Estimation of geostationary satellite-based hourly daytime and nighttime AOD using machine learning

Atmospheric aerosols not only scatter or absorb solar radiation and affect the Earth’s radiation balance, which plays an important role in climate change, but also react with air pollutants and affect public health. In East Asia, due to naturally occurring Asian dust and anthropogenic air pollution resulting from urbanization and industrialization, continuous aerosol monitoring is crucial. Atmospheric aerosols are quantified by satellite- or model-derived Aerosol Optical Depth (AOD), which is defined as the extinction of solar radiation due to aerosols integrated over the atmospheric columns. In this study, machine learning-based models were developed to estimate daytime and nighttime AODs in East Asia using a geostationary satellite Geo-KOMPSAT-2A (GK-2A). Two machine learning approaches, random forest (RF) and light gradient boosting machine (LightGBM), were used in this study. Top-of-atmosphere (TOA) reflectance and brightness temperature (BT) from visible and infrared channels of GK-2A, meteorological data, geographical information, and auxiliary variables were used as input features to the machine learning models. The estimated AOD by the model was evaluated with ground-based AOD data from Aerosol Robotic Network (AERONET) by 10-fold cross-validation methods. To consider the model continuity of day and night and the model performance, two schemes using different combinations of input variables from GK-2A were examined: scheme 1 uses the same composition of input variables of BT for both daytime and nighttime for day-and-night continuity, and scheme 2 additionally uses TOA reflectance only during the daytime based on scheme 1 for high model performance. The LightGBM model (R² = 0.78, RMSE = 0.1099 for scheme 1, R² = 0.82, RMSE = 0.0993 for scheme 2) showed higher performance than RF model (R² = 0.76, RMSE = 0.1213 for scheme 1, R² = 0.76, RMSE = 0.1214 for scheme 2). Especially in LightGBM model, scheme 2 showed higher performance than scheme 1, and it is supported by the SHapley Additive exPlanations (SHAP) feature importance showing that TOA reflectance of visible and NIR channels of daytime of scheme 2 played an important influence on the model result. The estimated AOD from machine learning-based models were compared with GK-2A level 2 AOD and Copernicus Atmosphere Monitoring Service (CAMS) AOD forecast products. The spatiotemporal distribution in East Asia and time series trend at ground-based stations of estimated AOD show similar patterns to CAMS AOD forecast product, and generally agreed well with AERONET AOD. In conclusion, using the machine learning-based models proposed in this study, it is expected to contribute to continuous satellite-based aerosol and air quality monitoring over a specific region including nighttime, when geostationary satellite-based AOD retrieval is not available.