Project No: 16212022

Title: Machine Learning Model Based on Mobile and Fixed Air Monitoring Network to Map Microscale NO2 Pollution in Complex Urban Environments

Principal Investigator: Prof. Zhi NING


Air pollution is one of the greatest environmental risks to health and ambient air pollution in both cities and rural areas and it is estimated to cause 4.2 million premature deaths worldwide in 2016 according to the World Health Organization (WHO). Recently, in September 2021, WHO tightened the Air Quality Guidelines (AQG) in response to rising health risks from various air pollutants. Among the different criteria pollutants, nitrogen dioxide (NO2) plays an important role in various respiratory diseases such as asthma, bronchial symptoms, lung inflammation and reduced lung function etc. Further, while progress has been made to reduce many of these pollutants, NO2 has proven resistant to control efforts in many places.

Understanding the spatial distribution of NO2 is critical to developing effective control strategies and for assessing public health risks. Several tools are available and described in the literature, and most of them provide estimates of pollutant concentrations over a regional or urban scale with a grid of 1-10 km, which is insufficient for estimating pollutant levels and exposure conditions encountered in the various built environments of cities. In many developed and developing countries and cities, regulatory monitoring networks have been set up to provide general pollutant concentration data. However, they are mostly sparsely distributed in cities, and there is limited knowledge of pollutant concentration to specific points of these cities. Recently, new monitoring practices have been developed and employed using stationary or mobile microsensor based monitoring platforms in networks with strict quality control and quality assurance protocols. These works have shown great potentials to improve the resolution of air quality information and characterize microscale pollution levels with considerable spatial details. However, the complex and large data sets from these networks have not been well utilized and there have been a lack of effective data fusion approaches to integrate the massive volume and often different dimensions of these data.

In the proposed study, we will develop and apply a machine learning-based method to achieve a microscale NO2 pollution mapping in sub-km resolution using data from multidimensional monitoring network measurements and well-developed modelling tools in the metropolitan Shanghai. The data includes 10 regulatory monitoring stations and 85 fixed air sensor stations, and 150 mobile sensors instrumented on taxis during the period from Jan 2020 to July 2021. Combined with satellite remote sensing data retrieved NO2 ground concentration at km scale grid, we propose a data assimilation-based machine model coupled with multi-source data to downscale the NO2 concentration map to a 100 m grid size on hourly time resolution.

A challenge for us to address is the fusion of data with widely different scales of time resolution, spatial coverage and frequency of data from these monitoring systems. Our team has developed and deployed microsensor based monitors in many air monitoring applications and developed protocols for quality assurance and quality control needed for effective use of the large data sets they produce. This study would offer an important extension in the use of big data collected from sensor networks with other more conventional air monitoring data to combat air quality problems. The success of the project will greatly improve the air quality information coverage in the urban places. The output will be of direct benefit to the air regulatory communities as well as atmospheric, health and exposure research. The approach can readily be applied in other megacities including Hong Kong as more and more such sensor networks are being established.