Rice blast, Crop diseases monitoring, Deep learning, Data imbalance, Image augmentation, Transfer learning, Agriculture (General), S1-972, Agricultural industries, HD9000-9495

Rice blast is an airborne disease which can spread rapidly from small disease foci, and result in severe yield loss. To monitor the disease foci in the rice field effectively and timely, deep learning is applied to recognize dual-scenario images of multi-symptomatic rice leaf blast. In this study, a benchmark dataset containing chronic type and acute type of rice leaf blast over different growth stages of plants, as well as two other common rice leaf diseases and healthy rice leaves was constructed and made publicly available. Firstly, the impact of different training methods on imbalanced datasets was compared. Then six state-of-the-art convolutional neural network models were trained with the dataset by transfer learning and the hyperparameters of the outperforming models were further optimized to improve the recognition accuracy of models. The results proved that the quantity and quality of images had great impacts on the model performance, and image augmentation could greatly alleviate the problem of imbalanced inter class recognition performance. According to the experimental results, the overall performance of InceptionV3 was the best among the six models, and its highest validation accuracy was 99.78 % after parameter adjustment, and its highest test accuracy reached 99.64 %. The research demonstrated that the use of computer vision and deep learning to identify symptoms of crop diseases and to locate disease foci through the feedback frequency of infected images would be an effective method for intelligent disease monitoring in the future.