Early Crops Diseases Identification Using Deep Transfer Learning

Most crops diseases first exhibit their symptoms on the leaves. By applying deep learning to classify disease leaves, what kind of disease occurs in corps can be identified. However, in the earliest stage of crop diseases, a lack of sufficient disease leaf image samples can degrade the performance of deep learning. We propose a sophisticated deep transfer learning model based on ResNet network which integrates transfer learning and image augmentation. With the final fine-tune model, we can still achieve an identification accuracy rate of over 99% even under extreme conditions where only a few dozen images are available. Experiments substantiate the reliability and robustness of our model. By identifying crop diseases as early as possible, an abundance of direct and derivative benefits can be obtained, including the reduction of farmer's losses, the minimization of pesticide usage to prevent soil contamination, and the enhancement of food safety, all of which meets the requirements of ecological agriculture.