Spatio-Temporal Analysis and Severity Prediction of Weather Events Using Machine Learning Models

A proposed machine learning framework for predicting the severity of weather events using spatio-temporal data on a large dataset of all one million weather events recorded across the United States. In this article, modern data preprocessing, exploratory data analysis, and feature engineering were used to improve our random forest and gradient boosting models. The random forest model produced an accuracy of 99.66%, an F1-Score of 0.9966, and an AUC-ROC of 1.0, demonstrating considerable predictive reliability. In contrast, though the gradient boosting model also displayed a marginally lower accuracy (94.86%), it excelled in terms of balanced performance, measured by an F1-score of 0.9476 and AUC-ROC of 0.9958, demonstrating its proficiency in handling diverse and imbalanced data scenarios. This suggests that our findings may prove to be vital in the field of meteorology, despite the fact that the systems are built on significantly different premises for analysis.