基于高光谱成像技术的红景天品种神经网络模式识别方法

Study on Neural Network Pattern Recognition Methods of Rhodiola Varieties Based on Hyperspectral Imaging Technology

采用高光谱成像技术采集大花红景天和狭叶红景天的近红外高光谱图像(935~1720 nm),并从中提取出感兴趣区域的平均光谱作为大花红景天和狭叶红景天的原始光谱.在采用多元散射校正(MSC)进行光谱预处理后,运用载荷系数法(X-LW)、连续投影算法(SPA)和竞争自适应重加权算法(CARS)分别提取了红景天近红外高光谱9、20和33个特征波长,最后基于全波长和特征波长建立偏最小二乘判别分析(PLS-DA)、概率神经网络(PNN)和广义回归神经网络(GRNN)分类判别模型.结果表明,基于CARS提取的特征波长建立的PLS-DA、PNN和GRNN模式识别模型优于基于X-LW、SPA提取特征波长建立的识别模型.而且,基于全波长和CARS提取的特征波长建立的PLS-DA、PNN和GRNN判别模型均能很好地区分大花红景天和狭叶红景天,对训练集和测试集样本分类的正确率全部达到100%.因此,高光谱成像技术结合PLS-DA与神经网络模式识别分析方法,能够实现大花红景天和狭叶红景天的无损、快速和准确的分类与鉴别,为红景天药材的质量控制、品种鉴别和临床应用奠定基础.

The near-infrared hyperspectral images(935-1720 nm)of Rholdiola crenulata and Rhodiola kirilowii samples were collected by hyperspectral imaging technology,and the average spectrum of the region of interest was extracted from the hyperspectral images as the original spectrum of the samples.After spectral preprocessing with the multiple scattering correction(MSC),using X-Loading Weights(X-LW),the successive projections algorithm(SPA),and the competitive adaptive reweight sampling method(CARS)extracted 9,20,and 33 characteristic wavelengths of Rhodiola near-infrared(NIR)spectra respectively.Finally,Partial least squares discriminant analysis(PLS-DA),probabilistic neural network(PNN),and generalized regression neural network(GRNN)pattern recognition models were established with full wavelength and characteristic wavelength.Results:PLS-DA,PNN,and GRNN pattern recognition models based on characteristic wavelength extracted by CARS is better than that based on characteristic wavelength extracted by X-LW and SPA.Moreover,PLS-DA,PNN,and GRNN discrimination models based on full wavelength and characteristic wavelength extracted by CARS can distinguish Rholdiola crenulata and Rhodiola kirilowii very well,and the classification accuracy of training sets and test sets could reached 100%.Conclusion:The established hyperspectral imaging technology combined with PLS-DA and neural network pattern recognition analysis method can realize the nondestructive,fast,and accurate classification and identification of Rholdiola crenulata and Rhodiola kirilowii,and lay the foundation for the quality control,variety identification and clinical application of Rhodiola.