Abstract:In order to analyze the distribution characteristics of ice crystals in cold clouds and reveal the evolution mechanism of their growth, ice crystals are classified into 8 categories and labeled according to their shapes and sizes. At the same time, one type of partition column is labeled for data quality control. Then, the 9 types of labeled images are integrated to build an image dataset. The transfer learning VGGNet16 model is adopted for identification training, and the classification accuracy of the trained model reaches 98%. The model is applied to the study of ice crystal characteristics of cold clouds in autumn. Three cumulus-stratiform mixed cloud precipitation processes and three stratiform cloud precipitation processes in 2019 are selected to analyze the proportion of ice crystal shapes in different temperature ranges and the change characteristics of ice crystal size distribution. The results show that the initial shape distribution of ice crystal is determined by the temperature’s affecting the ratio of base surface to edge surface of ice crystals. In the same temperature range, the proportion of spherical ice crystals and linear ice crystals in cumulus-stratiform mixed clouds is higher than that in stratiform cloud, and the proportion of various ice crystals is relatively fixed when the temperature is below -12℃. In cumulus-stratiform mixed clouds, the diameters of linear ice crystals are concentrated in 300-800 μm with the ice crystal size distribution being multimodal. The diameters of aggregated ice crystals are greater than 600 μm. The concentrations at the beginning and the end of the ice crystal spectrum are equal. The diameters of spherical ice crystals are concentrated in the range of 120-300 μm, and the ice crystal size distribution shows a monotonically decreasing trend.