Rapid detection of antibiotic dregs in feed protein materials using line-scan short-wave infrared hyperspectral imaging

Abstract

Antibiotic dregs contain antibiotic residues and pose safety risks if they are used illegally in animal feed protein materials. The objective of this study was to develop a shortwave infrared (SWIR) hyperspectral imaging system (HSI) to quantify and qualify antibiotic dregs for the adulteration of feed protein materials (FPMs). Three FPMs were adulterated with oxytetracycline dregs (OD) over a range of 2%–98% (w/w). Principal component analysis (PCA), partial least squares discriminant analysis (PLS-DA), and one-class partial least squares classifier (OCPLS) models were developed based on different spectral preprocessing techniques to predict adulterants. Furthermore, mean spectra with partial least squares (PLS) regression and images were used to predict adulterant concentrations. The results showed that FPMs and adulterated FPMs could not be completely distinguished using PCA. PLS-DA and OCPLS exhibited the highest classification accuracies (100%), and PLS-DA exhibited better recognition results for the pixel spectra. Using the mean spectra, PLS models were successfully established to predict adulteration levels in FPMs. The optimal parameters of residual predictive deviation (RPD) were 14.5, 6.1, and 7.6. Overall, the developed HSI system and optimized model demonstrated a high potential for discriminating antibiotic dregs in FPMs and allowed for quantitative evaluation.

Keywords

Feed safety antibiotic dregs hyperspectral imaging adulteration quantitative analysis qualitative analysis

Introduction

Currently, China is implementing the reduction and substitution of soybean meal in feed. In the context of food and feed safety, it has become an industry trend to reduce the amount of soybean meal and replace part of the soybean meal with unconventional feed protein.^1–3 However, the development and utilization of different types of feed protein materials (FPMs), can itself result in increasing feed and animal product safety issues. This will not only pose a threat to human and animal health but also considerably reduce the quality of feed. The nutritional and economic value of the products has seriously damaged their economic benefits and severely impacted the reputation of China’s livestock products in the international market. Therefore, it is imperative to effectively monitor the risks of substances in feed protein materials.

Antibiotic dregs are solid dregs after the fermentation of antibiotics, which are mainly composed of mycelium, remaining medium, trace antibiotics, and unknown fermentation intermediates.⁴ Antibiotic dregs contain high nutritional content, similar to the conventional chemical composition of feed protein materials,⁵ but may contain residual antibiotics and unknown secondary metabolites which pose considerable risk to the aquaculture industry.⁶ Of primary concern is the risk of the emergence of drug-resistant microorganisms, and secondly, there are various safety hazards due to the lack of safety test evaluation. Owing to the shortage of domestic feed protein materials, their market price continues to rise, and the cost of antibiotic dregs is extremely low. Driven by economic interests, some unscrupulous farmers add antibiotic dregs to feed protein materials to improve profitability.

A search of the literature revealed that the main method used to determine whether antibiotic residues were added to the feed was to detect the antibiotic residues. Analytical techniques based on liquid chromatography-tandem mass spectrometry (LC-MS-MS), gas chromatography-mass spectrometry (GC-MS), liquid chromatography-time-of-flight mass spectrometry, and other instrument methods are the most widely used analytical methods for the detection of antibiotic residues.^7,8 These methods rely heavily on high-end, expensive instruments, and high analysis costs and have many limitations in identifying nontarget risk substances. This research group studied the feasibility of discriminating antibiotic dreg adulteration in feed protein materials based on ATR-IR and microscopic infrared imaging and obtained ideal results.⁹ However, a more complex sample preparation process is still required, and the sample analysis amount is low, which cannot fully meet the current development needs of rapid and large-scale monitoring.

