Spatial-Spectral Cross-Domain Attention Network for Unsupervised Hyperspectral Image Classification

Unsupervised cross-domain hyperspectral image classification is one of the important challenges in remote sensing due to the difficulty of labeling hyperspectral image, and domain adaptation based on adversarial learning has received extensive attention. However, current methods only focus on the alignment of the two domains and do not fully integrate the spectral information. Therefore, we propose a spatial-spectral cross-domain attention network (SCANet) for unsupervised hyperspectral image classification. Our method first utilizes the adversarial learning process of feature extractor and two classifiers to align the feature distribution of source and target domain. Then we construct a spectral-spatial cross-domain attention module between feature extractor and classifiers, extract domain-related information and symmetrically focus each domain. Particularly, to improve the robustness of the model, we introduce a consistency penalty for attention features. Experimental results based on typical hyperspectral images verify the effectiveness of the proposed method.