Plant Disease Detection and Prevention Using Vision Transformer (VIT) and Large Language Model (LLM)

Abstract - In the worldwide production, the occurrence of plant diseases cause great economic losses. This paper introduces a plant disease prevention system that uses Vision Transformer model and Mistral-7B Large Language Model. Three structures were iteratively tested: ResNet18 (80% Accuracy), ViT-Small (~ 94%) and ViT-Base (99% overall accuracy on the PlantVillage dataset (38 diseases, 54,306 images). The ViT-Base model uses global self-attention to compute all 16×16 image patches, beating any kind of CNN even with thousands millions fine-grained categories in sight we were lucky only at first try resnet50 Dio One The ViT-Base model uses global self-attention to compute all 16×16 image patches,ening thousands millions fine-grain categories. Upon detection, Mistral-7B-Instruct generates structured prevention advice including organic and chemical treatments with real product recommendations. The system is made available as a Flask web application that supports file upload and real-time camera capture and is publicly accessible through an ngrok tunnel on a Google Colab GPU infrastructure. Keywords: PLANT DISEASE DETECTION, VISION TRANSFORMER (ViT), ViT BASE, ViT SMALL, RESNET 8, MISTRAL 7B, LLM, PLANT VILLAGE, FLASK, DEEP LEARNING, PYTORCH, AGRICULTURAL AI, TIMM .