Author(s) :

Rebecca Koleth.

Volume/Issue :

Abstract :

Ground Penetrating Radar (GPR) is an essential non-invasive method in archaeology for finding archaeological features below the surface. However, interpreting the data can be difficult due to its complexity and noise. This research investigates the application of deep learning models to enhance the detection and interpretation of subsurface archaeological features using GPR data. Due to the scarcity of publicly available annotated GPR datasets, simulated data generated using the gprMax software is utilized. A Convolutional Neural Network (CNN) is built with TensorFlow and Keras. It focuses on creating bounding boxes around hyperbolic reflection signatures in B-scan radargrams. These simulations feature a buried perfect electric conductor (PEC) cylinder in a dielectric half-space. The model showed promising results, with Intersection over Union (IoU) scores of 0.93 reflecting accurate localization on test samples. This study establishes the foundation for future applications of deep learning to archaeological GPR data analysis. It also demonstrates that simulation-based training can be effective and provides a basic model where annotated samples are often scarce. This contribution is important because it positions simulation-driven training as a cost-effective option for archaeological geophysics. Data shortages have often limited the use of AI in this field. By presenting a CNN framework that adapts well to new simulated conditions, the study emphasizes the future potential of hybrid methods that combine synthetic and real data.