Scaling from macro to micro: A novel approach to bridging gaps in multiple pavement texture scales using generative neural networks

Abstract

Both pavement macrotexture and microtexture impact skid resistance, which is vital to driving safety. Current laser-based texture measurement methods struggle to balance efficiency and accuracy in large-scale surveys. Static laser scanners offer highly precise texture data but slow in operation, while vehicle-mounted 3D lasers work at traffic speeds but are inferior in precision. To address this trade-off, a series of generative neural networks called Pavement Texture Scaling Networks (PTSNs) are introduced to scale pavement texture across both macro and micro scales. PTSNs feature a multi-layer invertible architecture where each layer doubles or halves the texture resolution, progressively upscaling lower-resolution data to the desired level. The model was trained on texture data from four asphalt surface types at ten resolutions and evaluated with six texture descriptors and wavelet coherence (WTC). At scaling factors of 8×, 64×, and 512×, PTSNs achieved mean profile depth errors of 2.98 %, 3.91 %, and 4.99 %, respectively. The actual and predicted texture power spectral densities coincide at macrotexture scales but diverge at finer microtexture scales (wavelength $q>{10}^5$ m⁻¹). Additionally, PTSNs’ performance varied over polishing levels, with the highest errors observed on unpolished surfaces and the lowest on highly polished surfaces. The WTC analysis found that actual and predicted textures correlated strongly across the lane at frequencies below 32 kHz. Overall, PTSNs effectively reconstruct multi-resolution texture across scales, bridging the resolution gap and offering a fast, cost-effective alternative for high-precision pavement texture measurement.