Aerosol Optical Depth Retrieval by Machine Learning Methods from Geostationary Environmental Monitoring Spectrometer

Globally, aerosols, emissions, and greenhouse gases have an impact on the environment. With the help of satellite data and field instruments we can understand and continue to study the atmosphere. Specifically, in terms of understanding air quality and aerosol optical depth (AOD), the radiative transfer model is traditionally used but can have unpredictability and is time consuming. With use of machine learning, one can improve accuracy and can be more time efficient. In this paper, we present machine learning methods to estimate AOD from the Geostationary Environmental Monitoring Spectrometer (GEMS). GEMS has a hyperspectral scanning spectrometer that monitors air pollutants over Asia by different observation nodes. Random Forest (RF), and Light Gradient Boosting Machine (LGBM) with auxiliary. meteorological, and ground-based observation data were used to estimate hourly AOD. Inclusion of meteorological data can support the model in performance and reflecting dynamic conditions in the atmosphere. The two machine learning models were evaluated by random, spatial, and temporal 10-fold cross validation to test the transferability and robustness. The results showed that random forest model performed lower than the light-gradient boosting model. LGBM produced R² of 0.286 – 0.680 and RSME of 0.025-0.057. Random forest produced R² of 0.283 – 0.643 and RSME of 0.028-0.057. Overall, the model was able to show that AOD can be retrieved from machine learning methods from the Geostationary Environmental Monitoring Spectrometer (GEMS).