Multimodal agricultural intelligence: Comparative analysis of large language models for precision nursery management systems

Authors: MANJUNATHA, H - The University Of Texas At Dallas; BORAH, S - University Of Texas; SUNDARAVADIVEL, P - University Of Texas; Torbert Iii, Henry; Kumpatla, Siva Prasad; KNIGHT, P - Mississippi State University

Submitted to: Meeting Abstract
Publication Type: Proceedings
Publication Acceptance Date: 10/13/2025
Publication Date: N/A
Citation: N/A

Interpretive Summary: Modern precision agriculture requires highly advanced computational platforms able to process heterogeneous sensor streams and deliver actionable advice for operational decisions. This research provides a comprehensive comparative evaluation of three state-of-the-art large language models (LLMs), Mistral, Gemma3, and Llama3, coupled with computer vision expertise designed to enable autonomous agriculture management. Through systematic experimentation based on high-definition UAV imagery from a 300-hectare rose plantation, we demonstrate significant variations in model performance, computational efficacy, and output advisory quality. This work sets out empirical baselines for LLMs' implementation in resource-poor agriculture environments while maintaining full data sovereignty, an essential constraint for agriculture enterprises protecting proprietary cultivation methods.

Technical Abstract: Modern precision agriculture requires highly advanced computational platforms able to process heterogeneous sensor streams and deliver actionable advice for operational decisions. This research provides a comprehensive comparative evaluation of three state-of-the-art large language models (LLMs), Mistral, Gemma3, and Llama3, coupled with computer vision expertise designed to enable autonomous agriculture management. Our work addresses the wide gap between AI theoretical possibilities and their eventual implementation in agriculture by designing annotation-free detection methods and evaluating multi-architecture scaling attributes of performance. Through systematic experimentation based on high-definition UAV imagery from a 300-hectare rose plantation, we demonstrate significant variations in model performance, computational efficacy, and output advisory quality. The analysis shows that the parameter quantity does not have a direct linear association with advisory quality in agriculture, as shown by Gemma3's parsimony-driven architecture realizing better performance for every computational unit. This work sets out empirical baselines for LLMs' implementation in resource-poor agriculture environments while maintaining full data sovereignty, an essential constraint for agriculture enterprises protecting proprietary cultivation methods.