Palacios Fernando, Diago Maria P, Tardaguila Javier
Televitis Research Group, University of La Rioja, 26006 Logroño (La Rioja), Spain.
Instituto de Ciencias de la Vid y del Vino, University of La Rioja, CSIC, Gobierno de La Rioja, 26007 Logroño, Spain.
Sensors (Basel). 2019 Sep 2;19(17):3799. doi: 10.3390/s19173799.
Grapevine cluster compactness affects grape composition, fungal disease incidence, and wine quality. Thus far, cluster compactness assessment has been based on visual inspection performed by trained evaluators with very scarce application in the wine industry. The goal of this work was to develop a new, non-invasive method based on the combination of computer vision and machine learning technology for cluster compactness assessment under field conditions from on-the-go red, green, blue (RGB) image acquisition. A mobile sensing platform was used to automatically capture RGB images of grapevine canopies and fruiting zones at night using artificial illumination. Likewise, a set of 195 clusters of four red grapevine varieties of three commercial vineyards were photographed during several years one week prior to harvest. After image acquisition, cluster compactness was evaluated by a group of 15 experts in the laboratory following the International Organization of Vine and Wine (OIV) 204 standard as a reference method. The developed algorithm comprises several steps, including an initial, semi-supervised image segmentation, followed by automated cluster detection and automated compactness estimation using a Gaussian process regression model. Calibration (95 clusters were used as a training set and 100 clusters as the test set) and leave-one-out cross-validation models (LOOCV; performed on the whole 195 clusters set) were elaborated. For these, determination coefficient (R) of 0.68 and a root mean squared error (RMSE) of 0.96 were obtained on the test set between the image-based compactness estimated values and the average of the evaluators' ratings (in the range from 1-9). Additionally, the leave-one-out cross-validation yielded a R of 0.70 and an RMSE of 1.11. The results show that the newly developed computer vision based method could be commercially applied by the wine industry for efficient cluster compactness estimation from RGB on-the-go image acquisition platforms in commercial vineyards.
葡萄串的紧实度会影响葡萄的成分、真菌病害发生率和葡萄酒质量。到目前为止，葡萄串紧实度评估一直基于经过培训的评估人员进行的目视检查，在葡萄酒行业中的应用非常有限。这项工作的目标是开发一种新的非侵入性方法，该方法基于计算机视觉和机器学习技术的结合，用于在田间条件下通过实时采集红、绿、蓝（RGB）图像来评估葡萄串紧实度。使用一个移动传感平台，在夜间利用人工照明自动捕捉葡萄树冠层和结果区域的RGB图像。同样，在收获前一周的几年时间里，对三个商业葡萄园的四个红葡萄品种的195串葡萄进行了拍照。图像采集后，由15名专家在实验室按照国际葡萄与葡萄酒组织（OIV）204标准进行评估，以此作为参考方法来评估葡萄串紧实度。所开发的算法包括几个步骤，包括初始的半监督图像分割，随后是自动葡萄串检测以及使用高斯过程回归模型进行自动紧实度估计。构建了校准模型（95串葡萄用作训练集，100串葡萄用作测试集）和留一法交叉验证模型（LOOCV；在整个195串葡萄数据集上进行）。对于这些模型，在测试集上，基于图像的紧实度估计值与评估人员评分平均值（范围为1 - 9）之间的决定系数（R）为0.68，均方根误差（RMSE）为0.96。此外，留一法交叉验证得出的R为0.70，RMSE为1.11。结果表明，新开发的基于计算机视觉的方法可被葡萄酒行业商业应用，用于在商业葡萄园中通过RGB实时图像采集平台高效估计葡萄串紧实度。
