For many horticultural crops, variation in quality (e.g., shape and size) contributes significantly to the crop’s market value. Metrics characterizing less subjective harvest quantities (e.g., yield and total biomass) are routinely monitored. In contrast, metrics quantifying more subjective crop quality characteristics such as ideal size and shape remain difficult to characterize objectively at the production-scale due to the lack of modular technologies for high-throughput sensing and computation. Several horticultural crops are sent to packing facilities after having been harvested, where they are sorted into boxes and containers using high-throughput scanners. These scanners capture images of each fruit or vegetable being sorted and packed, but the images are typically used solely for sorting purposes and promptly discarded. With further analysis, these images could offer unparalleled insight on how crop quality metrics vary at the industrial production-scale and provide further insight into how these characteristics translate to overall market value. At present, methods for extracting and quantifying quality characteristics of crops using images generated by existing industrial infrastructure have not been developed. Furthermore, prior studies that investigated horticultural crop quality metrics, specifically of size and shape, used a limited number of samples, did not incorporate deformed or non-marketable samples, and did not use images captured from high-throughput systems. In this work, using sweetpotato (SP) as a use case, we introduce a computer vision algorithm for quantifying shape and size characteristics in a high-throughput manner. This approach generates 3D model of SPs from two 2D images captured by an industrial sorter 90 degrees apart and extracts 3D shape features in a few hundred milliseconds. We applied the 3D reconstruction and feature extraction method to thousands of image samples to demonstrate how variations in shape features across SP cultivars can be quantified. We created a SP shape dataset containing SP images, extracted shape features, and qualitative shape types (U.S. No. 1 or Cull). We used this dataset to develop a neural network-based shape classifier that was able to predict Cull vs. U.S. No. 1 SPs with 84.59% accuracy. In addition, using univariate Chi-squared tests and random forest, we identified the most important features for determining qualitative shape type (U.S. No. 1 or Cull) of the SPs. Our study serves as a key step towards enabling big data analytics for industrial SP agriculture. The methodological framework is readily transferable to other horticultural crops, particularly those that are sorted using commercial imaging equipment.

对于许多园艺作物而言，质量（例如形状和大小）的变化会极大地影响作物的市场价值。常规监测表征较少主观收获量的指标（例如产量和总生物量）。相比之下，由于缺乏用于高通量传感和计算的模块化技术，量化量化主观作物质量特征（例如理想大小和形状）的指标仍然难以在生产规模上客观地进行表征。几种园艺作物在收获后被送到包装设施，在这里使用高通量扫描仪将它们分类到箱子和容器中。这些扫描仪捕获要分类和包装的每种水果或蔬菜的图像，但是这些图像通常仅用于分类目的，并被立即丢弃。通过进一步分析，这些图像可以提供有关作物质量指标在工业生产规模上如何变化的无与伦比的见解，并可以进一步了解这些特征如何转化为整体市场价值。当前，尚未开发使用通过现有工业基础设施生成的图像来提取和量化农作物质量特征的方法。此外，先前的调查园艺作物质量度量标准（特别是大小和形状）的研究使用的样本数量有限，没有包含变形的或不可销售的样本，也没有使用从高通量系统捕获的图像。在这项工作中，以Sweetpotato（SP）为用例，我们介绍了一种计算机视觉算法，用于以高通量方式量化形状和尺寸特征。这种方法从两幅2D图像中生成SP的3D模型，这些图像由相隔90度的工业分类器捕获，并在几百毫秒内提取3D形状特征。我们将3D重建和特征提取方法应用于成千上万个图像样本，以演示如何量化SP品种的形状特征变化。我们创建了一个SP形状数据集，其中包含SP图像，提取的形状特征和定性形状类型（美国1号或Cull）。我们使用该数据集开发了基于神经网络的形状分类器，该分类器能够以84.59％的准确度预测Cull与US No.1 SPs。此外，使用单变量卡方检验和随机森林，我们确定了确定SP的定性形状类型（美国1号或Cull）的最重要特征。我们的研究是实现工业SP农业大数据分析的关键一步。该方法框架易于转移到其他园艺作物，特别是那些使用商业成像设备分选的作物。