基于多模型决策融合的苹果产地判别及糖度含量预测

Discrimination of Apple Origin and Prediction of SSC Based on Multi-Model Decision Fusion

苹果产地溯源与苹果糖度含量预测具有非常重要的现实意义,通过建立模型达到产地判别与糖度预测目的。为了克服单个模型的局限性,通过将多个模型的预测结果综合,提高整体预测性能。采用近红外光谱结合多模型决策融合策略对苹果产地进行溯源鉴别,对苹果糖度值进行预测,验证理论方法的可行性。采用手持式近红外检测仪采集了苹果样本的光谱,使用样本光谱结合随机森林(RF)方法、偏最小二乘判别分析(PLS-DA)与支持向量机(SVM)方法建立了苹果产地判别模型。再对三种判别模型输出的预测结果使用投票制决策融合方法,输出新的判别结果。对所有苹果样本采集了糖度实际值,使用样本光谱与糖度实际值结合随机森林(RF)、偏最小二乘回归(PLSR)与支持向量回归(SVR)方法建立了糖度预测模型。采用三种回归模型输出的结果,通过加权法决策融合策略输出新的糖度预测结果。在不使用投票决策方法时,三种定性建模方法中使用RF方法建立判别模型效果最好,预测准确度达到88.71%。使用SVM方法预测效果最差,预测准确度为77.43%。使用投票决策方法后,对苹果产地鉴别的准确度达到93.42%,其预测的精确度与召回率也达到了双高,均在85%以上。在不使用加权的决策融合方法前提下,三种定量建模方法对苹果糖度的预测均有不错的效果。三种方法预测的决定系数均约0.87,预测均方根误差均约为0.78。使用了加权的决策融合方法,对糖度的预测效果有一定的提升。预测决定系数为0.91,预测均方根误差为0.66。通过在苹果产地的鉴别与苹果糖度的预测中,使用多模型决策融合方法提高了苹果产地判别的正确率,提升了对苹果糖度预测的准确性,证实了所提方法的可行性。同时,手持式近红外检测仪结合多模型决策融合方法也为现场无损检测分析提供了一种新的高精度预测手段。

Traceability of apple origin and prediction of apple SSC is of great practical significance,and the purpose of origin discrimination and SSC prediction is achieved by modeling.To overcome the limitations of a single model,the overall prediction performance is improved by combining the prediction results of multiple models.Near-infrared spectroscopy(NIRS)detection technology combined with a multi-model decision fusion strategy is utilized for traceability identification of apple origin and prediction of apple SSC to verify the feasibility of the theoretical method.The spectra of apple samples were collected using a handheld near-infrared detector,and apple origin discrimination models were established using the sample spectra in combination with the random forest(RF)method,the partial least squares discriminant analysis(PLS-DA)method,and the support vector machine(SVM)method.The predictions from the three discrimination models are then used in a voting system decision fusion method to generate new discriminant results.Actual values of SSC were collected for all apple samples,and SSC prediction models were developed using the sample spectra and actual values of SSC combined with the random forest(RF)method,the partial least squares regression(PLSR)method,and the support vector regression(SVR)method.Using the outputs of the three regression models,the new SSC prediction is output through the weighting method decision fusion strategy.When the voting decision-making method was not used,the discrimination modeling using the RF method was the most effective among the three qualitative modeling methods,with a prediction accuracy of 88.71%.The worst prediction was made using the SVM method,with a prediction accuracy of 77.43%.After using the voting decision method,the accuracy of apple origin identification reached 93.42%,and its prediction precision and recall also reached a double high,both above 85%.All three quantitative modeling methods gave good results in predicting apple SSC without using the weighted decision fusion method.All three methods predicted coefficients of determination around 0.87 and root mean square errors of prediction(RMSEP)around 0.78.The prediction of the SSC level was improved after using the weighted decision fusion method.The prediction coefficient of determination was 0.91,and the RMSEP was 0.66.The feasibility of the proposed method was confirmed by using the multi-model decision fusion method in the identification of apple origin and the prediction of apple SSC to improve the accuracy of apple origin discrimination and the precision of the prediction of apple SSC.Meanwhile,the handheld NIR detector combined with the multi-model decision fusion method provides a new high-precision prediction approach for on-site non-destructive testing analysis.