To overcome these limitations, line-scan-based hyperspectral imaging techniques that can accomplish rapid detection with high accuracy have been recently developed.^10–12 Hyperspectral imaging systems (HSI) are a nondestructive, nontargeted fingerprint identification method, which can quickly and cheaply provide reliable information about the physical properties and chemical composition of the sample.^13–15 The technique has demonstrated considerable capacity to detect food fraud and determine chemical composition, food quality and safety parameters, as well as monitor particular quality parameters during production, processing, or storage of food.^16,17 Compared to traditional shortwave infrared (SWIR) spectroscopy, the technique is characterized by high speed, accuracy, automation, and real-time monitoring, and is more suitable for automated quality evaluation and safety inspection of large sample sets.¹⁸ Further more, compared to Fourier-transform infrared spectroscopy, the quantitative performance of SWIR is even more prominent. Furthermore, the SWIR region is dominated by weak overtones, resulting in lower molar absorption and deeper penetration of SWIR waves inside the samples, making it more suitable for the analysis of heterogeneous samples, such as adulterated powders.¹⁹

In addition, the need for this nondestructive method has become more pronounced in the past 2 years due to the trend toward the increasing adoption of automation and artificial intelligence in the feed industry,²⁰ and the Vis/NIR range (especially 400–1000 nm) has been the most used mode.²¹ However, the team found that the spectral information of feed protein materials and antibiotic dregs in the range of 400–1000 nm was low, and the identification was not reliable. Until now, no studies have been conducted to develop optimal models using the HSI technique in the short-wave infrared range (1000–2500 nm) for rapid detection of antibiotic dregs in FPMs.

Hence, the main objective of this study was to investigate the feasibility of using SWIR-HSI combined with chemometrics to identify and quantify adulterants of oxytetracycline dregs in nonconventional feed protein materials such as cottonseed meal (CM), distiller-dried grains with solubles (DDGS), and nucleotide residues (NR). The specific aims were to (1) establish classification models principal component analysis (PCA), partial least squares discriminant analysis (PLS-DA), and one-class partial least squares classifier (OCPLS) methods to identify authentic and adulterated FPMs and (2) develop partial least squares (PLS) regression based on mean spectra and pixel count based on the optimal classification method to predict adulterated concentrations in feed protein materials.

Materials and methods

Sample preparation

A total of 238 samples were used in this study: 69 CMs, 84 DDGSs, 73 NRs, and 12 ODs, collected from 23 Chinese provinces. All samples used in this study were ground using a Retsch mill (Ultra centrifugal Mill ZM 100; Retsch GmbH, Haan, Germany) and passed through a 0.5-mm² mesh.

For adulterated samples: (1) three adulterated samples with known location of oxytetracycline dregs was prepared by selecting one from each type of FPMs. (2) adulterated samples with different OD concentrations, 2%–98% (increasing at a concentration gradient of 2%, w/w), were prepared, and two samples were prepared for each concentration. This resulted in 294 adulterated samples in total of three categories. A LabRAM benchtop ultrasonic mixer (Resodyn Acoustic Mixers, Resodyn Corp., Butte, Montana, USA) was used to uniformly mix the adulterated samples at room temperature (24 ± 1°C) for 5 min.

Instrumental measurements

All the prepared samples were scanned using a laboratory-based pushbroom hyperspectral imaging system (GaiaSorter, Zolix, Beijing, China). The system acquired the near infrared spectral range of 1000–2500 nm (288 data points) to capture hyperspectral images at room temperature (24 ± 1°C) with a spectral resolution of 5.2 nm. The spatial resolution was 400 μm × 400 μm, and the exposure time was set to 15 ms. The samples were placed in a glass dish (diameter: 85 mm; height: 10 mm), filled, and smoothed, and then the glass dish containing the sample was placed on the stage to obtain the image data.

After hyperspectral image acquisition, calibration was conducted with two reference images using the following equation:

R_{c} = (R_{o} ‐ D) / (W ‐ D) \times 100 %,

(1)

where Rc represents the calibrated relative reflectance hyperspectral image, Ro is the original acquired hyperspectral image, D is the dark reference hyperspectral image (GaiaSorter, Zolix, Beijing, China) with approximately 0% reflectance, and W is the white reference hyperspectral image (GaiaSorter, Zolix, Beijing, China) with approximately 99.9% reflectance. This procedure was performed directly using an image-calibration function in the data-acquisition software.

In this study, the dataset consisted of 265 wavelengths over the range of 1000–2380 nm after the last 23 bands were deleted because of high noise signals. A total of 5625 (75 × 75-pixel) spectra were extracted from each hyperspectral image, and then the mean and each sample were analyzed by extracting the mean spectrum from the pixels.

