Hail Identification And Forecasting Method Based On Dual Polarization Radar

Hail disaster is a kind of severe weather phenomenon caused by strong convective storm,which destroy crops and buildings and cause huge economic losses.Although some progress has been made in hail research,it is still challenging to accurately predict hail due to its rapid evolution and small spatial scale.In the field of meteorology,dualpolarization radar is the most advanced monitoring tool used for nowcasting.Compared with the traditional weather radar,the dual-polarization radar outputs a variety of polarization parameters and provides more information about the phase state and shape of precipitation particles.Based on dual-polarization radar data,this thesis studies the identification of convective cells that produce hail(hail cells)and the prediction of hail landing time by using computer vision technology and machine learning method.The main work of this thesis is as follows:1.To describing the morphological structure of hail cells,the three-dimensional interpolation method is used to construct three-deimesional hail cells respectively in terms of reflectivity,differential reflectivity,correlation coefficient and differential phase shift rate.At the same time,four kinds of dual polarization features,including typical value feature,gradient feature,intensity feature and texture feature,are designed by combining computer vision method and three-dimensional hail cell information.2.On the issue of hail recognition,a hail recognition model based on mechanism feature and principal component of dual polarization feature is proposed.Pearson correlation coefficient,standardized mutual information,L1-SVM and random forest are used to evaluate the importance of four types of dual polarization features.According to the evaluation results,the 86-dimensional dual-polarization feature is selected.After the principal component analysis,the first 2-dimensional principal of86-dimensional dual-polarization features component and the 6-dimensional mechanism features are used to form a comprehensive feature vector.Considering that a small number of hail samples may be mixed into the non-hail sample set,the class weighted support vector machine(CW-SVM)was selected as the hail recognition model.The experimental results show that the proposed algorithm is superior to the traditional hail recognition algorithm,with a hit rate of 86.2% and a critical success index of 64.1%.3.On the issue of hail landing time forecasting,an algorithm for hail landing time forecasting based on time series analysis is proposed in this thesis.This algorithm divides the time series into different periods and converts the hail landing time forecasting problem into a multi-classification problem of time period.Specifically,the hail cell disasters are split into sub-sequences by different time periods,and each subsequence is regarded as a sample.For each sample,the static features of cells at each moment and the dynamic features between adjacent moments are used as the basic information to correlate the sample and the information of time.This model of hail landing time forecasting is a multi-classification model,which forecasts time range of hail landing by classifying the categories of sample.Further,a comprehensive decisionmaking method based on Bagging CW-SVM,borrowed from the Positive and Unlabeled(PU)learning,is used to solve the sample imbalance problem and the negative impact of similar positive and negative samples.Experimental results show that the proposed algorithm is superior to other multi-classification algorithms in the prediction of hailfall time.The strict accuracy rate of predicting the hailfall period reaches 60.3%,and the loose accuracy rate reaches 84.3%.