Lesion-aware visual transformer network for Paddy diseases detection in precision agriculture

Abstract

Precision agriculture, driven by advancements in sensing technologies and data analytics, offers promising solutions for addressing challenges in paddy disease management. Paddy diseases have significant detrimental effects on crop yield and quality, necessitating timely and accurate detection for effective disease management. Deep learning has shown promise in identifying plant diseases from leaf images, including those in paddy crops. However, the presence of slight variations among different types of paddy diseases poses a significant generalizability challenge. In this study, for the first time, we introduce a lesion-aware visual transformer for accurate and reliable detection of paddy leaf diseases through identifying discriminatory lesion features. A Novel multi-scale contextual feature extraction network is presented to enable capturing a contextual local and global representation of disease features at different scales and channels. Then, a weakly supervised Paddy Lesion Localization (PLL) unit was presented to locate distinctive lesions in paddy leaves that provide the model with discriminative leaf regions that can guide the final classification decision. A feature tuning unit is presented to empower modeling the relations within the global and local latent spaces, thereby improving the spatial exchanges between visual semantics of paddy leaves. The exhaustive experimental comparison against state-of-the-art solutions on public paddy disease datasets demonstrated the efficiency and versatility of our system with an average of 98.74% accuracy and 98.18% f1-score.