Ü-Net: Deep-Learning Schemes for Ground Penetrating Radar Data Inversion

Abstract

Electromagnetic (EM) inversion is a quantitative imaging technique that can describe the dielectric constant distribution of a target based on the EM signals scattered from it. In this paper, a novel deep neural network (DNN) based methodology for ground penetrating radar (GPR) data inversion, known as the Ü-net is introduced. The proposed Ü-net consists of three parts: a data compression unit, U-net, and an output unit. The novel inversion approach, based on supervised learning, uses a neural network to generate the dielectric constant distribution from GPR data. The GPR data can be compressed and reshaped the size using data compression unit. The U-net maps the object features to the dielectric constant distribution. The output unit meshes the dielectric constant distribution more finely. A novel feature of the proposed methodology is the application of instance normalization (IN) to the DNN EM inversion method and a comparison of its performance to batch normalization (BN). The validity of this technique is confirmed by numerical simulations. The Mean-Square Error of the test data sets is 0.087. These simulations prove that the instance normalization is suitable for GPR data inversion. The proposed approach is promising for achieving quality dielectric constant images in real-time.