Quantitative detection of restructured steak adulteration based on hyperspectral technology combined with a wavelength selection algorithm cascade strategy

Abstract

This study aimed to create simplified models to rapidly and non-destructively predict the content of adulterated meat in restructured steak based on hyperspectral technology. The hyperspectral data for restructured steaks mixed with different proportions of pork and duck were collected, and then six pre-treatment methods were used to pre-process the spectral data. Importantly, the wavelength selection algorithm cascade strategy, that is, combined the wavelength range selection and wavelength band selection methods including successive projections algorithm (SPA), competitive adaptive reweighted sampling (CARS), variables combination population analysis (VCPA), interval random frog (iRF), iRF-SPA, iRF-CARS, and iRF-VCPA, were employed to establish a partial least squares (PLS) prediction model for adulterated content. The results showed that the best pre-treatment methods for beef adulterated with pork and duck were Mean centering (MC) and Savitzky-Golay (SG) respectively, and the corresponding best wavelength selection method was iRF-CARS.

Introduction

In recent years, the imports and consumption of steaks in China have increased significantly with the spread of Western food culture. The restructured steak is a conditioned meat product made from minced beef as the main raw material and added with food additives and auxiliary materials, which is widely favored by consumers because of its delicious taste, lower price than raw cut steak, and rich nutrients including protein, blood iron, vitamin B and 8 essential amino acids. However, some unscrupulous traders mix low quality, cheaper meat, such as pork and duck, in the beef to make huge profits during the production process (Saranwong et al., 2013). This behavior not only seriously harms the interests of consumers, but also triggers problems for some religious believers. Therefore, rapid, reliable, and non-destructive methods are urgently needed for adulteration detection of restructured steak.

Traditional methods including proteomics analysis (Mandli et al., 2018) and DNA analysis (Prusakova et al., 2017) have been used to identify the adulteration of meat products. Although these methods have accurate measurement results, they are expensive, cumbersome, time-consuming, and destructive to samples, could not satisfy the requirement of a fast, simple, and sensitive process for detecting adulteration of meat products. In recent years, hyperspectral technology has been increasingly utilized for the detection of adulteration of meat because of its speed, non-destructive nature, and high accuracy. Ropodi et al. (2015) and Kamruzzaman et al. (2015) successfully detected adulteration by pork and horse meat in beef using hyperspectral technology combined with chemometric methods. Garrido-Novell et al. (2018) used hyperspectral technology combined with partial least squares discriminant analysis (PLS-DA) to construct a fusion model of spectra and texture information to identify pork, poultry and fish proteins, and the final recognition rate reached 92%. However, there are still no previous studies reporting the feasibility of using hyperspectral technology to detect adulteration of reconstituted steaks.

The obtained hyperspectral data contains the spectral information in the continuous wavelength range, and the information contained between different wavelengths has collinearity and overlap, resulting in high data dimensions and more redundancy. Therefore, wavelength selection methods are necessary to pick out useful, contributing wavelengths and reduce the influence of noise and other interference wavelengths in the partial least squares (PLS) model, which can simplify the prediction model and improve its accuracy. Previous studies found that using wavelength selection methods in series can yield improved modeling results than using them alone. Li et al. (2020) demonstrated that a combination of Monte Carlo-uninformative variable elimination-successive projections algorithm (MC-UVE-SPA) to predict the firmness of multiple varieties of pears was superior to each method used alone, and MC-UVE-SPA with the least squares-support vector machine (MC-UVE-SPA-LS-SVM) model was the best among all developed models. Li et al. (2017) used SPA to select the wavelength bands of Particle Swarm Optimization (PSO) and greatly reduced the number of wavelength bands while ensuring that the prediction accuracy was not significantly decreased. Therefore, this paper focuses on the combination of wavelength interval selection and wavelength band selection methods to obtain a prediction model with less wavelength variation and high accuracy.

