Using Machine Learning Techniques for Supercell Tornado Prediction with Environmental Sounding Data
Loading...
Links to Files
Permanent Link
Author/Creator ORCID
Date
Type of Work
Department
Program
Citation of Original Publication
Brice Coffer et al., Using Machine Learning Techniques for Supercell Tornado Prediction with Environmental Sounding Data, http://hpcf-files.umbc.edu/research/papers/CT2020Team8.pdf
Rights
This item is likely protected under Title 17 of the U.S. Copyright Law. Unless on a Creative Commons license, for uses protected by Copyright Law, contact the copyright holder or the author.
Abstract
Tornadoes pose a forecast challenge to National Weather Service forecasters because of their quick development and potential for life-threatening damage. The use of machine learning in severe weather forecasting has recently garnered interest, with current efforts mainly utilizing ground weather radar observations. In this study, we investigate machine learning techniques
to discriminate between nontornadic and tornadic storms solely relying on the Rapid Update Cycle (RUC) sounding data that represent the pre-storm atmospheric conditions. This approach aims to provide for early warnings of tornadic storms, before they form and are detectable by weather radar observations. Two machine learning methods tested in our project are Random
Forest (RF) and Convolutional Neural Network (CNN). Performance testing of RF using various ranges of hyperparameters results in an overall accuracy score of 70.14%, but the accuracy of significantly tornadic class prediction is only 23.84%. The CNN model results in an overall accuracy score of 67.84%, but the accuracy for significantly tornadic storms is only 26.69%.
The higher accuracy in the RF and CNN models for the majority class of nontornadic supercells suggests that the imbalanced dataset is a meaningful contributor to the lower accuracy for tornadic storms. After applying the simple method of randomly undersampling (oversampling) the majority (minority) class, the accuracies of significantly tornadic class prediction of RF
and CNN are enhanced to 65.85% and 36.01%, respectively. Future work should investigate alternative methods of dealing with imbalanced datasets in a CNN, including more sophisticated undersampling/oversampling techniques.