A data-driven approach to predicting Severe Weather utilizing dynamic definition via dual-polarization products

As extreme weather events grow more frequent and severe, timely and accurate severe weather warnings are becoming more critical than ever before. With the complex dynamics of weather phenomena, it is primarily thanks to weather radars that operational meteorologists are able to monitor these situations and give out warnings for areas with high possibility of occurrence. With the growing availability of computational resources and the rise of data-driven methods, applying machine learning to severe weather nowcasting has become both feasible and increasingly effective. However, the classification of weather to be severe is often domain-specific and subjective, leading to a difficult objective definition that would stay true in any given geographical location. It is for this reason that we at Meteopress have trained a deep learning model that is able to work with dynamic definitions. When severe weather is defined as a combination of dual-polarization radar variables, such as copolar correlation coefficient or specific differential phase, paired with reflectivity, the model is able to adapt to these definitions and produce domain-specific outputs. Moreover, it enables the user to define what kind of weather they are interested in. It utilizes recent and current weather radar measurements to produce outputs 90 minutes into the future. The model’s output can be calibrated through targeted evaluation to meet specific performance criteria, such as maintaining a minimum expected recall or precision.