Grapevine inflorescence segmentation and flower estimation based on Computer Vision techniques for early yield assessment

Yield forecasting is of immeasurable value in modern viticulture to optimize harvest scheduling and quality management. The number of inflorescences and flowers per vine is one of the main components and their assessment serves as an early predictor, which can explain up to 85-90% of yield variability. This study introduces a sophisticated framework that integrates the benchmark of different advanced deep learning and classic image processing to automate the segmentation of grapevine inflorescences and the detection of single flowers, to achieve precise, early, and non-invasive yield predictions in viticulture. The YOLOv8n model achieved superior performance in localizing inflorescences ( F1-Score Box = 95.9%) and detecting individual flowers ( F1-Score = 91.4%), while the YOLOv5n model excelled in the segmentation task ( F1-Score Mask = 98.6%). The models demonstrated a strong correlation ( R 2 > 90.0%) between detected and visible flowers in inflorescences. A statistical analysis confirmed the robustness of the framework, with the YOLOv8 model once again standing out, showing no significant differences in error rates across diverse grapevine morphologies and varieties, ensuring wide applicability. The results demonstrate that these models can significantly improve the accuracy of early yield predictions, offering a non-invasive, scalable solution for Precision Viticulture. The findings underscore the potential for Computer Vision technology to enhance vineyard management practices, leading to better resource allocation and improved crop quality. • Robust dataset of grapevine inflorescences collected in real vineyard conditions. • Benchmarking of lightweight YOLO models for inflorescence detection and segmentation. • High accuracy in flower detection and inflorescence segmentation using YOLO models. • Models validated across varieties, ensuring generalisability and scalability.