A Nightshade-SWIN Transformer-Based Framework for Enhanced Multi-Crop Plant Disease Classification

Abstract: Nightshade crops (Solanaceae) face significant yield threats from phenotypically similar foliar diseases that can be difficult to distinguish during the early stages of infection. Accurate and timely disease identification is therefore essential for reducing crop losses, improving disease management, and supporting sustainable agricultural practices. While Convolutional Neural Networks (CNNs) have demonstrated strong performance in image-based plant disease classification due to their ability to extract spatial and local features, they may have limitations in capturing long-range dependencies and global contextual relationships within leaf images. These limitations become particularly important when disease symptoms exhibit subtle differences in texture, color, shape, and lesion distribution or when images are captured against complex field backgrounds. To address these challenges, we propose Nightshade-SWIN, an optimized deep learning framework based on the Swin Transformer architecture. The proposed framework utilizes a hierarchical shifted-window self-attention mechanism to effectively capture both minute lesion textures and global leaf morphology. By progressively learning multi-scale representations, Nightshade-SWIN can identify discriminative pathological characteristics while maintaining computational efficiency. The model was evaluated on a curated dataset of 5,000 multi- crop images collected under variable field conditions, including differences in illumination, background complexity, leaf orientation, and disease severity. Experimental results demonstrate that Nightshade-SWIN achieves a peak classification accuracy of 94.6%, outperforming ResNet50 and EfficientNet-B0 by 4.9% and 3.4%, respectively. These results indicate the effectiveness of transformer-based feature learning for distinguishing visually similar foliar diseases. Furthermore, interpretability analysis using LIME and Grad-CAM confirms that the model primarily focuses on biologically relevant pathological regions, such as lesions, discoloration, spots, and infected leaf areas. This enhances the transparency and reliability of the proposed framework and supports its potential application in precision agriculture, automated crop monitoring, and early disease diagnosis.

Keywords: Nightshade Crops, Leaf Disease Detection, SWIN-Transformer, Grad-CAM and LIME, Deep Learning