Tornado Speed Estimation Using Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM)-Based Video Processing Approach

Many natural processes in the universe occur in a rotational motion, such as the formation of drastic events including tornadoes and cyclones. For the past few decades, research has progressed to estimate the occurrence of such unpredictable small-scale meteorological events and their damage paths and to estimate the amount damage caused. In the case of short-lived yet disastrous tornadoes, it is possible to track the damage path. However, to estimate the damage through the Fujita Scale (F-Scale) or the Enhanced Fujita scale (EF-Scale), it is still necessary to rely on Radar Data Acquisition (RDA) systems working on Doppler effect to estimate the wind speed. For this, the equipment needs to be placed in the vicinity of where the tornadoes form, so they are often at risk of being damaged. Thus, in the current research, the tornado speed is estimated using video processing and AI techniques: Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) conjointly. A video model of a rotating tornado (without translational motion) is artificially generated with a tracking object inside it. The wind speed is estimated by tracking the speed of this object caught in the tornado’s whirl. CNN in combination with LSTM effectively predicts the shift of the object in each frame of the video in comparison with a reference frame.