MPA-Net: A Deep Learning Model for Estimating Rebar Parameters Based on GPR Data

Ground-penetrating radar (GPR) is an effective nondestructive technique for inspecting reinforced concrete structures. Estimating parameters related to reinforcing bars (rebar) based on GPR radargrams remains a challenge due to the strong correlation among different parameters when generating the reflection signature. To address this challenge, we introduce a neural network structure called the Multi-Polarimetric Aggregation Neural Network (MPA-Net) to simultaneously derive multiple rebar-related parameters from GPR radargrams. These parameters include concrete moisture content, rebar cover depth, radius, and orientation. Given multi-polarization radargrams as inputs, MPA-Net extracts informative features from each polarization radargram, mitigates the effects of parameter correlation, and simultaneously estimates the four rebar-related parameters. The high estimation accuracy of the well-trained MPA-Net has been demonstrated through numerical results.