Convolutional Neural Networks for Tomographic MIMO Ground Penetrating Radar Imaging

Abstract

Ground penetrating radar (GPR) represents a remote sensing modality which has been extensively exploited for the detection and characterisation of buried objects in a non-destructive way. To this aim, several algorithms have been developed to efficiently and automatically identify underground targets of interest. In this framework, approaches based on deep learning and convolutional neural networks (CNNs) have been proposed in the past years and recently gained a lot of attention by the scientific community. Despite their efficiency, these approaches require a large number of cases to perform the training step and improve their classification performance. In this paper, the use of multistatic GPR data is explored (via simplified numerical simulations) to automatically classify the kind of underground utility in areas in which both water and natural gas pipes can be considered. More in detail, some discussions on the classification performance obtained via considering more receivers in the measurement configuration close the paper, underlining the better results obtained via exploiting the multistatic data.