Prediction of tornado days in the United States with deep convolutional neural networks

Abstract

Tornado outbreaks in the United States have caused large numbers of human and economic losses. The analysis of gridded climate data to find spatial patterns of meteorological variables that produce tornadoes is often a time consuming and manual process requiring visual inspection. Therefore, the ability to train convolutional neural networks (CNNs) to recognize patterns in climate data that produce tornado outbreaks would be a huge benefit to climatologists. In this paper, three CNNs of increasing complexity, including LeNet-5, VGG-16, and Resnet-50 were trained to predict days with tornado outbreaks of varying size based on gridded values of convective available potential energy (CAPE), convective inhibition (CIN), storm relative helicity (HLCY), and geopotential height. All three models showed promising results in predicting large tornado outbreaks with greater than 20 tornadoes with VGG-16 using the variable CAPE having a slightly better performance, highlighting the importance of CAPE in the analysis of tornado environments.