Deep Convolutional Neural Networks for Hydrometeor Classification using Dual Polarization Doppler Radars

Description

Hydrometeor classification for dual polarization doppler radar is the process to identify the precipitation type based on the scattering properties of precipitation particle. Scattering properties of hydrometeor of different shape, size and orientation can be used to characterize and identify them. Deep learning methods like Convolutional Neural Networks (CNNs) have proven to be effective in image recognition and classification. We develop and apply deep CNNs, such as AlexNet, ResNet and VGG, to classify hydrometeors using dual polarization doppler weather radars at ARM SGP and NEXRAD Vance AFB facilities. We use four variables, (Horizontal Reflectivity (ZH), Differential Reflectivity (ZDR), Correlation Coefficient (ρHV) and Specific Differential Phase (KDP)) at the first elevation angle from January 1, 2015 to October 1, 2018 and combined as 3D timestack. We extracted 120,000 samples (4x60x60 kernels) the time series data divided into training, testing and validation set with four target hydrometeor categories (Ice Crystals (IC), Dry Snow (DS), Light and/or Moderate Rain (RA) and Big Drops (rain) (BD)). Preliminary experiments with a modified AlexNet with 5 convolutional layers and 3 fully connected layers shows a 70% accuracy after 1200 epochs. ResNet and VGG also reached similar accuracy level with even smaller epochs. Accuracy of above 81% was achieved when the kernel size was reduced to 4x30x30. We are applying the trained CNN models to doppler radar observations at ARM SGP facility.