Master thesis defense by Andreas Borgstad and Christoffer Thrysøe


Detecting Roads in Remote Sensing Data Using Deep Learning Method


Knowing the accurate locations of roads, and thereby having accurate maps, is vital to many applications, such as GPS systems, urban planning, and disaster relief. Obtaining road mappings is usually done by manual annotation of remote sensing data, resulting in a time-consuming and error-prone process. Such a task is becoming progressively infeasible to perform due to global urbanization, continually requiring the expansion and re-planning of cities, therefore increasing the need for an automated system capable of detecting roads. In this thesis, we aim to develop and analyze the components of an end-to-end pipeline capable of perform-
ing automatic detection of roads in satellite imagery, using deep learning methods. We show that our proposed network is capable of reliably detecting roads in imagery collected from the Sentinel-2 satellites, and show that the model achieves better performance than state-of-the-art methodologies within computer vision. We further show that it is possible to apply the model in a distributed framework for performing road detection on entire continents, using only a modest amount of hardware resources.

Supervisors: Fabian Gieseke, Marcos António Vaz Salles, Wouter Kouw

External Examiner: Ira Assent