A Low-Resource Transformer-Based Deep Learning Framework for Precise Localization and Classification of Visible Crop Diseases

Timely and accurate identification of crop diseases is essential for sustaining agricultural productivity and reducing economic losses, particularly in resource-constrained environments. This paper presents AgriVision Lite, a low-resource deep learning framework designed for precise localization and classification of visible crop diseases. The proposed hybrid model integrates YOLOv8-Nano for lightweight disease-region localization with DeiT-Tiny, a compact Vision Transformer, for efficient and high-fidelity classification. Using the Watermelon Disease Recognition Dataset, comprising images of Anthracnose, Mosaic Virus, Downy Mildew, and healthy samples, AgriVision Lite achieved an overall classification accuracy of 99%, outperforming conventional CNN and hybrid models. The model demonstrates robust generalization and can be extended to other crop types exhibiting similar symptoms without retraining, enabling cross-crop adaptability and few-shot learning. Its low computational cost and compatibility with embedded hardware such as Raspberry Pi make it suitable for real-time, field-deployable disease detection. The results establish AgriVision Lite as a scalable, affordable, and practical solution for advancing precision agriculture in low-resource settings.