GPR B-scan Data Classification through Deep Learning Approach

Ground Penetrating Radar (GPR) has proven indispensable in non-destructive geophysical investigations, offering a window into subsurface structures using electromagnetic waves. Its prowess in identifying underground utilities and hazards, like cavities, has been universally recognized. However, the intricate B-scan data generated by GPR poses significant interpretation challenges. Traditional methodologies, especially those reliant on dielectric constants and earlier machine learning models, often necessitate prior subsurface knowledge, inhibiting their universal applicability. Enter deep learning, which, in recent years, has changed the landscape of data interpretation across domains. Among the various architectures, Convolutional Neural Networks (CNNs) have shown immense potential in deciphering GPR data. This paper champions a novel methodology that harnesses the power of the ResNet18 architecture. Uniquely, our approach begins with a model pre-trained on the WiderFace dataset, renowned for facial recognition, and subsequently fine-tuned for GPR data. This merger of facial pattern recognition with geophysical data interpretation creates a synergy that promises unprecedented accuracy in subsurface detection. In summation, our research showcases how deep learning, especially the fusion of facial recognition insights with geophysical data, can redefine the boundaries of GPR data interpretation, offering enhanced accuracy and adaptability over conventional methods.