Exploring Deep Learning Models on GPR Data: A Comparative Study of AlexNet and VGG on a Dataset from Archaeological Sites

Abstract

This comparative study evaluates the performance of three popular deep learning architectures, AlexNet, VGG-16, and VGG-19, on a custom-made dataset of GPR C-scans collected from several archaeological sites. The introduced dataset has 15,000 training images and 3750 test images assigned to three classes: Anomaly, Noise, and Structure. The aim is to assess the performance of the selected architectures applied to the custom dataset and examine the potential gains of using deeper and more complex architectures. Further, this study aims to improve the training dataset using augmentation techniques. For the comparisons, learning curves, confusion matrices, precision, recall, and f1-score metrics are employed. The Grad-CAM technique is also used to gain insights into the models’ learning. The results suggest that using more convolutional layers improves overall performance. Further, augmentation techniques can also be used to increase the dataset volume without causing overfitting. In more detail, the best-obtained model was trained using VGG-19 architecture and the modified dataset, where the training samples were raised to 60,000 images through augmentation techniques. This model reached a classification accuracy of 94.12% on an evaluation set with 170 unseen data.