Pavement Patterns: An Unsupervised Learning Approach to Texture Analysis in Rigid Pavements

Abstract

Rigid pavement surface texturing plays a critical role in tire–pavement interactions, and although high-speed laser profiling has enabled cost-effective network-level texture measurement, methods for automatically identifying and characterizing surface texturing techniques from field data remain underdeveloped. In this study, two-dimensional (2D) texture profiles were collected over 176 km (109 mi) of rigid pavements across Texas. After processing the raw profile data, a set of 40 texture indices were evaluated from which only six proved to be sufficiently uncorrelated to quantify different characteristics of the pavement’s macrotexture. Principal component analysis (PCA) then reduced dimensionality, and three clustering algorithms—partition around medoids (PAM), agglomerative nesting (AGNES), and Gaussian mixture models (GMM)—were evaluated for their ability to reveal natural groupings. GMM produced the most coherent segmentation, yielding five clusters that correspond to deep-groove tining, shallow-groove tining, conventional diamond grinding, next-generation concrete surfaces, and exposed-aggregate/dragged and-polish surfaces. Although average silhouette scores (~0.40) indicated only moderate cluster cohesion, qualitative inspection confirmed that each cluster captured distinct physical characteristics shaped by wear, polishing, and original texturing methods. These findings demonstrate that network-level texture data, when combined with unsupervised learning, can objectively distinguish prevailing surface textures without relying on incomplete or outdated maintenance records. The proposed framework lays the groundwork for integrating texture data into performance modeling, network segmentation, maintenance decision-making, for developing supervised classifiers to predict surface texturing methods which can thereby enhance empirical models of skid resistance, hydroplaning potential, and tire–pavement noise.