Hyperspectral remote sensing image change detection based on tensor and deep learning

Considering the bottleneck in improving the performance of the existing multi-temporal hyperspectral remote sensing (HSRS) image change detection methods, a HSRS image change detection solution based on tensor and deep learning is proposed in this study. At first, a tensor-based information model (TFS-Cube) of underlying features change in HSRS images is established. The wavelet texture feature change, spectral feature change and spatio-temporal autocorrelation coefficient of different-temporal related pixels are combined with three-order tensor, so as to make full use of the underlying features change information of HSRS images, optimize the organization mode and maintain the integrity of constraints between different underlying features. Secondly, a restricted Boltzmann Machine based on three-order tensor (Tensor3-RBM) is designed. The input, output and unsupervised learning of TFS-Cube tensor data are realized by multi-linear operations in Tensor3-RBMs. A large number of unlabeled samples are trained layer by layer through multilayer Tensor3-RBMs. Finally, the traditional BP neural network on the top layer of deep belief network (DBN) is replaced with support tensor machine (STM), and a deep belief network with multi-layer Tensor3-RBM and STM (TRS-DBN) is constructed. A small number of labeled samples are used for supervised learning and TRS-DBN global parameters optimization to improve the accuracy of TRS-DBN change detection. Two types of HSRS images from different sensors, AVIRIS and EO-1 Hyperion, are used as the data sources (double-temporal). Four representative experimental regions are randomly selected from the two areas covered by AVIRIS and EO-1 Hyperion HSRS images respectively (two regions in each area) to detect the land use changes. Experimental results demonstrate that TRS-DBN has higher change detection accuracy than similar methods and a good automation level. (C) 2018 Published by Elsevier Inc.