基于卷积神经网络的飑线识别算法

为了探究深度学习用于飑线识别的可行性，基于2008—2020年河南省郑州和驻马店雷达数据，采用卷积神经网络（convolutional neural network，CNN）算法构建飑线识别模型，引用临界成功指数、公平风险评分、命中率和误判率定量评价模型的识别效果，对比不同样本组成比例和网络结构对飑线识别效果的影响。结果表明:建模所用的样本组成比例对飑线识别有一定影响，通过改变采样方式和优化网络结构均能够改善样本比例不平衡的问题，提高飑线识别效果，且后者提升的幅度更大，而两种方法的结合无明显提升。测试结果表明:该模型临界成功指数为0.66，公平风险评分为0.58，命中率为0.86，误判率为0.24。研究揭示了卷积神经网络能够提取并学习飑线和非飑线回波的图像特征，对飑线有一定识别能力。

Squall Line Identification Method Based on Convolution Neural Network

Squall line often leads to heavy rain, gale and hail, which is a difficult key problem in nowcasting. In order to explore the feasibility of deep learning for squall line identification, the training, validation and test set sample sets are established based on the radar data of Zhengzhou and Zhumadian in Henan Province during 2008-2020. The convolutional neural network (CNN) algorithm is used to construct a squall line identification model. The critical success index (CSI), equitable threat score (ETS), hit rate (POD) and false positive rate (FAR) are used to quantitatively evaluate the identification effect of the model. The influence of different sample composition and network structure on squall line identification effect are compared. The results show that the composition ratio of sample is imbalanced, because squall line accounts for very small proportion in all kinds of weather processes. This imbalance will degrade the classification performance of the identification model to squall line samples. The imbalance of sample composition can be improved by changing sampling mode and optimizing network structure, both can improve the identification efficiency, especially the latter. However, the combination of the two methods does not bring further improvement. The over fitting problem in network training can be alleviated by increasing the sparsity and randomness of the network structure. The validation set shows that CSI is 0.87, ETS is 0.82, POD is 0.96, and FAR is 0.10. Based on the test set, the echo can be correctly identified by network as non-squall line in the weak stage of convection development, and as squall line in the strong stage of squall line development. The echo intensity and spatial distribution of squall line cases differ greatly, and the samples in the test set have the image features which are not included in the training set, and therefore the identification effect reduces. The test set show that CSI is 0.66, ETS is 0.58, POD is 0.86, and FAR is 0.24. The research reveals that CNN can extract and learn the image features of squall line echo, and it has a certain ability to identify squall line.