Currently, pork and duck are the two most common ingredients used to adulterate restructured steak. Therefore, we employed a hyperspectral system to collect the spectral information of restructured steak mixed with different proportions of pork and duck, and the spectral information was pre-processed to establish a PLS prediction model. Because of the huge amount of hyperspectral data, effective selection of characteristic wavelengths to improve the efficiency and accuracy of the model was essential for real-world application of hyperspectral technique. variable selection can eliminate harmful variables dominated by noise, greatly reduce the amount of calculations, making the model simpler and more accurate. In addition, we can develop a multi-wavelength selection system through the selected wavelength to realize the exploration from the laboratory to the practical application. Thus, this study utilized the wavelength selection algorithm cascade strategy, that is, combined the wavelength range selection and wavelength band selection methods to establish a partial least squares (PLS) prediction model, including the successive projections algorithm (SPA), competitive adaptive reweighted sampling (CARS), variables combination population analysis (VCPA), interval random frog (iRF), iRF-SPA, iRF-CARS, and iRF-VCPA, to reduce model variability and improve accuracy. Between them, the wavelength interval selection method uses a set of neighboring wavelengths as the basic unit for wavelength screening, and final selected wavelength is a combination of several sets of continuous wavelength intervals. In contrast, the wavelength band selection method has fewer wavelength selections because a single wavelength is used as the basic unit.

Materials and Methods

Sample preparation    The steak raw materials and auxiliary materials, such as edible salt, white granulated sugar, and spices used in the experiment, were purchased from the Zhenjiang Metro supermarket. The food additives used in this experiment, including sodium bicarbonate, complex phosphates (sodium tripolyphosphate and sodium hexametaphosphate), TG enzyme, sodium caseinate, and carrageenan, were purchased from Henan Qianzhi Trading Co., Ltd.

The raw meat was trimmed, and the fascia and blood clots visible to the naked eye were removed before mincing. The auxiliary materials were weighed out in proportion to 1 kg of raw meat: 15 g of salt, 5 g of white sugar, 20 g of spices, 3 g of sodium bicarbonate, 3 g of complex phosphate, 3 g of TG enzyme, 8.5 g of sodium caseinate, 3 g of carrageenan, and 150 ml of water and mixed and dissolved as a pickling solution. The diced raw meat and marinade were placed in a vacuum tumbler for 1 h, drained of air by filling into a casing mold, refrigerated for 2 hours, and then frozen for 10 hours. After removal from the freezer, the meat was cut into 12-mm thick sheets using a bone saw and packed.

The restructured steak made entirely from beef was prepared first according to the aforementioned method, then the restructured steaks with different proportions of minced pork and duck minced meat were prepared separately. The adulterated meat accounted for 5%, 10%, 15%, 20%, or 25% of the total meat by weight. A total of 11 types of samples were obtained, with 30 samples in each type, yielding a total of 330 samples.

Hyperspectral system    The hyperspectral system used in this study, as shown in Figure 1, contains hardware and software components. The hardware includes the hyperspectral image spectrometer system consists of an imaging spectrometer (400–1100 nm, ImSpector V10E; Spectral Imaging Ltd., Oulu, Finland), a high performance CCD camera (Hamamatsu Photonics, Hamamatsu, Japan), 150W fiber halogen lamp (Fiber-Lite DC950 Illuminator, Dolan-Jenner Industries Inc, Boxborough, USA), three-axis precision electronically controlled translation stage (SC30021A, Zolix Instruments Ltd, Beijing, China), and computer. The software used was Spectral Cube (Spectral Imaging Ltd, Oulu, Finland). The entire system, except for the computer, was assembled in a dark chamber to minimize the effects of ambient light during the sample scanning.

Fig. 1.

Hyperspectral imaging system

Hyperspectral image data acquisition and calibration    Before experimenting, the spectrometer needed to warm up for approximately 30 minutes to remove any influence caused by baseline drift. We set the acquisition parameters through the Spectral Cube software: CCD camera exposure time was 45 ms, each scan had 1235 lines, with 1628 data points per line, corresponding to pixel sizes in the range of 431 to 962 nm, spectral wavelength interval was 0.858 nm, and a hyperspectral cubes of size 1628 pixel × 1235 pixel ×618 was finally obtained. The moving speed of the electronically controlled translation stage was 0.9 mm/s, and the fast forward displacement was 180 mm. At the time of collection, the restructured steak sample was placed on the electronically controlled translation stage, then we opened the translation stage device while clicking the save button, and the three-dimensional data cubes of the sample were obtained by scanning.

