Analysis of Deep CNN-based Ground Penetrating Radar (GPR) Image Classification Process using Explainable AI

Deep Learning techniques are increasingly being used for Ground Penetrating Radar (GPR)-based target detection and classification. However, DL models normally act as black-box systems leading to difficulties regarding interpretation, repeatability and robustness of target detection and classification processes. In this paper, we employ several variants of Class Activation Map (CAM) technique to superimpose colored coded heat maps on a particular output of GPR systems, namely B-Scan images, to expose the degree of contribution of specific spatial regions to the prediction of the DL model. As such, by comparing CAM results of B-Scan images that are subjected to different preprocessing stages, we aim to provide insight into the relationship between those pre-processing steps and the corresponding detection performance of the model. Thus, our approach facilitates more informed selection of signal processing techniques for the early phases of the GPR processing pipeline (i.e., earlier than the DL model) on one hand; and could lead to more accurate design of the higher level decision processes, which, in turn, would increase the overall operational effectiveness of the GPR systems.