Classifying Single versus Multiple Ground Strike Point Lightning Flashes from High-Speed Lightning Videos using Deep Learning

Inhalt:
Lightning flashes with Multiple Ground Strike Points (mGSPs) potentially pose a greater risk than Single Ground Strike Point (sGSP) flashes. However, the features influencing the number of strike locations within a lightning flash remain uncertain. A hindrance to advancing GSP lightning research is the labour-intensive task of analysing high-speed lightning videos. To address this challenge, we compared and evaluated the use of deep-learning techniques on high-speed videos of multistroke lightning for classifying lightning flashes that strike either a sGSP or mGSPs. Utilizing a dataset of 207 high-speed lightning videos captured in Johannesburg and São Paulo, we assessed the performance of three deep-learning models —specifically CNN-LSTM (uninitialised), InceptionResNetV2-LSTM and Xception-LSTM— in accurately classifying sGSP versus mGSP lightning flashes. These models are equipped to process both spatial and temporal data intricately. Our findings demonstrate improved classification accuracy when utilizing pre-trained deep-learning models compared to our CNN-LSTM model. An improvement was seen when utilizing an image-enhanced dataset with the Xception-LSTM model classifying 65.06% of sample videos correctly. Our results highlight the advantage of integrating pre-trained models with LSTM’s capability for analyzing lightning videos. While preliminary, the automation potential offered by these models facilitates a detailed and rapid characterization of lightning events, paving the way for new insights into the features influencing sGSP versus mGSP lightning flashes.