Lightweight CNN model for automatic detection and depth estimation of subsurface voids using GPR B-scan data

Abstract

Subsurface cavities pose significant risks, including structural instability, safety hazards, and environmental damage. Early detection of these cavities is crucial to prevent material losses and protect human lives. Investigation and manual processing of these structures using traditional methods can be difficult and time-consuming. Therefore, automated approaches using machine learning algorithms for identifying subsurface anomalies have recently emerged, providing promising pathways for real-time cavity detection. Consequently, this study proposes a Convolutional Neural Network (CNN)-based framework for the automated detection and depth estimation of subsurface cavities from Ground Penetrating Radar (GPR) B-scan images. The model was trained on 1408 augmented B-scans collected with 200 and 400 ​MHz antennas across various subsurface materials, ensuring exposure to a wide range of material types with different electromagnetic properties. Testing experiments were performed using eight profiles where cavity detection was confirmed by borehole data. The results demonstrate an impressive 100% success rate for cavity detection and over 95% accuracy in depth estimation. Comparing this model to other deep learning-based methods, our results show great remarkable performance tested in various subsurface environments. Furthermore, the model's lightweight design can be deployed on normal portable computing machines, enabling real-time cavity detection and depth estimation during the acquisition. The proposed approach in this study provides practical solutions that can have a significant impact in civil engineering applications, providing an efficient and reliable tool for subsurface challenging problems.