Validation of Support Vector Regression in deriving aerosol optical thickness maps at 1 km2 spatial resolution from satellite observations

Abstract

As a result of great improvements in satellite technologies, satellite-based observations have provided possibilities to monitor air pollution at the global scale with moderate quality in comparison with ground truth measurement. In tradition, the inversion process that derives atmospheric parameters from satellite-based data is replied on simulated physics models of matter interactions. Recently, the usage of machine learning techniques in this field has been investigated and presented competitive results to the physical approach. In this paper, we present validation of Support Vector Regression (SVR) technique in estimating Aerosol Optical Thickness (AOT), one of the most important atmospheric variables, from satellite observations at 1×1 km2 of spatial resolution. Validation by different European countries is carried out on a large amount of datasets collected in three years, which aims at investigating prediction quality of SVR data models built up on discrete and sparse data around ground measurement sites on continuous data domain presented by maps. The validation results obtained from 172 datasets showed good performance of SVR over most of the 31 countries that were considered.