Data processing

Spectral preprocessing

Spectral preprocessing is essential for establishing robust spectral models because of the interference caused by noise and uneven sample surfaces during image acquisition. The spectrum was preprocessed using autoscaling, standard normal variate (SNV), multiplicative scatter correction (MSC), and Savitzky-Golay derivatives. The SNV and MSC approaches were adopted to address the interference of light scattering and particle size. Derivative methods were used to eliminate baseline drifts and separate broad and overlapping SWIR bands. Each derivative pre-treatment was applied with a first or second order polynomial fitted over a window of five or 11 features.

Classification analysis

Principal component analysis (PCA) is an unsupervised multivariate analysis technique used to reduce spectral dimensions in principal components (PCs), which capture maximum data variability and maintain relevant spectral information.²² In this study, PCA was applied to the SWIR-HSI raw data as an exploratory analysis to visualize the clustering of feed protein materials and OD according to their spectral variations.

Partial least squares discriminant analysis (PLS-DA) is a multivariate classification technique (supervised model) based on the PLS regression model.²³ This widespread classification model correlates the X matrix (spectral data) mathematically with the Y matrix (class membership).²⁴ When there were two classes to be classified, the matrix Y was arranged in a single column, where 0 was used for the class of interest and one for that without interest. Bayes’ theorem was typically used to estimate the threshold (T).²⁵ A training set (including 67% pure FPMs and adulterated samples) and test set (including 33% pure FPMs and adulterated samples, ODs) were created using the Kennard–Stone (K–S) algorithm.²⁶ By establishing three PLS-DA models, the adulterated samples were discriminated. The number of latent variables was selected using the root mean square error of cross-validation (RMSECV).

Regarding chemometric classification tools, one-class modeling is a better option than the most commonly employed discriminant models for food authenticity problems. One-class modeling methods build a class that focuses on authentic/nonfraudulent samples. This is an appropriate approach because it aims to detect whether new samples belong to the authentic class regardless of the type of fraud being investigated. This strategy has several benefits in relation to the multiclass approach, in which, in addition to the nonadulterated class, one or more adulterant classes have to be modeled, because it is practically impossible to cover all possible adulterants in a representative way.²⁷ The one-class partial least squares (OCPLS) classifier is a one-class classifier based on the partial least squares method. It uses only one type of sample to establish the model. Score distance (SD) and absolute central residual (ACR) are the class thresholds of the OCPLS model. An extensive literature review describes the theoretical and practical aspects of OCPLS. Without being exhaustive in the references, a recent review that provides multiple references can be consulted.^28,29 In the current study, the K–S algorithm was used to select 2/3 of the spectra from each type of pure FPMs to establish three one-class models, and the remaining 1/3 spectra, adulterated sample spectra, and OD spectra were then discriminated.

To evaluate the quality of the classification models, the main performance parameters such as sensitivity, specificity, and efficiency were considered. These parameters were calculated from four well-known sample model assignment possibilities: true positive (TP), false positive (FP), true negative (TN), and false negative (FN).

Prediction of adulteration

The partial least square (PLS) regression method allows spectral information to be converted into latent variables or factors, which can describe the maximum covariance between the spectral data and the reference sample values.³⁰ Sixty-six samples from the 98 adulterated samples were used to develop prediction models and the remaining 32 samples were used to validate the developed models. The optimal number of latent variables is obtained using the cross-validation value with the lowest root-mean-square error. To evaluate the performance of the established calibration models, evaluation parameters, such as the correlation coefficients pertaining to calibration (R²c) and prediction (R²p), root mean square errors pertaining to calibration (RMSEC) and prediction (RMSEP), and residual predictive deviation (RPD), were calculated. An accurate prediction model should have high R²c, R²p, and RPD values and low RMSEC and RMSEP values.³¹

For the quantitative detection of OD levels in the FPMs, optimal classification models were applied to the hyperspectral images to calculate the number of OD pixels. The size of each image was 75 × 75 pixels, and the number of OD pixels without background pixels was counted relative to the total number of image pixels (approximately 5625). For example, in an OD-FPM mixture containing OD at a concentration of 2%, 113 of the 5625 total pixels should be detected.

