A machine learning ensemble model for predicting pavement conditions using automatic laser crack measurement data

Abstract

ABSTRACT Automated pavement survey (APS) systems (e.g. Laser Crack Measurement System) are increasingly adopted by transportation planners as a data-enriching and economical alternative to the traditional windshield pavement survey (WPS) systems. Application arises where transportation planners wish to use APS measurements to predict WPS distress scores, as such prediction will enable them to continue using the WPS-based pavement asset management system while using the newly collected APS data. In this paper, we develop a machine learning model to predict the WPS pavement distress scores using APS variables, by employing decision tree classifier based ensemble learning model. The novelty is threefold. First, instead of using the mean we use the distributions of APS variables in the prediction, by examining multiple statistics for every APS variable. Second, these statistics are then used to develop decision tree classifiers as base learners in predicting WPS distress scores. Third, base learners are ensembled using various weight assignment strategies. Our computational test uses the WPS and APS data for two wheel path crack distress from the Kentucky roadway systems, and suggests that conditional probability-based weight assignment provides the best prediction with approximately 79% accuracy.