Application of deep learning image-to-image transformation networks to GPR radargrams for sub-surface imaging in infrastructure monitoring

The corrosion of reinforced concrete sewer pipes in aging infrastructure is a serious ongoing issue and as such, research into technologies that allow for autonomous site assessments are of major priority to wastewater managing utilities. The use of Ground Penetrating Radar (GPR) is being investigated for providing sub-surface images of sewer crowns. Due to the nature of GPRs, the analysis of a radargram for identifying sub-surface features is non-intuitive and usually require the use of an expert. Traditional methods to help ease analysis involve the use of Synthetic Aperture Radar (SAR) and migration techniques. These techniques refocus dipping and point reflectors to be closer to their true shape but require an accurate velocity model to be effective. This is not always readily available and difficult to estimate especially in regards to sewer conditions. We instead provide an alternative and present a deep learning framework for transforming ground penetrating radargrams into sub-surface permittivity maps. An evaluation of various state-of-the-art deep learning architectures is also conducted, comparing the performance of different objective functions and identifying current limitations. This work provides the base for further exploration of the application of deep learning for use in infrastructure monitoring.