To reduce the influence of the uneven light intensity and dark currents on image quality in the acquisition system, flat-field correction was carried out to improve the accuracy of the prediction model (Sugiyama, 1999).

The spectral reflection intensity corresponding to the corrected hyperspectral image (R) was calculated using the following equation:   

where I is the spectral reflection intensity corresponding to the original hyperspectral image, B is the dark reference image, and W is the white reference image.

Then the corresponding absorbance value (A) was calculated according to the Beer's law:   

Spectral extraction and pre-treatment methods    Because there was a large area of background spectral information in the hyperspectral image, the effective area should be reasonably selected to extract the spectral signal of the sample, and the effective area was referred to in this article as the region-of-interest (ROI).

An ROI of 200 × 200 pixels near the center of the sample was selected by the ENVI software, and the spectral absorbance of all pixels in the ROI was extracted as the spectral data of this sample. Randomly divided the spectral data of 330 samples, 75% of which were partitioned into a training set and the remainder (25%) into a validation set.

Hyperspectral information was susceptible to environmental conditions and instrument operation during the acquisition process and therefore required pre-processing to reduce the effects of scattered light and noise (Rinnan et al. 2009). The pre-treatment methods used in this study were First derivative (1st Der), Second derivative (2nd Der), Mean centering (MC), Multiplicative scatter correction (MSC), Savitzky-Golay (SG), and Standard normal variate transformation (SNVT).

Selection method of characteristic wavelength    Hyperspectral data contained spectral information within the continuous wave band, and the information contained between different wavelengths overlapped and existed collinearly; therefore the data was high dimensional and redundant. When modeling with full-band spectral data, this redundant information increased the complexity, reduced the efficiency, and affected the prediction performance of the model. Therefore, algorithms were needed to select the wavelengths with higher correlation with the measurement index, which could simplify the prediction model and improve its accuracy.

The wavelength selection method can be divided into the following two parts. The wavelength interval selection method uses a set of neighboring wavelengths as the basic unit for wavelength screening. The final selected wavelength is a combination of several sets of continuous wavelength intervals, and this method is used to make a rough selection and remove wavelengths without losing information (Guo et al., 2019). In contrast, the wavelength band selection method uses a single wavelength as the basic unit. This method has received great attention due to its advantages such as fewer selection wavelengths and good modeling prediction effects (Yu et al., 2020). As each method has certain advantages and adaptability, a combination of these wavelength selection methods.

Hence, in this paper, three methods of wavelength band selection, SPA, CARS, and VCPA, were employed and combined with the wavelength interval selection method iRF for modeling.

Successive projections algorithm (SPA) SPA is a forward wavelength selection method that can minimize the collinearity between wavelengths and greatly reduce the number of wavelengths. To generate a wavelength combination, SPA first selected a wavelength band as a starting band, calculated its projection value on the remaining wavelength bands, and added the wavelength band with the largest projection value to this combination. The above steps were then repeated to finally obtain the wavelength combination with the minimum amount of redundant information (Cortes et al., 2017).

Competitive adaptive reweighted sampling (CARS)    CARS is a wavelength selection algorithm based on PLS regression coefficients. One part of the sample was randomly selected from the correction set by Monte Carlo sampling (MCS) to first establish a PLS model, and then the absolute value weight of the wavelength regression coefficient in this sampling was computed. Next, the exponentially decreasing function (EDF) was used to remove the wavelength bands with smaller absolute values, and the wavelength was competitively selected by the adaptive renewed samplings (ARS). After several iterations of the aforementioned process, we finally obtained a wavelength subset equal to the number of MCS sampling times and then built a PLS model and calculated the Root Mean Square Error of Cross-Validation (RMSECV). Eventually, the wavelength corresponding to the minimum value of RMSECV was selected as the characteristic wavelength of the spectral data (Guo et al., 2020).

