Pedestrian volume prediction with high spatiotemporal granularity in urban areas by the enhanced learning model

Abstract

Pedestrian volume prediction is a key strategy to explore the spatial patterns of pedestrian mobility and develop urban policies. However, due to the expensive costs of field sampling, most existing models are established on insufficient pedestrian samples and obtain limited prediction performance. Therefore, this study proposes an enhanced learning model for pedestrian volume prediction with high spatiotemporal granularity in urban areas. The enhanced learning model is applied to a case study in the central business district (CBD) of the city of Melbourne, Victoria, Australia. More than 1400 features are constructed for pedestrian estimation, covering macro aspects of transportation, socioeconomics, road networks, time, land use and place of interest. Compared with the optimal supervised learning model of LightGBM (Light Gradient Boosting Machine), the LightGBM-based enhanced learning model can significantly improve the performance of pedestrian estimation with root-mean-square error (RMSE) reduced by 41.75% and R-squared (R2) improved by 27.75%. The important parameters (i.e., spatial resolution, combination parameter) of the enhanced learning model are validated to significantly affect model performance. According to the spatiotemporal analysis of pedestrian volume in geographic information system (GIS) maps, different measures are proposed to optimize urban mobility and improve city management.