Aerosol optical depth prediction using machine learning techniques

Aerosol optical depth (AOD) describes adequately aerosols’s burden and extinction within an atmospheric column. AOD can be retrieved using remote sensing instruments such as ground-based sun photometers. Despite the very good quality of ground based AOD measurements, their spatiotemporal coverage is restricted. In this study, an alternative approach of AOD estimation is proposed with the synergy of ground-based measurements and machine learning (ML) techniques, in order to expand and complement the existing spatiotemporal capabilities of AOD data. The ML algorithms which are implemented are: Random Forests, Gradient Boosting Machines, Extreme Gradient Boosting Machines, Support Vector Regression, K-nearest Neighbors Regression, and Multivariate Adaptive Regression Splines. Each model receives as input information the Global Horizontal Irradiance (GHI) as well as water vapor (WV) content in hourly basis and under clear skies. A randomized cross-validation search scheme is implemented to obtain the optimal hyperparameters and avoid overfitting for each ML algorithm. GHI and WV are retrieved from Baseline Surface Radiation Network (BSRN) and NASA’s Modern-Era Retrospective Analysis for Research and Applications-2 (MERRA-2) reanalysis product respectively. AOD estimations are evaluated against AOD from AErosol RObotic NETwork (AERONET) inversion product, using the Level 2.0 Version 3 (L2V3) which provides cloud-screened and quality assured measurements. In total, 29 collocated AERONET-BSRN stations are used spanning from 2000 to 2019. Since, the aerosol pattern is different at each site, the effect of various aerosol types is further investigated. ML-based AOD predictions are adequately good, highlighting the feasibility of ML algorithms on producing AOD data. The results of this study could be useful for direct normal irradiance estimations as well as aerosols radiative effect calculations and climate projections.