All data were processed using MATLAB R2015a (The Mathworks, Inc., Natick, MA, USA) with the PLS_Toolbox version 7.1 (Eigenvector Research, Wenatchee, WA, USA).

Results and discussion

Spectral features

The sample spectra were compared visually to detect differences between origin-specific sample groups. The mean raw spectra of all samples are shown in Figure 1(a), with seven notable absorbance peaks. It can be seen that the information of the raw spectra was rich and mainly concentrated in the range 1780–2380 nm. Moreover, the at shorter wavelengths, the higher the absorbance of the OD spectrum and the greater the difference from feed protein materials, because the color of the OD was deeper.³² Compared with the microscopic NIR image data, the difference between the feed protein materials and OD SWIR-HSI data at the 1000–1500 nm band was more significant.³³ The first absorption band of feed protein materials, located at approximately 1185 nm, corresponds to the second overtone of C=C-H stretching.³⁴ The wavelengths at approximately 1698, 1890, 2005, 2115, and 2224 nm were associated with crude protein. The absorption peaks were assigned to the bands arising from the amide vibrations characteristic of proteins, that is, Amide A and Amide B (N–H stretch in Fermi resonance with N–H in-plane bend), Amide I (C=O stretch), Amide II (N–H in-plane bend), and Amide III (C–N stretch with N–H in-plane bend).^35,36 At 2005–2115 nm, the spectral curves of OD and DDGS were quite different because the raw material of DDGS was mainly corn and the crude protein content was low.

The shapes of the spectra after pretreatment with the first derivative (1-2-5) (Figure 1(b)) were significantly altered. This diminishes noise and enhances the visibility of features (e.g., the wavelengths at 1000–1390 nm and 2016–2146 nm).

PCA

As shown in Figure 1, the major absorption peaks of the raw spectra were similar for both feed protein materials and OD. Mathematical treatment of the spectra is usually required to exploit the information underlying these spectra and reveal the subtle differences that may exist between different types of samples.

PCA was used to conduct an exploratory analysis of pure FPMs, adulterated samples, and OD, and all spectra were preprocessed using the optimal combination of the 1^st Der (1-2-5) + Autoscale. As shown in Figure 2(a) and (b), the first two principal components (PC) explain 78.22% and 78.16% of the spectral information, respectively. PC1 and PC2 explain 65.47% of the spectral information in Figure 2(c). The best separation is observed between the feed protein materials and OD. The score points of the samples with a low percentage of adulteration are closer to the feed protein materials, whereas those of the samples with a high percentage of adulteration are closer to the OD. Low percentages of adulterated samples and feed protein materials partially overlap within the same quadrant. PCA is not effective in distinguishing samples with a low percentage of adulteration from the feed protein materials, particularly DDGS and NR. This may be related to the fact that DDGS, NR, and OD are byproducts of fermentation.

PLS-DA and OCPLS

PLS-DA and OCPLS models were utilized to predict OD adulteration in FPMs using different preprocessing methods with mean spectra. Table 1 presents the results of the PLS-DA and OCPLS models employing the mean spectra. Using mean spectra, the PLS-DA model performed well with the 2^nd Der (2-2-11) + Autoscale, with sensitivity, specificity, and accuracy of 100%, 100%, and 100%, respectively (Figure 3(a)–(c)). It was also observed that the optimal prediction of the OCPLS model was obtained with no spectral preprocessing, and the OCPLS modeling approaches produced outcomes that were comparable to those of the PLSDA modeling approaches, with sensitivity, specificity, and accuracy of 100%, 100%, and 100%, respectively (Figure 3(d)–(f)).

Table 1.

PLS-DA and OCPLS models outcomes of detecting OD adulteration in FPMs.

