Kasimati Aikaterini, Espejo-Garcia Borja, Vali Eleanna, Malounas Ioannis, Fountas Spyros
Natural Resources Management and Agricultural Engineering, Agricultural University of Athens, Athens, Greece.
Front Plant Sci. 2021 Jun 11;12:683078. doi: 10.3389/fpls.2021.683078. eCollection 2021.
The most common method for determining wine grape quality characteristics is to perform sample-based laboratory analysis, which can be time-consuming and expensive. In this article, we investigate an alternative approach to predict wine grape quality characteristics by combining machine learning techniques and normalized difference vegetation index (NDVI) data collected at different growth stages with non-destructive methods, such as proximal and remote sensing, that are currently used in precision viticulture (PV). The study involved several sets of high-resolution multispectral data derived from four sources, including two vehicle-mounted crop reflectance sensors, unmanned aerial vehicle (UAV)-acquired data, and Sentinel-2 (S2) archived imagery to estimate grapevine canopy properties at different growth stages. Several data pre-processing techniques were employed, including data quality assessment, data interpolation onto a 100-cell grid (10 × 20 m), and data normalization. By calculating Pearson's correlation matrix between all variables, initial descriptive statistical analysis was carried out to investigate the relationships between NDVI data from all proximal and remote sensors and the grape quality characteristics in all growth stages. The transformed dataset was then ready and applied to statistical and machine learning algorithms, firstly trained on the data distribution available and then validated and tested, using linear and nonlinear regression models, including ordinary least square (OLS), Theil-Sen, and the Huber regression models and Ensemble Methods based on Decision Trees. Proximal sensors performed better in wine grapes quality parameters prediction in the early season, while remote sensors during later growth stages. The strongest correlations with the sugar content were observed for NDVI data collected with the UAV, Spectrosense+GPS (SS), and the CropCircle (CC), during Berries pea-sized and the Veraison stage, mid-late season with full canopy growth, for both years. UAV and SS data proved to be more accurate in predicting the sugars out of all wine grape quality characteristics, especially during a mid-late season with full canopy growth, in Berries pea-sized and the Veraison growth stages. The best-fitted regressions presented a maximum coefficient of determination ( ) of 0.61.
确定酿酒葡萄品质特征的最常见方法是进行基于样本的实验室分析，这种方法既耗时又昂贵。在本文中，我们研究了一种替代方法，通过将机器学习技术与在不同生长阶段收集的归一化植被指数（NDVI）数据相结合，并采用当前精准葡萄栽培（PV）中使用的非破坏性方法，如近距离和遥感方法，来预测酿酒葡萄的品质特征。该研究涉及从四个来源获取的几组高分辨率多光谱数据，包括两个车载作物反射传感器、无人机（UAV）获取的数据以及哨兵 - 2（S2）存档图像，以估计不同生长阶段的葡萄树冠特性。采用了几种数据预处理技术，包括数据质量评估、将数据插值到100个单元格的网格（10×20米）以及数据归一化。通过计算所有变量之间的皮尔逊相关矩阵，进行了初始描述性统计分析以研究来自所有近距离和远程传感器的NDVI数据与所有生长阶段的葡萄品质特征之间的关系。然后将转换后的数据集准备好并应用于统计和机器学习算法，首先根据可用的数据分布进行训练，然后使用线性和非线性回归模型进行验证和测试，包括普通最小二乘法（OLS）、泰尔 - 森（Theil - Sen）和休伯回归模型以及基于决策树的集成方法。近距离传感器在季节早期对酿酒葡萄品质参数的预测中表现更好，而远程传感器在后期生长阶段表现更佳。在浆果豌豆大小和转色期，即树冠完全生长的中晚期，两年中使用无人机、光谱传感 + GPS（SS）和作物圈（CC）收集的NDVI数据与糖分含量的相关性最强。在所有酿酒葡萄品质特征中，无人机和SS数据在预测糖分方面被证明更准确，特别是在树冠完全生长的中晚期、浆果豌豆大小和转色生长阶段。拟合效果最好的回归模型的最大决定系数（ ）为0.61。
