Pipeline Detection with Satellite Images Using Machine Learning

Kurzfassung

This thesis has been written in the framework of the SciGRID_gas project, which has the goal of creating an open-source network model for the European gas transport network. For the creation of the model, public data sources such as OpenStreetMap or press articles are currently used. These conventional data sources tend to be inaccurate or incomplete. For the correction of these inaccuracies and incompleteness, satellite images could play an important role. The aim of this thesis is to prove that a pipeline dataset can be automatically generated from satellite images using machine learning methods. For this, a deep learning 1model is trained to recognize pipeline courses on satellite images. Training data will first be generated using gas network data from Great Britain. Open-access satellite data will be used which is easily accessible, of adequate resolution, and is well documented. A suitable model will be trained with these data. The trained model will then be tested with satellite images from Northern Germany. The results will be evaluated and a possible future large-scale application of the model to the whole European area will be discussed.