215 Severe Weather Identification Using Polarimetric Radar and Machine Learning Techniques

Thursday, 31 August 2017
Zurich DEFG (Swissotel Chicago)
Tulipa Silva, SIMEPAR - Parana Meteorological System, Curitiba, Brazil; and C. Beneti, P. H. Siqueira, M. F. Buzzi, and L. Calvetti
Manuscript (2.1 MB)

Handout (1.4 MB)

Severe Weather Events (SWE) cause human and financial damage worldwide. In the south of Brazil, a region frequently prone to these events usually associated with mesoscale convective systems, is covered by a S-Band dual polarization weather radar and a total lightning network which was used in this project to evaluate the usage of machine learning techniques to identify and forecast these SWE. In this way, using machine learning techniques Support Vector Machine (SVM) and Artificial Neural Network (ANN), we aim to identify convective storms that can become a SWE within the next 30 minutes. For this study we analysed a series of SWE which occurred in the state of Paraná, Brazil, from January 2015 to July 2016. The input variables of the model are those collected by the polarimetric weather radar, specifically: horizontal reflectivity, differential reflectivity, correlation coefficient and differential phase on propagation. In addition to the radar data, some derivatives were also used, such as altitude of the maximum reflectivity value per column, and azimuthal, vertical and radial shear. For the model obtained by the ANN technique, two events were not identified during the entire 30 minute period before the SWE occurrence. For the model using SVM, only one event of heavy rainfall was not identified during the period of study. The performance of the models may be directly related to the inaccuracy of location information, time and duration of events, but 80% of the events were fully identified during the 30 minutes preceding the SWE. In addition, some regions which were not previously identified as SWE, were indicated by the models as possible occurrence of severe weather. In regions where the models were not applied during training, a comparison was made with total lightning data. This comparison showed good correlation between the identified regions and high values of total lightning data. In this way, the study is efficient as a tool to support the decision, by meteorologists, in the identification and prediction of SWE. In order to improve these results, it is necessary to include new events to train the model, as well as to further investigate the regions with no previous classification of SWE but which were identified as possible severe events.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner