Publisher's Synopsis
Abstract Ground Penetrating Radar (GPR) Data Analysis deals with the problem of shallow subsurface imaging, which is motivated by the daily work of engineers, \eg those of municipalities.The concrete problem tackled in this thesis is motivated by the fact, that, at least in Germany, municipalities have knowledge about the existence of supply lines such as gas and water pipelines to cross and follow urban streets, while their actual position is often uncertain.The consequences are obvious: once a street undergoes maintenance works, pipes are easily broken.This also causes heavy problems to residents who are cut off from some supplies for a period of time. This thesis approaches a solution to the object detection problem in GPR data by means of (semi-)automated data analysis techniques, using Machine Learning methods.The problem is treated as a specialized problem for object detection in image data.In this application context, it is possible to integrate certain background knowledge and processing techniques in well-known Machine Learning methods. The thesis formalizes the problem first.A technical framework for the analysis of Complex Engineering Raw Data - CERD -, as a generalization of our current data at hand, will be used for all analysis methods developed.From a thorough data analysis, it becomes clear that our data labels are unsuitable for directly applying supervised Machine Learning methods.Therefore, we will be obtaining suitable ground truth data by semi-manually labeling more than 700 images by hand.The second part of the thesis presents both, supervised and unsupervised Machine Learning techniques for the detection of buried object locations.Techniques are introduced within the general context of object detection techniques within image data.The integration of geometrical background knowledge is shown to be feasible in all methods developed. This thesis will contribute in the followings: *The methodology and suitability of high-quality ground truth d
