Skip to main content
ARS Home » Southeast Area » Poplarville, Mississippi » Southern Horticultural Research Unit » Research » Research Project #445438

Research Project: Expanding Southern Highbush Blueberries To Underserved Regions of Southeastern U.S.

Location: Southern Horticultural Research Unit

Project Number: 6062-21000-011-006-R
Project Type: Reimbursable Cooperative Agreement

Start Date: Oct 1, 2023
End Date: Aug 31, 2026

I. Develop climate-resilient southern highbush blueberry cultivars for small- and mid-sized growers in Alabama and nearby regions through collaborative cultivar evaluation. II. Evaluate genotype-by-environment interactions of key traits to better allocate future cultivars to the appropriate environments, III. Enable accurate and high-throughput yield phenotyping through image-based analysis and machine learning.

Evaluate 30 advanced selections and 8 checks of southern highbush blueberry in central and south Alabama. Plants will be evaluated for dates of 50% bloom, dates of 50% fruit ripening, the ratio of flower buds damaged by spring frost (in the occasion of frost events), 50-berry weight (g), average berry diameter (mm), berry yield per plant (g), disease score (1-5), Brix (%), titratable acidity (TA, %), and firmness (g/mm) for three years. Data from all five locations will be integrated through pedigree and common checks to analyze breeding values for all environments. Per-plant yield data and image data will be collected to improve the accuracy of the existing deep-learning model for yield prediction. Berry count, yield, and other parameters such as genotype, plant age, and estimated canopy density will be used to train machine learning models for yield prediction, to better account for occlusion effects. The best yield prediction model will be implemented in a smartphone app to allow near real-time berry detection and yield prediction in the field and potentially other production systems. Images will be taken from two sides of a plant to test and improve the current yield prediction model. Field images will be manually annotated with the open-source image annotation software COCO Annotator to create a custom blueberry dataset for deep-learning research. Pretrained Mask R-CNN models and our 2022 data in Detectron2 (Facebook AI Research) will be further improved with the 2023 dataset to detect individual berries and estimate their maturity levels. Improved berry detection models will be applied to process single-plant images collected. Estimated berry count based on deep learning models, will then be used to train yield prediction models on the cultivars and selections in this multi-environment trial. Statistical models and machine learning models will be evaluated to predict the ground truth yield based on estimated berry count, average berry weight, plant age, genotype, and visually estimated canopy density to better account for occlusion effect. A smartphone app will be developed for blueberry yield prediction in the field setting. The generic object detection model will be replaced by a blueberry detection and yield prediction model developed from this study.