Target Decomposition Feature-Guided Multimodal Domain Generalization Learning for Unsupervised PolSAR Image Classification

Cross-domain (CD) models have been successfully applied to polarimetric synthetic aperture radar (PolSAR) image classification to alleviate the scarcity of labeled samples. However, these models often struggle to extract strongly generalizable domain-invariant features when encountering significant domain shifts. Furthermore, prevailing methods predominantly operate within the supervised domain adaptation paradigm, requiring access to target-domain data during training, which inherently limits their generalizability to entirely unseen target domains. Therefore, this paper proposes a target decomposition feature-guided multimodal domain generalization learning network (MMDGN), inspired by multimodality and domain generalization, for unsupervised classification of target-domain PolSAR images under the same frequency constraint. The core idea is to leverage the physical scattering mechanism consistency of the same frequency and categories to guide CD feature optimization, thereby enhancing the discriminative power and domain invariance of the features. The proposed MMDGN comprises three key modules: a multimodal feature extraction module that separately extracts image-modal and text-modal features, the latter derived from category-specific representative scattering components; a pattern memory-based space adjustment module that projects both modalities into a unified embedding space; and a dual-level semantic alignment module that integrates supervised contrastive learning with adversarial training to align multimodal features at both intra-class and inter-class levels. This promotes that the optimized image-modal features, guided by the text modality, exhibit intra-class compactness and inter-class separability. Experiments on six CD tasks constructed from three C-band PolSAR datasets demonstrate that MMDGN achieves state-of-the-art performance in unsupervised classification accuracy and CD generalization.