Tensor nested ring embedding used for domain adaptation of cross-source heterogeneous remote sensing feature tensors

Multisource remote-sensing data present richer information with various resolutions, angles, and physical attributes, while their heterogeneous data structures and diverse feature distributions bring significant challenges to the existing recognition methods. Although domain adaptation methods can reduce cross-source feature differences, they can only handle vector or homogeneous tensors and fail to process multisource data represented as heterogeneous tensors. Therefore, the Tensor Nested Ring Embedding (TNRE) method is proposed to achieve domain-adapted feature extraction using cross-source heterogeneous remote sensing tensors. To efficiently represent heterogeneous data and eliminate the structure difference, unlike the existing Tucker-based methods consuming an exponential storage cost with tensor order, the tensor-nested ring (TNR)-learning theory is built via rigorous mathematical analysis to learn shared feature space under a linear storage cost. Considering that the existing single-domain-based alignment strategy cannot adapt to the multisource data-containing diverse object characteristics, the latent domain centroid alignment strategy is established to break the multisource data into multiple latent domains and then guide alignment. Furthermore, an alternative optimization scheme is constructed to obtain the TNRE’s optimal solution. Experiments using multi-resolution, multi-angle, multi-scene remote-sensing datasets demonstrate that the proposed TNRE outperforms typical domain adaptation methods regarding recognition accuracy and storage cost.