Integrating LLM for Cotton Soil Analysis in Smart Agriculture System

Abstract

Cotton is a critical crop for the agricultural economy, with its productivity closely tied to soil quality, particularly soil nutrient levels and pH. Monitoring and optimizing these properties is essential for sustainable cotton cultivation. This study proposes using fine-tuned Large Language Models (LLMs)—specifically GPT-2 and LLaMA-2—to automate soil analysis and produce detailed soil reports with actionable recommendations, addressing the limitations of traditional machine learning models in this context. A custom dataset was created by extracting key information from cotton-specific resources, focusing on soil nutrient interpretation and recommendations across different growth stages. Fine-tuning was applied to GPT-2 and LLaMA-2 models (specifically, the Nous Research version LLaMA2-7b-hf from Hugging Face), enabling them to generate data-driven reports on cotton soil health. The fine-tuned GPT-2 model achieved a training loss of 0.093 and an evaluation loss of 0.086, outperforming LLaMA-2, which had a training loss of 0.033 and an evaluation loss of 0.25. Evaluation with BERT Score showed that GPT-2 scored average Precision, Recall, and F1 scores of 0.9284, 0.9308, and 0.9296, respectively, highlighting its superior report accuracy and contextual relevance compared to LLaMA-2. The generated reports included soil properties and actionable nutrient management recommendations, effectively supporting optimized cotton growth. Implementing fine-tuned LLMs for soil report generation enhances nutrient management practices, contributing to higher yields and more sustainable cotton farming.