Prediction of Long-Term Field Performance of Asphalt Concrete Overlays in a Southern State Using Supervised Machine-Learning

Abstract

Asphalt concrete (AC) overlays improve pavement condition by restoring the structural capacity of aged pavements. Several state agencies have reported significant budget savings by assessing the present condition and predicting the future performance of overlays. The performance of AC overlays is generally characterized by determining indices such as the Pavement Condition Index (PCI), International Roughness Index (IRI), and Random Cracking Index (RCI) during their service life. A review of the traditional pavement index prediction models showed that they are either mechanistic-empirical or purely empirical in nature. In hot and humid climates such as Louisiana, these prediction models are mostly linear, polynomial, or exponential, and are only a function of pavement age, thus masking the contribution of other key variables such as overlay thickness, pre-treatment PCI, etc. The prediction period is also limited to a maximum of five years. With the ever-increasing size of available pavement performance data, supervised machine learning algorithms can be utilized to develop prediction models that outperform traditional performance models. The objective of this study was to develop a machine-learning-based framework for states with a hot and humid climate to predict the long-term field performance (up to 11 years) of AC overlays based on key project conditions. PCI was used as the pavement performance indicator. Two machine learning algorithms, namely, random forest (RF) and CatBoost, were examined. A total of 892 log miles of AC overlay data were obtained from the Louisiana Department of Transportation and Development’s (LaDOTD) Pavement Management System (PMS) database. Based on the collected PMS data, six models were developed to sequentially predict PCI at AC overlay ages 1, 3, 5, 7, 9, and 11 based on highway function classification (C), overlay thickness (OT), overlay age (A), annual cumulative truck traffic (TT), annual cumulative rainfall (R), mean annual temperature (T), and PCI before overlay application (PCI-). Results indicated that the RF algorithm yielded higher accuracy than the CatBoost algorithm, with R2 values ranging from 0.91 at age 1 to 0.87 at age 11 in the training phase, and between 0.72 and 0.65 in the testing phase. Therefore, the RF-based models were considered in the proposed decision-making framework to predict the PCI of AC overlays based on PCI measurement in the pre-treatment stage and other project condition input data.