Pavement Crack Rating Using Machine Learning Frameworks: Partitioning, Bootstrap Forest, Boosted Trees, Naïve Bayes, and <i>K</i>-Nearest Neighbors
Abstract
Deteriorating highway pavement conditions have largely been evaluated through visual inspection, nondestructive evaluations, smart sensing technologies, and image analysis techniques. These techniques have been successful in rating the conditions of roadway segments, but have also been faced with the challenge of subjective uncertainties and errors, signal noise, electrical and electromagnetic interference, and other effects on a large-scale implementation for the forecasting of the future condition of pavement sections. The goal of this paper is to implement some machine learning methodologies in predicting the condition of highway pavements based on previous pavement condition ratings and selected time variant and invariant covariates. The paper commences with a terse introduction of partition, bootstrap forest, gradient boosted trees, K-nearest neighbors, naïve Bayes, and the traditional multivariable linear regression techniques. Predictive accuracies, relative deviations, and the level of exactness of forecasting made by each model for the response variable are estimated to assess the stability and robustness of the models. From the results, it was generally observed that the machine learning methodologies were promising in predicting the crack of pavement based on the R2 statistics (0.6–0.9), average absolute errors (0.2–0.4), and root-mean-square errors (0.4–0.9). Also, based on the learning rates, number of layers, and sampling rates chosen, the three major split contributors observed were the age of the pavement, the average daily traffic, and the time series crack rating for the immediate past year. Finally, it was observed that the computed loss estimators for the bootstrap forest, boosted trees, partitioning, and nearest neighbor algorithms related to the abilities of the models to sample from the bootstrap aggregate and out-of-bag observations, utilize the boosting gradient, and average over several splits, respectively.