International Roughness Index prediction for flexible pavements using novel machine learning techniques

Abstract

International Roughness Index (IRI) is an important pavement performance indicator that is widely used to reflect existing pavement condition and ride quality. Due to the importance of this significant index, the current research aims to develop a precise IRI prediction model using the Gaussian Process Regression (GPR) and Locally Weighted Polynomials (LWP). The long-term pavement performance (LTPP) datasets of pavement age, initial IRI, alligator, longitudinal and transverse cracks, standard deviation of rutting, and subgrade plasticity index variables are employed in predicting IRI. These datasets are collected from 126 different flexible pavement sections of the LTPP specific pavement studies (SPS-1) located in different climatic zones in the US. The total number of IRI measurements in the collected database is 925 which covers a wide range of IRI values. Multiple linear regression (MLR) is firstly applied to classify the input variables. Then the MLR model is compared with four machine learning techniques which are GPR, LWP, Particle Swarm Optimization-Adaptive Network based Fuzzy Inference System (PSO-ANFIS) and PSO-Artificial Neural Networks (PSO-ANN). The developed models’ performance is validated using different statistical indices, including the coefficient of determination (R2). The results demonstrate that the GPR (R2=0.93) and LWP (R2=0.90) outperformed the PSO-ANFIS (R2=0.65) and PSO-ANN (R2=0.52) in predicting IRI. Thus, the GPR model is found to be more accurate for IRI modeling compared to the hybrid investigated models.