VOODOO — A deep learning approach for reVealing supercOOled liquiD beyOnd lidar attenuatiOn

Abstract:

A novel approach for thermodynamic-phase classification of deep or multi-layer mixed-phase clouds (MPC) using ground-based cloud radar Doppler spectra and deep learning techniques is presented. The phase characterization (liquid vs. ice) in MPC is an observational challenge which can be addressed by synergistic profiling measurements with polarization lidars and Doppler cloud radars, as done within Cloudnet (Illingworth et al., BAMS 2007), a European cooperative effort to establish a network of ground-based remote sensing stations and data processing chains to improve the representation of clouds in weather and climate models. Cloudnet provides a target classification which distinguishes, amongst others, liquid water from ice hydrometeors. To date, liquid detection relies on lidar observations, which are sensitive to the numerous small cloud droplets and can thus detect supercooled liquid (SCL) layers with high accuracy. However, at a penetrated optical depth of about three, lidars suffer from complete signal attenuation. Cloud radars are able to penetrate multiple liquid layers and can thus be used to expand the identification of cloud phase to the entire vertical column beyond the lidar signal attenuation height, if morphological features in cloud radar Doppler spectra can be related to the existence of SCL. Relevant spectral signatures such as bimodalities or spectral skewness are extracted by a deep convolutional neural network (CNN) that relates to the cloud phase by training in a supervised scheme, using Cloudnets’ target classification as supervisor. Predictions are evaluated using several independent measurements such as liquid water path (LWP) detected by microwave radiometer, (liquid) cloud base detected by ceilometer, as well as relative humidity and temperature measured by radiosondes. An earlier machine learning approach has been applied successfully by Luke et al. (JGR 2010), on data from the U.S. Department of Energy (DOE) Atmospheric Radiation Measurement (ARM) instruments in Barrow, Alaska. VOODOO is first applied to a unique dataset from the Dynamics, Aerosol, Cloud And Precipitation Observations in the Pristine Environment of the Southern Ocean (DACAPO-PESO) and will become part of the Cloudnet processing toolbox in the future.