Skip to main content
ARS Home » Southeast Area » Auburn, Alabama » Soil Dynamics Research » Research » Publications at this Location » Publication #429551
Research Project: Sustaining Productivity and Ecosystem Services of Agricultural and Horticultural Systems in the Southeastern United States

Location: Soil Dynamics Research

Title: Ros-AI: An LLM-enhanced scalable multimodal framework for UAV-based rose bloom analysis in precision agriculture

Author
item SUNDARAVADIVEL, P - The University Of Texas At Dallas
item MANJUNATHA, H - The University Of Texas At Dallas
item NARASIMHAMURTHY, K - The University Of Texas At Dallas
item BORAH, S - The University Of Texas At Dallas
item ANAND, A - The University Of Texas At Dallas
item Torbert Iii, Henry
item Kumpatla, Siva Prasad
item KNIGHT, P - Mississippi State University

Submitted to: Smart Agricultural Technology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 5/31/2026
Publication Date: 6/1/2026
Citation: Sundaravadivel, P., Manjunatha, H., Narasimhamurthy, K., Borah, S.S., Anand, A., Torbert III, H.A., Kumpatla, S., Knight, P. 2026. Ros-AI: An LLM-enhanced scalable multimodal framework for UAV-based rose bloom analysis in precision agriculture. Smart Agricultural Technology. 14:102267. https://doi.org/10.1016/j.atech.2026.102267.
DOI: https://doi.org/10.1016/j.atech.2026.102267

Interpretive Summary: Accurate detection of rose blooms is important for nursery management, but it remains challenging for UAV images due to size of flowers, dense clustering, and frequent occlusion in images. This study presents a multimodal framework that combines supervised deep learning, unsupervised computer vision, and large language models (LLMs) to address these challenges. UAV flights collected high-resolution RGB, and multispectral imagery of rose nurseries analysis. The results demonstrate that supervised and unsupervised approaches provide complementary strengths: supervised YOLO ensures high accuracy, while the quantifier enables scalable, low-cost deployment. Together with LLM-driven advisory, the framework moves beyond detection to actionable decision support, offering a pathway toward robust, affordable, and field-ready AI systems for precision agriculture.

Technical Abstract: Accurate detection of rose blooms is important for nursery management, but it remains challenging due to the tiny size of flowers, their dense clustering, and frequent occlusion in UAV images. This study presents a multimodal framework that combines supervised deep learning, unsupervised computer vision, and large language models (LLMs) to address these challenges. UAV flights collected high-resolution RGB, and multispectral imagery of rose nurseries. Among the supervised methods, a tile-based YOLOv8-m model with sliding-window inference achieved strong validation performance (mAP@0.5 = 0.985, recall = 0.948), confirming the effectiveness of tiling strategies for small-object detection. But these approaches required heavy annotation and GPU resources. To overcome these limitations, we introduced the RGB Bloom Quantifier, an unsupervised algorithm that uses dynamic red-channel thresholding, morphological filtering, and contour analysis. This method achieved 93.1% accuracy while running entirely on CPU hardware, eliminating the need for training data and reducing deployment cost. Bloom counts from both pipelines were integrated with locally deployed LLMs (LLaMA 3 and Gemini), which generated practical recommendations for workforce allocation, irrigation scheduling, fertilizer application, and harvest timing. The results demonstrate that supervised and unsupervised approaches provide complementary strengths: supervised YOLO ensures high accuracy, while the quantifier enables scalable, low-cost deployment. Together with LLM-driven advisory, the framework moves beyond detection to actionable decision support, offering a pathway toward robust, affordable, and field-ready AI systems for precision agriculture.