Estimation of surface PM2.5 concentrations from atmospheric gas species retrieved from TROPOMI using deep learning: Impacts of fire on air pollution over Thailand

Abstract

Surface PM2.5 concentration is routinely observed at limited number of surface monitoring stations. To overcome its limited spatial coverage, space-borne monitoring system has been established. However, it also faces various challenges such as cloud contamination and limited vertical resolution. In this study, we propose a deep learning-based surface PM2.5 estimation method using the attentive interpretable tabular learning neural network (TabNet) with atmospheric gas species retrieved from the tropospheric monitoring instrument (TROPOMI). Unlike previous applications that primarily used decision tree-based algorithms, TabNet provides interpretable decision-making steps to identify dominant factors. By incorporating five TROPOMI products (i.e., NO2, SO2, O3, CO, HCHO), we have tested the system's capability to produce surface PM2.5 concentration without aerosol optical property, which was used more traditionally. The proposed model successfully captures spatiotemporal variations over Thailand in the period of 2018–2020, and it outperforms other leading machine learning models, particularly at high concentrations. The interpretable decision-making steps highlight that carbon monoxide is the most influential chemical component, which relates to the seasonal burning in southeast Asia. It is found that air quality impacts from fire are stronger in the northern part of Thailand and fires in neighboring countries should not be neglected. The proposed method successfully estimates surface PM2.5 concentration without aerosol optical property, implying its potential to advance monitoring air quality over remote regions.