Transfer Learning Based Crop Disease Identification Using State-of-the-art Deep Learning Framework

Agriculture production plays a crucial role in the growth of the world's economy. However, technologies and innovations in the Precision Agriculture (PA) domain are still in their infancy. Typically, the problem of incorrect identification of crop disease still remains unsolved, thus causes agricultural production losses worldwide. To solve this problem, we apply transfer learning based Convolutional Neural Networks (CNNs) to identify different diseases of the crops. Specifically, this study focuses on evaluating the identification performance of the four state-of-the-art deep learning architectures: Residual Network (ResNet), MobileNet GoogLeNet and AlexNet. The PlantVillage dataset used in this study is made publicly available for academic study at Kaggle repository, which consists of 54,323 images of 38 kinds of crop diseases, and 14 kinds of crop diseases from them are derived as dataset for this study. By applying transfer learning, we make a comprehensive performance comparison among the implemented architectures, and evaluation results show that the MobileNet performs better in identifying crop disease, achieving an accuracy of 99.27%, followed by the GoogLeNet achieving an accuracy of 99.14 %, AlexNet achieving an accuracy of 97.92%, and ResNet achieving an accuracy of 97.03 %. Experiment results in this study prove the effectiveness of transfer learning based CNN models in crop disease identification, as well as provide a guidance on choosing an appropriate network to identify different crop diseases.