A Deep Learning Model for Change Detection on Satellite Images

Abstract

Change detection is a classical problem in satellite imaging. The change detection problem aims at the analysis of changes between two images. In this work, we test a deep learning model proposed in 2019 by Caye Daudt et al. on data from the Sentinel-2 satellite with images between 10 m and 60 m of spatial resolution. The model uses the early fusion technique combined with a U-net architecture and can be used on color or multispectral images. The tests are performed on the Onera Satellite Change Detection (OSCD) dataset, which was already used for testing deep learning methods for the change detection problem. Here we propose some experiments to evaluate the performance and limits of the algorithm by Caye Daudt et al.