A COMPUTER VISION-BASED FRAMEWORK FOR CO-INFECTION DETECTION AND SEVERITY ASSESSMENT IN PLANT LEAVES

Abstract

Accurate quantification of plant disease severity is critical for early intervention and sustainable crop management. However, it is a challenging task due to the frequent co-occurrence of multiple pathologies on a single leaf, varying illumination conditions, and high interclass similarity among severity levels. In this paper, we present a hybrid feature representation framework for the simultaneous quantification of individual disease severity levels on a single leaf. It combines handcrafted texture descriptors with deep transformer-based visual features for robust multi-label severity analysis. Specifically, the Weighted Local Binary Pattern (WLBP) and Haralick texture features capture fine-grained local lesion variations and second-order statistical spatial relationships, while the EVA02 Vision Transformer models the global semantic context and long-range dependencies across the leaf surface. The extracted features are normalized and fused into a unified and discriminative representation. The model can estimate the exact percentage and severity grade for each identified disease. The framework was tested using images of cherry and pear leaves from the PlantCity dataset, which show complex symptomatic patterns in stone and pome fruits. Experimental results show that the proposed fusion strategy is able to achieve higher classification accuracy of 84.14% for cherry and 85.42% for pear leaves, able to classify signatures of disease conditions with overlapping features successfully and better than the individual feature extractors. The results demonstrate that the integration of these global features with local features extracted using texture descriptors greatly enhances the granularity of disease classification and ensures a reliable approach for accurate multi-symptom diagnosis in smart farming applications.