Sentinel-5P Based Estimation of PM_2.5 Concentrations Across Thailand Using Tabnet

The Fine particulate matter (PM2.5), concentration is a widely used indicator of air quality. The estimation of ground-level PM2.5 concentrations has been approximated by numerous models and machine learning (ML) approaches, utilizing satellite-derived datasets in their most recent advancements. However, past ML applications have used primarily tree-based algorithms, while neural network based methods are under-explored. In the current study we present a model via the Attentive Interpretable Learning neural network (TabNet), incorporating Sentinel-5P datasets and meteorological observations to estimate daily PM2.5 concentrations across Thailand. The proposed model achieves statistically reliable performances surpassing other tree-based leading ML models. Carbon Monoxide exhibits the highest significance on the model indicating agricultural burnings as the main emission source in Thailand. The proposed approach demonstrates the capacity of deep learning for PM2.5 estimations and expands its potential to improve the accuracy in air quality quantification.