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

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.