Indicator	Model	Processing	Factors	Sensitivity	Specificity	Accuracy
CM-OD	PLS-DA	2^nd Der (2-2-11) + Autoscale	4	100	100	100
CM-OD	OCPLS	None	10	100	100	100
DDGS-OD	PLS-DA	2^nd Der (2-2-11) + Autoscale	5	100	100	100
DDGS-OD	OCPLS	None	10	100	100	100
NR-OD	PLS-DA	2^nd Der (2-2-11) + Autoscale	4	100	100	100
NR-OD	OCPLS	None	10	100	100	100

Figure 4 shows the resultant images of adulterated samples with known OD positions obtained by applying the developed classification models to the hyperspectral images. As illustrated in Figure 4(b) and (c), the distribution of the OD in each image using the PLS-DA model is closer to the grayscale image (Figure 4(a), where the black area is the OD and the gray area is the FPMs), especially the OD mixed in NR. For the identification of pixel spectra, the OCPLS model had a higher sensitivity and lower specificity, which might be due to the low signal-to-noise ratio of the pixel spectrum. Consequently, the PLSDA model showed better identification potential for pixel spectra.

Quantitative analysis by PLS

The presence of OD in adulterated feed protein materials was successfully identified and accurately differentiated using the classification models. Subsequently, a quantitative calibration model using PLS was developed to predict the concentrations of the adulterated samples. Spectral preprocessing methods were optimized to improve the capabilities of the PLS model. A quantitative model was constructed to predict the additive ratios of the adulterants.

Table 2 lists the evaluation parameters of the quantitative analysis results of the three PLS models, and Figure 5 shows the results of the quantitative analysis. Using the mean spectra after treatment with the 1^st Der (2^nd order, 5-point window) and autoscale to establish the PLS models, the R²_p values of the three models were 0.995, 0.98, and 0.98, and the RPD were 14.5, 6.1, and 7.6, respectively, when using seven factors. In addition, the coefficients of determination obtained during calibration and prediction were close to each other, indicating that the established models were not underfitted or overfitted. The interpretation of the RPD was based on Williams for a complex feed matrix, which indicated the applicability of the prediction models based on the obtained RPD.³⁷ A prediction model with an RPD below 1.9 is not recommended to be used, an RPD between 2.5 and 2.9 could be applied for screening, an RPD above 3.0 could be adequate for quality control, and an RPD above 4.0 could be adequate for any application.³⁶ This shows that the quantitative correction model has a good predictive performance and can meet the needs of practical detection.

Table 2.

PLS models outcomes of detecting OD adulteration in FPMs.

Models	LV.s	Processing	RMSEC/%	RMSEP/%	R²_c	R²_p	RPD
CM-OD	7	1^st Der (1-2-5) + Autoscale	2.3	1.9	0.994	0.995	14.5
DDGS-OD	7	1^st Der (1-2-5) + Autoscale	2.9	3.5	0.991	0.98	6.1
ND-OD	7	1^st Der (1-2-5) + Autoscale	4.2	3.4	0.98	0.98	7.6

Quantitative analysis by images

PLS-DA models were applied to hyperspectral images of 0%–10% adulteration concentrations. Based on the number of OD pixels counted in the images, a scatter plot was produced with the actual and predicted OD concentrations based on the images (Figure 6). As the OD concentration increased, the number of detected OD pixels also increased; however, there was a large difference between the actual and predicted OD concentrations. The number of OD pixels for 10% CM-OD, DDGS-OD, and NR-OD was calculated to be approximately 563. However, the actual numbers of pixels counted for 10% CM-OD and DDGS-OD were 5400, 4782, and 2532, respectively.

Although the FPMs containing the OD used in this experiment were well mixed, the detected OD pixels varied in size. As mentioned above, the size of the OD particle used in this study was controlled at approximately 500 µm. If the OD particle size is larger than the pixel size, the adjacent pixels form a mixed spectrum, that is, an adulterated spectrum. Consequently, the performance of the developed classification model for detecting OD depends on the size of the OD particles used.

Conclusion

The study proposes optimal qualitative and quantitative models developed in the short-wave infrared region and its hyperspectral imaging system to detect OD in PFMs. The results show that PCA could only distinguish between FPMs and OD. PLS-DA and OCPLS were used to classify OD adulteration, and the accuracy of each optimal model was 100%. The PLS-DA showed better recognition results for the pixel spectra. Subsequently, the R²_p values of the PLS models for determining adulterated levels of OD using the mean spectra were 0.995, 0.98, and 0.98, and the RPD were 14.5, 6.1, and 7.6, respectively, when using seven factors. The OD concentrations in FPMs obtained by applying the developed PLS-DA models to hyperspectral images were not satisfactory. Overall, SWIR-HSI combined with chemometrics is a useful tool for classifying and quantifying feed protein materials and their adulterants.

