Retrieving aerosol properties using signals from an All-Sky camera and a random forest model

Abstract

The aim of this study is to predict the main aerosol properties in the atmosphere, Aerosol Optical Depth (AOD) and Angstrom Exponent (AE), with the aid of machine learning techniques and images from an All-Sky camera. Two different machine learning techniques have been used in this work: a random forest (RF) and an artificial neural network (ANN) with target values furnished by AERONET database. HDR images from the All-Sky camera sited in Burjassot (Spain) have been used. All of them have been taken in a clear-sky condition (without clouds) and with different aerosol depth. Selected images come out with a range from 0 to 0.5 of AOD at 500 nm as reference. The data in the groundbased station are available since the 10th of February of 2020 to the 31th March of 2021 in almost one year of samples. We have developed two ways of building signals combined with the two machine learning methods. Firstly, a signal generated from scattering angles in a single image which is obtained as the average of relative irradiance (RGB) using 100 random points in each scattering angle isoline, obtaining 29 values for each signal. Secondly, the signal has been generated in the same way but from zenith angles isolines of a single image. The main result obtained is that we improve significantly the state of art results of not calibrated images. For example, the red channel improves the percentage of predicted AOD values within the AERONET uncertainties from 62% to 90%-93% using an ANN and the zenith method.