Variables combination population analysis (VCPA)    VCPA is a novel and effective algorithm used to select the optimal variables based on Exponentially Decreasing Function (EDF) and Binary matrix sampling (BMS) in an iterative manner. EDF was first employed to select useful variables and eliminate variables that have a negative impact on the sample discrimination. Then, in each EDF run, BMS was used to select the variables from the variables space by considering the interaction effect among variables through random combinations. Next, the model population analysis (MPA) strategy was employed to select the optimal variable subset based on the RMSECV values from a large population of sub-models obtained by BMS. Finally, after N running times, the RMSECV of all the combinations among the remaining variables was calculated, and the subset with the lowest RMSECV value was chosen as the final modeling data (Ouyang et al., 2020).

Interval random frog (iRF) iRF is a new wavelength interval selection method based on Random frog (RF). This method simulates a reversible jump Markov Chain Monte Carlo (RJMCMC)-like strategy to iteratively calculate the selected frequency of each wavelength, establishes the PLS model for the wavelength with the highest selection possibility, and finally selects the characteristic wavelength with the smallest model error. The whole spectra were first divided into sub-intervals through moving windows of a fixed width, denoted as w. If the spectral data has p wavelength bands, p-w + 1 intervals can be obtained. Next, the importance of the interval combination was evaluated by the sum of the absolute regression coefficients of each wavelength band in the interval. The remaining steps were the same as for RF (Sun et al., 2020).

Quantitative prediction model    PLS can extract the effective information of data and solve the problem of variable collinearity and has been widely applied in spectral data modeling (Trivittayasil et al., 2018; Wold et al., 2001). In this paper, this model was chosen as the quantitative model. The evaluation indicators of the prediction effect of the PLS model were Correlation Coefficient of Calibration (R_C), Correlation Coefficient of Prediction (R_P), RMSECV, and Root Mean Square Error of Prediction (RMSEP). Generally, the closer R_C and R_P are to 1, and the closer the RMSECV and RMSEP are to 0, the higher performance of the predictive model.

Results and Discussion

Spectral data pre-processing    Figures 2 (A) and (B) show the original spectra and the average spectra of restructured steak samples adulterated with pork, respectively. The original spectra and average spectra of the restructured steak samples adulterated with duck are shown in Figures 3 (A) and (B) respectively. The trends of the spectral curves of samples with different adulteration contents were similar, but the absorbance intensity was different. Among the adulterated samples in this study, the greater the amount of adulteration, the lower the absorbance intensity.

Fig. 2.

(A) The original spectrum of conditioning steaks samples adulterated with pork; (B) The average spectrum of conditioning steaks samples adulterated with pork.

Fig. 3.

(A) The original spectrum of conditioning steaks samples adulterated with duck; (B) The average spectrum of conditioning steaks samples adulterated with duck.

The difference in spectral absorbance values in the visible region (430–780 nm) was because the color value (L*) of pork was higher than that of beef, while in the near-infrared region (780–960 nm) the difference was caused by chemical composition factors such as moisture. Spectral patterns of adulterated samples had obvious absorption peaks at 550 nm, 575 nm, 760 nm, and 960 nm. Among them, the peaks at 550 nm and 575 nm were primarily associated with oxygenated myoglobin (Van Beers et al., 2018), the peak at 760 nm was the stable absorption peak of myoglobin (Huang et al., 2018), and the peak at 960 nm was related to the second overtone of the O-H stretching vibration in the moisture of the restructured steak (Talens et al., 2013). The overall change in the spectral absorbance of beef adulterated with duck was greater than that of beef adulterated with pork because the difference between duck and beef was greater than that of pork and beef. Figure 2(B) and Figure 3(B) were the average spectra of samples, which characterized the change trend between the real sample and different adulterated pork and duck meat. This change trend was not only the change of few wavelength points, but also the overall spectral curve. It was not suitable to use these wavelengths to distinguish the difference of adulteration. In addition, the spectral signals of different samples will cross and overlap. Therefore, it was necessary to use chemometrics to combine multiple spectral points, and to eliminate adverse effects by selecting a part of the information wavelength from the full spectrum. As shown in Figure 4 and 5, we used six pre-processing methods, including 1st Der, 2nd Der, MC, MSC, SG, and SNVT to preprocess the original spectral datasets of adulterated pork and adulterated duck, respectively.

