Detection of Plant Diseases Using AI

Abstract — This project introduces an AI-powered agricultural assistance system that integrates text-based consultancy, voice interaction, and image-based disease detection into a cohesive platform. Unlike cloud-reliant solutions like Kissan AI, which uses GPT-3.5 Turbo’s outdated 2021 dataset, our system employs the Llama 3.2 3B model - a lightweight, locally deployable language model updated through December 2023. This ensures access to the latest agricultural practices and policies. By combining Retrieval-Augmented Generation (RAG) with semantic search, it surpasses generic chatbots in domain-specific accuracy, processing queries on-device to eliminate API costs and safeguard data privacy. The system’s architecture leverages Llama 3.2 3B for its optimal performance (63.4% MMLU score) and efficiency (3.4GB memory footprint), paired with a Chroma vector database for swift retrieval of farming guidelines. This contrasts with GPT-3.5 Turbo’s resource-intensive 175B parameter design, unsuitable for rural hardware. Queries are processed via SpaCy’s NLP pipeline for entity recognition and noun chunking, while Llama 3.2’s advanced long-context handling, validated on BABILong datasets ensures precise synthesis of technical documents like crop rotation manuals, improving on GPT-3.5 Turbo’s generic outputs. For disease diagnosis, a CNN model that has been trained on 61,486 plant images from 39 categories works in the RAG system. Locally processed images cause immediate retrieval of treatment plans regardless of the internet, which is critical for remote areas. With localized processing, computer vision, and voice interaction, the system provides real-time, scalable crop care assistance, responding 62% quicker than GPT-3.5 Turbo-based systems while being accurate in low-resource environments. Index Terms—Artificial Intelligence, Agriculture, Plant Disease Detection, Retrieval-Augmented Generation, Voice Interaction.