Publications / 2025 Proceedings of the 42nd ISARC, Montreal, Canada

GPR-Former: Context-Aware Moisture Detection

Kevin Lee, Chen Feng
Pages 1-8 (2025 Proceedings of the 42nd ISARC, Montreal, Canada, ISBN 978-0-6458322-2-8, ISSN 2413-5844)
Abstract:

Moisture damage in roofing systems poses a significant challenge to building operation and maintenance, with far-reaching implications for energy efficiency, structural integrity, health, safety, and sustainability. This study introduces GPR-former, a novel transformer-based architecture designed to leverage spatial context in ground-penetrating radar (GPR) data, enhancing the accuracy of moisture detection. Unlike our previous approach that analyzes isolated B-scans, GPR-former incorporates spatial groupings of A-scans with coordinate-based positional encoding, enabling the model to detect complex, nonlinear moisture patterns. Comprehensive experiments across diverse roofing materials demonstrate that GPR-former outperforms our previous method by up to 6% across performance metrics and achieves improved calibrated results. These findings highlight the potential of GPR-former as an improved, transformative tool for sustainable building maintenance, contributing to embodied carbon reduction and extending roof lifecycle management.

Keywords: Moisture detection, ground-penetrating radar, transformers, spatial context, sustainability