Deep convolutional neural network for estimation of depth and radius from GPR raw signals

Precise location and identification of underground utility networks became critical in terms of preventing strikes on the network during excavations. In this context, Ground Penetrating Radar (GPR) has emerged in practice for its ability to detect both metallic and non-metallic buried objects. Furthermore, depth and radius of the pipes are two key geometrical parameters needed to be estimated from GPR signals for accurate mapping. All existing methods use either physical models or supervised machine learning techniques and based on a prior extraction of local features from the data through series of signal processing steps. Nevertheless, extraction of the correct local features remains a challenge.

In the recent years, supervised Deep Convolutional Neural Network (DCNN) approaches brought great attention due to its proven characteristics to automatically extract features from the input pictures to perform classification or regression precisely and rigorously in various fields. In terms of GPR, signals to picture rescaling can lead to loss of information. To benefit from DCNN while conserve the information, this research work is focused on implementation DCNN on GPR 2D raw signals instead of pictures, for the estimation of depth and radius. Hence, the model was numerically validated with the data generated from gprMax 2D.