Accurately assessing the health risks of air pollution remains challenging owing to the limited spatial coverage of ground-based monitoring stations. Geostationary satellites, such as the Geostationary Environment Monitoring Spectrometer (GEMS), provide unprecedented opportunities for real-time large-scale air quality surveillance. This study integrated multi-source data, including GEMS satellite observations, ground monitoring data, meteorological variables, and geographic information to predict daytime hourly concentrations of six key air pollutants (particulate matter with an aerodynamic diameter ≤ 2.5 μm PM₂.₅, particulate matter with an aerodynamic diameter ≤ 10 μm PM₁₀, O₃, NO₂, CO, and SO₂) across Taiwan. A multi-output categorical gradient boosting (CatBoost) model was developed to simultaneously estimate the concentrations of all six pollutants within a single framework. To enhance the robustness and operational relevance of the model, a rolling prediction approach was employed for training and validation. The model demonstrated strong predictive performance, with R² values of 0.86 (O₃), 0.84 (PM₁₀), 0.81 (PM₂.₅ and NO₂), 0.79 (CO), and 0.52 (SO₂), and mean absolute error (MAE) values ranging from 0.06 to 9.40. O₃ predictions were particularly stable, exhibiting relatively low root mean squared error (RMSE) and MAE values. A spatial comparison of predicted and observed concentrations from July‒December 2023 revealed strong concordance, although some localized discrepancies were observed. Overall, the model effectively captured complex spatiotemporal pollutant dynamics. These findings demonstrated that integrating GEMS observations with multisource data in a machine-learning framework enabled accurate, real-time, hourly predictions of multiple air pollutants and offers a scalable solution to support public health policy and exposure risk assessment.
Lin et al. (Sun,) studied this question.