Thunderstorm prediction model using SMOTE sampling and machine learning approach.

Abstract

Thunderstorms are one of the most destructive phenomena worldwide and are primarily associated with lightning and heavy rain that cause human fatalities, urban floods, and crop damage. Therefore, predicting thunderstorms with reasonable accuracy is one of the crucial requirements for the planning and management of many applications, including agriculture, flood control, and air traffic control. This study extensively applied the historical lightning and meteorological data from 2011 to 2018 of the southern regions of Peninsular Malaysia to predict thunderstorm occurrence. Positive CG lightning rarely occurs compared to negative CG lightning and also due to the non-linear and complex characteristics of the thunderstorm and lightning itself, leading to an imbalance in the dataset. The resampling technique called SMOTE is introduced to overcome the imbalance of the training dataset. Then the dataset is trained and tested with five Machine Learning (ML) algorithms, including Decision Trees (DT), Adaptive Boosting (AdaBoost), Random Forest (RF), Extra Trees (ET), and Gradient Boosting (GB). The results have shown a good prediction with accuracy (74% to 95%), recall (72% to 93%), precision (76% to 97%), and F1-Score (74% to 95%) with SMOTE. The SMOTE and GB model prediction model is the best algorithm for thunderstorm prediction for this region in terms of performance metrics. In the future, the prediction results based on the lightning pattern and weather dataset will likely alert the related authorities to make an early strategy to handle the occurrence of thunderstorms.