A Deep Neural Network Model for Improving Hourly Precipitation Estimates from Numerical Models

Precipitation prediction is fundamental to better understand the land-atmosphere water budget, improve water resources management, and prepare for increasingly extreme hydrometeorological events. However, the precipitation process is generally considered to be poorly represented in current numerical weather models. To improve the accuracy of precipitation estimates, we develop a Deep Neural Network model that estimates hourly precipitation based on the numerical model resolved circulation data. The network combines the Convolution Mechanism and Attention Mechanism to explicitly capture the spatiotemporal evolution of the precipitation-related circulation patterns. We train and validate the network model using the MERRA-2 (Modern-Era Retrospective analysis for Research and Applications, Version 2) hourly products and their dynamical downscaled products. Compared to 1) the precipitation estimates from the raw MERRA-2 model, 2) the MERRA-2 bias-corrected precipitation products and 3) the dynamical downscaled precipitation products, the precipitation estimations using the Deep Neural Network model better match the hourly gauge observations. Our model offers implications for improving relevant precipitation parameterization schemes and enhancing precipitation estimations at fine temporal scales.