Bruce Robert W, Rajcan Istvan, Sulik John
Department of Plant Agriculture, University of Guelph, Guelph, ON, Canada.
Plant Phenomics. 2021 Aug 4;2021:9806201. doi: 10.34133/2021/9806201. eCollection 2021.
The accurate determination of soybean pubescence is essential for plant breeding programs and cultivar registration. Currently, soybean pubescence is classified visually, which is a labor-intensive and time-consuming activity. Additionally, the three classes of phenotypes (tawny, light tawny, and gray) may be difficult to visually distinguish, especially the light tawny class where misclassification with tawny frequently occurs. The objectives of this study were to solve both the throughput and accuracy issues in the plant breeding workflow, develop a set of indices for distinguishing pubescence classes, and test a machine learning (ML) classification approach. A principal component analysis (PCA) on hyperspectral soybean plot data identified clusters related to pubescence classes, while a Jeffries-Matusita distance analysis indicated that all bands were important for pubescence class separability. Aerial images from 2018, 2019, and 2020 were analyzed in this study. A 60-plot test (2019) of genotypes with known pubescence was used as reference data, while whole-field images from 2018, 2019, and 2020 were used to examine the broad applicability of the classification methodology. Two indices, a red/blue ratio and blue normalized difference vegetation index (blue NDVI), were effective at differentiating tawny and gray pubescence types in high-resolution imagery. A ML approach using a support vector machine (SVM) radial basis function (RBF) classifier was able to differentiate the gray and tawny types (83.1% accuracy and kappa = 0.740 on a pixel basis) on images where reference training data was present. The tested indices and ML model did not generalize across years to imagery that did not contain the reference training panel, indicating limitations of using aerial imagery for pubescence classification in some environmental conditions. High-throughput classification of gray and tawny pubescence types is possible using aerial imagery, but light tawny soybeans remain difficult to classify and may require training data from each field season.
准确测定大豆茸毛对植物育种计划和品种登记至关重要。目前，大豆茸毛是通过目视分类的，这是一项劳动强度大且耗时的工作。此外，三类表型（黄褐色、浅黄褐色和灰色）可能难以通过目视区分，尤其是浅黄褐色类别，经常与黄褐色发生误分类。本研究的目的是解决植物育种工作流程中的通量和准确性问题，开发一套区分茸毛类别的指标，并测试机器学习（ML）分类方法。对高光谱大豆地块数据进行主成分分析（PCA）确定了与茸毛类别相关的聚类，而杰弗里斯-马图西塔距离分析表明所有波段对茸毛类别可分性都很重要。本研究分析了2018年、2019年和2020年的航空图像。对2019年60个已知茸毛基因型的地块测试用作参考数据，而2018年、2019年和2020年的全场图像用于检验分类方法的广泛适用性。两个指标，即红/蓝比和蓝色归一化差异植被指数（蓝色NDVI），在高分辨率图像中能够有效区分黄褐色和灰色茸毛类型。使用支持向量机（SVM）径向基函数（RBF）分类器的ML方法能够在有参考训练数据的图像上区分灰色和黄褐色类型（基于像素的准确率为83.1%，kappa = 0.740）。测试的指标和ML模型在不含参考训练面板的年份图像中不能普遍适用，表明在某些环境条件下使用航空图像进行茸毛分类存在局限性。使用航空图像对灰色和黄褐色茸毛类型进行高通量分类是可能的，但浅黄褐色大豆仍难以分类，可能需要每个田间季节的训练数据。
