Hyperbola Detection with RetinaNet and Comparison of Hyperbola Fitting Methods in GPR Data from an Archaeological Site
Hyperbolic diffractions in Ground Penetrating Radar (GPR) data are caused by a variety of subsurface objects such as pipes, stones, or archaeological artifacts. Supplementary to their location, the propagation velocity of electromagnetic waves in the subsurface can be derived. In recent years, it was shown that deep learning tools can automatically detect hyperbola in radargrams using data measured over urban infrastructure, which are relatively clear. In contrast, in this study, we used an archaeological dataset with diverse underground structures. In the first step we used the deep learning network RetinaNet to detect hyperbola automatically and achieved an average precision of 0.58. In the next step, 10 different approaches for hyperbola fitting and thus velocity determination were applied. The derived information was validated with manually determined velocities and apex points. It was shown that hyperbola extraction by using a threshold and a column connection clustering (C3) algorithm followed by simple hyperbola fitting is the best method, which had a mean velocity error of 0.021 m/ns compared to manual determination. The average 1D velocity-depth distribution derived in 10 ns intervals was in shape comparable to the manually determined one, but had a systematic shift of about 0.01 m/ns towards higher velocities.