An interpretable fusion model integrating lightweight CNN and transformer architectures for rice leaf disease identification

Abstract

Swift identification of leaf diseases is crucial for sustainable rice farming, a staple grain consumed globally. The high costs and inefficiencies of manual identification underline the requirement of prompt disease detection. Traditional approaches for identifying leaf diseases in crops, particularly rice, are laborious, and and often ineffective. Given the significant impact of leaf diseases (such as Rice Blast, Brown Spot, and Rice Turgor) on rice quality and yield, computer-assisted detection can be an effective method of ensuring the long-term sustainability of rice production. This study utilizes advanced artificial intelligence (AI) as the optimized bidirectional encoder representations from the transformers for images(BEiT) model along with pre-trained CNNs (Convolutional Neural Networks), to build a comprehensive study for detecting rice leaf diseases. We train and validate two extensive datasets, featuring healthy and various types of unhealthy plant and rice leaf images respectively.Our optimized model demonstrates high accuracy, outperforming other deep learning and transformer-based models such as ViT, Xception, InceptionV3, DenseNet169, VGG16, and ResNet50. The proposed model achieves a precision of 0.97, a recall of 0.96, and an F1-score of 0.97.The explainability of our proposed model is achieved through the use of segmentation techniques in conjunction with the Local Interpretable Model-agnostic Explanations (LIME) method.

Keywords

• Plant disease detection
• BEiT model
• Attention mapping
• Deep learning
• Process innovation
• Food productivity