Air pollutant PM$ _{10} $ estimation in Saudi Arabia using machine learning

Air pollution, specifically PM$ _{10} $, is a critical challenge in Saudi Arabia, where levels often exceed World Health Organization (WHO) guidelines due to industrial activities and arid conditions, posing serious risks to human health. A key barrier is obtaining accurate PM$ _{10} $ data, as estimations are limited by the few and unevenly distributed air quality stations. Notably, despite its severity, research on PM$ _{10} $ estimation in the region remains scarce. Atmospheric reanalysis datasets like MERRA-2 offer complementary data, but their model-based nature, lacking actual measurements, introduces potential biases. To bridge this gap, this study developed a machine learning framework to estimate daily and monthly PM$ _{10} $ concentrations in three climatically distinct Saudi cities. The framework integrates ground-based PM$ _{10} $ data, meteorological parameters, and MERRA-2 reanalysis data. To our knowledge, this study represents the first application of MERRA-2 for PM$ _{10} $ estimation in Saudi Arabia. The proposed AtmoStack is a stacked machine learning model, and we compared it against individual models (RF, HGB, CatBoost, and MLP) and state-of-the-art models, including LightGBM, ANN, and LSTM. Moreover, the framework incorporates feature-importance analysis to identify the most influential factors, helping to interpret the model. AtmoStack outperformed all baselines; in the dust-dominated environment of Buraidah, it achieved a daily $ R^2 $ of 0.73 and a monthly $ R^2 $ of 0.96. In Taif, it achieved a daily $ R^2 $ of 0.63 and a monthly $ R^2 $ of 0.94, indicating that AtmoStack effectively captures realistic distribution characteristics. These results support effective air-quality management and public health decisions.