Fig. 4.

Spectrogram of samples adulterated with pork after preprocessing

Fig. 5.

Spectrogram of samples adulterated with duck after preprocessing.

Establish PLS model using full band    Before establishing the PLS prediction model for adulterated content, the spectra of beef adulterated with pork and duck samples with different adulterated amounts were divided into a calibration set and a prediction set at a ratio of 2:1 by random grouping. The prediction results of the PLS model established using different pre-processing methods are shown in Table 1.

Table 1. Predictive results of adulterated content (%) based on different pre-processing methods.

Adulterated species	Pre-treatment method	RC	RMSECV	RP	RMSEP
pork	1st Der	0.9369	2.85	0.9341	3.15
	2nd Der	0.9211	3.45	0.9164	3.56
	MC	0.9381	2.69	0.9375	2.76
	MSC	0.9366	3.01	0.9219	3.47
	SG	0.9372	2.59	0.9344	3.15
	SNVT	0.9265	3.36	0.9263	3.38
duck	1st Der	0.9557	2.48	0.9542	2.53
	2nd Der	0.9492	2.61	0.9433	2.66
	MC	0.9513	2.56	0.9476	2.61
	MSC	0.942	2.67	0.939	2.82
	SG	0.9664	2.43	0.9639	2.46
	SNVT	0.9543	2.52	0.9525	2.58

The most efficient pre-treatment method for the prediction model of beef adulterated with pork was MC; the R_C and R_P of the model were 0.9381 and 0.9375, and the RMSECV and RMSEP were 2.69% and 2.76%, respectively. Meanwhile, the most efficient pre-treatment method for the prediction model of beef adulterated with duck was SG. The R_C and R_P of this model were 0.9664 and 0.9639, and the RMSECV and RMSEP were 2.43% and 2.46%, respectively. The prediction accuracy of the beef adulterated with duck model was higher than that of beef adulterated with pork model because the difference between duck and beef was greater; therefore it was easier to identify the duck.

Establish PLS model using wavelength selection method    A sample of beef adulterated with pork was provided to introduce the process of screening characteristic wavelengths. When using SPA to select feature wavelengths, we set the number of selected wavelengths to range from 1 to 25 and then selected the wavelengths according to the root mean square error (RMSE) of cross validated (CV). As illustrated in Figure 6, with an increase in the number of wavelengths, the value of the RMSE declined rapidly and then slowly, and 8 feature wavelengths were finally selected.

Fig. 6.

Process of selecting characteristic variables by SPA.

The sampling times of CARS was set to 100, and its running process is shown in Figure 7. The number of selected wavelengths gradually decreased as the number of sampling increased, and RMSECV first decreased slowly and then rose stepwise. The lowest RMSECV was obtained when the number of samplings was 50, and 36 characteristic wavelengths were finally selected.

Fig. 7.

Process of selecting characteristic wavelengths by CARS.

The parameters for VCPA were set as follows: the best subset with a ratio of 0.1, 1 000 BMS runs, 50 EDF runs, and 14 remaining wavelengths. First, the data of the correction set was sampled 1 000 times by BMS to obtain 1 000 sets of wavelength combinations, and the probability of each wavelength being selected was the same in this process. These wavelength combinations were then used to build the PLS model, in which the 100 wavelength combinations with the smallest RMSECV values were employed for the EDF calculation. The change of RMSECV during the operation of EDF is shown in Figure 8. The feature space gradually shrank with EDF repeated runs, and the RMSECV showed a downward trend. At this time, the wavelengths that had little correlation with beef adulterated with pork content were removed, and the remaining wavelengths were added to the optimal subset. After EDF, we calculated the RMSECV of all possible combinations and finally selected the 12 characteristic wavelengths with the smallest RMSECV values.

Fig. 8.

Changes in RMSECV with the number of EDF runs.

The parameters for iRF were set as follows: mobile window width of 20, 10 000 iterations, 50 initial variable sets, and 3 variables extracted at the node. After the iteration, the 599 intervals were sorted according to the selected probability, and then a total of 590 RMSECV values were calculated for the top 1–10 to 1–599 combinations. Finally, the wavelength interval combination with the smallest RMSECV value was selected. The final selected wavelength (blue band) is shown in Figure 9.

