Unsupervised deep learning based change detection in Sentinel-2 images

Change Detection (CD) is an important application of remote sensing. Recent technological evolution resulted in the availability of optical multispectral sensors that provide High spatial Resolution (HR) images with many spectral bands. Such characteristics allow for new applications of CD, however present new challenges on the proper exploitation of the information. HR multitemporal data processing is challenging due to spatial correlation of pixels and spatial context information needs to be exploited to benefit from multitemporal HR images. Moreover most of the state-of-the-art CD methods exploit single or couple of spectral channels from the optical sensors to derive CD map. To overcome these challenges, this paper presents a novel unsupervised deep-learning based method that can effectively model contextual information and handle all the bands in multispectral images. In particular, we focus on the Sentinel-2 images provided by the European Space Agency (ESA) that provides both higher spatial and temporal resolution optical images with 13 spectral bands with respect to previous generation sensors. Experimental results on the urban Onera satellite CD (OSCD) dataset and on agricultural multitemporal images from Barrax, Spain confirms the effectiveness of the proposed method.