citation

If you find MineNetCD useful for your study, please kindly cite us:

@ARTICLE{10744421,
  author={Yu, Weikang and Zhang, Xiaokang and Gloaguen, Richard and Zhu, Xiao Xiang and Ghamisi, Pedram},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={MineNetCD: A Benchmark for Global Mining Change Detection on Remote Sensing Imagery}, 
  year={2024},
  volume={},
  number={},
  pages={1-1},
  keywords={Data mining;Remote sensing;Feature extraction;Benchmark testing;Earth;Transformers;Annotations;Graphical models;Distribution functions;Sustainable development;Mining change detection;remote sensing;benchmark;frequency domain learning;unified framework},
  doi={10.1109/TGRS.2024.3491715}}