Abstract: To improve the identification rate of maize diseases, an identification method of diseases is established based on a deep convolutional neural network under natural environment condition. Ten common maize diseases were taken as the study objects, and the image preprocessing was carried out, the Triplet loss double convolution neural network structure was applied to study the features of maize images. Then, SIFT algorithm was adopted to extract textural features. By this way, the labeling and classification of images through the Softmax were conducted. The normal maize images and the diseased ones were combined by the training set. Deep learning on similarity network check was used to explore the features of normal maize images, and furthermore, transfer learning was used to delve into the features of diseased maize images. The features were classified and identified. The result shows that this method could accurately identify the ten common maize diseases, and the classification accuracy is over 90%.