CMCD: A Consistency Model-Based Change Detection Method for Remote Sensing Images
Change detection is a key research area in remote sensing, focusing on identifying differences between images captured at different time points and generating change maps. While denoising diffusion probabilistic models have shown preliminary success in this area, the quality of the generated change...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10938396/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Change detection is a key research area in remote sensing, focusing on identifying differences between images captured at different time points and generating change maps. While denoising diffusion probabilistic models have shown preliminary success in this area, the quality of the generated change maps remains unsatisfactory. Furthermore, these methods utilize diffusion networks to extract key features from dual-temporal remote images and generate change maps, yet they often overlook the model's parameter size and the time cost associated with iterative sampling. To address these challenges, we propose a novel consistency model-based change detection method (CMCD), which directly generates high-quality change detection maps in one or a few steps. Specifically, we employ dynamic time interval to prioritize the modeling of challenging image data distributions, enhancing the perception of dual-temporal remote sensing images. Then, we introduce a novel joint loss function to prevent the training collapse of the consistency model caused by errors accumulated from exponential moving average updates. In addition, we propose a new strategy for noise injection that concatenates with one remote sensing image rather than two, thereby reducing noise interference with feature information. We also develop a pruning strategy of skip connections and a top–down feature aggregation module to improve feature utilization efficiency. Extensive experiments demonstrate that CMCD significantly reduces computational complexity and inference time compared to existing diffusion model-based methods. Through extensive experiments on the LEVIR, WHU-CD, and SYSU datasets, our method achieved competitive results, with F1 scores of 91.60%, 92.66%, and 82.26%, respectively. |
|---|---|
| ISSN: | 1939-1404 2151-1535 |