Prediction of Pavement Overall Condition Index Based on Wrapper Feature-Selection Techniques Using Municipal Pavement Data

Abstract

Municipalities make significant efforts with limited resources to collect pavement condition data. Overall Condition Index (OCI), which uses the pavement surface evaluation and rating manual to identify roadways needing repair, is a convenient and common way of pavement condition assessment. Many data used in assessing the OCI are collected from fieldwork. Some data features give little insight into road conditions, and one feature may provide similar information to another; thus, effective data-collection resources can be optimized by selecting which data feature to keep and which to discard. In addition, the OCI reflects how local agencies highlight the important variables driving their pavement-management investments. It is also a reflection of the triggers that they use to propose various treatment strategies. This research aimed to evaluate pavement distresses in West Des Moines, Iowa, using machine-learning methods, and determine which combination of distresses and their distress proportions can accurately predict the OCI class of a particular pavement type. The wrapper feature-selection methods were used, fitting their results to classification-tree models. Automatic feature selection, using Featurewiz, was also considered to select the desired number of variables. Feature parameters were screened for OCI prediction using the mean decrease impurity, and the results could be used to model classification that may be used for year-long predictions. Results showed that an effective OCI estimate methodology could be determined with significant accuracy with fewer features.