A Comparative Study on Machine Learning Approaches to Thunderstorm Gale Identification.
Haifeng Li,Yan Li,Xutao Li,Yunming Ye,Xian Li,Pengfei Xie
DOI: https://doi.org/10.1145/3318299.3318317
2019-01-01
Abstract:In this paper, we make a comparative study to examine the performance of different machine learning approaches for the thunderstorm gale identification. To this end, a thunderstorm gale benchmark dataset is constructed, which comprises radar images in Guangdong from 2015 to 2017. The corresponding wind velocities recorded by the automatic meteorological observation stations are utilized to offer the ground-truth. Based on the dataset, we evaluate the performance of Decision Tree Regressor (DT), Linear Regression (LR), Ridge regression, Lasso regression, Random Forest Regressor (RFR), K-nearest Neighbor Regressor (KNNR), Bayesian Ridge Regressor (BR), Adaboost Regressor (AR), Support Vector Regressor (SVR), Gradient Boosting Regressor (GBR), and Convolutional Neural Network (CNN). Ten important features are extracted to apply these approaches, except CNN, which include radar echo intensity, radar reflectivity factor, radar combined reflectivity, vertical integrated liquid, echo tops and their changes with respect to (w.r.t.) time. Experimental results demonstrate the machine learning approaches can effectively identify the thunderstorm gale, and the CNN model performs the best. Finally, a thunderstorm system is developed based on CNN model, which help meteorologists to identify thunderstorm gales in terms of radar images.