Learning about the vertical structure of radar reflectivity using hydrometeor classes and neural networks in the Swiss Alps

Abstract

Abstract. The use of radar for precipitation measurement in mountainous regions is complicated by many factors, especially beam shielding by terrain features, which, for example, reduces the visibility of the shallow precipitation systems during the cold season. When extrapolating the radar measurements aloft for quantitative precipitation estimation (QPE) at the ground, these must be corrected for the vertical change of the radar echo caused by the growth and transformation of precipitation. Building on the availability of polarimetric data and a hydrometeor classification algorithm, this work explores the potential of machine learning methods to study the vertical structure of precipitation in Switzerland and to propose a more localised vertical profile correction. It first establishes the ground work for the use of machine learning methods in this context: from volumetric data of 30 precipitation events, vertical cones with 500 m vertical resolution are extracted. It is shown that these cones can well represent the vertical structure of different types of precipitation events (stratiform, convective, snowfall). The reflectivity data and the hydrometeor proportions from the extracted cones constitute the input for the training of artificial neural networks (ANNs), which are used to predict the vertical change in reflectivity. Lower height levels are gradually removed in order to test the ANN's ability to extrapolate the radar measurements to the ground level. It is found that ANN models using the information on hydrometeor proportions can predict from altitudes between 500 and 1000 m higher than the ANN based on only reflectivity data. In comparison to more traditional vertical profile correction techniques, the ANNs show less prediction errors made from all height levels up to 4000 m a.s.l., above which the ANNs lose predictive skill and the performance levels off to a constant value.

AI-derived 3D cloud tomography from geostationary 2D satellite data

“…While the results indicate an improvement in resource efficiency, they are predominately restricted to horizontal processes of the cloud field. Reconstructing the cloud vertical column can deliver insights into 3D dynamics (van den Heuvel et al, 2020;Leinonen et al, 2019). Current studies by Hilburn et al (2020) and Leinonen et al (2019) use AI techniques such as convolutional neural networks (CNN) and conditional generative adversarial networks (CGAN) to address this issue.…”

Using artificial neural networks to predict riming from Doppler cloud radar observations

“…Their potential to extract useful relations from ground-based radar observations has been demonstrated in work published e.g. by Luke et al (2010) and van den Heuvel et al (2020).…”