Interpretable machine learning tools to analyze PM2.5 sensor network data so as to quantify local source impacts and long-range transport

Abstract

Sensor networks provide spatially resolved information about the time variation of PM2.5 concentrations in urban areas around the world. With relatively simple improvements to the control of the temperature and humidity of incoming air, and with proper quality assurance and calibration protocols, a low cost monitor was developed that provides measurements that were highly correlated with a reference PM2.5 monitor. These have sufficient resolution and reliability to quantify small differences within an urban area. Nevertheless, many sites report similar concentrations and it can be difficult to interpret the results or distinguish local from regional effects. Generalized Additive Models are an effective Machine Learning method to distinguish the impact of factors across very different scales. As a type of Interpretable Machine Learning / eXplainable Artificial Intelligence, they provide direct information on the link between specific factors and PM2.5 concentrations. GAM simulations were developed for sensors located around Dhaka, Bangladesh, for both the dry and the wet seasons. The simulations show that the largest contributor to high PM2.5 concentration variations across both urban and peri-urban sites is the boundary layer height which represent the vertical mixing of the urban plume. Using Trajectory Cluster Concentration Impacts, the simulations showed that robust estimates of long-range transport could be obtained from measurements located within a polluted environment, and the model further showed that enhancements of more than 40μg/m3 were associated with air transport from the Indo-Gangetic Plain in the dry season. Finally, interaction maps of the effect of horizontal wind speed and direction showed that these could be associated with up to +/−20 % variation in PM2.5 from site to site. Most of the enhancements are related to very calm winds and appear to be more strongly associated with road emissions than with point sources. Overall, the sensor network shows that air is polluted throughout the Dhaka area and into the peripheries, and that a multipronged approach will be needed to improve air quality for its inhabitants.