基于迁移CNN的江淮持续性强降水环流分型

利用新建的1981—2018年区域持续性强降水个例集、1981—2018年中国逐日降水量及NCEP/NCAR全球再分析资料，运用江淮地区持续性强降水典型模态个例样本及残差神经网络（CNN），通过迁移学习分步训练建立针对江淮强降水的环流客观分型模型；并运用该模型对1981—2015年全国持续性强降水个例的环流进行客观分型，比较其与相似量（R）分型、余弦相似系数（COS）分型的效果，且对2016—2018年逐日环流进行客观识别与分型。结果表明:迁移CNN在拟合准确率达到100%后，测试集损失函数很快稳定，准确率较高，比R分型、COS分型效果好。在强降水客观分型中，迁移CNN所得各型与典型模态降水之间的相关系数远高于R分型、COS分型，其中不一致型个例分析表明迁移CNN所得各型与典型模态降水间的相关系数明显高于R分型、COS分型。在独立样本分型中，迁移CNN所得各型与典型模态降水的相关系数也均高于R分型、COS分型，且对非持续性强降水环流分型也存在一定的识别能力。

Circulation Pattern Classification of Persistent Heavy Rainfall in Jianghuai Region Based on the Transfer Learning CNN Model

Newly reconstructed dataset of regional persistent historical heavy rain events in 1981-2018, corresponding daily rainfall data of 2474 observational stations in China, and NCEP/NCAR global reanalysis data of daily geopotential height field are used to study the persistent heavy rain events in Jianghuai Region.Based on 72 persistent heavy rainfall cases, typical rain patterns and circulation fields are refined by empirical orthogonal function(EOF). And the corresponding time coefficient is obtained by projecting rainfall of individual days to the typical rain patterns, and the training and test dataset samples are determined by the time coefficient. Using residual neural network(CNN), a transfer learning CNN classification model of Jianghuai persistent heavy rainfall is established by three transfer learning processes. Compared with the analog quantity(R) and Cosine similarity coefficient(COS) methods, the transfer learning CNN model has the highest classification accuracy on the test dataset.CNN, R and COS methods are used to objectively classify the circulation of all persistent heavy rain cases and to synthesize the distribution of various types of rainfall and circulation during 1981-2015. The statistical analysis shows that the transfer learning CNN model is better at classification. By comparing the correlation coefficients between rain distribution of each type and typical rain patterns, it shows that the transfer learning CNN model performs better than the R and COS methods. The variance between different types of geopotential height fields at 500 hPa obtained by the CNN model is the largest and the CNN model can better distinguish the circulation fields of different types of heavy rainfall.The analysis of samples with inconsistent objective classification of three methods shows that the correlation coefficients of various patterns of rainfall of the transfer learning CNN model are significantly higher than those of R typing and COS typing methods. The spatial distribution of various rainfall patterns of CNN model can clearly show the characteristics of the three typical heavy rain patterns, while the results obtained by R typing and COS typing methods are almost opposite to the typical rain patterns except for type Ⅱ. Considering classification of independent samples in 2016-2018, the correlation coefficients between the rain distribution of each type and typical rain patterns obtained by the transfer learning CNN model are much higher than the R and COS methods. The transfer learning CNN model has certain advantages over R typing and COS typing methods in classification and also has a certain ability to distinguish the non-continuous heavy rainfall circulation pattern.