Retrieving High Temporal Resolution Aerosol Layer Height From EPIC/DSCOVR Using Machine Learning Method

Aerosol altitude is one of the key parameters for radiative forcing estimation and environment studies. In order to obtain wide coverage and multiple times aerosol layer height (ALH) per day, we developed a machine learning (ML)-based algorithm for retrieving ALH from the Earth polychromatic imaging camera (EPIC) observations. The reflectance of all bands from EPIC, solar and view angles are used as predictors. High-accuracy ALH from cloud-aerosol LIDAR with orthogonal polarization (CALIOP) is used as response. The algorithm was applied to observations over western sea of Africa (Area 1) and north plain of China (Area 2). The data from January to November 2021 are used for the XGBoost model construction, and independent 10-fold cross-validation (CV) results show very good performance of retrieval algorithm with correlation coefficient (R) larger than 0.9, root mean square error (RMSE) less than 0.55 km and the proportion falling within the expected error (EE) is over 80%. The independent EPIC ALH retrievals of whole 2016 year also show consistency with CALIOP observations. The RMSE is 0.78 (0.9) km and the proportion falling within EE is 58% (45%) for Areas 1 and 2, respectively. The comparisons between XGB retrievals and the operational aerosol optical centroid height (AOCH) product from EPIC for dust and haze cases show that the XGB retrievals are much more consistent with CALIOP observations than the operational EPIC AOCH. ALH retrievals from this study could capture the spatial and temporal variations both for the high aerosol layer (dust) and the low aerosol layer (haze).