ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Multi-Class Instance Segmentation of Ornamental Vine Canopies: A Comparative Study of Modern Deep Learning Architectures for Field-Based Phenotyping

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2600651.(doi:10.13031/aim.202600651)
Authors:   Saroj Burlakoti, Utsav Bhandari, Aaron Etienne, Bruce Bugbee, Youping Sun, Wei Zhang, Alessandro Ossola, Shital Poudyal
Keywords:   Canopy cover, deep learning, instance segmentation, ornamental vines, precision horticulture, RGB imaging, YOLO

Abstract. Accurate estimation of canopy area is critical for optimizing irrigation, pest management, and overall production efficiency. Traditional approaches based on manual measurements (sampling and handheld sensors) or semi-automated image analysis (ImageJ) are labor-intensive and impractical at large-scale applications. Although deep learning (DL) models such as U-Net and DeepLabv3 have advanced agricultural image segmentation, they still struggle with irregular growth patterns, overlapping foliage, and environmental variability inherent in field-grown ornamental vines. In this study, we develop and compare deep learning models for multiclass instance segmentation of seven ornamental vine species using high-resolution RGB images. A total of 1,320 high-resolution RGB images were collected under field conditions at the Utah Agricultural Experiment Station and manually annotated in Roboflow; after augmentation and quality control, the final dataset comprised 2,307 images across eight classes (background plus seven vine species). Models evaluated included U-Net, DeepLabv3, SOLOv2, YOLO11-seg, and YOLO11-PGSS. Performance was assessed using precision, recall, F1-score, and mean Intersection over Union (mIoU), and canopy area estimates were validated against ImageJ measurements. SOLOv2 achieved the highest overall classification performance (Precision = 91.27%, Recall = 91.57%, F1-score = 91.42%), whereas YOLO11-PGSS achieved the highest segmentation quality (mIoU = 79.13%) and the most accurate canopy area estimates (MAE = 1,217 cm², RMSE = 1,895 cm², R² = 0.720). YOLO11-seg also provided competitive performance with real-time inference capability, while U-Net and DeepLabv3 were less robust under complex canopy conditions. The results demonstrate that modern instance segmentation architectures can accurately delineate vine canopies and distinguish species, providing a scalable alternative to manual canopy area estimation workflows in ornamental vine production systems.

(Download PDF)    (Export to EndNotes)