ResRadNet: A 3D-Residual Neural Network Approach for Temporal Super-Resolution and Ground Adjustment of Weather Radar Rainfall Estimates

Weather radars are advanced tools for atmospheric observations that provide QPE with a high spatial representativeness and a high temporal resolution (e.g. 5-minutes). However, due to their indirect measurement aloft, strong systematic errors as well as temporal sampling errors compared to rain gauge measurements at even higher resolution (e.g. 1-minute) persist. As a solution, bias and advection correction techniques are used. Residual neural networks have proven to be efficient tools to approximate the behavior of dynamical systems. Here, we present ResRadNet, a 3D-residual neural network (3D-RNN), that is capable of correcting biases and increasing the temporal resolution of weather radar based quantitative precipitation estimates (QPE). ResRadNet is trained to correctly reproduce 1-minute rain gauge data from sequences of 5-minute radar images and information about the orography. The dataset used in this study consists of 8 years of country-wide rainfall observations in Germany. The weather radar composite used as model input is based on reflectivity derived rainfall information from 17 C-band radars. The rain gauge reference consists of 1066 rain gauges with a 1-minute resolution used to train and test ResRadNet. An additional 1138 rain gauges with a daily resolution are used for long-term evaluation of remaining biases. The results showed that ResRadNet can significantly increase the linear correlation and reduce the root mean squared error of the QPE field compared to rain gauge data at 1- and 5-minute, as well as daily resolutions. A qualitative analysis also showed that ResRadNet is a suitable optical flow estimator and that the provided rainfall fields are not subject to temporal or spatial inconsistencies even though spatio-temporal consistency was not enforced during training. Therefore, our study shows how using 3D-RNNs can provide accurate 1-minute, ground-adjusted, and advection-corrected QPE.