CSDNet: Synergy of Content and Style: Enhanced Remote Sensing Change Detection via Disentanglement and Refinement

Official PyTorch implementation of CSDNet, a novel bitemporal change detection network that leverages content-style disentanglement and contextual refinement to achieve robust and high-precision remote sensing change detection. The paper has been accepted at IEEE TGRS!!

Paper

Title: Synergy of Content and Style: Enhanced Remote Sensing Change Detection via Disentanglement and Refinement

Authors: Sijun Dong, Changxin Lu, Siming Fu, Xiaoliang Meng*
School of Remote Sensing and Information Engineering, Wuhan University, Wuhan, China

Abstract:

Bitemporal change detection is often hindered by significant style discrepancies in images, stemming from variations in acquisition time and conditions. To mitigate this, we introduce CSDNet (content–style disentanglement network), a novel bitemporal feature interaction network that leverages content–style disentanglement and a channel gating mechanism. In the feature encoding stage, our Content-Style Disentanglement Module (CSDM) disentangles multi-scale features into content and style components using instance normalization. It then employs a dynamic gating mechanism to selectively preserve style information beneficial for change detection while suppressing background noise. A subsequent feature-level swapping strategy enhances information flow and further aligns the style representations between the bitemporal images. In the decoding stage, the Contextual Content Refiner Module (CCRM) uses a joint channel and spatial gating mechanism to attentively filter and refine the style features. These refined features are then recombined with the content features, enabling a fine-grained delineation of change regions. Extensive experiments on five public datasets—LEVIR-CD, SYSU-CD, S2Looking, WHUCD, and MSRSCD—demonstrate that CSDNet significantly surpasses various state-of-the-art methods in F1-score, IoU, and precision.

The source code and pre-trained weights are available at https://github.com/dyzy41/CSDNet.

Quantitative Results (Test Set Performance)

Dataset	OA	IoU	F1	Recall	Precision
LEVIR-CD	99.16	84.47	91.58	90.00	93.23
SYSU-CD	92.39	71.16	83.15	79.60	87.03
S2Looking	99.24	50.72	67.31	64.34	70.55
WHUCD	99.56	90.88	95.22	95.12	95.33
MSRSCD	93.07	62.01	76.55	76.38	76.73

CSDNet consistently achieves top-tier or competitive results across diverse scenarios, particularly excelling in datasets with strong style discrepancies (e.g., seasonal changes in S2Looking) and high-resolution details (e.g., WHUCD).

running steps

bash install_env.sh
bash run.sh

Citation

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

@ARTICLE{11396066,
  author={Dong, Sijun and Lu, Changxin and Fu, Siming and Meng, Xiaoliang},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Synergy of Content and Style: Enhanced Remote Sensing Change Detection via Disentanglement and Refinement}, 
  year={2026},
  volume={64},
  number={},
  pages={1-16},
  keywords={Feature extraction;Remote sensing;Adaptation models;Accuracy;Noise;Computational modeling;Robustness;Decoding;Change detection algorithms;Buildings;Channel–spatial gating;content–style disentanglement;contextual content refinement;remote sensing change detection},
  doi={10.1109/TGRS.2026.3664457}}

Some other change detection repositories

ChangeCLIP
PeftCD
EfficientCD
Open-CD