Few-Shot Target Recognition of Remote Sensing Images Based on Causal Consistency and Domain-Sensitive Channel Regularization

Remote sensing target recognition is significantly impacted by domain shifts between the source and target domains, especially under few-shot conditions, where models struggle to extract effective features. Current methods emphasize extracting features that are invariant across domains, yet they often neglect the influence of domain-sensitive structures within convolutional networks, which hinders the overall generalization performance of the model. To address this issue, a novel few-shot domain generalization method named causal consistency and domain-sensitive channel regularization is proposed for remote sensing target recognition. Within the proposed framework, causal consistency learning is introduced to capture key causal features of target categories, mitigating the interference of non-causal features during inference. Additionally, a domain-sensitive channel regularization module is designed to identify and suppress convolutional kernels with unstable activations caused by domain shifts, thereby preserving the structural robustness of the network and enhancing cross-domain feature extraction. To further stabilize training, a structural bias attenuation loss is incorporated to mitigate the instability introduced by kernel suppression, ensuring the model maintains robust generalization across different domain environments. Experimental results confirm the effectiveness of the proposed method in suppressing the impact of domain-sensitive structures, improves domainin-variant feature extraction, and achieves superior recognition performance under domain shift conditions compared to existing approaches.