Unsupervised and Self-Supervised Tensor Train for Change Detection in Multitemporal Hyperspectral Images

• DOI: 10.3390/electronics11091486
• Publication Date: 2022-05-06T18:49:38Z
• Authors: Muhammad Sohail; Haonan Wu; Zhao Chen; Guohua Liu
• Keywords: Discriminative model; Benchmark (surveying)

Remote sensing change detection (CD) using multitemporal hyperspectral images (HSIs) provides detailed information on spectral–spatial changes and is useful in a variety of applications such as environmental monitoring, urban planning, disaster detection. However, the high dimensionality low spatial resolution HSIs do not only lead to expensive computation but also bring about inter-class homogeneity inner-class heterogeneity. Meanwhile, labeled samples are difficult obtain reality field investigation expensive, which limits application supervised CD methods. In this paper, two algorithms for based tensor train (TT) decomposition proposed called unsupervised (UTT) self-supervised (STT). TT uses well-balanced matricization strategy capture global correlations from tensors can therefore effectively extract low-rank discriminative features, so curse spectral variability be overcome. addition, methods learning, where no manual annotations needed. ket-augmentation (KA) scheme used transform low-order into high-order while keeping total number entries same. Therefore, features with richer texture extracted without increasing computational complexity. Experimental results four benchmark datasets show that outperformed their counterpart, tucker (TD), higher-order singular value (HOSVD), some other state-of-the-art approaches. For Yancheng dataset, OA KAPPA UTT reached 98.11% 0.9536, respectively, STT were at 98.20% 0.9561, respectively.