AI-BASED LEAF DISEASE DETECTION SYSTEM USING VISION-LANGUAGE MODELS

Waste management is a pressing issue worldwide, but the threat of plant diseases to agricultural productivity is equally formidable. According to studies published by the Food and Agriculture Organization (FAO), crop diseases account for approximately 20-40% of total global crop losses annually, translating to billions of dollars in economic damage and threatening food security in vulnerable regions, particularly across Asia and Africa.

Traditionally, the identification of plant diseases has relied on visual inspection by trained agronomists and pathologists. This approach, while effective, is expensive, time-consuming, and not scalable to the vast acreage of farmlands particularly in developing nations. The proliferation of smartphones with high-resolution cameras and affordable mobile data has opened a compelling opportunity: enabling farmers to photograph diseased leaves and receive an instant, AI-generated diagnosis.

Prior approaches to automated plant disease detection have predominantly employed custom-trained convolutional neural networks (CNNs) on large labelled datasets such as PlantVillage (Hughes & Salathe, 2015). While these approaches achieve high accuracy in controlled laboratory settings, they generalize poorly to real-world in-field photographs due to variations in lighting, background, leaf orientation, and image quality. Furthermore, the development of such systems requires significant GPU compute infrastructure, months of data collection and labelling, and domain-specific expertise in model architecture selection.

This paper proposes an alternative paradigm: leveraging a pre-trained, hosted vision-language model (VLM) via API, guided by a carefully engineered prompt, to produce structured disease analysis reports from raw leaf images. Recent advances in large multimodal models, including GPT-4V (OpenAI, 2023) and Meta's Llama series, have demonstrated that such models encode substantial embedded knowledge about plant biology from pre-training on internet-scale data, making them viable candidates for agricultural diagnostic applications without domain-specific fine-tuning.