Change Detection Method for High Resolution Remote Sensing Images Using Deep Learning

Abstract

A novel change detection method is proposed based on deep learning to improve the accuracy of change detection in very high spatial resolution remote sensing images. On the base of image pre-processing, spectral and texture changes are extracted by modified change vector analysis and grey level co-occurrence matrix respectively, both concerning spatial-contextual information. Most likely changed and unchanged pixel-pairs are obtained by an adaptive threshold for selecting the labeled samples. The proposed model based on Gaussian-Bernoulli deep Boltzmann machines with a label layer is built to learn high-level features and is trained for determining the change areas. Experimental results on WorldView-3 and Pléiades-1 show that the proposed method out performs the compared methods in the accuracy of change detection.