PROTOINCNET: A PROTOTYPICAL NETWORK-BASED DEEP LEARNING APPROACH FOR CLASSIFYING RICE LEAF DISEASES

Abstract

Rice disease is always a challenge worldwide, particularly with the harmful threats of Leaf Blast, Brown Spot, and Hispa to rice crops. This study proposes a comprehensive meta- learning framework, ProtoIncNet, that integrates InceptionV3 for feature extraction and Prototypical Networks for classification, combining advanced deep learning techniques to effectively address the challenges of rice leaf disease classification. The proposed model was trained on the Rice Crop Dataset, which consists of images representing four rice leaf categories: Leaf Blast, Brown Spot, Hispa, and Healthy, capturing intricate patterns and features associated with each disease. The model was comprehensively trained on this complex dataset to identify subtle distinctions that determine the presence of these paddy diseases. This approach shows that combining transfer learning with few-shot learning enables accurate disease detection even with limited data, offering a scalable path toward explainable and farmer-friendly digital agriculture systems. To evaluate its effectiveness, we compared ProtoIncNet with ResNet50, VGG16, and EfficientNet under the same experimental conditions. The proposed approach achieved the highest accuracy of 98.3%, significantly outperforming these baseline models. Its deployment can help large-scale farmers mitigate crop losses and promote the adoption of digital agriculture systems in developing regions.