Evaluation of road condition indices methods and applicability for use in machine learning

Abstract

Road maintenance is a crucial process for pavement management systems. South African local roads managed by lower road authorities (municipality, etc) are in critical condition, and their management is not at optimum level which is evident from their poor condition. The aim of this paper is to provide a Machine-Learning algorithm to assist road authorities to provide optimal maintenance strategies. The objective of the study was to determine the most effective condition index for management of flexible pavements. This is achieved by conducting descriptive and inferential statistical analysis of two case studies (Low volume roads and High-volume roads). Statistical analysis indicated that the visual condition index (VCI) has inconsistencies compared to the deduct point surface condition index (CISURF) and deduct point pavement condition index (CIPAVE) found in TMH 22. Four machine learning models were created which included the Gradient Boosting Classifier, Random Forest Classifier, Support Vector Machine Classifier, and Decision Tree Classifier. Of the four models explored, the model with the greatest potential for deployment was the Gradient Boosting Classifier (GBC) model. The GBC model had an accuracy of 74 %, 85 % and 93 % in relation to the VCI, CISURF & CIPAVE respectively. The CISURF and CIPAVE was identified as the most effective index for use in flexible pavements.