Historical similar typhoons are important references for typhoon prediction and decisionmaking besides conventional methods. However, it is timeconsuming and laborious to retrieve similar information from a large number of historical typhoons. In this paper, a similarity retrieval method of subtropical high based on improved visual geometry group model (VGG16) is proposed to query historical similar typhoons based on subtropical high similarity. Through the image extraction, data enhancement, model learning and optimization of 19 〖KG-*5〗736 subtropical high corresponding times in the typhoon season from 1979 to 2020, taking the learned perceptual image patch similarity (LPIPS) as the measurement index of subtropical high similarity, we establish an improved VGG16 model finally. The test results show that greatly similar historical typhoons can be found by using this model. The similarity between the historical similar typhoon ranked first retrieved by this model and the target typhoon is 92.55%, which provides a very meaningful reference for typhoon forecasters. At the same time, compared with the traditional manual recognition, this model takes shorter recognition time and has higher retrieval efficiency, and can be applied in typhoon forecasting and research.