OPTICAL REMOTE SENSING CHANGE DETECTION THROUGH DEEP SIAMESE NETWORK

This paper presents a change detection approach for optical remote sensing images based on deep learning. Due to the excellent performance of Convolutional Neural Network (CNN) in feature learning, two models are explored in this work, where the proposed algorithms show how to learn, directly from images, a similarity function to compare bitemporal images. Two-stream network named as Siamese network is presented. First, bi-temporal images are fed directly into the proposed network. Second, a combination of the aforementioned model with a perceptual loss is presented, this combination focus on high representational features that are extracted from a pre-trained network on large dataset of natural images (ImageNet) rather than opting directly on remote sensing images. Experimental results on real dataset show the effectiveness and the superiority of the proposed framework.