Causal Intervention-Enhanced Feature Disentanglement Network for SAR ATR in Data-Scarce Environments
IEEE Transactions on Aerospace and Electronic Systems · 2026 · 62 · 13332-13347
DOI: 10.1109/taes.2026.3708075 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Aerospace and Electronic Systems, 2026. Verfasst von Haohao Ren und 6 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Fundamental Research Funds for the Central Universities.
Abstract (Originalsprache)
Synthetic aperture radar (SAR) automatic target recognition (ATR) remains challenging under limited training samples and complex scene interference. Due to the coherent imaging mechanism, SAR images often entangle intrinsic target scattering characteristics with scene-dependent artifacts, such as clutter and speckle noise. Under data-scarce conditions, deep models tend to overfit these scene-related confounders, leading to spurious correlations and degraded generalization in cluttered environments. Existing methods mainly alleviate this problem through data augmentation or feature regularization. However, they usually treat sample scarcity and scene variation separately and lack an explicit mechanism to suppress scene-induced confounding effects. To address this issue, this article proposes a causal intervention-enhanced feature disentanglement network for SAR ATR under limited-sample settings. Specifically, a frequency domain-based causal augmentation strategy is developed to enrich training diversity and reduce selection bias caused by insufficient observations. In addition, a global-local feature disentanglement architecture is designed to separate target-discriminative information from scene-related interference. Finally, an intervention-driven invariant strategy mitigates scene-driven spurious correlations and improves representation robustness via cross-sample feature alignment and hierarchical contrastive learning. Experimental results on benchmark SAR datasets demonstrate that CI-FDN consistently outperforms state-of-the-art methods and achieves superior generalization under limited-sample conditions.
Förderung
- National Natural Science Foundation of China
- Fundamental Research Funds for the Central Universities
Zitieren
APA
Ren, H., Miao, L., Qian, L., Zou, L., Zhou, Y., Wang, X., & Tang, H. (2026). Causal Intervention-Enhanced Feature Disentanglement Network for SAR ATR in Data-Scarce Environments. IEEE Transactions on Aerospace and Electronic Systems, 62, 13332-13347. https://doi.org/10.1109/taes.2026.3708075
Vancouver
Ren H, Miao L, Qian L, Zou L, Zhou Y, Wang X, et al. Causal Intervention-Enhanced Feature Disentanglement Network for SAR ATR in Data-Scarce Environments. IEEE Transactions on Aerospace and Electronic Systems. 2026;62:13332-13347. doi:10.1109/taes.2026.3708075
BibTeX
@article{Ren2026,
author = {Ren, Haohao and Miao, Lei and Qian, Lu and Zou, Lin and Zhou, Yun and Wang, Xuegang and Tang, Hao},
title = {Causal Intervention-Enhanced Feature Disentanglement Network for SAR ATR in Data-Scarce Environments},
journal = {IEEE Transactions on Aerospace and Electronic Systems},
year = {2026},
volume = {62},
pages = {13332--13347},
doi = {10.1109/taes.2026.3708075}
}
EUVIMED