Although SWIR-HSI obtains a spectrum for each pixel, the concentrations of adulterants in individual pixels are unknown. To achieve more realistic results, it is necessary to focus on the pixel spectra, which can be directly used to build chemometric models and predict the pixel spectra of unknown samples.

Footnotes

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was financially supported by the Science and Technology Project Plan of the Jiangsu Vocational College of Agriculture and Forestry (Grant Nos. 2022kj17 and 2022kj04).

ORCID iD

Wei Ding

References

Kim

Less

Wang

, et al. Meeting global feed protein demand: challenge, opportunity, and strategy. Annu Rev Anim Biosci 2019; 7: 221–243.

Harris

Mellencamp

Johnston

, et al. Effectiveness of different corn dried distillers grains with solubles feeding strategies and increasing the time intervals between the second Improvest dose and slaughter of immunologically castrated pigs on belly and pork fat quality. Meat Sci 2018; 135: 62–73.

Mottet

de Haan

Falcucci

, et al. Livestock: on our plates or eating at our table? A new analysis of the feed/food debate. Glob Food Secur 2017; 14: 1–8.

Guo

Gong

Duan

, et al. Characteristics of penicillin bacterial residue. J Air Waste Manag Assoc 2012; 62: 485–488.

Mohring

SAI

Strzysch

Fernandes

, et al.

Degradation and elimination of various sulfonamides during anaerobic fermentation: a promising step on the way to sustainable pharmacy?

Environ Sci Technol 2009; 43: 2569–2574.

Biao

Lijun

Xinyan

, et al. Prevalence of antibiotic resistance genes of wastewater and surface water in livestock farms of Jiangsu Province, China. Environ Sci Pollut Res Int 2015; 22(18): 13950–13959.

Zhang

, et al. Alkaline thermal pretreatment at mild temperatures for biogas production from anaerobic digestion of antibiotic mycelial residue. Bioresour Technol 2016; 208: 49–57.

Lin

Yang

Liang

, et al. Identification of antibiotic mycelia residue in protein rich feed using on near-infrared microscopy imaging. Food Addit Contam 2018; 35(5): 818–827.

Fan

, et al. A novel analytical strategy for discriminating antibiotic mycelial residue adulteration in feed based on ATR-IR and microscopic infrared imaging. Spectrochim Acta Mol Biomol Spectrosc 2021; 261: 120060.

10.

Moghaddam

Tamiji

Lakeh

, et al. Multivariate analysis of food fraud: a review of NIR based instruments in tandem with chemometrics. J Food Compos Anal 2022; 107: 104343.

11.

Mei

. An overview on optical non-destructive detection of bruises in fruit: technology, method, application, challenge and trend. Comput Electron Agric 2023; 213: 108195.

12.

Pan

Zhang

Chen

, et al. Applications of hyperspectral imaging technology combined with machinelearning in quality control of traditional chinese medicine from the perspective of artificial intelligence: a Review. Crit Rev Anal Chem 2024; 54(8): 2850–2864.

13.

Kim

Lee

Baek

, et al. Quantitative detection of benzoyl peroxide in wheat flour using line-scan short-wave infrared hyperspectral imaging. Sens Actuators, B 2022; 352: 130997.

14.

Saha

Senthilkumar

Singh

, et al. Quantitative detection of metanil yellow adulteration in chickpea flour using line-scan near-infrared hyperspectral imaging with partial least square regression and one-dimensional convolutional neural network. J Food Compos Anal 2023; 120: 105290.

15.

Zhong

Zhang

Tang

, et al. Feasibility study of combining hyperspectral imaging with deep learning for chestnut-quality detection. Foods 2023; 12: 2089.

16.

Nagy

Wang

Farag

. Quality analysis and authentication of nutraceuticals using near IR (NIR) spectroscopy: a comprehensive review of novel trends and applications. Trends Food Sci Technol 2022; 123: 290–309.

