Skip to main content
ARS Home » Plains Area » Fargo, North Dakota » Edward T. Schafer Agricultural Research Center » Sugarbeet Research » Research » Publications at this Location » Publication #428187
Research Project: Improving Sugarbeet Productivity and Sustainability through Genetic, Genomic, Physiological, and Phytopathological Approaches

Location: Sugarbeet Research

Title: A comparative study of U-Net and YOLOv11 for weed segmentation

Author
item ZHANG, XIAMENG - University Of Washington
item Kim, James
item BALAMURUGAN, RIDHANYA SREE - North Dakota State University
item TIDA, UMAMAHESWARA RA - North Dakota State University
item VEMURI, MADHAVA SARMA - University Of Washington

Submitted to: Applied Engineering in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 2/3/2026
Publication Date: N/A
Citation: N/A

Interpretive Summary: This paper discusses an Artificial Intelligence (AI)-based classification approach to identify weeds in images. Prior studies showed potential of AI-driven weed identification but are limited to separation of weeds from crop canopies. In addition, herbicide resistant (HR) weeds are spreading and difficult to identify. We propose and compare two different AI models to identify weeds, including HR weeds. To improve detection accuracy of the model, Meta's segmentation model called Segment Anything Model (SAM) was used to generate mask images of target weeds from manual rectangular notation. The trained models were tested with four weed types (kochia, horseweed, ragweed, and redroot pigweed). This study provides a performance analysis of the AI models for weed segmentation and facilitates site-specific weed management in agricultural crop fields.

Technical Abstract: In agriculture, weeds compete with crops for essential resources, often leading to significant reductions in crop yields. Accurate and efficient discrimination between crops and weeds is therefore essential for improving agricultural productivity and enabling effective weed management. Deep learning-based segmentation models such as U-Net and YOLOv11 have shown promise for weed classification tasks, but they differ in terms of segmentation accuracy and computational efficiency. In this study, we use pixel-level masks generated by the Segment Anything Model (SAM) to train both U-Net and YOLOv11, and conduct a comparative evaluation of their performance in weed segmentation. We assess the models with respect to segmentation accuracy and inference time. The results offer insights into the trade-offs between these approaches and provide guidance for selecting suitable models for real-time weed segmentation in agricultural applications.