This repo contains official implementation of the paper Cross-modal change detection flood extraction based on convolutional neural network.

The code is based on MMSegmentation (version 0.20.0).

1. Install mmcv and other dependencies following the MMSegmentation instructions, mmsegmentation/get_started.md at v0.20.0 · open-mmlab/mmsegmentation (github.com).

2. Clone this repo.

   git clone https://github.com/CAU-HE/CMCDNet.git
   

3. Create the data directory to hold the CHU-Flood dataset.

   mkdir data
   

4. Download CAU-Flood from https://pan.baidu.com/s/1i5yxdfwjP-oTyiRmq6FZHQ (rnx6), extract the train.tar.gz and test.tar.gz to the data folder.

5. The code and data should be organized like this:

|- data
| |- train
| | |- flood_vv # the ground gruth flood map
| | |- vv       # the post-event SAR images
| | |- opt      # the pre-event optical images
| |- test
| | |- flood_vv 
| | |- vv       
| | |- opt      
|- cmcdnet

CMCDNet was implemented in mmseg/models/backbone/cmcd.py. We also created a new dataset named WCDataset to read samples from the CHU-Flood dataset.

The configuration files are in my_scripts/water_change, alter batch size, normalization type and other parameters as you need.

Change to the code directory:

cd cmcdnet

Single GPU train:

python tools/train.py my_scripts/water_change/opt_sar_cmcd_r50-effb2_30e.py

Multi-GPU train:

bash tools/dist_train.sh my_scripts/water_change/opt_sar_cmcd_r50-effb2_30e.py {num_gpus}

If you find this repo useful for your research, please consider citing our paper:

@article{HE2023103197,
title = {Cross-modal change detection flood extraction based on convolutional neural network},
journal = {International Journal of Applied Earth Observation and Geoinformation},
volume = {117},
pages = {103197},
year = {2023},
issn = {1569-8432},
doi = {https://doi.org/10.1016/j.jag.2023.103197},
url = {https://www.sciencedirect.com/science/article/pii/S1569843223000195},
author = {Xiaoning He and Shuangcheng Zhang and Bowei Xue and Tong Zhao and Tong Wu},
}