A Consideration of Heavy Rainfall Detection Method Using Multi-Parameter Phased Array Radar, Convolutional Neural Network, and Long Short-Term Memory Network

Abstract

In order to mitigate weather disasters caused by heavy precipitations, it is important to observe 3-dimensional precipitation structure in a storm with high temporal resolution. In recent years, the development of phased array weather radar is being promoted for high-speed precipitation observations. We propose an algorithm for predicting heavy rainfall using machine learning for the novel phased array weather radar (Multi-Parameter Phased Array Weather Radar: MP-PAWR) observation data. The algorithm predicts localized convective rainfall by extracting the vertical structure of storms observed by MP-PAWR for each precipitation cell. The proposed method with the combination of convolutional neural networks and long short-term memory networks were applied to various observation data from MP-PAWR with high spatial and temporal resolution to predict heavy rainfalls a few minutes later. The results showed that the use of specific differential phase data gave particularly accurate predictions for heavy rainfall compared to radar reflectivity factor and differential reflectivity data.