Fig. 9.

Wavelengths selected by iRF.

The SPA, CARS, VCPA, iRF, iRF-SPA, iRF-CARS, and iRF-VCPA methods were applied to select characteristic wavelengths. For each method, the majority of wavelengths selected were between 440 nm and 630 nm. The spectra around 440 nm were related to the respiratory pigment of hemoglobin, which was consistent with the experimental results of Kamruzzaman et al. (2016) when using hyperspectral technology to detect chicken adulteration in beef. The spectra around 630 nm were related to the absorption of methemoglobin (Sivertsen et al., 2012), and the methemoglobin content of beef was higher than adulterated meat.

The model prediction results based on different wavelength selection methods are shown in Table 2. For the prediction of two types of adulterated meats, although VCPA selected a small number of wavelengths, the accuracy of the model has not been significantly improved and even declined. This is because VCPA might choose the wavelength with lower signal-to-noise while minimizing the collinearity between variables, which affected the prediction accuracy of the model. For the prediction of beef adulterated with pork, iRF-CARS had the best prediction accuracy, and the R_C and R_P of this model were 0.9862 and 0.9849 respectively. For the prediction of beef adulterated with duck, SPA, iRF, and iRF-CARS had high prediction accuracy and R_P value. Among them, iRF-CARS had the highest R_C value while selecting few characteristic wavelengths, which can make the prediction model more efficient. Hence, iRF-CARS was chosen as the best characteristic wavelength selection methods for the two detection models of beef adulteration. The results showed that the wavelength selection method cascade strategy prominently reduced the number of wavelengths required while improving the accuracy of modeling.

Table 2. Predictive results of adulterated content (%) based on different wavelengths selection methods.

Adulterated species	Selection method of characteristic wavelength	Number of wavelengths	RC	RMSECV	RP	RMSEP
pork	SPA	8	0.9618	2.38	0.9607	2.3
	CARS	36	0.9859	1.42	0.9843	1.51
	VCPA	12	0.9825	1.62	0.9765	1.78
	iRF	472	0.9803	1.74	0.9795	1.6
	iRF-SPA	32	0.9759	1.88	0.9692	2.12
	iRF-CARS	28	0.9862	1.39	0.9849	1.51
	iRF-VCPA	10	0.981	1.68	0.9806	1.64
duck	SPA	28	0.9856	1.44	0.9819	1.66
	CARS	48	0.9845	1.35	0.9865	1.42
	VCPA	10	0.9825	1.61	0.9781	1.73
	iRF	241	0.9855	1.47	0.9773	1.84
	iRF-SPA	24	0.9838	1.52	0.9809	1.69
	iRF-CARS	17	0.9888	1.66	0.9821	1.67
	iRF-VCPA	8	0.9818	1.59	0.9783	1.81

Conclusions

In this paper, hyperspectral technology was employed to detect restructured steak adulterated with pork and duck. Different pre-treatment methods were first used to establish a PLS model based on the full band, and then the best pre-treatment method was determined through comparison. Next, different selection methods were used to select characteristic wavelengths to establish the PLS model. Our results show that:

(1) The best of the six pre-treatment methods, 1st Der, 2nd Der, MC, MSC, SG, and SNVT, for the PLS prediction model for beef adulterated with pork and duck were MC and SG, respectively.

(2) Among the seven characteristic wavelength selection methods, iRF-CARS was best for beef adulterated with pork and duck. The iRF-CARS method for pork selected 28 characteristic wavelengths (R_C and R_P of the PLS model were 0.9862 and 0.9849, and RMSECV and RMSEP were 1.39% and 1.51%, respectively). The iRF-CARS method for duck selected 17 characteristic wavelengths (R_C and R_P of the PLS model were 0.9888 and 0.9821, and RMSECV and RMSEP were 1.66% and 1.67%, respectively).

