Interpretable machine learning approach in estimating traffic volume on low-volume roadways

Abstract

Abstract Many state and local agencies are currently facing challenges concerning the collection and estimation of traffic volumes, particularly regarding the collection of annual average daily traffic (AADT) on low-volume roads. To overcome these challenges, there is a need to develop new affordable methods to collect data and estimate traffic volume on low-volume roadways. In this study, the research team developed an innovative interpretable machine learning framework and applied it to low-volume roads in Vermont to estimate traffic volumes. This study used several databases (e.g., U.S. Census, the American community survey) to prepare the final dataset for the model development. The findings show that population density and work area characteristic (WAC) density are the best predictors in estimating AADT. The model outcomes show that the machine learning models yield better estimates than the conventional parametric statistical methods. By improving the accuracy of AADT estimations, this study contributed to traffic monitoring and safety improvement, and it can help reduce costs of data collection. This study developed the top five decision rules for three types of low-volume roadways. Stakeholders can use the findings of this study to meet the new requirements pertaining to availability of AADT estimates for low-volume roads. Additionally, the best fit estimates and the developed rules from the current study could enhance the predictive power of the SPF development for the low-volume roadways in Vermont and therefore improve the decision-making process.