高光谱成像技术的三文鱼多品质指标的预测与分布可视化研究

Prediction and Distribution Visualization of Salmon Quality Based onHyperspectral Imaging Technology

采用颜色、剪切力和K值评价冰鲜与冻融三文鱼的品质,利用高光谱成像技术结合化学计量学方法对三个品质指标进行预测,并讨论了不同波长选择算法所建模型的预测效果。准备不同冻融次数三文鱼样本,进行高光谱数据采集和品质指标真实值的测定。采用六种预处理方法减少光谱数据中暗电流以及噪声的干扰,采用竞争性自适应重加权算法(CARS)、区间变量迭代空间收缩法(iVISSA),iVISSA-CARS筛选出与待测指标相关的变量,通过比较三种波长选择算法筛选的特征变量所建偏最小二乘(PLS)模型的预测结果,优选出三个品质指标最佳的变量选择方法。结果表明1^(st)Der-CARS-PLS模型对颜色中的a^(*)预测效果最好,筛选出的51个变量建立模型的R_(c)和R_(p)分别为0.9316和0.9297,RMSECV和RMSEP分别为0.716和0.735;2 nd Der-CARS-PLS模型对剪切力的预测效果最好,筛选出的61个特征变量建立模型的R_(c)和R_(p)分别为0.8921和0.8873,RMSECV和RMSEP分别为0.67 N和0.80 N;模型N-CARS-PLS取得了K值最好的预测效果,筛选出的51个特征变量所建模型的R_(c),R_(p),RMSECV和RMSEP分别为0.9514,0.9500,1.33,1.53。说明CARS变量筛选方法能够有效提取与特征指标相关的变量,提高模型的预测性能。除此之外,特征变量筛选联合算法iVISSA-CARS-PLS对三个指标的预测也取得了较好的结果,对三个指标测试集的R p分别为CARS-PLS预测模型的97.48%,97.02%,98.98%,而所用变量数仅为CARS-PLS的60.78%,62.29%,60.78%,说明变量筛选组合算法极大的减少了建立模型所用的数据量。三个指标的CARS-PLS以及iVISSA-CARS-PLS模型取得的预测效果均高于iVISSA-PLS,说明对于三文鱼三个品质指标的预测,CARS波长点筛选策略优于iVISSA波段选择策略。将优选出来的PLS模型分别用于构建三个品质指标的可视化分布图,清楚的展示了不同冻融次数三个品质指标的大小以及空间分布。因此,高光谱成像技术结合化学计量学方法可以较好的表征三文鱼的品质指标,为三文鱼多品质指标的同时快速检测提供了部分理论参考。

In this study,color,shear force and K value was used to evaluate the quality of salmon with different freeze-thaw times,and then predicted by hyperspectral imaging technology combined with chemometric methods.Besides,the prediction performance of the PLS model developed with characteristic variables was compared and discussed to select the optimal variable selection method for color,shear force and K value.The prepared salmon samples with different freeze-thaw times were scanned and analyzed to obtain hyperspectral data and the true values of quality indicators(color,shear force,K value).Afterwards,six different pretreatment methods were used to reduce dark current and noise interference in the spectral data.The competitive adaptive reweighting algorithm(CARS),interval variable iterative space shrinkage approach(iVISSA),and iVISSA-CARS algorithms were applied to screen out variables related to the indicators to improve the prediction performance of the model.The optimal variable selection method was determined according to the prediction performance of the PLS model built by the characteristic variables screened by the three-wavelength selection algorithms.The result exhibited that the 1 st Der-CARS-PLS model developed by 51 characteristic variables related to a*possessed the best prediction with R c of 0.9316,R p of 0.9297,RMSECV and RMSEP of 0.72 and 0.74,respectively.Similarly,in shear force prediction,2 nd Der proved to be the best pretreatment method and 2nd Der-CARS-PLS model developed by 61 characteristic variables displayed the best prediction with R c of 0.8853,R p of 0.8609,RMSECV and RMSEP of 0.69 N and 0.90 N respectively.Besides,the N-CARS-PLS model built by 51 characteristic variables achieved the best predictive effect on K value and obtained R c of 0.9513,R p of 0.9460,RMSECV and RMSEP of 1.33 and 1.53,respectively.It indicates that CARS can effectively extract variables related to feature indicators and improve the prediction performance of the PLS model.Besides,the combined algorithm iVISSA-CARS-PLS also achieved a significant results in the prediction of the three indicators.The R p of the test set was 97.48%,97.02%and 98.98%of the CARS-PLS prediction model.In comparison,the number of variables used was only 60.78%,62.29%and 60.78%of CARS-PLS,indicating that the variable selection combined algorithm greatly reduces the amount of data.The CARS-PLS and iVISSA-CARS-PLS models of the three indicators show higher prediction performance than iVISSA,which indicates that the feature variable selection strategy of CARS is more advantages than iVISSA in predicting of the above three quality indicators of salmon.Using the optimized PLS model,the visual distribution map of salmon quality indexes with different freezing and thawing time was constructed in the form of a pseudo color images,which provided more detailed and intuitive information for understanding the quality of salmon.In general,the combination of hyperspectral imaging combined with chemometrics,can accurately and non-destructively determine the quality indicators in salmon.This study can provide the same theoretical reference for the simultaneous rapid detection of multiple quality indicators of salmon.