17.

Yan

, et al. Application of near-infrared spectroscopy and hyperspectral imaging combined with machine learning algorithms for quality inspection of grape: a Review. Foods 2023; 12: 132.

18.

Hassoun

Ojha

Tiwari

, et al. Monitoring thermal and non-thermal treatments during processing of muscle foods: a comprehensive review of recent technological advances. Appl Sci 2020; 10(19): 6802.

19.

Gholizadeh

Borůvka

Saberioon

, et al. Visible, near-infrared, and mid-infrared spectroscopy applications for soil assessment with emphasis on soil organic matter content and quality: state-of-the-Art and key issues. Appl Spectrosc 2013; 67(12): 1349–1362.

20.

Khaled

Parrish

Adedeji

. Emerging nondestructive approaches for meat quality and safety evaluation-a review. Compr Rev Food Sci Food Saf 2021; 20: 1541–4337.

21.

Hassoun

Jagtap

Garcia-Garcia

, et al. Food quality 4.0: From traditional approaches to digitalized automated analysis. J Food Eng 2023; 337: 111216.

22.

Wang

Kamruzzaman

. Portable NIR spectroscopy and PLS based variable selection for adulteration detection in quinoa flour. Food Control 2022; 138: 108970.

23.

Wang

Chen

Dai

, et al. Recent advances of chemometric calibration methods in modern spectroscopy: algorithms, strategy, and related issues. Trends Anal Chem 2022; 153: 116648.

24.

Rodionova

Titova

Pomerantsev

. Discriminant analysis is an inappropriate method of authentication. Trends Anal Chem 2016; 78: 17–22.

25.

da Silva

Murphy

Amigo

, et al. Classification and quantification of microplastics (<100 µm) using a focal plane array-Fourier transform infrared imaging system and machine learning. Anal Chem 2020; 92: 13724–13733.

26.

Kennard

Stone

. Computer aided design of experiments. Technometrics 1969; 11: 137–148.

27.

Rovira

Miaw

CSW

Martins

MLC

, et al. One-class model with two decision thresholds for the rapid detection of cashew nuts adulteration by other nuts. Talanta 2023; 253: 123916.

28.

Oliveri

Malegori

Mustorgi

, et al. Qualitative pattern recognition in chemistry: theoretical background and practical guidelines. Microchem J 2021; 162: 105725.

29.

Yan

Cai

, et al. One-class partial least squares (OCPLS) classifier. Chemometr Intell Lab Syst 2013; 126: 1–5.

30.

Féret

J-B

François

Gitelson

, et al. Optimizing spectral indices and chemometric analysis of leaf chemical properties using radiative transfer modeling. Remote Sens Environ 2011; 115: 2742–2750.

31.

Wang

Cui

Jin

, et al. Tea Analyzer: a low-cost and portable tool for quality quantification of postharvest fresh tea leaves. Lebensm Wiss Technol 2022; 159: 113248.

32.

Arianna

Marica

Federico

, et al. Application of a handheld near-Infrared spectrometer to predict gelatinized starch, fiber fractions, and mineral content of ground and intact extruded dry dog food. Animals 2020; 10: 1660.

33.

Fan

Mei

, et al. Identification of antibiotic mycelia residues in cottonseed meal using Fourier transform near-infrared micro spectroscopic imaging. Food Chem 2019; 293: 204–212.

34.

Schütz

Riedl

Achten

, et al. Fourier-transform near-infrared spectroscopy as a fast screening tool for the verification of the geographical origin of grain maize (Zea mays L.). Food Control 2022; 136: 108892.

35.

Chen

Yang

Han

. A review on the use of near-infrared spectroscopy for analyzing feed protein materials. Appl Spectrosc Rev 2013; 48(7): 509–522.

36.

Daszykowski

Wrobel

Czarnik-Matusewicz

, et al. Near-infrared reflectance spectroscopy and multivariate calibration techniques applied to modelling the crude protein, fibre and fat content in rapeseed meal. Analyst 2008; 133: 1523–1531.

37.

Williams

. The RPD statistic: a tutorial note. NIR News 2014; 25: 22–26.