Our data demonstrated that the combination of hyperspectral technology and the wavelength selection algorithm cascade strategy not only effectively simplified the prediction model, but also achieved accurate prediction of the content of beef adulterated with pork and duck in restructured steak. In addition, the time spent in this study to collect images using a band-pass filter based system might be no more than 60 s, which is shorter than the time required to use the hyperspectral system to control the translation stage (200 s). Thus, the proposed strategy has some advantages including simple structure, less data, low price, faster and more convenient processing, has broad application prospects in the field of rapid and non-destructive detection of adulterated meat products.

Acknowledgements    The authors gratefully acknowledge the financial support provided by the National Key Research and Development Program of China (2016YFD0401104), Open Project Program of National Engineering Laboratory for Agri-product Quality Traceability, Beijing Technology and Business University (AQT-2019-YB7), Open Research Fund of Beijing Key Laboratory of Big Data Technology for Food Safety, Beijing Technology and Business University (BTBD-2020KF09), Key Research and Development Program of Zhenjiang City (SH2019019). The authors thank AiMi Academic Services (www.aimieditor.com) for English language editing and review services.

Conflict of interest    The authors declare that they have no conflict of interest.

Compliance with ethics requirements    This article does not contain any studies with human or animal subjects.

References

•  Cortes,  V.,  Rodriguez,  A.,  Blasco,  J.,  Rey,  B.,  Besada,  C.,  Cubero,  S.,  Salvador,  A.,  Talens,  P., and  Aleixos,  N. (2017). Prediction of the level of astringency in persimmon using visible and near-infrared spectroscopy. J. Food Eng., 204, 27-37.
•  Garrido-Novell,  C.,  Garrido-Varo,  A.,  Pérez-Marín,  D., and  Guerrero  J. E. (2018). Using spectral and textural data extracted from hyperspectral near infrared spectroscopy imaging to discriminate between processed pork, poultry and fish proteins. Chemom. Intell. Lab. Syst., 172, 90-99.
•  Guo,  Z. M.,  Wang,  M. M.,  Agyekum,  A. A.,  Wu,  J. Z.,  Chen,  Q. S.,  Zuo,  M.,  El-Seedi,  H. R.,  Tao,  F. F.,  Shi,  J. Y.,  Ouyang,  Q., and  Zou,  X. B. (2020). Quantitative detection of apple watercore and soluble solids content by near infrared transmittance spectroscopy. J. Food Eng., 279, 109955.
•  Guo,  Z. M.,  Wang,  M. M.,  Wu,  J. Z.,  Tao,  F. F.,  Chen,  Q. S.,  Wang,  Q. Y.,  Ouyang,  Q.,  Shi,  J. Y., and  Zou,  X. B. (2019). Quantitative assessment of zearalenone in maize using multivariate algorithms coupled to Raman spectroscopy. Food Chem., 286, 282-288.
•  Huang,  C. P.,  Zhu,  X. R.,  Zhang,  C. L.,  Qiao,  N.,  Hu,  S. S., and  Zhang,  L. F. (2018). Pork freshness spectral feature index: development and sensitivity analysis. Spectrosc Spect Anal, 38, 552-558.
•  Kamruzzaman,  M.,  Makino,  Y.,  Oshita,  S., and  Liu,  S. (2015). Assessment of visible near-infrared hyperspectral imaging as a tool for detection of horsemeat adulteration in minced beef. Food Bioproc Tech, 8, 1054-1062.
•  Kamruzzaman,  M.,  Makino,  Y., and  Oshita,  S. (2016). Rapid and non-destructive detection of chicken adulteration in minced beef using visible near-infrared hyperspectral imaging and machine learning. J. Food Eng., 170, 8-15.
