Study On Hail Detection Method Of Dual-polarization Radar Based On Machine Learning

In China,hail damage to agriculture,electric power and other industries is up to billions of yuan every year.Accurate detection and safety warning of hail play an important role in hail prevention and disaster reduction.Dual-polarization weather radar can not only obtain the macroscopic information of the cloud cluster,but also obtain the polarization information related to the particle microphysical characteristics(shape,size,orientation)in the cloud cluster,which provides strong support for hail detection and warning.Fuzzy logic describes the classification system with simple and flexible rules rather than fixed formulas,which can reasonably divide the boundary of polarization parameters.However,the membership function of fuzzy logic needs to be determined by expert experience value,which is highly subjective.When the parameters of membership function are biased,the classification results will be affected.At present,the hail detection model mainly has some problems,such as parameter dependence on empirical value determination and single characteristics of input data and other problems.Machine learning is characterized by data feature extraction,automatic analysis,and finding rules.It can adjust model parameters adaptively according to data and predict unknown data by using rules,and has been successfully used in data analysis,image processing,feature recognition etc.Therefore,it is of great research significance to construct the hail detection model of dual-polarization weather radar by using machine learning correlation algorithm.The work content of this thesis is as follows:Firstly,the generation conditions,physical characteristics and corresponding polarization parameters of ice crystal,snow,rain and hail particles are introduced.Then,the physical significance and calculation formula of polarization parameters of dual polarization weather radar are described.Finally,the data and simulation polarization parameters of WSR-88 D are introduced.Secondly,aiming at the problem that the parameters of the traditional hail detection model are not easy to determine and the input characteristics are single,a dual-polarization weather radar hail detection method based on convolutional neural network was proposed.Firstly,the polarization data is interpolated to make the data form match the input of the convolutional neural network and improve the resolution of the data,then the convolutional neural network is built,mainly including forward propagation and back propagation.Forward propagation includes the Settings of convolution layer,pooling layer and full connection layer.Back propagation is mainly used to adjust model parameters.Then,learning rate adaptive,regularization,moving average and other measures are added to the established model to optimize the network model.After offline training of the network.Finally the trained model is used to realize hail detection.The test results of measured data and simulation data show that,hail detection method of dual-polarization weather radar based on convolutional neural network can extract polarization parameter characteristics,and the selection of network parameters does not rely on expert experience value,which can effectively detect hail.Thirdly,aiming at the hail detection problem when there is only a small amount of precipitation particle sample label,a dual-polarization weather radar hail detection method based on MS-CS is proposed.The method first builds a mean shift clustering model,and then the model is trained using unlabeled polarization parameters.Contour coefficient,CalinskiHarabaz scores,DBI were used to evaluate the clustering model,after the model reaches the standard,the model is used to cluster the polarization parameters without labels.Then cosine similarity is used to analyze the correlation between the cluster and a small number of labeled polarization parameters.Finally by judging the similarity between the data to complete the hail detection,the experimental results show that the method can be implemented under the background of small samples of label hail detection,it also has good detection results.