Stacking ensemble model of deep learning for plant disease recognition

Diverse plant diseases have a major impact on the yield of food crops, and if plant diseases are not recognized in time, they may spread widely and directly cause losses to crop yield. In this work, we studied the deep learning techniques and created a convolutional ensemble network to improve the capability of the model for identifying minute plant lesion features. Using the method of ensemble learning, we aggregated three lightweight CNNs including SE-MobileNet, Mobile-DANet, and MobileNet V2 to form a new network called Es-MbNet to recognize plant disease types. The transfer learning and two-stage training strategy were adopted in model training, and the first phase implemented the initialization of network weights. The second phase re-trained the network using the target dataset by injecting the weights trained in the first phase, thereby gaining the optimum parameters of the model. The proposed method attained a 99.37% average accuracy on the local dataset. To verify the robustness of the model, it was also tested on the open-source PlantVillage dataset and reached an average accuracy of 99.61%. Experimental findings prove the validity and deliver superior performance of the proposed method compared to other state-of-the-arts. Our data and codes are provided at https://github.com/xtu502/Ensemble-learning-for-crop-disease-detection. © 2022, The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.