NATIONWIDE ANNUAL AVERAGE DAILY TRAFFIC (AADT) ESTIMATION ON NON-FEDERAL AID SYSTEM (NFAS) ROADS BY MACHINE LEARNING WITH DATA MINING OF BUILT-IN ENVIRONMENT

Abstract

This study aims to address the nationwide gap in AADT data on NFAS roads in U.S. With a Spatial Autoregressive Model as a benchmark, two machine-learning approaches, i.e. Artificial Neural Network and Random Forest, show notable improvement in the accuracy of estimating AADT according to five measures, i.e. MSE, RSQ, RMSE, MAE, and MAPE. A data-mining of the built-in environment from three perspectives, i.e. on-road and off-road features, network centralities, and neighboring influences, paves the way for AADT estimation, which covers 87 variables in centrality, neighboring traffic, demographics, employment, land-use diversity, road network density, urban design, destination accessibility, etc. Data integration using different buffering sizes and statistical analysis of linearity and monotonicity promote the variable selection for estimation. When implementing the machine-learning approaches, not only the estimation performance is analyzed, but also the relationship between each variable and AADT, the interplays among variables, variable importance measures are thoroughly discussed.