Detection of Changes in Built-Up Areas with a Fully Convolutional Network in the Context of the European Settlement Map

This work presents a novel method for detecting newly built-up areas in Very High Resolution satellite imagery using Fully Convolutional Neural Networks. The architecture builds on Early Fusion concept where the pairs of image patches are concatenated before inputting into the network as different color channels. The model was trained on different sensors (Pleiades, Kompsat and Superview) using the labels from the Urban Atlas change maps (2012-2018). The results demonstrate the potential of the proposed method for a pan-European mapping of changes in built-up areas at a spatial resolution of 2 meters in the framework of the European Settlement Map of the Joint Research Centre.