An Edge-Deployable Lightweight Ensemble Framework For Grape Leaf Disease Detection Using Vision Transformers And Cnns

Abstract

Grape leaf diseases significantly impact crop yield and quality in viticulture, necessitating rapid and accurate identification for timely intervention. While deep learning has shown promise in plant disease detection, most existing models are either computationally intensive or limited in generalization across real-world conditions. This paper proposes a novel ensemble framework that combines lightweight Convolutional Neural Networks (CNNs) with a Vision Transformer (ViT) to detect multiple grape leaf diseases efficiently.

The system integrates MobileNetV3, EfficientNet-B0, and a fine- tuned ViT model using weighted hard voting to improve prediction robustness.An optimized preprocessing pipeline using contrast enhancement and color-space transformations is used to improve model sensitivity to subtle visual features. Furthermore, the trained model is compressed and deployed to an edge device (Raspberry Pi 4) to validate inference latency and performance under constrained computing environments.

Experimental results on a custom grape leaf disease dataset demonstrate an average classification accuracy of 96.4%, outperforming individual models and previous state-of-the-art methods. The proposed architecture balances accuracy, speed, and model size, making it suitable for real-time disease monitoring in vineyard settings. This work contributes toward accessible, AIdriven agricultural solutions for small and medium-scale grape farmers.