Enhanced GPR signal interpretation via deep learning fusion for unveiling road subsurface conditions

• Published in: Measurement : journal of the International Measurement Confederation Vol. 249; p. 117007
• Main Authors: Zhong, Shan, Wu, Difei, Du, Yuchuan, Yan, Yu, Liu, Chenglong, Weng, Zihang, Wang, Guoqing, Xu, Fei
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 31.05.2025
• Subjects: Deep learning; Electromagnetic signal display; GPR data fusion; Ground penetrating radar; Road structure monitor; Deep learning; Ground penetrating radar; GPR data fusion; Road structure monitor; Electromagnetic signal display
• ISSN: 0263-2241
• DOI: 10.1016/j.measurement.2025.117007

[Display omitted] •Fusion framework based on deep learning proposed for integrating multiple GPR images.•Three fusion cases for unveiling different road subsurface condition tasks are proposed.•Two fusion network architectures are discussed to assess their impact on fusion process.•The fusion results significantly enhance the image quality compared to the original profile.•The proposed method outperforms competitors in the GPR image fusion domain. Ground Penetrating Radar (GPR) is an indispensable tool for assessing the internal condition of roads. However, the high dynamic range of electromagnetic wave signals often surpasses the capacity of conventional image representations. During the imaging process, electromagnetic wave signals are subject to compression and clipping. This limitation is particularly acute for weak amplitude signals, as they are compressed into narrow color ranges and become challenging to detect. To overcome this limitation, this study introduces a local mapping method that partitions raw data into multiple patches, map each patch independently, and seamlessly stitches the results. This approach ensures that weak signals remain unaffected by the influence of strong amplitudes in other regions. Furthermore, a fusion framework, GPRFusion, is proposed to integrate complementary information from traditional GPR images and local mapped GPR images. The fused images preserve the traditional amplitude distribution while enhancing the visibility of weak features, minimizing the risk of critical information being overlooked by expert. Experimental results reveal that GPRFusion enables the clear visualization of weak signals with amplitudes up to 20 times lower than dominant signals. Moreover, it outperforms other fusion methods in terms of SSIM and PSNR metrics, establishing a new standard for GPR image fusion.