Blueberry Yield Estimation Through Multi-View Imagery with YOLOv8 Object Detection

Abstract. Accurately estimating blueberry yields is crucial for farmers aiming to optimize crop management practices and enhance agricultural profitability. However, the inherent challenges posed by blueberries growing in clusters and being frequently occluded by leaves or other fruits make direct counting of individual blueberries nearly impossible. Existing approaches rely on sampling a few plants or clusters to estimate yields based on expert knowledge or employ indirect regression analysis using yield-related features as inputs for predictive models. With recent advancements in deep learning technologies, there has been a growing interest in leveraging machine vision techniques to directly count fruits for yield estimation. In this paper, we propose a novel approach for blueberry yield estimation utilizing multi-view imagery in conjunction with the state-of-the-art YOLOv8 object detection framework. Our methodology involves a customized mobile platform equipped with a multi-camera sensing system that captures images of blueberry plants from three distinct views (top, left, and right) to ensure comprehensive coverage. We train a YOLOv8x model as the detector to accurately detect and localize individual blueberries within the images. Accounting for the overlapping information from the three views, we employ a regression model to estimate the total number of blueberries per plant. To evaluate the effectiveness of our approach, we compare single-view and multi-view methodologies and assess their estimation performance on 12 individual blueberry plants with varying genotypes. The multi-view imagery approach demonstrates promising results, exhibiting a mean absolute percentage error of 24.6% and an R2 value of 0.77. These figures represent a substantial improvement of 5.2% to 15.7% when compared to single-view approaches. Additionally, leveraging the predicted bounding boxes of blueberries, we are able to generate density maps that facilitate further phenotyping analysis. The methodology presented in this study holds significant potential for accurately and autonomously estimating blueberry fields, enabling the generation of high-resolution yield density maps, even at the individual plant level, with the aid of mobile robots.