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

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