Hydrometeor Classification Method Of Dual-polarization Radar Based On Transfer Learning

My country’s climate is complex,and it is easy to form meteorological disasters,which pose a threat to transportation,property safety,and means of production.Using weather radar to classify precipitation particles can greatly reduce such threats.Compared with the traditional weather radar,the dual-polarization weather radar can not only detect the echo intensity information of the meteorological target,but also obtain the difference information of the meteorological target in the horizontal and vertical directions,so as to determine the phase,shape,and space of the meteorological target.It is very helpful to complete the task of precipitation particle classification.In the process of establishing a precipitation particle classification model,the traditional precipitation particle classification method requires sufficient label information or relies on expert experience values.Therefore,the establishment of a precipitation particle classification model when the radar data has no labels or only a few labels requires in-depth research on the above problems.When researching the precipitation particle classification method for the band to be classified that has no label or insufficient label information,the transfer learning method can use the precious label information of other bands to assist in establishing the precipitation particle classification model in this band,which can effectively complete the classification of the precipitation particles in the band to be classified.Precipitation particle classification task with label information or only a small amount of label information.Therefore,it is of great significance to study the precipitation particle classification method of dual-polarization meteorological radar based on transfer learning for disaster warning,meteorological prediction and avoidance of dangerous meteorological targets on aircraft routes.In this thesis,the transfer learning method is studied for the problem that the data has no label or the label information is insufficient,and the accurate classification of the dual-polarization radar precipitation particles based on the transfer learning is realized.The main contents are as follows:Firstly,a TrAdaBoost precipitation particle classification method based on Weighted Support Vector Machine(W-SVM)is proposed for the small amount of data in the band to be classified.First,combine the band data to be classified and other band data with rich label information to construct an input matrix and initialize the weights of the training data;next,combine the weight coefficients and the constructed input data to train a W-SVM multi-classifier And return a hypothetical model,predict the training data through the hypothetical model,adjust and update the weight coefficients corresponding to the two band samples respectively according to whether the prediction category is correct or not,and repeat the training process continuously;finally,the iteration is completed,Returns a trained W-SVM based TrAdaBoost classifier.The experimental results based on the measured data and simulation data show that when there is a small amount of label data of the band to be classified,the method can effectively use the labeled data of other bands to establish the classification model of the band to be classified,and obtain an accurate classification effect.Secondly,a precipitation particle classification method based on Balanced Distribution Adaptation(BDA)is proposed for the situation that the band data to be classified has no label information.First,initialize the model parameters,obtain the transformation matrix and its eigenvectors,project the samples into a regenerated kernel Hilbert space,use the labeled data of other bands to train to obtain a basic classifier,and use the classifier to treat the classified band data Mark the pseudo-label;then,use the Maximum Mean Discrepancy(MMD)method to minimize the distance between the two parts of the data;finally,iteratively update the pseudo-label and transformation matrix until convergence,and obtain the final transformation matrix and final The basic classifier,that is,a precipitation particle classification model based on equilibrium distribution adaptive.The experimental results based on the measured data and simulation data show that: when the band data to be classified has no labels,the balanced distribution adaptive can effectively classify the precipitation particles in the band to be classified with the help of the data of other bands,and obtain good classification results.And in the case of unbalanced category distribution,the Weighted Balanced Distribution Adaptation(W-BDA)method can effectively improve the classification effect for categories with a small number of samples.