Spectral-Spatial Ensemble Low-Rank Domain Adaptation for Hyperspectral Image Classification

Domain adaptation has been proven effective for addressing cross-domain hyperspectral image (HSI) classification, especially when the target domain has no labeled samples. Current domain adaptation algorithms focus on finding domain-invariant subspaces to reduce domain shift, but there is still room to improve sample discriminability. Moreover, inappropriate projection between source and target domains can also affect the transfer effect. To tackle these issues, this article proposes a spectral-spatial ensemble low-rank domain adaptation (S2ELRDA) method. In S2ELRDA, low-rank constraints are imposed to align samples with the same labels in the two domains, enhancing sample discriminability while reducing domain shift. Furthermore, ensemble learning is utilized to select samples based on confidence as supervised information for the target domain, further improving subspace robustness. Additionally, the inherent characteristics of HSIs are considered, and spectral-spatial features are extracted for alignment to enhance sample discriminability in the subspace. Experimental results on three HSI datasets demonstrate that the proposed S2ELRDA shows better performance than existing algorithms.