On Deep Learning Models for Detection of Thunderstorm Gale

The purpose of this paper is to perform a comprehensive study on the performance of different deep learning models for detection of thunderstorm gale. We construct a benchmark dataset from the radar echo images in Guangdong province of China. Each radar image is partially labeled according to the wind velocities recorded by meteorological observation stations. We design four deep learning models to address the thunderstorm gale detection problem, including a simple convolution neural network (CNN), a recurrent neural network (S-RCNN), a time context recurrent convolutional neural network (T-RCNN), and a spatio-temporal recurrent convolutional neural network (ST-RCNN). Ten traditional machine learning algorithms are selected as comparison baselines. Experimental results demonstrate that four deep learning models can achieved better detection performance than traditional machine learning algorithms.