An LLM-Powered Crop Disease Detection System for Potato Agriculture Using Deep Learning Comparative Study

Abstract

Crop diseases represent a major challenge to global food security, especially in countries where agriculture is the main economic sector, such as Algeria exactly Elouad region. The potato crop is particularly vulnerable to two of the most destructive diseases, early blight and late blight, two of the most aggressive and fast-spreading fungal diseases. These pathogens can devastate crops in a short period, causing significant yield losses, reducing market quality, and increasing production costs. The lack of early and accurate detection methods further exacerbates the impact, leaving farmers with limited options for timely intervention and disease. This thesis proposes an advanced deep learning-based system for the automated de- tection and classification of early blight and late blight in potato crops. The aim is to provide a scalable, efficient, and accurate solution for identifying these diseases at an early stage, enabling timely intervention and reducing crop losses. A combined dataset of healthy and diseased potato leaf images was used to train multiple CNN models. To boost accuracy and generalization, we applied ensemble learning with techniques such as weighted averaging. The resulting model showed strong performance across standard metrics, including accuracy and precision. To bridge the gap between predictions and practical use, the system integrates a large language model (LLM) to generate clear, context-specific advice for farmers, including treatments, prevention methods, and best practices. Finally, the proposed solution provides a scalable, intelligent tool for accurate disease detection and practical support, highlighting the potential of deep learning and language models in advancing sustainable, data-driven agriculture.