Research On Wavelength Selection Method In Near-Infrared Spectra And Its Application For The Rapid Detection Of Freshness In Tilapia Fillets

Near-infrared(NIR)spectroscopy has been a maturely fast and nondestructive determination technique which has been widely applied in the fields of food.NIR hyperspectral imaging(HSI)technique is the combination of imaging and spectroscopy technique,which could gain a three-dimensional image by capturing not only the spectral information of samples but image information of corresponding wavelengths.The key achieving rapid quality determination is to build the calibration model between the attribute parameter and spectral information of food.Thus,a series of chemometrics methods such as outlier elimination,spectral preprocessing and variable(wavelength)selection algorithms could be used to optimize the calibration model continuously.Tilapia fillets were investigated in this study.A portable NIR spectrometer and HSI systems were applied to acquire the spectral information of tilapia fillets and then combined with chemometrics,image processing methods and self-proposing variable selection strategy to explore the variation of between spectral information and traditional freshness indicators.Moreover,the quantitative model and chemical information visualization of total volatile basic nitrogen(TVB-N)was also studied.The whole work could provide technical support for rapidly and nondestructively detecting tilapia fillet freshness.The results were listed as follows.Three-step hybrid strategy was proposed for wavelength selection to improve the prediction performance of NIR spectral quantitative models.Hundreds of wavelength points are usually included in NIR spectral data which could exist many noisy and interfering wavelengths.Wavelength selection algorithms could be employed to conduct wavelength selection for the elimination of uninformative and interfering variables and improvement of predictive accuracy of the calibration model.Three-step hybrid strategy could be used to solve the high dimensional issue in spectral data by three stages containing rough selection,fine and optimal selection.In the first stage,rough selection applying interval selection method(i PLS)remained several informative intervals and greatly reduced the number of variables,which reached the effect of rough selection.In the second stage,fine selection utilized wavelength point selection methods(VIP and m VCPA)to finely select important variables from the remained variables in the first stage and continuously shrink variable space.In the third stage,optimal selection using optimization algorithms(GA and IRIV)kept optimizing variable space established in the previous step and gained a small and optimal variable subset.When NIR spectral beer datasets that were published on the internet were employed to evaluate the performance of three-step hybrid strategy,the quantitative model of four three-step hybrid methods(i PLS-VIP-GA,i PLS-VIP-IRIV,i PLS-m VCPA-GA and i PLS-m VCPA-IRIV)performed better than that of corresponding single methods as well as two-step hybrid methods.A portable NIR spectrometer combined with chemometrics methods were employed to establish the quantitative model for the prediction of tilapia fillet freshness.After the three steps(including outlier elimination,spectral preprocessing and wavelength selection)processing spectral data were implemented respectively,model performance was improved greatly.Results indicated that,among all spectral preprocessing methods,the hybrid method of smoothing with 3 of window size-first-order derivative(SM3-D1)performed best.The models built by three-step hybrid methods had better prediction performance.In addition,i PLS-m VCPA-IRIV obtained the optimal model,with R~2p and RMSEP of 0.9201 and2.0907,respectively.HSI systems(Visible-NIR(Vis-NIR)and NIR)combined with data fusion analysis were employed to build the quantitative model for the prediction of tilapia fillet freshness.With two single data blocks of Vis-NIR and NIR,the model optimized by the wavelength selection methods of GA and IRIV obtained the best predicted performance,respectively.In comparison with single data blocks,corresponding wavelength selection methods with low-level fusion data formed by simply jointing Vis-NIR and NIR data blocks,gained better prediction results.Principal components(PC)and wavelength extraction was conducted to create mid-level fusion data.The model which was built on the strength of first ten principal components performed better than single data blocks.The model built on extracted wavelengths performed better than single Vis-NIR data block,but worse than single NIR data block.In addition,the model on mid-level fusion data established by CARS possessed best predicted ability,with R~2p and RMSEP of 0.8813 and 2.1954.Finally,visualization distribution of TVB-N contents,by which the degree of tilapia fillets freshness could be observed according to the variation of color,could be established using respective optimal models in Vis-NIR and NIR data.