Application of Convolutional Neural Networks for the Detection of Lightning in Weather Radar Data

Abstract:
Machine learning is increasingly being used in different fields of weather forecasting. Due to the difficulty of predicting lightning discharges, there are also approaches to use machine learning in lightning forecasting. As a subset of machine learning, Convolutional Neural Networks (CNN), are one of the methods used to automatically classify images. Thus, CNNs can be used to recognize patterns in weather radar data. The present study, shows a first approach of training five CNN models with weather radar data from Austro Control GmbH for Austria. The models are trained to decide whether lightning occurred in the respective radar image or not. The weather radar data of the years 2020, 2021 and 2022 were processed and merged from provided raw data. Lightning location system data from the Austrian Lightning Detection and Information System were used to categorize the weather radar data into two sub groups (Flash and No Flash). All models are pre-trained on the ImageNet dataset or used in the untrained state. The SqueezeNet model delivered the highest results in the pre-trained case and without feature extraction with a prediction accuracy of 89.84%.