This is FAST: multivariate Full-permutAtion based Stochastic foresT method—improving the retrieval of fine-mode aerosol microphysical properties with multi-wavelength lidar

Abstract

Despite the small size and ubiquitous presence, fine-mode aerosols play a vital role in climate change and human health, especially in highly populated regions. Hitherto, some studies have been carried out to retrieve fine-mode aerosol microphysical properties from multi-wavelength lidar measurements. However, there has been a dearth of lidar-based methods with quality retrieval and high time efficiency to adequately quantify the impacts of fine-mode aerosols on Earth's energy budget. The reasons for the gap involve the limitations in current approaches and the nature of the under-determined problem with insufficient information of three backscattering coefficients (β) and two extinction coefficients (α), typically known as the 3β + 2α configuration. Furthermore, the latest lidars, especially for spaceborne and airborne applications, are inherently difficult to perform routine diurnal 3β + 2α observations with high resolution and require high processing speed for massive data. To overcome these conundrums, we developed a novel unsupervised machine learning method—multivariate Full-permutAtion based Stochastic foresT method (dubbed the FAST method)—to improve the retrieval of fine-mode aerosol microphysical properties. To the best of knowledge, this work is the first time that machine learning algorithms are employed in attempts to retrieve the aerosol microphysical properties with stand-alone multi-wavelength lidar data. The major merits of the FAST method include 1) high accuracy of fine-mode aerosol products with the typical 3β + 2α configuration, 2) acceptable performances for fewer input optical channels, and 3) high processing speed for large volume data. Comprehensive simulations have been conducted to investigate the error characteristic of the FAST method under different conditions. We also applied the FAST method to the airborne lidar data acquired during the NASA DISCOVER-AQ field campaign. The retrievals of the FAST method provide high resolution time-height distributions of fine-mode aerosol microphysical properties at 20-s temporal resolution and 45-m vertical resolution. In situ measurements are compared with multi-wavelength lidar retrievals showing good agreements. We achieved 0.010, 0.014 and 0.016 in terms of mean absolute difference for retrieved 532-nm single-scattering albedo with 3β + 2α, 3β + 1α, and 2β + 1α configurations, respectively. The proposed method is expected to represent an important step toward improving microphysical retrievals from multi-wavelength lidar data, especially for airborne and spaceborne lidar missions.