Grafting is beneficial in enhancing the adaptability to biotic and abiotic stresses， and improving the yield and quality of grapes. There are varieties of grape rootstocks， while their precise clarification become complex and difficult. Deep learning， capable of rapidly capturing deep features from images， has been extensively applied in the field of plant image classification and recognition. In this study， the mature leaf images of 30 grape rootstocks were deployed to construct a dataset， comprising 13547 grape rootstock leaf images. Four convolutional neural networks， GoogleNet， ResNet-50， ResNet-101 and VGG-16， were used for image recognition. The results show that the classification network with the highest accuracy is ResNet-101， and the accuracy reaches 97.5 % under the optimal model parameters （learning rate：0.005， mini-batch size：32， Max epochs：50）. Among the 30 varieties， the average prediction precision rate was 92.59%， and the prediction precision reaching 100% was observed in seven varieties. The recall rate of eight varieties reached 100%， and the average recall rate was 91.08 %. The leaf surface texture， leaf vein and leaf margin were major factors that influence variety classification. This study confirmed the application capacity of deep learning network models in real-time automatic identification of grape rootstocks， thus providing reference for the protection， utilization， clasification research of grape rootstock varieties and the variety recognition of other crops.