Reconstruction of Severe Storms Observed by Weather Radars Using Recurrent Neural Networks

Severe storms events with tornadoes can cause destruction and significant loss of life, mainly through injuries from flying debris and collapsing structures. Analyze the formation and structure of these systems helps to identify the favorable conditions for their generation, contributing to specific techniques for prediction. One way to observe these storms is through weather radar. They have one of the most accurate and used technologies within meteorology, and it is the only instrument capable of observing the three-dimensional structure of the clouds, providing this data with high spatial and temporal resolution. The State of Paraná has two meteorological radars operated by the Parana Meteorological System (SIMEPAR). One of them, located in Cascavel - PR, is an S-Band Radar with Dual Polarization and performs measurements in an area up to 480 km range, monitoring rainfall in the west region of Paraná. However, the scans performed by this radar have intervals of 5 minutes and, as a storm can dissipate quickly, it may not be possible to observe its complete evolution through these data. The purpose of this paper is to use recurrent neural networks using the long short-term memory approach (LSTM) to create a continuous visualization of weather radar data. This approach will allow us to study the dynamics of these systems by focusing on the analysis of their physical parameters such as life cycle and volume.