可见-近红外光谱的鸭梨黑心缺陷在线检测AdaBoost集成模型研究

Study on Online Detection Method of“Yali”Pear Black Heart Disease Based on Vis-Near Infrared Spectroscopy and AdaBoost Integrated Model

黑心病是鸭梨贮藏期间发生的生理病害,其病变初期表现在内部果核处出现褐色斑块,而在果实外观上与正常果几乎没有任何差异,严重影响鸭梨的贮藏时间和品质,亟需一种快速无损的检测方法为鸭梨质量保驾护航。采用可见-近红外光谱法对鸭梨黑心缺陷进行在线检测和识别,结合平滑(Smoothing)、标准正态变量变换(SNV)、多元散射校正(MSC)、SG一阶导数(SG 1 st-Der)以及小波变换(WT)预处理方法和主成分分析(PCA)、k近邻(kNN)、朴素贝叶斯(NBC)、支持向量机(SVM)以及基于Adaboost的集成学习等方法对鸭梨黑心病进行判别研究。Adaboost集成了kNN、NBC和SVM三个独立学习器。将120个健康鸭梨和165个黑心鸭梨共计285个样品划分为训练集和测试集进行模型的构建和评价,采用训练集的查准率/查全率的调和平均值(F-measure)和正确识别率(Accuracy)对分类模型进行优化和评价。研究结果表明:不同属性(正常和黑心)鸭梨样品光谱的前三主成分分布图相互交错,很难直观地对黑心鸭梨进行区分。样品光谱经小波变换(小波基为“Haar”)预处理的kNN模型训练集的F-measure和Accuracy分别为78.98%和82.62%;经过SG一阶导数预处理后的NBC模型训练集的F-measure和Accuracy分别为80.90%和82.11%;经过小波变换预处理后的SVM模型训练集的F-measure和Accuracy分别为90.24%和91.58%;经小波变换预处理的AdaBoost模型训练集的F-measure和Accuracy分别为91.46%和92.63%。通过测试集对模型进行验证可知:光谱经小波变换预处理后建立的Adaboost分类模型最优,分类的F-measure达到90.91%,较WT-kNN,SG 1 st-Der-NBC和WT-SVM模型分别提高了11.39%,15.23%和2.30%;Accuracy达到92.63%,分别提高了10.52%,11.58%和2.10%;模型对测试集样品预测时的计算时间约为0.12s,满足在线分选要求。可见-近红外光谱结合AdaBoost分类方法,可以为鸭梨黑心病的在线检测提供一种快速简便的分析方法。

Black heart disease is a physiological disease that occurs during the storage of“Yali”pears.The initial stage of the disease manifests itself in brown plaques on the inner core,but there is no difference in the appearance of the fruit from normal fruits,which seriously affects the storage time and quality of“Yali”pears.A fast and non-destructive testing method is urgently needed to escort the quality of“Yali”pears.The vis-near infrared spectroscopy method was used to explore the feasibility of online detection of“Yali”pear black heart disease,combined with principal component analysis(PCA),k-nearest neighbor(kNN),naive Bayes classifier(NBC),support vector machines(SVM),and integrated learning based on Adaboost modeling were used to identify“Yali”pear black heart disease.Standard normal variable(SNV),multiplicative scatter correction(MSC),Savitzky Golay first-derivative derivative(SG 1 st-Der)and wavelet transform(WT)were used to preprocess the spectra.Adaboost integrates three base learners:kNN,NBC and SVM.A total of 285 samples,including 120 normal pears and 165 black hearted pears,divided into the training set and test set for model construction and evaluation.The harmonic average of the precision/recall rate(F-measure)and accuracy were used to optimize and evaluate the classification model.The results of the study show that the first three principal components of the spectrum of the samples of different attributes(normal and black heart)“Yali”pears were interlaced with each other,and it was difficult to distinguish the black heart pears visually.The F-measure and accuracy of the training set of the kNN model,in which the spectra of the samples were preprocessed by wavelet transform(the wavelet basis is“Haar”),were 78.98%and 82.62%,respectively.The F-measure and accuracy of the training set of NBC model after the Savitzky Golay first-derivative pretreatment were 80.90%and 82.11%,respectively.The F-measure and accuracy of the training set of SVM model after the wavelet transform pretreatment were 90.24%and 91.58%,respectively.The F-measure and accuracy of the training set of AdaBoost model after the wavelet transform pretreatment were 91.46%and 92.63%respectively.By verifying the model through the test set,it can be known that:the Adaboost classification model after the wavelet transform pretreatment was the best,and the F-measure reached 90.91%,which was 11.39%,15.23%and 2.30%higher than that of WT-kNN model,SG 1 st-Der-NBC model and WT-SVM model,respectively.Accuracy reached 92.63%,improved by 10.52%,11.58%and 2.10%respectively.The calculation time of the model for the prediction of test set samples was about 0.12 s,which meets the requirements of online sorting.The combination of vis-near infrared spectroscopy and the AdaBoost classification method can provide a quick and easy analysis method for online detection of“Yali”pear blackheart disease.