Non-invasive grapevine inflorescence detection using YOLOv11 under field conditions

Abstract

Accurate and early yield estimation in vineyards is essential for the effective management of resources and informed decision-making in viticulture. Winemakers and vineyard managers rely on these predictions to optimize agricultural practices and efficiently manage winemaking processes (Laurent et al., 2021). Furthermore, more efficient management based on precise predictions contributes to the sustainability of the sector, helping to mitigate the effects of climate change and promoting a more responsible use of natural resources.

Traditionally, estimation techniques such as manual bunch counting or destructive sampling have been the primary tools. However, these practices, aside from being costly and labour-intensive, fail to capture the inherent spatial and temporal variability of vineyards (Martin et al., 2003).

In recent years, the development of emerging technologies has transformed precision viticulture, offering new opportunities for yield estimation. Computer vision (Nuske et al., 2014) and deep learning (Iñiguez et al., 2024) have proven to be effective tools for non-invasive crop monitoring, enabling the automated detection and counting of grape bunches.

The use of object detection algorithms has become increasingly prominent in grape bunch estimation, particularly at stages closer to harvest when bunches are fully developed (Iñiguez et al., 2024). For instance, Sozzi et al. (2022) showcased the effectiveness of various YOLO versions in detecting white grape bunches during these late phenological stages, achieving notable accuracy under controlled conditions.

However, as vines approach harvest, challenges such as leaf occlusion and overlapping bunches significantly complicate estimations, limiting the reliability of these methods in less ideal scenarios (Iñiguez et al., 2021). Many studies addressing these difficulties have relied on images captured under optimal conditions, where minimal foliage or occlusion ensures better bunch visibility, simplifying detection tasks (Santos et al., 2020). The phenology of grapevines plays a pivotal role in yield estimation, as conducting earlier sampling during stages with reduced foliage can mitigate occlusion and provide more timely and actionable predictions. These advancements highlight the potential to optimize vineyard management and decision-making throughout the growing season.