卷积神经网络联合SVM的降水粒子分类

Hydrometeor Classification Based on Convolutional Neural Network Combined with SVM

针对传统模糊逻辑降水粒子识别算法存在过度依赖专家经验来设置参数的缺陷,提出一种基于卷积神经网络(Convolutional Neural Network,CNN)和支持向量机(Support Vector Machine,SVM)联合结构的降水粒子分类方法。本文首先搭建了适用于双偏振雷达数据矩阵传播结构的4种卷积神经网络模型,通过对KOHX雷达各极化参数进行分块和堆叠操作,制作模型所需数据集并训练模型。然后根据各CNN模型对目标块的分类特点,结合SVM分类器,搭建能够识别5类目标粒子的联合结构。最后,对KOHX雷达0.5°仰角扫描数据进行测试,得到该联合结构模型的分类准确率达94.92%。并且对于不同仰角、不同雷达的扫描数据均能进行有效分类,表现出较好的鲁棒性。

Aiming at the defect of traditional fuzzy logic hydrometeor classification algorithm that excessively re⁃lies on expert experience to set parameters,a hydrometeor classification method based on convolutional neural network(CNN)combined with support vector machine(SVM)is proposed.Firstly,we build four convolutional neural network models applicable to the propagation structure of dual polarization radar data matrix,and the dataset required by the model is made through blocking and stacking the polarization parameters of KOHX radar.Then we train the model.Ac⁃cording to the classification characteristics of each CNN model on the target block and combining with the SVM classifi⁃er,a joint model that can recognize five types of target hydrometeors is built.Finally,the 0.5°elevation scanning data of KOHX radar are tested,and the classification accuracy of the joint model is 94.92%.It can effectively classify the scan⁃ning data of different elevations and different radars,which shows good robustness.