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 #413015
Research Project: Improving Sugarbeet Productivity and Sustainability through Genetic, Genomic, Physiological, and Phytopathological Approaches

Location: Sugarbeet Research

Title: Weed identification using U-Net machine learning model and SAM segmentation

Author
item Kim, James
item BALAMURUGAN, RIDHANYA - North Dakota State University
item TIDA, UMAMAHESWARA - North Dakota State University
item VEMURI, MADHAVA - North Dakota State University

Submitted to: ASABE Annual International Meeting
Publication Type: Abstract Only
Publication Acceptance Date: 3/19/2024
Publication Date: 3/19/2024
Citation: Kim, J.Y., Balamurugan, R., Tida, U., Vemuri, M. 2024. Weed identification using U-Net machine learning model and SAM segmentation. ASABE Annual International Meeting. https://doi.org/10.13031/aim.202401312.
DOI: https://doi.org/10.13031/aim.202401312

Interpretive Summary: This paper discusses an artificial intelligence (AI)-based classification approach to identify weeds in the images. There are many prior studies conducted for weed identification using machine learning algorithm but limited on separation of weeds from crops. Herbicide-resistant (HR) weeds are spreading and difficult to identify and control. We propose machine learning model U-Net to identify weeds including HR weeds by using weed image database that contains five different types of weeds (kochia, horseweed, ragweed, redroot pigweed, and waterhemp). To improve detection accuracy of the model, output mask images are obtained through Segment Anything Model (SAM) that is Meta AI to classify any object. The trained model is tested with UAV images to created a prescription weed map for site-specific weed control.

Technical Abstract: Weeds are unwanted plants that compete with crops for water and nutrients and can cause the yield loss in sugarbeet and other crop fields. Accurate classification of weeds and crops in the images is critically important to implement integrated weed control. Learning-based models are evaluated for weed identification. Good quality training dataset is essential for the development of accurate classification model. Towards this, Segment Anything Model (SAM) from Metal artificial intelligence (Meta AI) is used to generate accurate pixel-wise segmentation masks for weed identification task. Using this approach, pixel-wise training data for weed classification is obtained. We then develop a U-Net model by training the model with the generated dataset. With this approach, fine-grained accuracy of the weeds in a given image can be obtained. The trained model is validated and then tested with a dataset of UAV images. Successful weed identification ensures the accuracy of a prescription weed map for effective site-specific weed control.