Praf-Net: A Pathology Relational Attention Fusion Network for Fine-Grained Rice Leaf Disease Classification and Severity-Aware Analysis

Abstract

The reduced crop yield and grain quality due to rice leaf diseases have a great impact on agricultural productivity and food security. Due to complex lesion texture, intra-class variability, inter-class similarity and different severities of infected leaf regions, the presence of visually similar rice leaf diseases is still a challenging problem. To overcome these challenges, this paper proposes a novel Pathology Relational Attention Fusion Network (PRAF-Net) for fine-grained rice leaf disease classification and severity-aware analysis. This proposed framework combines handcrafted pathology descriptors, convolutional texture representations, transformer-guided contextual embeddings within a unified pathology-aware fusion architecture for learning graph-based disease relational and severity-aware auxiliary representation separately. Here, rice leaf images are preprocessed with data augmentation and normalization to help generalize well and avoid overfitting. To detect and extract disease specific structural features PRAF-Net employ some handcrafted pathological features such as Gray-Level Co-occurrence Matrix (GLCM), Local Binary Pattern (LBP), color statistics, lesion shape descriptors. On the other hand, EfficientNetB0 captures local texture-aware features and MobileViT gathers global contextual dependencies and disease spread patterns. The combination and optimization of the extracted multimodal features are performed using feature normalization and PCA. In order to distinguish between visually similar diseases, a Graph Attention Network (GAT)-based relational learning module is presented to study the semantic relationships for disease categories. And moreover, a severity-aware auxiliary learning is introduced to model disease progression characteristics and improve pathology representation learning. The Pathology Relational Attention Fusion proposed in this work, combines the favourable characteristics of multi-head attention and adaptive gating-based mechanisms to combine all texture, contextual, relational as well as severity-aware embeddings learned in a dignostic task for better robustness against unseen disease classification. Experimental evaluation on a six-class rice leaf disease dataset shows the efficacy of PRAF-Net, providing an accuracy of 97.44%, precision of 97.43%, recall of 97.44% and f1-score of 97.43%, which outperforms several competing state-of-the-art CNN, transformer and hybrid deep learning models in literature. In addition, large-scale ablation studies and explainability analyses confirm that the corresponding sites of proposed modules also contribute to performance improvements. Moreover, a real-time GUI-based deployment system is created to support practical disease diagnosis and severity visualization for intelligent precision agriculture applications.