A Machine Learning Approach to Retrieving Aerosol Optical Depth Using Solar Radiation Measurements

Abstract

Aerosol optical depth (AOD) constitutes a key parameter of aerosols, providing vital information for quantifying the aerosol burden and air quality at global and regional levels. This study demonstrates a machine learning strategy for retrieving AOD under cloud-free conditions based on the synergy of machine learning algorithms (MLAs) and ground-based solar irradiance data. The performance of the proposed methodology was investigated by applying different components of solar irradiance. In particular, the use of direct instead of global irradiance as a model feature led to better performance. The MLA-based AODs were compared to reference AERONET retrievals, which encompassed RMSE values between 0.01 and 0.15, regardless of the underlying climate and aerosol environments. Among the MLAs, artificial neural networks outperformed the other algorithms in terms of RMSE at 54% of the measurement sites. The overall performance of MLA-based AODs against AERONET revealed a high coefficient of determination (R2 = 0.97), MAE of 0.01, and RMSE of 0.02. Compared to satellite (MODIS) and reanalysis (MERRA-2 and CAMSRA) data, the MLA-AOD retrievals revealed the highest accuracy at all stations. The ML-AOD retrievals have the potential to expand and complement the AOD information in non-existing timeframes when solar irradiances are available.