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

Image-based blueberry yield prediction by integrating object detection and vision foundation models

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

Citation:  2026 ASABE Annual International Meeting  202601368.(doi:10.13031/aim.20202601368)
Authors:   Puranjit Singh, Jeffrey Neyhart, John H. Price, Sushan Ru, Yin Bao
Keywords:   Blueberry yield prediction, vision foundation models, deep learning, machine learning, YOLO11.

Abstract. Accurate blueberry yield estimation is critical for optimizing crop management, improving profitability, and supporting breeding decisions; however, traditional sampling-based approaches are labor intensive and lack scalability. This study presents a multi-model framework that integrates computer vision-based deep learning and machine learning techniques for blueberry detection, maturity assessment, and yield prediction across diverse genotypes and growing environments. A maturity-based blueberry detection model was developed using YOLO11 models trained on 8,024 annotated images collected over three growing seasons (2022-2024) from research and commercial farms in Alabama, Georgia, and New Jersey. The YOLO11-l model achieved best detection performance with a precision of 0.813, recall of 0.786, and mAP@0.5 of 0.846 for maturity-based visible blueberries detection. To estimate Genotype-specific yield rank, vision foundation models including DepthAnythingV2 and the DINOv3 Vision Transformer were employed to segment foreground images and extract corresponding feature embeddings, respectively. These features were then used to train machine learning models, achieving up to 80% accuracy in distinguishing Genotype-specific yield rank across different cultivars in our dataset. The yield prediction models were subsequently trained using combinations of image-derived berry counts, average berry weight, and genotype or vision-based genotype specific yield rank, with the highest accuracy achieved when genotype information was included. Notably, yield prediction models using genotype-specific yield rank showed comparable performance, indicating that genotypic differences reflected in plant architecture can be effectively captured through vision foundation model embeddings. This integrated framework demonstrates the potential of combining deep learning, vision foundation models, and machine learning to develop scalable decision-support tools for blueberry producers and breeders, with broader applicability to other fruit crops.

(Download PDF)    (Export to EndNotes)