Machine Learning Approach to Predict International Roughness Index Using Long-Term Pavement Performance Data

Abstract

On-time pavement maintenance and rehabilitation are required to maintain or improve pavement roughness, which is an indicator of pavement performance and road safety. Better prediction of maintenance time can help in budget planning and allocation for highways as well as a better and safer driving experience for drivers. In this research, the International Roughness Index (IRI) for asphalt concrete pavement is predicted based on the 12,637 observations in the Long-Term Pavement Performance (LTPP) data set for 1,390 roads and highways in 50 states of the US and the District of Columbia from 1989 to 2018. To identify the research gaps and to better understand the state-of-the-art research in IRI prediction, a systematic literature review (SLR) has been performed to develop a comprehensive view of machine learning techniques used for IRI prediction. We used a machine learning algorithm that can handle missing data in the LTPP data set. Extreme gradient boosting (XGBoost) was used to predict the IRI. Also, the support vector regression (SVR) and random forest (RF) models were used to compare the results. Our results show that XGBoost provides a better model fit in terms of mean absolute error and coefficient of determination. Moreover, our results show that No.-200-passing, hydraulic conductivity, and equivalent single-axle loads in thousands (KESAL) are the most important factors in predicting the IRI.