Global calibration for non-targeted fraud detection in quinoa flour using portable hyperspectral imaging and chemometrics

Abstract

Quinoa is one of the highest nutritious grains, and global consumption of quinoa flour has increased as people pay more attention to health. Due to its high value, quinoa flour is susceptible to adulteration. Cross-contamination between quinoa flour and other flour can be easily neglected due to their highly similar appearance. Therefore, detecting adulteration in quinoa flour is important to consumers, industries, and regulatory agencies. In this study, portable hyperspectral imaging in the visible near-infrared (VNIR) spectral range (400–1000 nm) was applied as a rapid tool to detect adulteration in quinoa flour. Quinoa flour was adulterated with wheat, rice, soybean, and corn in the range of 0–98% with 2% increments. Partial least squares regression (PLSR) models were developed, and the best model for detecting the % authentic flour (quinoa) was obtained by the raw spectral data with R²p of 0.99, RMSEP of 3.08%, RPD of 8.77, and RER of 25.32. The model was improved, by selecting only 13 wavelengths using bootstrapping soft shrinkage (BOSS), to R²p of 0.99, RMSEP of 2.93%, RPD of 9.18, and RER of 26.60. A visualization map was also generated to predict the level of quinoa in the adulterated samples. The results of this study demonstrate the ability of VNIR hyperspectral imaging for adulteration detection in quinoa flour as an alternative to the complicated traditional method.

Graphical abstract

Highlights

• •Portable HSI was used to build global calibration to detect fraud in quinoa flour.
• •Pre-treatment did not improve predictive ability compared to raw spectra.
• •Important wavelengths selected with BOSS improved the model performance.
• •Image processing algorithm was developed for prediction maps.

1. Introduction

Quinoa is a remarkably nutritious grain, and it is one of the few plants that contains all the amino acids the human body needs, along with a high percentage of other essential nutrients such as lipids, fiber, vitamins, and minerals (Navruz-Varli and Sanlier, 2016). Multiple bioactive compounds including but not limited to saponins, phenolics, phytosterols, and bioactive peptides were identified in quinoa that have been proven to benefit skin, immune and neuron system (Vilcacundo and Hernández-Ledesma, 2017). These compounds can also protect people against diabetes, obesity, and cancer (Vilcacundo and Hernández-Ledesma, 2017). In this increasingly health-conscious society, people around the world are showing more interest in food with high nutritional content, and the international demand for quinoa has increased dramatically over the past few decades (Schmidt et al., 2021). With greater global demand, food safety protection has become a top priority. However, grains like quinoa, especially flour, can be easily contaminated intentionally or unintentionally. Numerous flour adulteration events around the globe have been reported recently due to possible economic benefits that resulted in chaos and fatalities (Kwesiga et al., 2019; Petcu et al., 2019). As a result, it is urgent and crucial to developing an efficient and accurate way to identify possible adulteration in quinoa flour. Several methods have been developed to test the adulteration level in different flour (Arslan et al., 2020; Ndlovu et al., 2021b). These methods are mainly based on targeted analysis such as high-performance liquid chromatography, gas-chromatography mass-spectrometry, and enzyme-linked technology etc. (Aghili et al., 2022; Li et al., 2021a, Li et al., 2021b; Li et al., 2019; Mota et al., 2021). All of these targeted analytical tools for adulteration detection are well established and very accurate. However, they are expensive, slow, complicated, and laboratory-based. On the other hand, these targeted analytical tools authenticate specific marker(s) based on predefined target(s). Therefore, these targeted methods provide limited information and insufficient protection against highly complex food fraud or authentication issues. Since hundreds of potential adulterants can be added for food fraud, detecting one or more specific maker(s) is not enough to properly authenticate the food and food products unless one particular adulterant is suspected. Therefore, non-targeted analytical tools are required for food authentication to replace targeted analytical tools. Fingerprinting techniques provide detailed profiles that allow the detection of many markers or analytes by comparing the previously collected profiles (Esteki et al., 2019, van de Steene et al., 2022; Wu et al., 2021). Meanwhile, chemometrics or machine learning tools can extract the required analytes from the profiles.

Fingerprinting techniques, particularly spectroscopic techniques, are frequently used for authenticating food and detecting adulteration (Chen et al., 2019; Hou et al., 2019; Lima et al., 2020; Truzzi et al., 2021; Varnasseri et al., 2022; Weng et al., 2020). However, most of the studies for adulteration detection are based on detecting specific adulterants. For instance, Wang et al. (2022) used portable NIR spectroscopy to detect wheat flour adulteration in quinoa flour. Spectroscopic techniques are indirect methods based on calibration. That means calibration is required for each adulterant. This is not a convenient solution to fight against food fraud in the supply chain because fraudulent are always one step ahead of the scientists. Once a specific test is developed to detect fraud, fraudsters become aware of it and may start adding unknown adulterants to the food to try and evade detection. Although the authors presented a model with good results, it is not the best option to be used in reality. Developing models for each possible adulterant is laborious and impractical. Therefore, with the complex environment of the food industry, a non-targeted global calibration model is needed that will be able to predict adulteration in a specific authentic food product.

Recently, hyperspectral imaging technology has received ample attention for authentication and detecting adulteration in food and food products (He et al., 2023; Kim et al., 2022; Lohumi et al., 2019). Hyperspectral imaging scans every pixel of the sample to create a complete spectral image of the whole sample that increases the accuracy and integrity of detection or prediction (ElMasry and Sun, 2010). However, due to complicated high dimensional data processing, it is difficult to apply hyperspectral imaging technology directly for real-time applications (Burger and Gowen, 2011). Nonetheless, the prospects for this technology are certainly promising if some of the most important wavelengths can be selected properly to reduce the burden of measurements and analysis to design a multispectral imaging system (Burger and Gowen, 2011).

