Retrieval of Intensive Aerosol Microphysical Parameters from Multiwavelength Raman/HSRL Lidar: Feasibility Study with Artificial Neural Networks

Abstract

We present results of a feasibility study that uses Artificial Neural Networks (ANN) for the retrieval of intensive microphysical parameters of atmospheric pollution from combinations of backscatter (β) and extinction coefficients (α) that can be measured with multiwavelength Raman and high-spectral resolution lidar at 355, 532, and 1064 nm. We investigated particle effective radius, and the real and imaginary part of the complex refractive index. ANN could be a useful alternative or supplementary method over the traditional approach of retrieving microphysical particle properties with classical inversion algorithms because data analysis with ANN is significantly faster and allows for investigating the information content of the optical input data. We investigated the data combinations 3β+2α, 3β+1α (355 and or 532 nm), 2β (532, 1064 nm) +1α (532 nm), and 3β with Feedforward Backpropagation Multilayer Perceptron Neural Networks. The synthetic optical data were computed with a Mie-scattering algorithm for monomodal particle size distributions. Mean radii of the size distributions ranged between 0.01 and 0.5 µm, and mode widths ranged between 1.4 and 2.5 resulting in effective radii between 0.13 and 4.1 µm. We tested real parts between 1.2 and 2, and imaginary parts between 0.0i and 0.1i. The complexity of developing the networks did not allow us to test the influence of measurement errors of the optical data but the error produced by the ANN can be quantified. From the five basic data combinations, our current network design allows us to derive effective radius with an accuracy of approximately ±16 to ±35 %, and ±17 to ±39 % if the true mean radii is in the range from 110−250 nm, and 260−500 nm, respectively. The real part can be derived with an accuracy of approximately ±7 to ±10 %. We find retrieval errors of approximately ±31 to ±38 % for the imaginary part. We show that ANN can potentially estimate some particle parameters with various levels of uncertainty not only from what we denote as 3β+2α information but also from data combinations of 3β+1α (355 or 532), 2β (532, 1064) +1α (532), and 3β. We hypothesize that the ANN carries out first a pre-selections of various values of extinction-based Ångström exponents with regard to effective radius and then uses this information to create the strong correlation between particle effective radius and lidar ratios in all particle size distributions (PSDs) we investigated.