Traffic volume data are essential for policymakers in traffic management and for constructing a high-resolution greenhouse gas emissions inventory within the road transport sector. However, traffic volume is measured only on select roads, and coverage is limited. This study aims to construct traffic volume estimation models based on traffic speed using machine-learning approaches, including linear regression, random forest, gradient boosting models, a deep neural network (DNN), and a combined long short-term memory and DNN (LSTM-DNN) model. Among these, the LSTM-DNN model demonstrated the best performance, achieving an R 2 of 0.9404, a mean squared error of 94,331, and a mean absolute error of 199. While performance varied across different road grades—with somewhat higher errors for national highway, special metropolitan city road, and local road—these variations did not substantially affect downstream applications such as CO 2 emissions estimation. Validation using CO 2 emissions calculated from the estimated traffic volumes showed similar levels to those from other institutions, confirming the appropriateness of the traffic volume estimation. Notably, achieving high performance using only traffic speed and road information highlights the significance and the practical potential of this study’s approach for scalable traffic volume estimation.
Kim et al. (Wed,) studied this question.