基于t-SNE法对寒地水稻种质资源品质的分类研究

Quality Classification of Cold-land Rice Germplasm Resources Based on the t-SNE Algorithm

为探究数据降维后可视化呈现并进行图像分割聚类的方法,本研究利用160份粳稻种质资源的品质指标,引入t-SNE算法将11项品质指标转化为两个相互独立的成分并进行可视化呈现,进一步采用K-means聚类,将图像聚类分割为具有不同特征的三个簇,分别利用非线性对数主成分分析,根据每个指标在主成分中平均贡献率进行排序,得到具有不同品质特征的三个类别。结果表明,160份寒地水稻种质资源被分成具有不同特征的簇0(62份资源)、簇1(47份资源)、簇2(51份资源),其中簇0强调蛋白质、直链淀粉等营养品质,簇1注重光泽、味道和香气等感官品质,而簇2更加突出糙米率、精米率和完整性等加工品质特征。本方法解决了聚类时数据维度过高、信息冗余等问题,并较完整地保留了高维数据的分布特征,而非线性对数主成分分析又可以进一步明晰每一类的特征。

In order to explore the method of visual presentation of data after dimensionality reduction and image segmentation clustering,this study used the quality indicators of 160 japonica rice germplasm resources,introduced t-SNE algorithm to convert 11 quality indicators into two independent components and perform visual presentation,and further adopted K-means clustering to divide the image cluster into three clusters with different characteristics,and used nonlinear logarithmic principal component analysis to sort each index according to its average contribution rate in the principal component to obtain three categories with different quality characteristics.The results showed that 160 rice germplasm resources were divided into cluster 0(62 resources),cluster 1(47 resources)and cluster 2(51 resources)with different characteristics,of which Cluster 0 emphasized nutritional qualities such as protein and amylose,cluster 1 focused on sensory qualities such as gloss,taste and aroma,and cluster 2 highlighted processing quality characteristics such as brown rice rate,white rice rate and integrity.This method solved the problems such as too high data dimension and information redundancy in clustering,and preserved the distribution characteristics of high dimensional data completely,and the nonlinear logarithmic principal component analysis could further clarify the characteristics of each class.