Various shgape of small particles are exist in clouds. HYVIS is the one of the direct and insitu observation insturument. HYVIS takes the movies which has approximately 110,000 frames for an obserbation. Although, the movies were manually classified and measured size of particles every several frames, it was tough to applied statistical works because of massive bserved frames. Therefore, The purpous of this study is to classify cloud particles objectively and effectively by deep learning, then investigate the statistical microphysical structures in clouds.

In this study, YOLO whchi is one of the deep lerning and computer vision and several types of training data was used fore training and detection. The result of trained by only HYVIS images was low validity because not enough particle numbers in taining data. So, we aimed to improve validities by added typical particle shapes, data augumented images and chouse gradient desent mesod. Validities of most of the particle type without irregurler type is grately improved by changing gradient decent to Adam. Reasons of this is we chouse aprropleate iigradient decent and it directed to effective training. In the case of Cloud Scope, its validity was not equal to the manual classification. Because Cloud Scope take much more shrap and dark figures and those are different to that of HYVISs. To solve this problem,we propose make Cloud Scope images resemble to HYVISs sharpness and brightness.