Application of Shannon Entropy for Predicting Pavement Macrotexture Using Machine Learning Technique

Abstract: The Safe System approach, the guiding paradigm of the U.S. Department of Transportations (DOT) National Road Safety Strategy, recognizes an essential aspect overlooked in the transition towards proactive pavement surface management on the Nation’s roads to reduce crashes and improve safety is the integration of pavement surface macrotexture into asphalt mix design and aggregate properties. According to the Federal Highway Administration (FHWA), inadequate pavement texture, particularly macrotexture, can increase skidding crashes, compromising safety, and potentially leading to serious injuries and fatalities, especially in high-speed traffic. While pavement safety is achieved through pavement surface characteristics, agencies must ensure adequate levels of friction and macrotexture to maximize contact and traction at the pavement-tire interface to minimize crash risk potential. Studies show that pavement surface macrotexture significantly influences wet crashes by playing an essential role in facilitating water drainage and improving friction between pavement and tire. Macrotexture is defined by the deviation of a pavement surface and refers to the larger-scale surface texture, determined by the size, shape, gradation, and spacing of aggregate in asphalt concrete pavement (AC).
Predicting pavement macrotextures using mix design and aggregate properties can involve elements of multi-criteria decision-making (MCDM). MCDM techniques are methodologies used to make decisions in situations where multiple, often conflicting criteria need to be considered simultaneously. Although conventional modeling techniques such as linear regression are frequently utilized, have often not comprehensively capturing the complex non-linear relationship between macrotexture, aggregate properties, and asphalt mix characteristics. Accordingly, further exploration is needed to comprehensively compare the performance in obtaining complex relationships in predicting pavement safety metrics, especially pavement surface macrotexture. Furthermore, analysis confirms that Machine Learning (ML) techniques demonstrate strong generalization and proficiency in predicting future data.
The objective of this research is to examine the effectiveness of ML techniques, especially Artificial Neural Network (ANN) and Random Forest (RF) algorithms, in modeling the expected field-macrotexture values of hot mix asphalt surfaces based on aggregate and mix properties. This study utilizes production mix properties and as-constructed macrotexture data within a year of construction for DGAC and SMA surface mixes. The primary parameters for the investigation include Nominal Maximum Aggregate Size (NMAS) and sieve analysis for gradation characterization, as well as binder content, air voids, Void in Mineral Aggregate (VMA), Void in Total Mix (VTM), and particle size distribution (Cc and Cu). In addition, the integration of Shannon Entropy (SE) as a feature selection and solution to the MCDM will be investigated within RF and ANN models for predicting pavement macrotexture. The ML models are trained and evaluated using varying subsets of predictors selected based on the entropy weights for the dataset from six state DOTs. The RF model produced the most accurate predictions, achieving the highest R2 values and the lowest RMSE. Furthermore, the graphical analysis of predicted macrotexture validates the significance of the preferred model. The findings of this study contribute to advancing predictive modeling techniques in pavement evaluation, offering a novel approach to feature selection and model assessment.