Counting grape bunches using deep learning under different fruit and leaf occlusion conditions

Introduction

Yield estimation is very important for the wine industry since provides useful information for vineyard and winery management. The early yield estimation of the grapevine provides information to winegrowers in making management decisions to achieve a better quantity and quality of grapes. In general, yield forecasts are based on destructive sampling of bunches and manual counting of berries per bunch and bunches per vine. This traditional approach does not provide accurate estimations because the sample of the vineyard cannot represent all the variability that may be present in the plot. These techniques are time-consuming, expensive, and labour-intensive (Martin et al., 2003). The number of bunches per vine is the most important of the yield components, explaining 60% of average field yield variability, while the number of berries per bunch explains 30% and berry weight explains 10% (Laurent et al., 2021). In this regard, precision viticulture has brought new opportunities for yield monitoring and prediction, taking advantage of the new sensors, platforms, and modelling techniques.

Nowadays one of the most common and successful techniques for monitoring the amount of fruit in viticulture has been computer vision. Several applications and methods have been reported in the scientific literature (Mohimont et al., 2022). Computer vision systems have been used to estimate grapevine yield at different phenological stages, such as budbreak (Liu et al., 2017), flowering (Palacios et al., 2020), pea-size (Palacios et al., 2022), and harvest (Xin et al., 2020). The computer vision techniques used for bunch detection are mainly classified into three classes: i) colour-based thresholding and colour features (Hacking et al., 2020), ii) active contour segmentation (Xiong, 2018), and ii) pixels segmentation (Íñiguez et al., 2021). In general, computer vision has shown good results for bunch detection; however, the results of these techniques are highly influenced by image acquisition conditions such as background effects and light conditions and intrinsic conditions of the grape canopies such as bunch occlusion (Íñiguez et al. 2021). In this context, new artificial intelligence techniques can help us to solve these problems. Deep learning methods have proved to be very effective in object detection (Fuentes, 2017). This novel technique has shown promising results for bunch detection and counting in grapevines (Sozzi et al., 2022).