Map Creation from Semantic Segmentation of Aerial Images Using Deep Convolutional Neural Networks - Utilizing publicly available spatial data to make an aerial image labeling dataset

Abstract

For centuries cartographers have segmented and labeled the surface of the earth onto analog and digital maps. Today, map-making is still a time-consuming, manual process that requires large amounts of work.

In this thesis, we investigate the possibility of using publicly available spatial data to train deep convolutional neural networks into performing accurate semantic segmentation of aerial images. We collect spatial data from the Norwegian Mapping Authority and propose a dataset consisting of aerial photographs and high-quality labels from five cities in Norway. The dataset consists of four different spatial classes: Buildings, Roads, Water and Vegetation. We present detailed statistics and highlight issues in the dataset and show that the technique for creating the dataset can be expanded to include data from all of Norway.

To investigate the possibility of creating maps automatically, we adopt two deep convolutional neural networks and train them on the proposed dataset. We show that training one network for each semantic class yield better results than training one network on all the classes simultaneously.

Finally, we present maps produced by the networks and assess its quality and usability. The results show that our method can produce useful maps without the need for pre- and postprocessing of the data.