FARM PREDICT 360: AN AI-POWERED WEB PLATFORM FOR AGRICULTURAL CROP ADVISORY, MARKET PRICE FORECASTING, AND CNN-BASED PLANT DISEASE DETECTION

Abstract

FarmPredict 360 is a full-stack agricultural intelligence web application developed on the MERN stack — MongoDB, Express.js, React.js, and Node.js — extended with a Python FastAPI microservice that integrates LangChain and the Groq large language model API. The platform delivers three core capabilities to farmers, traders, and agribusinesses across Telangana, India. The Crop Advisor module accepts soil nutrient parameters and GPS coordinates, retrieves live weather data from the OpenWeather API, and applies the Llama3-70b large language model to recommend the top three most suitable crops with agronomic justifications, yield estimates, and fertilizer guidance. The Price Forecasting module predicts tomorrow’s minimum and maximum market prices and generates a seven-day forward forecast for any crop across ten Telangana districts and their APMC markets, accompanied by actionable sell or hold recommendations. The Plant Disease Detection module accepts a leaf photograph, identifies the disease using a Convolutional Neural Network (CNN) trained on the PlantVillage dataset, and delivers a comprehensive treatment plan covering chemical treatments, organic remedies, and preventive measures. Experimental results show that the CNN achieves 92.4 percent overall accuracy across 38 disease classes, while LLM-based advisory responses are generated in under three seconds on average, confirming the system’s practical viability for real-world agricultural decision support.

Plant phenotyping relevance

農業意思決定支援全体の一部だが、葉画像から植物病害をCNNで推定する機能が中核モジュールとして明示され、38病害クラスで精度評価されているため、植物表現型（病害状態）の取得・推定手法として採録する。

Code and data availability

The paper describes a CNN trained on PlantVillage and an LLM-based platform, but provides no public dataset, image, code, model checkpoint, or supplement of its own. PlantVillage is cited prior work, and all listed URLs are generic documentation or external services, not authors' assets.

No evidence-backed public reproduction asset is currently recorded.