Convolutional Neural Network (CNN) supported urban design to reduce particle air pollutant concentrations

PM2.5 has become a significant factor contributing to the haze outbreak in mainland China, which has negative impacts for public health. The current agility of CFD-based modelling to reveal in real-time the changes in PM2.5 concentrations in response to (proposed) changes in urban form limits its practical applications in the design processes. To support urban design for better air quality (AQ), this study presents a machine learning approach to test: (1) that the spatial distribution of PM2.5 concentrations measured in an urban area reflects the area’s capacity to disperse particle air pollution; (2) that the PM2.5 concentration measurements can be linked to certain urban form attributes of that area. A Convolutional Neural Network algorithm called Residual Neural Network (ResNet) was trained and tested using the ChinaHighPM2.5 and urban form datasets. The result is a ResNet-AQ predictor for the city centre area in Beijing which had one of the highest air pollution levels within the Beijing-Tianjin-Hebei region. The urban area covered by the ResNet-AQ predictor contains 4,000 grid cells (approx. 25.3 km x 25.3 km), of which 1,200 (30%) cells were selected randomly for testing. The ResNet-AQ prediction accuracy achieved 87.3% after 100 iterations. An end-use scenario is presented to show how a social housing project can be supported by the AQ predictor to achieve better urban air quality performance.