Rice Leaf Disease Stages Classification Using Few-shot Learning Techniques

Plant disease can significantly reduce crop yield, affect biodiversity, farmers, and the agricultural industry. Identifying diseases early allows farmers to mitigate financial losses by incorporating timely control measures and minimizing crop damage. Few-shot learning methods can be used for identifying plant diseases; it is advantageous in situations that include only a few samples of data for a particular disease and cannot be applied to other methods for identifying the affected region. Fewshot learning methods like the Siamese model and prototypical networks are used in the current research to identify the disease stages of two diseases affecting rice crops. A few-shot dataset for the two diseases was developed and evaluated using the proposed Siamese model and the prototypical networks. The dataset included the different early, advancing, severe, and healthy stages of the two diseases considered in the research. The Siamese model gave an average accuracy of above 85% for both diseases. The proposed prototypical network with a custom convolutional neural network as the embedding network gave an accuracy of above 60% for both diseases. The model with the pretrained models as the embedding network gave an accuracy above 65%. The Siamese model performed best on the disease stages of the Rice Bacterial Blight disease, providing an accuracy of 95.014%.