# SaDL_CD Here, we provide the pytorch implementation of the paper: Semantic-Aware Dense Representation Learning for Remote Sensing Image Change Detection. For more ore information, please see our published paper at [IEEE TGRS](https://ieeexplore.ieee.org/document/9874899/) or [arxiv](https://arxiv.org/abs/2205.13769). ![overview](images/overview.png) ## Requirements ``` Python 3.7 pytorch 1.10.1 torchvision 0.11.2 einops 0.3.0 kornia 0.6.3 ``` ## logs 20230407: update the download link of the cropped Inria data. ## Installation Clone this repo: ```shell git clone https://github.com/justchenhao/SaDL_CD.git cd SaDL_CD ``` ## Quick Start You can simply run `python main_ssl.py` to train our model (`sadl_fpn_m2_resnet18_sample16_syn1`) on the given small samples (in the folder `samples`). ## Training You can find the training script `train_ssl.sh` in the folder `scripts`. You can run the script file by `sh scripts/train_ssl.sh` in the command environment. The detailed script file `train_ssl.sh` is as follows: ```shell #!/usr/bin/env bash gpus=0 checkpoint_root=checkpoints dataset_type=SegDataset img_size=256 batch_size=64 optim_mode=sgd lr_policy=poly lr=0.01 max_epochs=200 net_G=sadl_fpn_m2_resnet18_sample16_syn1 data_name=inria256 split=pos0.1_train split_val=pos0.1_val project_name=SSLM_${net_G}_${data_name}_b${batch_size}_lr${lr}_${split}_${split_val}_${max_epochs}_${lr_policy}_${optim_mode} python main_ssl.py --dataset_type ${dataset_type} --img_size ${img_size} --optim_mode ${optim_mode} --checkpoint_root ${checkpoint_root} --lr_policy ${lr_policy} --split ${split} --split_val ${split_val} --net_G ${net_G} --gpu_ids ${gpus} --max_epochs ${max_epochs} --project_name ${project_name} --batch_size ${batch_size} --data_name ${data_name} --lr ${lr} ``` ## Dataset Preparation ### Pretraining dataset We leverage image-mask pairs from the existing Inria building segmentation dataset. We cut the original samples into small patches of size 256 × 256. We additionally obtain the coregistered image patch of the corresponding geospatial region. The original Inria building segmentation dataset can be found at: https://project.inria.fr/aerialimagelabeling/ Our processed pretraining dataset can be accessed by Baidu yun (code: 2p1b): [link](https://pan.baidu.com/s/1ObAu_4Xm3SWCdMvAfiN7yw) Note that you need only the image-mask pairs in `A` and `label` to train our model. We also provide the spatially registered image of another temporal in `B` for possible usage. #### Data structure ``` """ The pretraining data set with bitemporal images and building mask for one temporal. Note that the masks in the folder 'label' are aligned with the corresponding images in folder 'A'; ├─A ├─B ├─label └─list """ ``` #### Data Download ### Downstream Datasets We test our pretrained model at three downstream change detection datasets. #### Data structure ``` """ Change detection data set with pixel-level binary labels; ├─A ├─B ├─label └─list """ ``` #### Data Download LEVIR-CD: https://justchenhao.github.io/LEVIR/ WHU-CD: https://study.rsgis.whu.edu.cn/pages/download/building_dataset.html GZ-CD: https://github.com/daifeng2016/Change-Detection-Dataset-for-High-Resolution-Satellite-Imagery ## License Code is released for non-commercial and research purposes **only**. For commercial purposes, please contact the authors. ## Citation If you use this code for your research, please cite our paper: ``` @Article{chen2022, title={Semantic-Aware Dense Representation Learning for Remote Sensing Image Change Detection}, author={Hao Chen, Wenyuan Li, Song Chen and Zhenwei Shi}, year={2022}, journal={IEEE Transactions on Geoscience and Remote Sensing}, volume={}, number={}, pages={1-18}, doi={10.1109/TGRS.2022.3203769} } ```