A data-driven method of traffic emissions mapping with land use random forest models

Abstract

The development of intelligent approaches to quantify and mitigate on-road emissions is essential for urban and transportation sustainability for global megacities. Here, we utilize high-density traffic monitoring data and land use data to train random forest models capable of accurately predicting dynamic, link-level vehicle emissions. A total of 272 predicting indicators, including road features, population density, and land use information, were included in model training. Our model performed well, with a spatial generalization R2 > 0.8 for both volume and speed simulations. Dynamic link-based emissions of major air pollutants and carbon dioxide (CO2) were estimated for the whole road network of Chengdu, a populous city with the second greatest vehicle population in China. We adopted a generalized additive model to identify the drivers of spatial heterogeneity of on-road emissions and energy consumption, and nonlinear relationships between emissions, demographic and land use variables were found. Fine-grained assessments of emission reductions from potential Low Emission Zone policies are explored based on the high-resolution vehicle emission mapping tool. With high computational efficiency, the method is promising for handling traffic data streams in a real-time fashion, thus offering the potential for more precise vehicle emission management and carbon footprint tracking.