Automatic Layer Extraction of Ground-Penetrating Radar Based on Deep Learning

Summary Subsurface layers, which record the spatio-temporal distribution characteristics of geological formations, are important indicators for inferring the historical evolution of the geological environment. Ground-penetrating radar (GPR) provides high-resolution data which are essential for identifying and analyzing subsurface layers, enabling detailed insights into geological conditions. However, traditional layer extraction methods, primarily manual or semi-automated, are time-consuming when processing large volumes of radar data, particularly for extracting continuous layers with low contrast. Therefore, we propose an automatic method to extract GPR layers, integrating radar attribute analysis with the U-Net semantic segmentation model. To train the extraction network, we design a novel synthetic method for radar profiles based on characteristic textures that are similar to those in the observation data. Our results show significant improvements in both accuracy and efficiency compared to traditional methods and existing automatic methods that rely on radar amplitude profiles, thereby advancing the automatic processing of subsurface layer extraction.