A Convolutional Neural Network for Classification of Lightning LF/VLF Waveform

With the accumulation of raw lightning data in various fields, it's more important to automatically analyze the large amounts of natural lightning data. In recent years, the neural network has made great process in pattern recognition, which makes it possible to efficiently classify the raw lightning data. In this study, a convolutional neural network is proposed for classification of lightning waveform in Low Frequency/ Very Low Frequency (LF/VLF) bands. The network is trained by a six-year (2012-2017) data set including over 50,000 lightning waveforms. By using a series of one-dimensional convolutions layer with rectified linear unit (ReLU) as its activation function, we use the labeled lightning time series directly as the input and train the network to get the classifier. The output of the network is a vector containing the results that an event belongs to each lightning type. We classify the types of lightning into five categories, including the cloud-to-ground (N/PCG) flash, ordinary intracloud (IC) flash, preliminary breakdown pulse (PB), narrow bipolar event (NBE) and the classifier has an average accuracy more than 95%. Furthermore, we apply the classifier to five isolated thunderstorm and give the actual accuracy of the NCG by manual checking.