This is the code of D. Zhang , “C2F2:Cross-Task Cross-Domain Feature Fusion for Semi-Supervised Change Detection” IEEE Geoscience and Remote Sensing Letters.
In this study, we propose to enhance traditional semi-supervised learning frameworks by leveraging cross-task cross-domain (CTCD) models, which generate complementary features that differ from standard hidden features. The procedure is as follows. Firstly, the standard features obtained from a traditional encoding-decoding structure are fused with attention-augmented complementary features. Secondly, a secondary decoder maps the fused heterogeneous features into the label space to obtain high-quality pseudo-labels, offering more precise guidance for semi-supervised learning on traditional structures. This approach improves pseudo-labels by leveraging the strength of CTCD models, including large pre-trained models, to enhance the semi-supervised learning process of domain-specific and task-specific models. Experimental results on benchmark datasets demonstrate that our proposed approach surpasses state-of-the-art methods.
git clone https://github.com/Undefinedx-Zhang/CCFF.git
pip install -r requirements.txt
Please download the following ".pth" files :
"3x3resnet50-imagenet.pth" to the folder "CCFF/models/backbone/pretrained".
"resnet101-63fe2227.pth" to the "CCFF/CTCD/backbone/pretrained".
CTCD pre-training weight can be trained by yourself from the repo or directly downloaded from here (code:7721).
Additionally, optimal parameters for our model can be downloaded here (code:7721).
The processed LEVIR-CD dataset, and supervised-unsupervised splits can be downloaded here
The processed WHU-CD dataset, and supervised-unsupervised splits can be downloaded here
to the folder "CCFF/datasets".
Then simply run:
python train.py --config configs/config_LEVIR.json
The following table summarizes the required changes in config file to train a model with different percentage of labeled data.
| Setting | Required changes |
|---|---|
| Semi-upervised - 5% labeled data | sup_percent= 5 |
| Semi-upervised - 10% labeled data | sup_percent= 10 |
| Semi-upervised - 20% labeled data | sup_percent= 20 |
| Semi-upervised - 40% labeled data | sup_percent= 40 |
If you find this repo useful for your research, please consider citing the paper as follows:
@ARTICLE{10693550,
author={Zhang, Dongjie and Hong, Yuting and Qiu, Xiaojie and Dong, Li and Yan, Diqun and Peng, Chengbin},
journal={IEEE Geoscience and Remote Sensing Letters},
title={C2F2: Cross-Task Cross-Domain Feature Fusion for Semi-Supervised Change Detection},
year={2024},
volume={},
number={},
pages={1-1},
keywords={Standards;Feature extraction;Decoding;Training;Accuracy;Semisupervised learning;Supervised learning;change detection;remote sensing;semi-supervised learning;cross-task cross-domain models},
doi={10.1109/LGRS.2024.3467260}}
This code is based on "https://github.com/zxt9/FPA-SSCD" and "https://github.com/xiaoqiang-lu/LSST".