基于近红外光谱和机器学习的大豆种皮裂纹识别研究

Identification of Soybean Seed Coat Crack Based on Near Infrared Spectroscopy and Machine Learning

针对目前大豆种皮裂纹检测主要依靠人工、检测效率低、误差大的问题,提出一种基于近红外光谱技术和机器学习的大豆种皮裂纹自动识别方法。采用FTNIR光谱仪采集150粒大豆样品(裂纹大豆75粒,正常大豆75粒)的近红外光谱,采用原始光谱、标准正态变量变换(Standard normal variate,SNV)、多元散射校正(Multiple scatter correction,MSC)、一阶导数结合SG平滑、二阶导数结合SG平滑等5种方法对获得的光谱进行预处理,分别采用偏最小二乘判别分析法(Partial least squares discriminant analysis,PLSDA)、k-近邻法(k-nearest neighbor,KNN)、支持向量机法(Support vector machine,SVM)、随机森林法(Random forest,RF)、随机梯度提升法(Stochastic gradient boosting,SGB)、极端梯度提升法(Extreme gradient boosting,XGBoost)等6种机器学习方法建立了大豆种皮裂纹识别模型,研究了不同光谱预处理方法对6种机器学习方法分类效果的影响,对比分析了不同建模方法的分类效果。结果表明,光谱预处理方法对不同机器学习方法的分类效果差别较大。在合适的光谱预处理条件下,6种不同的机器学习算法的验证集准确率均不低于80.00%。PLSDA的分类效果最好,验证集最优准确率达到90.00%;XGBoost的分类效果次之,验证集最优准确率达到86.67%,接下来依次是SVM、KNN、SGB和RF。利用近红外光谱技术和机器学习方法识别大豆种皮裂纹是可行的,在原始光谱条件下,PLSDA是大豆种皮裂纹识别的最佳方法。

At present,the detection of soybean seed coat crack mainly depends on visual inspection,which has low detection efficiency and large error,a method for automatic identification of soybean seed coat cracks based on near infrared spectroscopy and machine learning was proposed.The near infrared spectra of 150 soybean samples(75 cracked and 75 normal)were collected by FTNIR spectrometer.The original spectra,standard normal variable(SNV),multiple scatter correction(MSC),the first derivative and the second derivative with SG smoothing were used to process the obtained spectra.Then partial least squares discriminant analysis(PLSDA),k-nearest neighbor(KNN),support vector machine(SVM),random forest(RF),stochastic gradient boosting(SGB)and extreme gradient boosting(XGBoost)were used to establish soybean seed coat crack identification models.The effects of different spectral preprocessing methods on the classification results of the six machine learning methods were compared and analyzed.Under the appropriate spectral preprocessing conditions,the accuracy of validation set of six different machine learning algorithms was not less than 80.00%.PLSDA had the best classification result,and the optimal accuracy rate of validation set reached 90.00%;the next was XGBoost,the optimal accuracy rate of validation set reached 86.67%,followed by SVM,KNN,SGB and RF.The results showed that near infrared spectroscopy combined with machine learning was feasible to identify soybean seed coat cracks,and PLSDA was the best method to identify soybean seed coat cracks under the original spectral conditions.The research result can provide a method for automatic identification of soybean seed coat cracks.