Abstract

The study aimed to evaluate the most predictive traits of fresh maize and the most appropriate multivariate approach for estimating silage fermentation quality. The use of near infrared (NIRs) instruments allowed rapid, accurate and cheap analysis. Samples of fresh maize plant (n = 822) from hybrids (Class Cultivar) of early and late classes, were harvested at three maturity stages: early, medium and late, in three areas (level input field) of ‘low’, ‘medium’ and ‘high’ soil fertility, along three consecutive years. Several algorithms of feature selection, regression, classification and machine learning, were tested. Maize silage fermentative quality was summarised through a Fermentative Quality Index (FQI). We found the most predictive traits as dry matter (DM), starch, and acid detergent lignin (ADL), with negative coefficients, or water-soluble carbohydrates (WSC) with a positive coefficient. FQI was significantly (p < 0.0001) affected by year (negatively for 2018), level input field (positively for high level) and maturity stage (negatively for the late harvest). The most satisfying results were attained using a stepwise regression algorithm (R² = 0.48), improved by the introduction of fixed effects (R² = 0.55) and partial least square discriminant analysis (PLS-DA), which was assessed through the Mattew Correlation Coefficient (MCC) in validation (MCC = 0.57). Concluding, among the tested approaches, the use of linear regression after stepwise algorithm or the use of PLS could be of practical help for the farmers to the effective management of the ensiling process of maize plants, even though environmental conditions should be considered to improve the predictions.

• HIGHLIGHTS

• The prediction of FQ at harvest would allow the farmer to tune up the ensiling process

• The prediction of FQ through the use of portable NIRs instruments was successful

• DM, starch and ADL were negatively related to FQ index

摘要

该研究旨在评估新鲜玉米最具预测性的性状和最合适的多变量方法来评估青贮发酵质量。使用近红外 (NIR) 仪器可以进行快速、准确和廉价的分析。鲜玉米植株样品 ( n = 822) 来自早晚班的杂交种（栽培品种），在三个成熟阶段收获：早、中、晚，在“低”、“中”和“高”土壤肥力三个区域（水平输入田） ，连续三年。测试了特征选择、回归、分类和机器学习的几种算法。玉米青贮发酵质量通过发酵质量指数 (FQI) 进行总结。我们发现最具预测性的特征是干物质 (DM)、淀粉和酸性洗涤剂木质素 (ADL)，系数为负，或水溶性碳水化合物 (WSC) 为正系数。FQI 显着 ( p < 0.0001) 受年份（对 2018 年不利）、水平输入田（对高水平有利）和成熟阶段（对晚收不利）的影响。最令人满意的结果是使用逐步回归算法 ( R ² = 0.48) 获得的，通过引入固定效应 ( R ² = 0.55) 和偏最小二乘判别分析 (PLS-DA) 进行改进，该算法通过 Mattew 相关性进行评估验证中的系数 (MCC) (MCC = 0.57)。结论是，在经过测试的方法中，使用逐步算法后的线性回归或使用 PLS 可以为农民有效管理玉米植株的青贮过程提供实际帮助，即使应考虑环境条件来改善预测。

• 强调

• 收获时 FQ 的预测将允许农民调整青贮过程

• 通过使用便携式 NIR 仪器预测 FQ 成功

• DM、淀粉和ADL与FQ指数呈负相关