Change detection in satellite imagery using deep learning

Details

Supervisors

Schaus, Pierre

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique

Abstract

Change detection in satellite imagery is an excellent way to monitor the evolution of geographical areas. Although object detection in that field is already well-developed, change detection by semantic segmentation isn’t, but is a more direct way of seeing the changes in the infrastructure of a region over time. With two picture of a region at different times, we can build a segmentation mask outlining the contours of new buildings. This thesis aims to develop an architecture that performs change detection in buildings in urban areas. After building a dataset, we first study the performances of the UNet architecture. There is a loss of localisation precision with this model. To remedy to this we study two modular UNets with different depths and a UNet++, both evolutions of the UNet. The performances of 4 different models are compared. Our results show that even with a small architecture of a UNet++, we obtain segmentation masks that allow for an efficient detection of new buildings. The F1 score obtained on the test set is of 42% and visibly neater contours on the masks. Lead for further improvement of the best-performing model are included. The implementation of this project can be found on the LamboiseNetGitHub repository : https://github.com/hbaudhuin/LamboiseNet[2].