Aerosol optical depth retrieval using ground-based solar irradiance measurements and machine learning

Aerosol optical depth (AOD) is a key parameter for many scientific fields like air pollution and climate modelling. Ground-based AOD measurements outperform the other datasets in terms of data accuracy but still lack spatiotemporal resolution, while satellite/reanalysis encompasses the opposite behaviour. The overreaching goal of this study is to present a retrieval technique for cloud-free conditions that can be used to reproduce the AOD at the high temporal resolution of radiometric instruments. For the purposes of this analysis, reference measurements of aerosols (AERONET) and solar irradiance (BSRN) have been used. The retrieval technique uses various machine learning (ML) algorithms (e.g., gradient boosting machines, random forests, neural networks, etc.) by testing different components of ground-based solar radiation measurements in order to retrieve AOD at the 1-min temporal resolution of AERONET retrievals. In particular, each ML algorithm includes as an input parameter the optical air mass (m), water vapor (WV), and the direct or global clearness index. The WV is accessed through the Copernicus Atmosphere Monitoring Service Reanalysis, operated by ECMWF. For each ML algorithm, a randomized cross-validation searching method is applied to retrieve the optimal ML model architecture.

The ML-derived AOD retrievals are compared against reference ground-based (AERONET), satellite (MODIS), and reanalysis (MERRA-2, CAMS) AOD retrievals at 26 AERONET-BSRN stations under different aerosol and climatic conditions from 2004 to 2019. We validated the MLA-based AODs against reference AERONET retrievals, finding root mean square error (RMSE) values ranging from 0.01 to 0.15, irrespective of the underlying climate and aerosol environments. Among the different ML algorithms, the artificial neural networks outperformed the other algorithms in terms of RMSE at 54% of the measurement sites. The overall performance of ML-based AODs against AERONET revealed a high coefficient of determination (R2 = 0.97), a mean absolute error of 0.01, and an RMSE of 0.02.