Box-Behnken法冷鲜滩羊肉蛋白质的高光谱模型优化

高光谱成像技术是一种将成像与光谱相结合的新型无损检测技术，属于间接分析方法；光谱模型的建立非常关键，需综合考察各建模因素间的交互作用。应用Box-Behnken法设计响应面试验优化冷鲜滩羊肉蛋白质含量的可见/近红外高光谱定量检测模型。使用可见/近红外高光谱成像系统采集冷鲜滩羊肉样本的高光谱图像，分析肉样反射光谱特性。采用二维相关光谱技术（2DCOS），以冷鲜滩羊肉中蛋白质含量为“外界扰动”，研究扰动条件下光谱信号的动态变化，解析二维相关光谱谱图特征，寻找与微扰相关的敏感变量。分别采用多元散射校正（multiplicative scatter correction，MSC）和标准正态变量变换（standard normalized variate，SNV）提取有用信号，优化所选特征波段光谱质量。为实现数据快速降维，减少大量光谱数据处理负担，采用变量组合集群分析法（variable combination population analysis，VCPA）和应用竞争性自适应加权算法（competitive adaptive reweighted sampling，CARS）对2DCOS范围内特征波段进行二次优选。根据Design-Expert软件中Box-Behnken法设计响应面试验，以特征优选、光谱预处理及多元校正方法为考察因素，各因素中3种不同方法为水平，建立冷鲜滩羊肉蛋白质含量分析的优化检测体系。结果表明，波长473，679，734和814 nm处存在较强的自相关峰，473~814 nm范围内的特征波段为冷鲜滩羊肉蛋白质检测的敏感区域；MSC和SNV能够消除肉样自身散射作用的干扰，CARS和VCPA对特征波段进行二次优选，分别优选出了16和9个特征波长；各因素对蛋白质可见/近红外光谱模型预测性能的影响顺序为特征优选方法>预处理方法>多元校正方法，优选出2DCOS-SNV-LSSVM模型具有较高的运行速率和预测能力，其R_c=0.858 8，RMSEC=0.005 8；R_p=0.860 4，RMSEP=0.005 7。研究表明，Box-Behnken法在可见/近红外高光谱（400~1 000 nm）建模参数优化选择中的应用，可以有效地实现滩羊肉品质智能监控与质量安全快速无损分析，为分析对象光谱模型的优化及提高预测结果的准确性提供理论参考。

Hyperspectral Model Optimization for Protein of Tan Mutton Based on Box-Behnken

Hyperspectral imaging is a new non-destructive testing technology which combines imaging and spectrum.It is an indirect analysis method.The establishment of the analytical model is critical,which needs to comprehensively consider the interaction among various modeling factors.This paper aimed to investigate the optimization of visible/near-infrared hyperspectral quantitative detection model for protein content in chilled Tan mutton based on the Box-Behnken design.The hyperspectral images of meat samples were collected by the visible/near-infrared hyperspectral imaging system.The reflectance spectral characteristics of chilled Tan mutton were analyzed.The protein contents were regarded as an external disturbance.The dynamic change of spectral signal was studied by two-dimensional correlation spectra under disturbance conditions.The synchronization spectra and autocorrelation spectra were analyzed to find the sensitive variables related to protein contents.Multiplicative scatter correction(MSC)and standard normalized variate(SNV)were used to extract useful signal and optimize the spectral quality of selected characteristic bands.In order to achieve data dimensionality reduction and reduce the burden of processing a large number of spectral data,competitive adaptive reweighted sampling(CARS)and variable combination population analysis(VCPA)were used to perform secondary extracted feature wavelengths.Extraction method,spectral pretreatment and multivariate calibration methods were factors,and each factor had 3 different levels.The response surface experimental design was used to build an optimal detection system for protein content analysis of chilled Tan mutton.The results indicated that there were strong autocorrelation peaks at 473,679,734 and 814 nm.The feature bands in the range of 473~814 nm were a sensitive area of protein detection in mutton.MSC and SNV could effectively eliminate the interference of scattering.Sixteen and nine characteristic wavelengths were selected by CARS and VCPA from 2 DCOS,respectively.The factors in descending order affecting the predictive performance of the model were detection band,preprocessing method and modeling method.The 2 DCOS-SNV-LSSVM model was selected with a high operating rate and prediction capability(R_c=0.8588,RMSEC=0.0058;R_p=0.8604,RMSEP=0.0057).The results showed that the application of the box-behnken method in the optimization of visible/near-infrared hyperspectral(400~1000 nm)modeling parameters could effectively realize the intelligent monitoring and fast non-destructive analysis of Tan mutton quality.It could also provide a theoretical reference for the optimization of the model and improving prediction accuracy.