RainNet v1.0: a convolutional neural network for radar-based precipitation nowcasting

• Main Authors: G. Ayzel, T. Scheffer, M. Heistermann
• Format: Article
• Language: English
• Published: Copernicus Publications 2020-06-01
• Series: Geoscientific Model Development
• Online Access: https://www.geosci-model-dev.net/13/2631/2020/gmd-13-2631-2020.pdf

<p>In this study, we present RainNet, a deep convolutional neural network for radar-based precipitation nowcasting. Its design was inspired by the U-Net and SegNet families of deep learning models, which were originally designed for binary segmentation tasks. RainNet was trained to predict continuous precipitation intensities at a lead time of 5&thinsp;min, using several years of quality-controlled weather radar composites provided by the German Weather Service (DWD). That data set covers Germany with a spatial domain of <span class="inline-formula">900 km×900</span>&thinsp;<span class="inline-formula">km</span> and has a resolution of 1&thinsp;<span class="inline-formula">km</span> in space and 5&thinsp;<span class="inline-formula">min</span> in time. Independent verification experiments were carried out on 11 summer precipitation events from 2016 to 2017. In order to achieve a lead time of 1&thinsp;<span class="inline-formula">h</span>, a recursive approach was implemented by using RainNet predictions at 5&thinsp;<span class="inline-formula">min</span> lead times as model inputs for longer lead times. In the verification experiments, trivial Eulerian persistence and a conventional model based on optical flow served as benchmarks. The latter is available in the <i>rainymotion</i> library and had previously been shown to outperform DWD's operational nowcasting model for the same set of verification events.</p> <p>RainNet significantly outperforms the benchmark models at all lead times up to 60&thinsp;<span class="inline-formula">min</span> for the routine verification metrics mean absolute error (MAE) and the critical success index (CSI) at intensity thresholds of 0.125, 1, and 5&thinsp;<span class="inline-formula">mm h⁻¹</span>. However, <i>rainymotion</i> turned out to be superior in predicting the exceedance of higher intensity thresholds (here 10 and 15&thinsp;<span class="inline-formula">mm h⁻¹</span>). The limited ability of RainNet to predict heavy rainfall intensities is an undesirable property which we attribute to a high level of spatial smoothing introduced by the model. At a lead time of 5&thinsp;<span class="inline-formula">min</span>, an analysis of power spectral density confirmed a significant loss of spectral power at length scales of 16&thinsp;<span class="inline-formula">km</span> and below. Obviously, RainNet had learned an optimal level of smoothing to produce a nowcast at 5&thinsp;<span class="inline-formula">min</span> lead time. In that sense, the loss of spectral power at small scales is informative, too, as it reflects the limits of predictability as a function of spatial scale. Beyond the lead time of 5&thinsp;<span class="inline-formula">min</span>, however, the increasing level of smoothing is a mere artifact – an analogue to numerical diffusion – that is not a property of RainNet itself but of its recursive application. In the context of early warning, the smoothing is particularly unfavorable since pronounced features of intense precipitation tend to get lost over longer lead times. Hence, we propose several options to address this issue in prospective research, including an adjustment of the loss function for model training, model training for longer lead times, and the prediction of threshold exceedance in terms of a binary segmentation task. Furthermore, we suggest additional input data that could help to better identify situations with imminent precipitation dynamics. The model code, pretrained weights, and training data are provided in open repositories as an input for such future studies.</p>