Residual Neural Network applied to lightning classification in a modern lightning location network

Abstract:
In general, natural lightning flashes can be distinguished based on the charge that they neutralize (positive or negative), and if they reach the ground (Cloud-to-Ground – CG) or remain in the cloud (Intracloud – IC). On average, about 100 to 200 lightning strikes occur worldwide per second. This equates to roughly 8 million lightning strikes per day. Accurate lightning classification is crucial for identifying various types of thunderstorms and assessing their severity. Modern lightning location systems effectively capture a substantial portion of all lightning that occurs on Earth. However, due to the vast amount of data gathered, accurately classifying each lightning pulse poses a significant challenge. In this study, we evaluate the use of Earth Networks Total Lightning Network (ENTLN) data to develop a lightning classification model based on a Residual Neural Network (ResNet). We develop a new methodology to retrieve lightning electric field wave-forms from ENTLN non-uniformly sampled data. The model was optimized to work in real-time, and its performance scores are accuracy = 95,2%, and F1-score = 92,6%.