Deep Learning-Based Early Yield Prediction via Automatic Inflorescence Detection in Vineyard

Yield estimation is a key component of vineyard management, as it allows for the prediction of production and, indirectly, offers valuable insights into grape quality. Traditional approaches, based on cluster counts, berry number, and berry weight, are labor-intensive, time-consuming, and subject to operator variability. Estimation at the flowering stage is particularly advantageous, since leaf occlusion is minimal; however, conventional survey methods remain inefficient. Based on this premise, this study presents a deep learning-based computer vision system for automated grapevine inflorescence detection and counting. The YOLOv11s architecture, implemented with Ultralytics, was fine-tuned on 1,628 annotated images from diverse viticultural conditions. Validation in the Cadriano experimental vineyard (Italy) achieved a mean Average Precision at IoU 0.5 (mAP50) of 0.86 and a recall of 0.79 under complex canopies. In the future, the integration of early-stage detection through computer vision with yield modeling holds strong potential for advancing early yield forecasting in viticulture. Moreover, the proposed system represents a scalable, accurate, and sustainable tool to support data-driven decision-making in vineyard management.