Deep learning-based investigation of the impact of urban form on the particulate matter concentration on a neighborhood scale

Despite the importance of urban forms to the dispersion of particulate matter (PM), only a few studies exist on the relationship between them due to the limitations of the data and methodology. Thus, this study used a deep learning-based investigation of the impact of urban form on PM 2.5 concentration. Autoencoder, long short-term memory (LSTM), and the random forest model were used to analyze their relationship. The random forest model showed that urban form variables predict PM 2.5 concentration with a 95.66% accuracy, confirming that urban form characteristics significantly impact PM 2.5 concentrations. Among the urban form variables, floor area ratio turned out to be the most important, suggesting the need for more detailed efforts to reduce PM 2.5 in locations with high floor areas. The effect on PM 2.5 prediction accuracy was evaluated with root mean square error. It was difficult to accurately predict PM 2.5 in areas with large building coverage areas and low-rise residential areas. This study improved the accuracy of results on the influence of urban form by using PM 2.5 non-aggregated data measured hourly over 4Â years, which expanded the applicability of deep learning-based urban analysis.