Deep learning–based high-throughput phenotyping for tiller quantification in interspecific bentgrass hybrids using YOLOv8

Authors: FERM, DENNIS - Orise Fellow; KIM, YONGHYUN - Orise Fellow; Barnaby, Jinyoung

Submitted to: Frontiers in Plant Science
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 3/23/2026
Publication Date: 5/1/2026
Citation: Ferm, D.W., Kim, Y., Barnaby, J.Y. 2026. Deep learning–based high-throughput phenotyping for tiller quantification in interspecific bentgrass hybrids using YOLOv8. Frontiers in Plant Science. 17. Article 1810220. https://doi.org/10.3389/fpls.2026.1810220.
DOI: https://doi.org/10.3389/fpls.2026.1810220

Interpretive Summary: Grass tillers are the primary productive structures in many plants within the grass family. In bentgrass, tillers are produced in extremely high numbers, often reaching thousands per plant. Breeding programs that address tillering in bentgrass require counting of these tillers, which is time-consuming and labor-intensive. To address this, ARS scientists developed multiple methods for tiller detection and counting using artificial intelligence (AI) methods. These methods provided fast and accurate estimations of tiller numbers. This method allows breeders to use tillers as a yield metric in large-scale analyses of bentgrass species, which was previously difficult due to inefficiencies in labor. The data and methods from this study will aid in the genetic mapping of interspecific bentgrass populations, helping to identify genomic regions associated with potential yield improvements. This system is valuable to turfgrass breeders and researchers interested in yield metrics across large populations.

Technical Abstract: Grass tillers are the primary productive structures in many plants within the grass family Poaceae. In Agrostis grasses, commonly known as Bentgrass, tillers are produced in extremely high numbers, often reaching thousands per plant. Counting these tillers is both time-consuming and labor-intensive. To address this, we developed three methods for tiller detection and counting: an edge segmentation model using conventional computer vision algorithms with OpenCV, and two deep learning models based on You Only Look Once version 8 (YOLOv8) and Faster R-CNN (Regions with Convolutional Neural Networks). We evaluated these models using an interspecific Bentgrass population comprising 300 hybrids and two parent plants. Among the models, the YOLOv8-based CNN model (R² = 0.97) demonstrated the highest precision and accuracy in estimating tiller numbers, outperforming both the edge segmentation algorithm (R² = 0.86) and the Faster R-CNN model (R² = 0.85). Additionally, we assessed the accuracy of tiller detection based on tiller number groupings (large, medium, or small). For small-size samples (fewer than 150 tillers), both the YOLOv8-based CNN model and the edge segmentation algorithm yielded highly accurate results (R² = 0.97). However, for large-size samples (more than 400 tillers), the edge segmentation algorithm's accuracy dropped significantly (R² = 0.38), while the YOLOv8 model maintained strong performance (R² = 0.83). This study highlights the growing role of neural network technology in agricultural research, with ongoing advancements in methodologies aimed at improving efficiency and accuracy.