A Generalizable Deep Learning Approach to Classifying Ground Radar Images

Ensuring the safety of underground structures depends on the rapid and reliable detection of potential ground defects. Ground Penetrating Radar (GPR) is one of the widely used non-destructive testing methods for this purpose. However, traditional GPR data analysis has limitations in terms of both time and accuracy due to operator dependency and high workload. In this study proposes a generalizable deep learning approach for the automatic classification of GPR images. Within the scope of the study, EfficientNet, MobileNetV2, and VGG16 architectures were comparatively evaluated using two independent GPR datasets with different structural features and reflection characteristics. To objectively examine the generalization capabilities of the models, 5-fold cross-validation was applied and their performance was analyzed using Accuracy, Precision, Recall, and F1-Score metrics. The experimental results demonstrate that deep learning architectures can accurately distinguish structural anomalies in GPR data. MobileNetV2 model showed the highest performance, achieving 100% accuracy on the Underground Utilities Dataset (UUD) and 97.46% accuracy on the TIGPR dataset. The findings demonstrate that GPR images can be successfully classified using deep learning architectures, and this approach has strong potential in applications such as infrastructure health monitoring, tunnel lining assessment, and subsurface defect detection.