Study On Hydrometeor Classification Method Of Dual-Polarization Radar Under Small Samples

As a powerful tool for atmospheric monitoring,dual-polarization radar is widely used in atmospheric physics research and meteorological disaster prevention and control,as it can obtain a variety of additional polarization parameters.The detection and classification of hydrometeors in the cloud is one of the important applications of dual-polarization radar.By comprehensive analysis of different polarization parameters,the accuracy of hydrometeor classification can be improved.The dual-polarization radar obtains a large variety of polarization parameters and a large amount of data,which puts high demands on the data processing,storage and transmission capabilities of the radar system.In order to reduce the requirements on hardware devices and save radar manufacturing costs,the reduction in data volume can be achieved by reducing the sampling rate.But this will lead to problems such as low data resolution and insufficient data.In addition,the polarization parameter data is easily interfered by various factors such as radar system error,electromagnetic wave signal attenuation,non-weather echo and human error,resulting in data loss.For different polarization parameters,the location and scale of data loss are also different,which will have a great impact on hydrometeor classification.In view of the above problems,this thesis studies the hydrometeor classification method of dual-polarization radar under small samples.The main contents include:Firstly,the definition,physical meaning and calculation method of each polarization parameter obtained by dual-polarization weather radar are introduced.The relationship between the microphysical properties of hydrometeors and the values of polarization parameters is clarified.The polarization characteristics of different hydrometeors are analyzed according to the statistical results,which provides a theoretical basis for the subsequent selection of polarization parameters and data processing.Secondly,aiming at the classification problem of dual-polarization radar polarization parameter small sample data obtained under low sampling rate,a small-sample hydrometeor classification method based on modified wavelet transform-Tree-Augmented Naive Bayesian(TAN)is proposed.In the method,the low-resolution data after bilinear interpolation is transformed by wavelet to obtain high-resolution low-frequency and high-frequency components,and the correction coefficient matrix is calculated according to the similarity between the wavelet components at each level at first.Then,the high-resolution wavelet component is corrected by the correction coefficient matrix,and the high-resolution polarization parameter data is obtained by inverse wavelet transform.The naive Bayesian network for hydrometeor classification is established and the structural and parameter training are performed to obtain the TAN network,according to the mutual information theory.At last,the hydrometeor classification of interpolated data is realized with the trained TAN network.The relatively complete high-frequency information of the polarization data is retained and the classification parameters are calculated by network training in the method.Compared with the traditional fuzzy logic method,in which the classification parameters follow the empirical values summarized by the predecessors,the robustness of the classification effect is improved.Thirdly,aiming at the classification problem of dual-polarization radar polarization parameter small sample data with random missing and outliers,a small sample hydrometeor classification method based on Singular Value Thresholding(SVT)-TAN is proposed.According to the matrix filling theory,the optimization problem of the polarization parameter low rank matrix filling is constructed and further transformed into a convex optimization problem that is easy to solve.Then,on the basis of the Lagrangian multiplier method,the operation of the convex optimization problem is simplified by the singular value thresholding iteration to obtain the reconstructed polarization parameter data matrix.The TAN network for hydrometeor classification is obtained by structural and parameter training performed on the naive Bayesian network.At last,the hydrometeor classification of interpolated data is realized with the trained TAN network.The method realizes the reconstruction of data missing samples with small error and has good classification performance.