A comparative analysis using machine learning approach for thunderstorm prediction in southern region of Peninsular Malaysia

Abstract

Thunderstorms are one of the most destructive natural phenomena on the planet, as they are predominantly associated with lightning and heavy rainfall that result in human deaths, urban flooding, and agricultural damage. Thus, accurate thunderstorm prediction is essential for planning and managing agriculture, flood control, and air traffic control. This study utilized historical lightning and meteorological data from 2011 to 2018 in the southern regions of Peninsular Malaysia to predict thunderstorm occurrences. The lightning dataset is classified into three class ranges, where the high range of lightning rarely occurs in this region compared to the low and medium ranges of lightning because of the nonlinear and complex characteristics of the thunderstorm and lightning itself, leading to an imbalanced dataset. The k-fold and stratified cross-validation (CV) methods and a resampling technique called SMOTE are introduced to overcome the imbalance in the training dataset. Then the dataset is trained and tested using 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 that the GB ML model using stratified k-fold CV and SMOTE is the best algorithm for thunderstorm prediction for this region, with accuracy ranging from 74% to 95%, recall ranging from 72% to 93%, precision ranging from 76% to 97%, and F1-Score ranging from 74% to 95%. Future thunderstorm predictions based on lightning patterns and meteorological datasets are expected to establish an early strategy to address the presence of thunderstorms by notifying the relevant authorities, to prevent any damage that may be caused by the thunderstorms.