PMDA-Net: a polarization-conditioned multi-scale dynamic attention network for polarimetric hyperspectral camouflage target classification

To effectively fuse heterogeneous polarimetric hyperspectral data and improve the classification performance of camouflage targets in complex backgrounds, this paper proposes a polarization-conditioned multi-scale dynamic attention network (PMDA-Net). PMDA-Net adopts a dual-stream framework for modality-specific feature extraction, where spectral and polarimetric information are separately modeled. A multi-scale dynamic AoLP attention module is designed to transform geometric information encoded in the angle of linear polarization into dynamic weights. A polarization-conditioned spatial-channel attention module is designed to adaptively calibrate spectral features using polarimetric information, facilitating cross-modal interaction and complementary feature fusion. Experimental results on three self-collected datasets and one public dataset show that PMDA-Net outperforms mainstream hyperspectral deep learning models, traditional polarization-feature-based methods, and recent multimodal fusion methods on all four datasets. Compared with the model using only hyperspectral information, PMDA-Net improves the average accuracy by 5.87, 5.38, 25.71, and 5.01 percentage points on the four datasets, respectively. These results indicate that polarization information can effectively complement spectral representations and help suppress aliasing responses caused by spectral similarity in complex scenes. This work provides a promising approach for polarimetric hyperspectral image classification and camouflage target recognition.