Thunderstorm Areas Discrimination Based On Feature Fusion

Thunderstorm cloud clusters develop rapidly,which have many changes and complex features that are difficult to fully describe.At present,the thunderstorm region discrimination method is mainly based on the threshold method of a single feature.The feature description is insufficient;the discrimination method is expensive;the accuracy is not high.Aiming at the above problems,in order to excavate the characterization of thunderstorm regions accurately,a more reasonable and efficient method for identifying thunderstorms is designed.From the perspective of image features and nonimage features of thunderstorm areas,this paper designs a set of feature-based thunderstorm area discrimination schemes.This scheme combines the characteristics of infrared channel brightness temperature,channel-to-channel brightness temperature difference,and deep image convolution characteristics in a thunderstorm area,combined with multi-core support vector machines,and uses multi-core learning ideas to fully learn the characteristics of thunderstorm areas to realize the identification of thunderstorm areas.The main work and innovations are as follows:(1)Mining and analyzing the non-image characteristics of thunderstorm area,and proposed a thunderstorm area discrimination algorithm based on infrared brightness temperature and brightness temperature difference between channels.Aiming at the problem of densely distributed brightness temperature data and irrelevant brightness temperature data,brightness temperature histogram equalization,typical brightness temperature extraction and gradient extraction are used to optimize the brightness temperature data.On this basis,the optimized brightness temperature,brightness temperature difference and support vector machine are combined to generate a thunderstorm region discrimination algorithm.Experimental results show that,compared with the traditional brightness temperature threshold discrimination method,this method avoids setting the threshold manually and improves the discrimination accuracy.(2)Mining and analyzing the image features of thunderstorm area,and proposed a thunderstorm area discrimination algorithm based on convolutional neural network.Firstly,texture and edge features are used to describe and analyze the shallow image features of the thunderstorm area.Secondly,a convolutional neural network framework is constructed to extract the deep convolution features of the thunderstorm area image.At the same time,the thunderstorm image data set is constructed and labeled for network training and test.The experimental results show that the convolutional neural network can fully extract the deep image features of the thunderstorm area,and the discriminative precision reaches 85%.(3)Exploring the fusion of image features and non-image features in thunderstorm areas,and proposed a thunderstorm area discrimination algorithm based on the fused features.Combine the image features(deep convolution features)and non-image features(bright temperature features,bright temperature difference features)of thunderstorm areas to increase the diversity of sample features.On this basis,the fused features are combined with a multi-core support vector machine to assign different kernel functions to the new features and to learn local features in a targeted manner.The experimental results show that the scheme can better integrate the characteristics of thunderstorm regions,and the accuracy rate is greatly improved,and the accuracy reaches 95% compared with the single feature thunderstorm region discrimination method.