Multi-task learning for segmentation of building footprints with neural networks

Benjamin Bischke; Patrick Helber; Joachim Folz; Damian Borth; Andreas Dengel

In: AI4SocialGood. International Conference on Learning Representations (ICLR-2019), May 6-9, New Orleans, Louisiana, USA, arXiv, 2019.


The increased availability of high-resolution satellite imagery allows to sensevery detailed structures on the surface of our planet and opens up new direc-tions in the analysis of remotely sensed imagery. While deep neural networkshave achieved significant advances in semantic segmentation of high-resolutionimages, most of the existing approaches tend to produce predictions with poorboundaries. In this paper, we address the problem of preserving semantic seg-mentation boundaries in high-resolution satellite imagery by introducing a novelmulti-task loss. The loss leverages multiple output representations of the seg-mentation mask and biases the network to focus more on pixels near bound-aries. We evaluate our approach on the large-scale Inria Aerial Image Label-ing Dataset. Our results outperform existing methods with the same architec-ture by about 3% on the Intersection over Union (IoU) metric without additionalpost-processing steps. Source code and all models are available under

Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence