Acta Agrestia Sinica ›› 2026, Vol. 34 ›› Issue (11): 4172-4180.DOI: 10.11733/j.issn.1007-0435.2026.11.024

Previous Articles    

Biomass Prediction and High-Yield Germplasm Screening of Forage Oats Based on Machine Learning

LU Mao-jin1, ZHAO Jia-xin1, CHEN Wen1, ZHONG Zhi-hua1, SUN Cheng-xing1, MA Zi-chen1, DENG Bo-si-tian1, LIU Yi-bo1, MA Yu-wen1, KANG Ji-ping2, GUO Huan1, YAN Ming-ke1, WANG Ya-fang1   

  1. 1. College of Grassland Agriculture, Northwest A&F University, Yangling, Shaanxi Province 712100, China;
    2. Tianshui Institute of Agricultural Sciences, Tianshui, Gansu Province 741000, China
  • Received:2026-08-14 Revised:2026-09-18 Published:2026-11-04

基于机器学习法的饲用燕麦生物量预测及高产种质筛选

卢茂金1, 赵嘉欣1, 陈雯1, 钟志花1, 孙承星1, 马紫晨1, 邓博偲田1, 刘一菠1, 马钰雯1, 康继平2, 郭欢1, 闫明科1, 王亚芳1   

  1. 1. 西北农林科技大学草业与草原学院, 陕西 杨凌 712100;
    2. 甘肃省天水市农业科学研究所, 甘肃 天水 741000
  • 通讯作者: 王亚芳,E-mail:yafang.wang@nwafu.edu.cn
  • 作者简介:卢茂金(2002-),汉族,男,福建龙岩人,硕士研究生,主要从事牧遗传育种研究,E-mail:chenjialuooo@nwafu.edu.cn

Abstract: Forage oat (Avena sativa L.) is an important forage source in cool regions. During high-yield germplasm breeding, there is a conflict between biomass determination and the preservation of seed propagation materials. Therefore, a yield prediction model based on agronomic traits of forage oat at the milk-ripe stage was constructed. In this study, eighty-seven M4 lines, obtained after EMS chemical mutagenesis of the space mutation-bred cultivar ‘Zhongtian No. 4’, were used as materials. Six regression algorithms were employed to construct non-destructive prediction models for per-plant dry weight under a repeated nested cross-validation framework, and model-assisted screening of high-yield germplasm was evaluated. Results showed that the Support Vector Machine (SVM) model performed best, with a mean outer-repeat coefficient of determination (R2) of 0.91 and a consensus out-of-fold prediction R2 of 0.92. Permutation importance analysis ranked effective tiller number (relative importance 100.00), plant height (52.14), and stem diameter (19.42) as the leading predictors. Using per-plant dry weight ≥ 1.3 times the control as the threshold, a total of 26 high-yield lines were identified by measurement, with a mutation efficiency of 5.33‰; the model predicted 20 high-yield lines, with a precision of 1.00 and an overall agreement of 0.91. The prediction model realized accurate prediction of per-plant biomass in forage oats under non-destructive conditions, providing an efficient tool for the early screening of high-yield germplasm.

Key words: Forage oat, Machine learning, Biomass prediction, Non-destructive phenotyping, High-yield germplasm

摘要: 饲用燕麦(Avena sativa L.)是冷凉地区重要饲草来源。高产种质选育中,其生物量测定与留种材料保存之间存在矛盾,因此本研究构建了一种基于乳熟期饲用燕麦农艺性状的无损产量预测模型。以空间诱变育成的品种‘中天4号’经EMS化学诱变后获得的87个M4代株系为材料,采用6种回归算法在重复嵌套交叉验证框架下构建单株干重预测模型,评估模型筛选高产种质的准确性。结果表明,支持向量机模型最优,外层重复平均决定系数(R2)为0.91,共识折外预测R2为0.92。置换重要性分析显示,有效分蘖(相对重要性100.00)、株高(52.14)和茎粗(19.42)是生物量预测的核心因子。以单株干重≥对照均值1.3倍为标准,实测筛选出26个高产株系,诱变效率为5.33‰;模型预测筛选出20个高产株系,精确率为1.00,整体一致性为0.91。该模型可在无损条件下实现饲用燕麦单株生物量的精准预测,为早期高产种质的快速筛选提供了高效工具。

关键词: 饲用燕麦, 机器学习法, 生物量预测, 无损表型, 高产种质

CLC Number: