Explainable Vision Transformers for Early Crop Disease Localization in Smart Farming

Abstract

Early detection and localization of crop diseases are critical for improving agricultural productivity and ensuring food
security. This study introduces a novel Centroid Entropy Localization Explainable Vision Transformer (CEL-EViT)
framework designed for accurate and interpretable identification of plant diseases in smart farming applications. The proposed approach integrates the classification capabilities of Vision Transformers (ViTs) with a Centroid Entropy Localization (CEL) module, which highlights disease-affected regions, providing explainable visual insights alongside predictions. The model divides input leaf images into patches, extracts deep feature representations, and employs attention-based mechanisms to capture subtle disease patterns. The CEL module computes entropy-weighted centroids to precisely localize diseased areas, enabling both high accuracy and interpretability. Experiments conducted on multiple crops—including tomato, mango, papaya, and maize—demonstrate that CEL-EViT significantly outperforms baseline models such as CNN, ResNet50, and hybrid ViT-CNN. The proposed model achieves an average classification accuracy of 96.6%, Intersection over Union (IoU) of 87.9%, and explainability score of 87.1%, along with high localization consistency and minimal centroid distance errors. These results confirm that CEL-EViT effectively identifies early disease symptoms while providing reliable visual explanations for the affected regions. The framework offers a scalable and practical solution for deployment in real-time precision agriculture systems, including drones and edge devices, facilitating timely interventions, reducing crop losses, and supporting sustainable farming practices. Overall, the proposed method demonstrates the potential of combining deep learning and explainable attention mechanisms for robust, interpretable, and early crop disease management.