Deep learning for surface precipitation estimation using multidimensional polarimetric radar measurements

Traditionally, polarimetric radar-based rainfall estimates are derived through empirical parametric relations obtained from nonlinear regression between rain rates and simulated radar data. The performance of such empirical relations is highly dependent on the variations of raindrop size distribution. In real applications, such empirical algorithms often need to be adjusted for different climate regimes and/or different rainfall types, and we do not have a simple parametric expression linking radar observables to rainfall intensity. Even if one can eliminate all the random errors associated with radar measurements, the parameterization uncertainty inherent in the empirical relations is hard to reduce. In addition, it is difficult to estimate surface rain rates with the radar measurements aloft, especially during the precipitation events characterized by dramatically changing vertical structures (i.e., precipitation particle sizes and distributions are changing during the falling process). In this paper, a convolutional neural network-based machine learning approach is developed to estimate surface rainfall rate from multidimensional polarimetric measurements. This deep learning algorithm can extract the complex relation from high dimensional input space (i.e., radar data) to the target space (i.e., surface rain rate). Polarimetric radar observations and rain gauge data collected near Melbourne, Florida are utilized for demonstration. The trained model is also extended to the whole radar coverage domain to provide complete rainfall mapping. Independent verification results show that this deep learning model has superior performance to conventional fixed-parameter rainfall relations based on either single- or dual-polarization radar measurements.