Achieving fast and accurate recognition of plant leaf diseases in natural environments is crucial for plants’ growth and agricultural development. The deep learning technique has been broadly used in recent years in the area of plant leaf disease classification. However, existing networks with large number of parameters are not easily deployed to farms with limited end devices and cannot be effectively utilised in natural agricultural environments. This paper proposes a data augmentation-based knowledge distillation framework for plant leaf disease recognition. We improve the traditional knowledge distillation method based on a single image by using mixed images generated from data augmentation and label annotation, significantly enhancing the recognition accuracy of the model. We have experimented on the PlantDoc dataset. The experimental results demonstrate that our approach improves recognition accuracy by up to 3.06% compared to the traditional knowledge distillation method and up to 7.23% compared to the baseline model. This study shows that the method provides a viable resolution for the diagnosis of plant foliar diseases in realistic scenarios.