Joint Tensor Subspace Alignment on Multi-Angular Remote Sensing Image

High-resolution remote sensing satellites produce many images with different angles during the shooting process. Some of the angular images are labeled, but more are not labeled. How to use these labeled images to classify images of other different angles is a problem that needs to be solved. In view of this, we proposes an unsupervised DA approach by aligning tensor subspace and combining multi-angular images. The source subspace is aligned to the target subspace so that the source domain sample probability distribution is close to the target domain sample. Using images from different angles as source domains can make full use of their similarities and differences to achieve better classification results. To verify the effectiveness of this method, this paper conducts a set of five angular images taken by WorldView-2 satellites. The experimental results show that the proposed method can effectively distribute the source image group close to the target image distribution and achieve better classification.