Automatic Detection of Underground Objects in Ground Penetrating Radar Images using Machine Learning

  • Author / Creator
    Martoni Amaral, Leila C.
  • The population increase has stimulated the need for the creation and expansion of existing urban infrastructures such as sewer, water, power and telecommunication lines. In order to support this need, multiple efforts to find sustainable solutions that support the urbanization trend have been studied. As a solution, construction methods such as trenchless techniques—ranging from maxi horizontal directional drilling for large-scale pipe installations to microtrenching for the installation of telecommunications infrastructure, that offer reduced social and environmental impacts compared to conventional opencut construction methods, are becoming more accessible and increasingly used. Despite the advantages of underground construction, challenges exist in knowing the locations of existing subsurface utilities or other underground objects (such as rocks) an that uncertainty can impact new installations. Existing non-destructive technologies, such as ground penetrating radar (GPR), can be used to map extensive areas in a fast and accurate manner and support the construction of new infrastructure. However, GPR data is difficult to interpret, and requires an experienced person to be able to locate the features within the image that correspond to objects. In order to overcome the issue of data interpretation, multiple studies aiming towards the automation of GPR data interpretation have been conducted—however, the methods proposed can still improve in terms of detection speed and accuracy.
    The objective of this research is to automate GPR data interpretation to support underground construction. To achieve this, an extensive database of GPR images due to commonly encountered underground objects, such as rock-sized boulders and PVC and metal pipes, was collected based on GPR measurements in a laboratory setting. The database compiled from these measurements was used alongside different machine learning algorithms, including YOLO v3 and R-CNN, to determine the methodology that has higher accuracy of detection and classification for the automated object detection based on the correspondent features in GPR images. The limitations of the various algorithms are considered, and recommendations are proposed for future studies.
    A study on the long-term performance evaluation of micro-trenching backfill materials was continued, with GPR surveys performed to precisely determine conduit depth within the micro-trench. Updated measurements of conduit depth were compared to previously collected data to determine the conduit displacement four to seven years after installation.

  • Subjects / Keywords
  • Graduation date
    Fall 2021
  • Type of Item
  • Degree
    Master of Science
  • DOI
  • License
    This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.