Pavement roughness level classification based on logistic and decision tree machine learnings

Abstract

Two supervised machine learning methods, logistic regression and classification and regression trees (CART) are used to predict pavement roughness level using the data collected from the Long Term Pavement Performance (LTPP) database. Analysis results showed that the most significant factors for treatment roughness are pre-treatment roughness, overlay thickness, followed by overlay age, milling treatment and structural number are two marginal significant factors. In addition, high pre-treatment roughness and long service time increase the probability of high roughness; whereas thick overlay, deep milling and strong pavement structural capacity tend to reduce the roughness level. According to the machine learning performance measures including precision, recall, accuracy, and F1_score, the CART decision tree obtain much better classification results than that of the logistic regression. The decision tree uses multiple decision rules and therefore can also generate complex classification. Logistic regression is a linear classification but can provide explicit model form and parameter estimates.