Using artificial neural networks to predict riming from Doppler cloud radar observations

• Published in: Atmospheric measurement techniques Vol. 15; no. 2; pp. 365 - 381
• Main Authors: Vogl, Teresa, Maahn, Maximilian, Kneifel, Stefan, Schimmel, Willi, Moisseev, Dmitri, Kalesse-Los, Heike
• Format: Journal Article
• Language: English
• Published: Katlenburg-Lindau Copernicus GmbH 24-01-2022; Copernicus Publications
• Subjects: Accretion; Aerosols; Artificial neural networks; Case studies; Cloud systems; Clouds; Configurations; Convective clouds; Convective systems; Data acquisition; Datasets; Deposition; Doppler sonar; Experiments; Feature extraction; Freezing; Graupel; Gravity waves; Ground-based observation; Ice; Ice particles; Neural networks; Precipitation; Precipitation (Meteorology); Precipitation formation; Predictions; Radar; Radar data; Radar reflectivity; Radiation; Reflectance; Remote sensing; Rime; Skewness; Snowfall; Training; Variables; Velocity; Water; Width; United States > US; Finland
• ISSN: 1867-8548; 1867-1381; 1867-8548
• DOI: 10.5194/amt-15-365-2022

Riming, i.e., the accretion and freezing of supercooled liquid water (SLW) on ice particles in mixed-phase clouds, is an important pathway for precipitation formation. Detecting and quantifying riming using ground-based cloud radar observations is of great interest; however, approaches based on measurements of the mean Doppler velocity (MDV) are unfeasible in convective and orographically influenced cloud systems. Here, we show how artificial neural networks (ANNs) can be used to predict riming using ground-based, zenith-pointing cloud radar variables as input features. ANNs are a versatile means to extract relations from labeled data sets, which contain input features along with the expected target values. Training data are extracted from a data set acquired during winter 2014 in Finland, containing both Ka- and W-band cloud radar and in situ observations of snowfall by a Precipitation Imaging Package from which the rime mass fraction (FRPIP) is retrieved. ANNs are trained separately either on the Ka-band radar or the W-band radar data set to predict the rime fraction FRANN. We focus on two configurations of input variables. ANN 1 uses the equivalent radar reflectivity factor (Ze), MDV, the width from left to right edge of the spectrum above the noise floor (spectrum edge width – SEW), and the skewness as input features. ANN 2 only uses Ze, SEW, and skewness. The application of these two ANN configurations to case studies from different data sets demonstrates that both are able to predict strong riming (FRANN > 0.7) and yield low values (FRANN ≤ 0.4) for unrimed snow. In general, the predictions of ANN 1 and 2 are very similar, advocating the capability of predicting riming without the use of MDV. The predictions of both ANNs for a wintertime convective cloud fit with coinciding in situ observations extremely well, suggesting the possibility to predict riming even within convective systems. Application of ANN 2 to an orographic case yields high FRANN values coinciding with observations of solid graupel particles at the ground.