Doppler processing in weather radar using deep learning

Abstract

A deep learning approach to estimate the mean Doppler velocity and spectral width in weather radars is presented. It can operate in scenarios with and without the presence of ground clutter. The method uses a deep neural network with two branches, one for velocity and the other for spectral width estimation. Different network architectures are analysed and one is selected based on its validation performance, considering both serial and parallel implementations. Training is performed using synthetic data covering a wide range of possible scenarios. Monte Carlo realisations are used to evaluate the performance of the proposed method for different weather conditions. Results are compared against two standard methods, pulse-pair processing (PPP) for signals without ground clutter and Gaussian model adaptive processing (GMAP) for signals contaminated with ground clutter. Better estimates are obtained when comparing the proposed algorithm against GMAP and comparable results when compared against PPP. The performance is also validated using real weather data from the C-band radar RMA-12 located in San Carlos de Bariloche, Argentina. Once trained, the proposed method requires a moderate computational load and has the advantage of processing all the data at once, making it a good candidate for real-time implementations.