Advanced cavity detection in ground penetrating radar B-scan image using fully convolutional networks

Detecting underground cavities and voids is critical for ensuring structural safety in sectors such as civil engineering and environmental studies. Ground Penetrating Radar (GPR) B-scan imaging is a valuable tool for this purpose, yet traditional methods often struggle with precise cavity characterization, especially as cavities develop over time. Addressing this gap, this study introduces an advanced methodology using Fully Convolutional Networks (FCNs) to improve cavity detection accuracy across four progressive stages: Initial, Intermediate, Critical, and Damaged. The GUI based KIT-GPR model, trained on Finite Difference Time Domain (FDTD) simulated data, can identify cavities as they grow from small initial voids to significant structural threats. This method influences GUI programming, enabling non-experts to interpret B-scan images more intuitively. Key findings indicate that while the KIT-GPR model demonstrates potential in cavity detection across different developmental stages, it faces challenges in accurately identifying and classifying cavities, particularly in complex scenarios. These limitations highlight the need for further refinement to improve detection reliability in GPR analysis and enhance its applicability in subsurface imaging and infrastructure monitoring.