Leaf water stress degree of real-time diagnosis was one of the methods of scientific irrigation. A kind of leaf water stress degree classification method was putted forward, for feature extraction based on ResNet101 convolution Mask R-CNN networks, the blade was firstly divided, through the study of the migration Mask R-CNN on pre training to get the weight of COCO data set for instance segmentation of tomato leaves, the original convolution network training parameters were retained, and only the connection layer was adjusted. By using the features extracted from the convolutional network, tomato leaf segmentation could be regarded as a dichotomy problem to distinguish the leaf from the background, so as to segment tomato leaf images under different water stresses. Then after using the fine-tuning DenseNet169 leaf water stress degree classification image classification model, through the study of the migration DenseNet169 ImageNet data set for the training to get the weight for the classification of tomato leaf water stress degree, remain unchanged, the parameters of DenseNet169 convolution only trained the last fully connection layer, and modified the original DenseNet169 fully connection layer, amended the classification number from 1000 to 3. In the experiment, a total of 2000 images were collected of leaves of greenhouse tomatoes with obvious characteristics, including no water stress, moderate stress and severe stress. A data set was established and the model was trained and tested. Experimental results showed that the average Mathews correlation coefficient (MCC) of the Mask R-CNN blade instance segmentation model after training was 0.798 for single and multiple leaves on the test set, and the average accuracy (ACC) could reach 94.37%. After training of DenseNet169 leaf water stress, the degree of accuracy of classification model on the test set was 94.68%, and compared with that of the VGG-19 and AlexNet, the classification model accuracy was increased by 5.59 percentage points and 14.68 percentage points, respectively, and the average operation time of method to detect a 2 million-pixel image was 1.2s, but it had good effect on greenhouse tomato leaf water stress degree real-time diagnosis, which could provide reference for building intelligent technology to water stress analysis.