Radar-based ground truth for deep learning Towering Cumulus (TCU) nowcasting: towards a minimum definition for aviation

Towering Cumulus clouds (TCUs) pose a significant challenge to aviation because they are associated with strong vertical motion and should therefore be avoided during all flight phases. Beyond the meteorological challenges, TCU nowcasting is further complicated by the large variety of existing definitions, which differ across user groups and operational purposes. In the context of TCU nowcasting for aviation, it is important to identify a minimum viable definition that satisfies the needs of different users, ranging from air traffic controllers to pilots and meteorological forecasters. A common distinction is that TCUs do not produce lightning, while storms are classified as Cumulonimbus as soon as lightning activity begins.  However, current operational convection forecasts at MeteoSwiss are trained on lightning-based Cumulonimbus/thunderstorm labels and therefore miss TCUs entirely. 

In this work, we explore an impact-oriented nowcasting approach that explicitly targets TCUs as observed by aircraft onboard weather radars. Our approach is to retrain an existing recurrent convolutional neural network (RCNN), originally designed for lightning and hail prediction, on radar-derived TCU labels in a transfer-learning framework.  The training dataset is generated by a TCU detection algorithm that is already used operationally by MeteoSwiss for METAR (METeorological Aerodrome Report) products at the airports of Geneva and Zurich. In this product, TCUs are identified from Swiss radar images using two-dimensional maximum radar reflectivity fields combined with spatial Fourier bandpass filtering and, optionally, three-dimensional radar information. 

To obtain spatially consistent training data, we extend the currently airport-centered automatic METAR processing to cover the entire Swiss domain. In an initial step, the computationally expensive 3D component of the algorithm is omitted, allowing us to focus on a more efficient 2D radar-based classification approach. This approach enables the generation of multi-year TCU datasets from radar archives that can serve as training data for machine learning models. 

Based on these radar-derived TCU fields, we retrain the RCNN to produce spatially explicit probabilistic forecasts of TCU occurrence over Switzerland, with a temporal resolution of 10 minutes and lead times of up to 4 hours. We present the full data pipeline, from radar and lightning archives to TCU ground-truth generation and model training, and we discuss the potential of combining radar-based classification with deep recurrent networks. This approach bridges the gap between radar observations, operational aviation definitions of convection, and probabilistic machine learning-based impact forecasting.