Change detection of buildings with the utilization of a deep belief network and high-resolution remote sensing images

To enhance the accuracy results of the change detection of buildings utilizing high-resolution remote sensing (HRRS) images, a novel method was proposed by combining both tensor and deep belief network (DBN). To optimize the description of the essential characteristics for changes in buildings, a tensor-based structure covering time-space-spectrum-shadow features integrated into the model (TSSS-Cube) is proposed. The changes occurring as a combination of shadow and spectral features and spatio-temporal autocorrelation at each pixel are represented by a third-order tensor to maintain the structural information and the constraint integrity between them. Then, a restricted Boltzmann machine (TC-RBM) that can be directly used to process TSSS-Cube data is designed, and the support tensor machine (STM) is used to replace the conventional backpropagation neural network at the top of the DBN to construct a multi-tensor deep belief network (MTR-DBN) composing of multi-layer TC-RBMs and an STM classifier. Finally, the multi-layer TC-RBMs in MTR-DBN are trained layer by layer, and the global parameters of the MTR-DBN are optimized by combining a limited number of labeled data and fine-tuning the STM classifier. The implementation of both supervised and unsupervised learning methods comprehensively provides advantages to increase the accuracy result of the MTR-DBN network to detect changes. Three representative different sub-regions are selected from the whole original experimental area respectively for building change detection experiments, and a dataset composed of double-temporal HRRS images in 2012 and 2016 is used as the related experimental dataset. The experimental results show that both a change detection accuracy result with a higher average and better detection efficiency is attained by the proposed method called the MTR-DBN when compared with other similar methods.