Prediction of high-level cloud parameters based on polarization lidar observations using machine learning tools

Abstract

The paper focuses on machine learning methods in atmospheric laser remote sensing. Atmospheric lidar data have been stored since 2009, when high-altitude matrix polarization lidar measurements became systematic. Light detection and ranging (lidar) data are used to determine optical (backscattering phase matrix (BSPM), thickness, and scattering ratio) and geometrical (altitude of lower and upper boundary, vertical extent) parameters of higher-level clouds. The data storage is then added by meteorological parameters derived from radiosonde observations and ERA5 reanalysis data (temperature, relative and specific humidity, wind speed and direction). Methodology includes machine learning models, random forest, and modified version incorporating principal component analysis for dimensionality reduction to probe the complex relationship between meteorological parameters and specific BSPM elements. For the first time, different lidar measurement techniques identify two distinct maxima in the distribution of BSPM elements.