Multi-Agent Edge-Fog Intelligence With Explainable Vision Transformers For Sweet Lemon Crop Disease Analytics

Abstract—Edge Artificial Intelligence (Edge AI) has recently emerged as a crucial enabler of real-time, decentralized decision-making in smart agriculture. However, many existing models lack interpretability and scalability across large-scale farms. This paper proposes a Multi-Agent Edge-Fog Intelligence System using Vision Transformers (ViT) for crop disease diagnosis, specifically targeting sweet lemon leaf infections. A federated learning mechanism facilitates secure model training across edge nodes, while a centralized fog node performs global aggregation. Furthermore, explainable AI (XAI) via SHAP highlights the critical leaf areas influencing predictions. A novel Crop Disease Severity Index (CDSI) is introduced for severity assessment. Experimental results reveal 98.9% classification accuracy, inference latency below 80ms, and energy consumption 10× lower than cloud models. This approach bridges the gap between performance, transparency, and practicality in edge-deployed precision agriculture