27 NOV 2025 (THU) 16:35 - 17:05

High-resolution NO2 estimation and forecasting using deep explainable neural networks

Ms XIAO Man    

( Supervisor: Prof Bo Huang )

Abstract:

NO2 is a significant air pollutant that poses serious risks to human health and contributes to various environmental issues. Accurate estimation and forecasting of NO2 concentrations are crucial for informed decisions regarding travel modes and preventive measures. Although numerous monitoring stations have been established to provide timely and precise NO2 concentrations, their sparse spatial distribution—due to high manpower and resource costs—limits the ability to generate comprehensive coverage. Remote sensing offers extensive earth observation capabilities, however, satellite data related to NO2, such as that from the Tropospheric Monitoring Instrument (TROPOMI), only reflects vertical column density rather than ground-level concentrations. As a result, these measurements do not directly convey the health impacts on populations. The strong relationship between ground-level NO2 concentrations and satellite data, alongside various geographical variables, suggests that mapping surface NO2 is feasible. Nevertheless, developing an interpretable, efficient, and effective model remains a big challenge. 

Carbon emission is an important contributor to climate change and global warming, raising urgent concerns among researchers. However, the scarcity of monitoring data makes it challenging to achieve accurate estimation and verification of carbon emissions. NO2 is generally recognized for its close relationship with carbon emissions, as they share common sources. This suggests that NO2 can serve as a valuable supplement to enhance the precision of CO2 emission estimations. Existing studies primarily focus on calculating data correlations while lacking in-depth exploration of spatial scales, land covers, and various processes, such as emissions and consumption. This limitation hinders further advancement in this field. 

Therefore, this research aims to achieve accurate estimations and predictions for ground-level NO2 concentrations and explore the detailed correlations with carbon emissions. Specifically, the objectives of this study are: 1) to estimate the high-resolution ground-level NO2 concentrations through data from monitoring stations, TROPOMI and auxiliary variables by combing model-based and deep learning-based methods; 2) to investigate the correlation mechanism between NO2 concentrations and carbon emissions varied with spatial scales, land covers, and different processes; 3) to generate the long-term NO2 forecasting based on historical observations and analyze their changing trends.