A Method of Forecasting Thunderstorms and Gale Weather Based on Multisource Convolution Neural Network

Abstract

The local thunderstorms and gale weather occurring frequently has brought huge losses to the agriculture and transportation industries. This paper presents a method of forecasting the local thunderstorms and gale weather, in which a multisource convolution neural network is constructed to extract the features of weather-related data with multiple types from Doppler radar. To improve the discriminative power of features, Center-Loss and Softmax were jointly used as loss function in the training, and then the features obtained are combined with SVM for classification. Furthermore, a comparative experiment of multisource convolution neural network based on CNN-4, ResNet30, ResNet50, and VGG16 is designed, in which the ResNet30 achieves the highest accuracy. The experimental results show that the multisource convolution neural network avoids the limitation of using one type of data and improves the accuracy of forecasting local thunderstorms and gale.

Keywords

• Feature extraction
• Doppler radar
• forecasting
• center_loss