Exploring Self-Supervised GPR Representation Learning for Building Rooftop Diagnostics

Ground Penetrating Radar (GPR), known for its applications in diverse domains, demonstrates potential for nondestructive diagnostic assessments on building rooftops. This study delves into the unique characteristics and data structure of GPR, investigating the novel approach of processing GPR as a “contextual neighborhood” of A-scans within their respective B-scans as opposed to the typical pixel-based approach. Given the challenge of obtaining a large corpus of annotated rooftop GPR data, we employ self-supervised deep learning methods for GPR representation learning. Experiments include training a vanilla Autoencoder, Variational Autoencoder, and a Transformer-based Autoencoder on GPR A-scans. Additionally, we extend our analysis by fine-tuning a pre-trained Masked Autoencoder on image based GPR B-Scans to investigate the differences between the conventional pixel-based approach and our proposed A-scan-based approach. Through a meticulous analysis of the learned latent spaces across these methods, we assess the viability of self-supervised deep learning in encoding meaningful GPR representations for downstream tasks. This research contributes to the exploration of GPR’s applicability in building rooftop diagnostics and underscores the potential of self-supervised deep learning for efficient representation learning in the absence of annotated data.