•  Li,  J.,  Zhang,  H.,  Zhan,  B.,  Zhang,  Y.,  Li,  R., and  Li,  J. (2020). Nondestructive firmness measurement of the multiple cultivars of pears by Vis-NIR spectroscopy coupled with multivariate calibration analysis and MC-UVE-SPA method. Infrared Physics & Technology, 104, 103154.
•  Li,  Y.,  Guo,  Y. J.,  Liu,  C.,  Wang,  W.,  Rao,  P. F.,  Fu,  C. L., and  Wang,  S. Y. (2017). SPA combined with swarm intelligence optimization algorithms for wavelength variable selection to rapidly discriminate the adulteration of apple juice. Food Anal Methods, 10, 1965-1971.
•  Mandli,  J.,  El Fatimi,  I.,  Seddaoui,  N., and  Amine,  A. (2018). Enzyme immunoassay (ELISA/immunosensor) for a sensitive detection of pork adulteration in meat. Food Chem., 255, 380-389.
•  Ouyang,  Q.,  Yang,  Y. C.,  Wu,  J. Z.,  Chen,  Q. S.,  Guo,  Z. M., and  Li,  H. H. (2020). Measurement of total free amino acids content in black tea using electronic tongue technology coupled with chemometrics. LWT, 118, 108768.
•  Prusakova,  O.,  Glukhova,  X.,  Afanas'eva,  G.,  Trizna,  Y.,  Nazarova,  L., and  Beletsky,  I. (2017). A simple and sensitive two-tube multiplex PCR assay for simultaneous detection of ten meat species. Meat Sci., 137, 34-40.
•  Rinnan,  S.,  Berg,  F., and  Engelsen,  S.B. (2009). Review of the most common pre-processing techniques for near-infrared spectra. Trends Anal. Chem., 28, 1201-1222.
•  Ropodi,  A. I.,  Pavlidis,  D. E.,  Mohareb,  F.,  Panagou,  E. Z., and  Nychas,  G. J. E. (2015). Multispectral image analysis approach to detect adulteration of beef and pork in raw meats. Food Res. Int., 67, 12-18.
•  Saranwong,  S.,  Kawano,  S.,  Ikehata,  A.,  Gou,  N., and  Haff,  R. (2013). Development of a low-cost nir instrument for minced meat analysis: spectrophotometer and sample presentation. Am.J.Exp.Agri., 1, 61-68.
•  Sivertsen,  A. H.,  Heia,  K.,  Hindberg,  K., and  Godtliebsen,  F. (2012). Automatic nematode detection in cod fillets (Gadus morhua L.) by hyperspectral imaging. J. Food Eng., 111, 675-681.
•  Sugiyama,  J. (1999). Visualization of sugar content in the flesh of a melon by near-infrared imaging. J Agric Food Chem., 47, 2715-2718.
•  Sun,  J.,  Yang,  W.,  Feng,  M.,  Liu,  Q., and  Saleem,  K. (2020). An efficient variable selection method based on random frog for the multivariate calibration of NIR spectra. RSC Adv, 10, 16245-16253.
•  Talens,  P.,  Mora,  L.,  Morsy,  N.,  Barbin,  D. F.,  Elmasry,  G., and  Sun,  D. W. (2013). Prediction of water and protein contents and quality classification of Spanish cooked ham using NIR hyperspectral imaging. J. Food Eng., 117, 272-280.
•  Trivittayasil,  V.,  Tsuta,  M.,  Kasai,  S.,  Matsuo,  Y.,  Sekiyama,  Y.,  Shoji,  T.,  Aiyama,  R.,  Kokawa,  M., and  Sugiyama,  J. (2018). Classification of 1-methylcyclopropene treated apples by fluorescence fingerprint using partial least squares discriminant analysis with stepwise selectivity ratio variable selection method. Chemom. Intell. Lab. Syst., 175, 30-36.
•  Van Beers,  R.,  Kokawa,  M.,  Aernouts,  B.,  Watté,  R.,  De Smet,  S., and  Saeys,  W. (2018). Evolution of the bulk optical properties of bovine muscles during wet aging. Meat Sci., 136, 50-58.
•  Wold,  S.,  Sjöström,  M., and  Eriksson,  L. (2001). Pls-regression: a basic tool of chemometrics. Chemom. Intell. Lab. Syst., 58, 109-130.
•  Yu,  H. D.,  Yun,  Y. H.,  Zhang,  W. M.,  Chen,  H. M.,  Liu,  D. L.,  Zhong,  Q. P.,  Chen,  W. X., and  Chen,  W. J. (2020). Three-step hybrid strategy towards efficiently selecting variables in multivariate calibration of near-infrared spectra. Spectrochim. Acta A Mol. Biomol. Spectrosc., 224, 117376.