Deep Learning Prediction of Thunderstorm Severity Using Remote Sensing Weather Data

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing Vol. 15; pp. 4004 - 4013
• Main Authors: Essa, Yaseen, Hunt, Hugh G. P., Gijben, Morne, Ajoodha, Ritesh
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Ash; Convolutional neural network (CNN); Data models; Datasets; Deep learning; Feature recognition; Information processing; Lightning; Lightning detection; Lightning flashes; long short-term memory (LSTM); Meteorological data; Meteorology; Neural networks; Predictions; recurrent neural network (RNN); Remote sensing; Risk management; Severe weather; Surface acoustic waves; Temporal variations; thunderstorm severity; Thunderstorms; Weather; Weather stations; South Africa
• ISSN: 1939-1404; 2151-1535
• DOI: 10.1109/JSTARS.2022.3172785

Lightning is one of the leading causes of electrical outages in South Africa, and the most severe weather-related killer in the country. Unfortunately for risk management, quantitative lightning prediction remains challenging. In this study, we evaluate the accuracy of LSTM neural network model variants on thunderstorm severity using remote sensing weather data. These LSTM model variants are LSTM-FC, CNN-LSTM, and ConvLSTM variants. The CNN-LSTM and ConvLSTM models recognize spatio-temporal features, which assist processing. The data used consists of lightning detection network (LDN) data from the SALDN and weather-feature information from the network of weather stations operated by the SAWS. We forecast thunderstorm severity every hour, as quantified by lightning flash frequency, between December-2013 and March-2016 for North-Eastern South Africa. Models were trained on data between July-2008 and November-2013. All models minimized MSE but evaluated on mean absolute error (MAE flashes.hr −1 ). We also varied models based on input datasets: SALDN-only, SAWS-only, and SALDN+SAWS datasets. We found the CNN-LSTM model (MAE=51) performed best among LSTM model variants (LSTM-FC MAE=67; ConvLSTM MAE=86). When models were evaluated between input datasets, we found that SALDN only (MAE=59) outperformed SAWS only and SALDN+SAWS (SAWS MAE=74; SAWS+SALDN MAE=70). We conclude that CNN-LSTM models outperform prediction accuracy compared with ConvLSTM and LSTM-FC models but consideration on input data is required.