|
A semantic prior-enhanced semi-supervised method for on-board change detection
|
Authors: Kexin Zhu, Zizhen Zhang, Jun Pan
Status: Final
Date of publication: 30 September 2026
Published in: ITU Journal on Future and Evolving Technologies, Volume 7 (2026), Issue 3, Pages 409-429
Article DOI : https://doi.org/10.52953/BCLF1365
|
Abstract:
Change Detection (CD) plays a crucial role in Earth observation applications such as environmental assessment, land use analysis and disaster monitoring. As space information systems increasingly demand on-board data processing capabilities, CD models are required to deliver reliable performance under conditions where labelled samples are scarce and data distributions are unknown. Conventional supervised approaches, however, rely heavily on large-scale annotated datasets and often generalize poorly to unseen on-board data, resulting in noticeable performance degradation. To address these limitations, this paper proposes a semantic prior-enhanced semi-supervised method for on-board CD. The proposed approach incorporates general semantic priors derived from vision foundation models, thereby alleviating the limited semantic representation capacity caused by insufficient annotations and improving the detection of complex real-world changes. To balance the integration of rich semantic priors with the strict computational constraints of on-board scenarios, a semantic prior-enhanced encoder based on DINOv3 fine-tuning is proposed, alongside a lightweight frequency-separated feature adaptation mechanism. This design preserves transferable visual knowledge while substantially reducing fine-tuning overhead, enabling efficient adaptation to the domain discrepancies of remote sensing imagery and yielding more discriminative feature representations. In addition, a complementary cross-temporal feature perturbation strategy is proposed to better exploit latent information within unlabelled data and encourage the learning of robust change features. Experiments conducted on three public datasets and real-world scenarios demonstrate that the proposed method consistently outperforms existing state-of-the-art approaches under limited annotation settings. |
Keywords: Change detection, deep learning, remote sensing, semi-supervised learning, vision foundation models
Rights: © International Telecommunication Union, available under the CC BY-NC-ND 3.0 IGO license.
|
|
| ITEM DETAIL | ARTICLE | PRICE | |
|---|
ENGLISH 
Full article (PDF) |
|
| 0
| Free of charge | DOWNLOAD |
|
|
| |