Improving the Predictability of Heavy Precipitation Events during the Nowcast Period using Convolutional Neural Networks

Given the substantial risk to life and property that heavy precipitation events pose to the public, heavy rainfall events, particularly those that lead to flash flooding, are of great importance to atmospheric scientists. Accurate forecasting allows for emergency managers and the public to prepare for such dangers. Substantial improvements have been made to the short-term forecasting of such events through the implementation of the MetWatch Desk within the Weather Prediction Center (WPC). However, despite advances over the past several years in numerical weather prediction forecasting models, improvements can still be made to forecasting during the nowcast (0-6 hours) period. Meteorologists staffing the MetWatch Desk have identified limitations in forecast models that include inaccurate determination of sub-hourly rainfall rates as well as delayed convection initiation within Convective-Allowing Models (CAMs). This project aims to utilize machine learning techniques to better predict heavy precipitation events. The machine learning techniques in this project are using the Multi-Radar/Multi-Sensor (MRMS) System dataset. Convolutional neural networks will be utilized using this data with the aim of this research project to develop a tool that can better predict the location and magnitude of high-QPF events that may lead to flash flooding. This tool can then be used by meteorologists at the WPC MetWatch Desk to assist in nowcast forecasting and to issue products such as Mesoscale Precipitation Discussions.