长江中下游粳稻食味品质综合评价方法研究  

Study on comprehensive evaluation method of Japonica rice eating quality in the middle and lower reaches of Yangtze River

作　　者：卞金龙 许方甫 刘国栋 徐栋 朱盈 胡群 张洪程[1] 魏海燕[1] BIAN Jinlong;XU Fangfu;LIU Guodong;XU Dong;ZHU Ying;HU Qun;ZHANG Hongcheng;WEI Haiyan(College of Agriculture,Yangzhou University,Yangzhou 225009,China;Erperimental Farm,Yangzhou University,Yangzhou 225000,China)

机构地区：[1]扬州大学农学院,江苏扬州225009 [2]扬州大学实验农牧场,江苏扬州225000

出　　处：《扬州大学学报（农业与生命科学版）》2023年第5期1-11,共11页Journal of Yangzhou University：Agricultural and Life Science Edition

基　　金：国家重点研发计划项目(2022YFD2301401);山东省重点研发计划项目(2021LZGC020-03)。

摘　　要：针对长江中下游稻区不同粳稻品种,以不同纬度多个地点的860份稻米样本为材料,基于11项食味品质指标,运用主成分分析、多元线性回归、偏最小二乘法、判别分析、分类回归树(classification and regression free,CART)决策树与反向传播(back propagation,BP)神经网络分别建立稻米食味品质等级预测模型,通过验证模型的准确性与稳定性,筛选出适合稻米食味品质综合评价的方法或模型。结果表明:BP神经网络模型的2次验证准确度与稳定性均最高,预测准确率分别为92.68%与92.31%;偏最小二乘法模型的2次预测准确率均在80%以上(分别为80.49%、87.18%),但2次验证结果相差较大(6.69%);判别分析与多元线性回归模型的平均预测准确率相近(分别为80.12%、78.77%),但判别分析模型的稳定性优于多元线性回归模型;主成分分析模型的平均预测准确率最低(67.32%),且2次验证结果差异也较大(8.94%);CART决策树模型的稳定性最差,2次验证准确率分别为53.66%与89.74%,相差达36.08%。因此,利用BP神经网络模型预测稻米食味品质等级具有较高的准确性与稳定性,可为长江中下游稻区稻米食味品质综合评价与优质食味水稻品种的筛选提供理论与方法支持。860 rice samples from various locations in the middle and lower reaches of the Yangtze River were used as materials in this study. Different comprehensive evaluation methods were used to establish the prediction model of rice eating quality, including principal component analysis, multiple linear regression, partial least squares method, discriminant analysis, classification and regression free(CART) decision tree and back propagation(BP) neural network, which were used to predict rice eating quality grade based on 11 indicators of rice eating quality. This article also clarified the applicability of different comprehensive evaluation methods in rice eating quality evaluation. By verifying the accuracy and stability of the model, the method or model suitable for comprehensive evaluation of rice taste quality was selected. The results showed that the accuracy and stability of BP neural network model were the highest, and the prediction accuracy was 92.68% and 92.31%, respectively. The accuracy of two predictions of partial least square model was above 80%(80.49%, 87.18%), but there was a big difference between the two verification results(6.69%). The average prediction accuracy of discriminant analysis and multiple linear regression model was similar(80.12% and 78.77%), but the stability of discriminant analysis model was better than that of multiple linear regression model. The average prediction accuracy of principal component analysis model was the lowest(67.32%), and the difference between the two verification results was also large(8.94%).The stability of CART decision tree model was the worst,and the accuracy of two validations was 53.66%and 89.74%,respectively,with a difference of 36.08%.Therefore,it was more accurate and stable to use BP neural network model to predict rice eating quality grade.The results provide theoretical and methodological support for comprehensive evaluation of rice eating quality and selection of high eating quality in japonica rice varieties in the middle and lower reaches of the Y

关 键 词：水稻 稻米食味品质 BP神经网络 主成分分析 多元线性回归 偏最小二乘法 判别分析 CART决策树 综合评价 

分 类 号：S511[农业科学—作物学]