A method to estimate missing AERONET AOD values based on artificial neural networks

Abstract

In this work, we present a method to predict missing aerosol optical depth (AOD) values at an AERONET station. The aim of the method is to fill gaps and/or to extrapolate temporal series in the station datasets, i.e. to obtain AOD values under cloudy sky conditions and in other situations where there is a temporary or permanent lack of data. To accomplish that, we used historical AOD values at two stations, air mass trajectories passing through both of them (calculated by using the HYSPLIT model) and ANN calculations to process all the information. The variables included in the neural network training were the station numbers, parameters representing the annual average trend of meteorological conditions, the number of hours and the distance traveled by the air mass between the stations, and the arrival height of the air mass. The method was firstly applied to predict AOD at 440 nm in 9 stations located in the East Coast of the US, during the years 1999–2012. The coefficient of determination r2 between measured and calculated AOD values was 0.855, which show the good performance of the method. Besides, this result represents a remarkable improvement compared to three simple approaches. To further validate the method, we applied it to another region (Iberian Peninsula) with different characteristics (lower density of AERONET stations, different meteorology, and lower wind field spatial resolution). Although the results are still good (r2 = 0.67), the performance of the method was affected by these characteristics. Considering the obtained results, this method can be used as a powerful tool to predict AOD values in several conditions. The methodology can also be easily adapted to predict AOD values at other wavelengths or other aerosol optical properties.