Accurately predicting air pollutants, especially in urban areas with well-defined spatial structures, is crucial. Over the past decade, machine learning technologies have been widely applied to forecast urban air quality. However, traditional machine learning methods have certain limitations in dealing with the spatial resolution of pollutants and complex nonlinear relationships. The research team proposed a convolutional neural network (CNN) model to predict the spatial distribution of CO concentration in Nanjing urban area at an ultra-high spatial resolution 10 m. Our model incorporates various factors as input, such as building height, topography, emissions, and is trained against the outputs simulated by the parallelized large-eddy simulation model (PALM) (Fig. 1). The high correlation coefficient (R2 = 0.96) between mean CO concentrations predicted by both PALM and CNN model indicates a high degree of accuracy in urban pollutant prediction offered by the CNN model (Fig. 2).
The CNN model also revealed the significant impact of input features such as topography and emissions on the CO concentration distribution. Fig. 3 shows different patterns of topography weight contribution learned by the CNN model, such as the proximity effect, indicating that the CO concentration at a certain location is more influenced by nearby topography and building heights. Additionally, the feature maps revealed the contributions of topography weights in different directions, such as horizontally (Figs. 3e-3h) and in other directions (Figs. 3i-3l), showing that CO concentration is influenced by a combination of factors including building heights and terrain. These features learned by the CNN model are consistent with the results of the PALM model, indicating that the model can effectively learn features related to the fluid dynamics laws within the PALM model on a fine scale, demonstrating strong interpretability. The model has also been successfully applied to pollutant prediction in Dongguan urban area, showing high predictive accuracy and proving its potential application in different cities. The research team will further integrate observational data and other predictive models, such as Long Short-Term Memory (LSTM) models, to achieve more accurate and real-time air quality predictions in the future studies.
Fig. 1. Workflow of the CNN model.
Fig. 2. Spatial distribution of CO concentration predicted by the PALM (left) and CNN models (right) under different scenarios.
Fig. 3. The weight contribution from the first layer of the CNN model for topography, which are normalized to a specific range [-0.025, 0.025]. Typically, red indicates higher positive weight contribution values, while blue denotes higher negative values. Arrows indicate the direction of increasing weight contribution.
This study, entitled “Predicting high-resolution air quality using machine learning: Integration of large eddy simulation and urban morphology data” was published on January 23, 2024 in Environmental Pollution (https://doi.org/10.1016/j.envpol.2024.123371). The first author of the paper, Shibao Wang, is a PhD student directed by Professor Yanxu Zhang. Dr. Jeremy McGibbon from the Allen Institute for Artificial Intelligence in the United States is the collaborative author. Prof. Yanxu Zhang is the corresponding author. This research has been supported by the National Key R&D Program of China (grant no. 2019YFA0606803), National Natural Science Foundation of China (grant no. 71974092), Start-up fund of the Jiangsu Innovative and Entrepreneurial Talents Plan, and Collaborative Innovation Center of Climate Change of Jiangsu Province. Special thanks to Professor Chris Bretherton from the University of Washington for his guidance.
References:
Wang, S., Jeremy, M., Zhang, Y*., (2024). Predicting high-resolution air quality using machine learning: Integration of large eddy simulation and urban morphology data, Environmental Pollution, 344, 123371, https://doi.org/10.1016/j.envpol.2024.123371.
Zhang, Y*., Ye, X., Wang, S., et al., (2021). Large-eddy simulation of traffic-related air pollution at a very high resolution in a mega-city: evaluation against mobile sensors and insights for influencing factors, Atmos. Chem. Phys., 21, 2917–2929, https://doi.org/10.5194/acp-21-2917-2021.
Wang, S., Ma, Y., Wang, Z., et al., (2021). Mobile monitoring of urban air quality at high spatial resolution by low-cost sensors: impacts of COVID-19 pandemic lockdown, Atmos. Chem. Phys., 21, 7199–7215, https://doi.org/10.5194/acp-21-7199-2021.