Evaluation for the machine learning based PM_2.5 estimation using different spatial resolution of geostationary satellite AOD over megacities in Korea

Using machine learning skills, this study evaluated the performance of surface PM_2.5 estimation based on the consideration of geostationary satellite AOD having two different spatial resolution. This study targeted two megacities in Korea: Seoul and Busan. We used PM_2.5 data from hundreds of low-cost sensors installed in these megacities over a 19-month period (June 2018 to December 2019), combined with 6 × 6 km² (officially released version) and 0.5 × 0.5 km² (research-purpose product) AOD data from the Geostationary Ocean Color Imager (GOCI). Both cross-validation and independent validation against national surface PM_2.5 observations (AirKorea) showed improved performance at finer spatial resolution. When our machine-learning PM_2.5 estimation using low-cost sensor data was compare to AirKorea data, the 0.5 × 0.5 km² resolution yielded higher R² and lower RMSE compared to the 6 × 6 km² resolution in both cities; R² slightly increased from 0.32 (0.44) to 0.38 (0.45) and RMSE decreased from 11.36 (10.18) to 10.82 (10.13) μg/m³ in Seoul (Busan) at 0.5 × 0.5 km² resolution. When the variable importance is evaluated, 6 × 6 km² AOD contribution is large, but 0.5 × 0.5 km² AOD contribution is not significant in the machine learning process. This finding shows that the usage of satellite data having higher spatial resolution results in better performance of PM_2.5 estimation in spite of larger uncertainty. In other words, it is expected to achieve better results when more qualified AOD is ready in the future with higher spatial resolution. We also provided estimated PM_2.5 in a hourly scale, which is another advantage to use the geostationary satellite AOD.