Unsupervised Domain Adaptation With Dense-Based Compaction for Hyperspectral Imagery

Enormously hard work of label obtaining leads to the lack enough annotated samples in hyperspectral imagery (HSI). The mentioned reality inferred unsupervised classification performance barely satisfactorily. Unsupervised domain adaptation is exploited for knowledge delivery from a labeled source boost on an unlabeled target domain. In this paper, we propose architecture with dense-based compaction (UDAD) HSI (HSIC). processes spectral-spatial feature compaction, adaptation, and classifier training are incorporated integrated framework complete cross-scene classification. core proposed utilize adversarial learning reduce discrepancy. To end, trained would accomplish well HSIC. Besides, extract discriminative domains, network applied semi-symmetric mapping. Our experiments illustrate that UDAD model yields more effective than other state-of-the-art HSIC methods.