Currently, hyperspectral imaging in the visible near-infrared range (400–1000 nm) and near-infrared range (900–1700 nm) is usually used for application in foods and agricultural products (Hashemi-Nasab et al., 2023; Khan et al., 2021; Zhao et al., 2019). Low-cost detectors such as CCD and CMOS are used in the VNIR range, whereas expensive detector such as InGaAs is used for the NIR range. Due to low-cost detectors and availability, CCD or CMOS is more advantageous for industrial applications (Stanley, 1992; Taghizadeh et al., 2009) than expensive InGaAs dectectors. VNIR hyperspectral imaging has been applied recently to many areas including but not limited to the detection of DON levels in Fusarium head blight wheat kernels and wheat flour (Liang et al., 2020), the prediction of deoxynivalenol contamination in whole wheat flour (Zhao et al., 2020), and the simultaneous detection of fusarium and ergot damage in wheat (Erkinbaev et al., 2022). In the field of fraud detection for powder/flour products, hyperspectral imaging has been gaining more attention in recent years. The technology has been applied to identify fraud in various powder/flour products, including wheat flour (He et al., 2023), almond powder (Faqeerzada et al., 2020), and high value spices such as cinnamon (Cruz-Tirado et al., 2023), cumin (Florián-Huamán et al., 2022), and black pepper (Orrillo et al., 2019). All of these studies used laboratory-based, large, and heavy instruments and it is impossible to take these devices to the food supply chain to detect food fraud. Thus, developing portable technology for detecting food fraud in the food supply chain is crucial. To the best of the authors’ knowledge, no research has investigated the potential of portable hyperspectral imaging systems to detect levels of adulteration in quinoa flour.

Therefore, the main objective of this study was to develop a global calibration model for adulteration detection in quinoa flour using VNIR hyperspectral imaging. The specific objectives of this study were: (a) to create a global calibration model to detect the levels of pure quinoa flour in adulterated samples, (b) to select important wavelengths to simplify and improve the efficiency of the calibration model, and (c) to develop image processing algorithms and generate a visualization prediction map for the tested samples.

2. Material and methods

2.1. Sample preparation

Quinoa, wheat, rice, soybean, and corn seeds were collected from various local supermarkets to prepare the flour samples. Quinoa seeds were collected from four different sources. Wheat, rice, soybean, and corn were used as potential adulterants for quinoa flour. All seeds were dried in a mechanical dryer before grinding for uniform moisture content. Seeds were separately grinded with a mechanical grinder and filtered through a sieve with 250-μm openings (standard Tyler equivalent No. 60 mesh) to ensure the same particle size for all the flour particles. The adulteration levels were set in the range of 0–98% with an increment of 2% (each adulteration level corresponds to the percentage of the contaminant presented in quinoa flour). The type of contaminant corresponding to each adulteration rate was chosen in a random order determined by the random function in Microsoft Excel. A total mass of approximately 15g was maintained throughout the mixing process for each sample. A total of 150 samples (3 samples at each adulteration level) were prepared. Among them, 36 samples in 12 adulteration levels (approximately 8%, 12%, 24%, 32%, 40%, 52%, 62%, 66%, 76%, 80%, 82%, 86%) were randomly selected to be excluded from the calibration model and used as validation samples. Therefore, 114 samples were included in the calibration sample set, and 36 samples were used for validation.

2.2. Image acquisition and correction

All the imaging data were acquired using the Specim IQ hyperspectral imaging system in reflectance mode. The system consists of a battery-based portable 12-bit Specim V10E IQ line-scan hyperspectral imaging camera with a CMOS sensor, a lens, and an imaging spectrograph (Specim, Spectral Imaging Ltd, Oulu, Finland); an illumination set with two lamps as the light source (Oulu, Finland); a tripod for supporting and fixing the camera at a constant height; and an output analyzing computer with Specim IQ Studio software installed for extracting the image and spectral information. Each scan of the system produces 204 spectral bands in the wavelength range between 400 and 1000 nm as well as a 2D square image with a resolution of 512 × 512 pixels. As a result, a 3D hypercube was created with dimensions of 512 × 512 × 204 with over 53 million data points for each scan. The system was operated in a completely dark environment to ensure that the only light sources were from the pre-set illumination lamps so that the effect of unintended stray light could be eliminated. The camera was fixed on the tripod facing down, while samples were placed on a blackboard below the camera for scanning. The blackboard was put on a flat surface with a fixed height to maintain a uniform distance from the camera to the samples throughout the data gathering process. Before the data acquisition process, a random sample on the blackboard with a white reference panel (∼99% reflectance) was scanned for system calibration. This calibration data were saved and used for the reflectance transformation for every scan taken for the samples. After proper system calibration, the prepared samples were put into a small black container, approximately 4 cm in diameter and 2 cm in depth. One sample was put exactly below the camera for image acquisition. The dark reference data (∼0% reflectance) were captured inside the system automatically by blocking the incoming light completely. The transformation can be symbolized in the following equation:

$$Reflectance=\frac{R_0^{t1}-D^{t1}}{W^{t2}-D^{t2}}\times \frac{t2}{t1}$$

where R₀ represents the raw reflectance data, W and D represent the white and dark references, respectively. t1 and t2 are different when the target sample has low reflectance, so different integration times are used for the white reference and the target sample. In this study, since flour samples have relatively high reflectance, the integration times were the same for both the reference and the measured sample (t1 = t2). The acquired hyperspectral images were saved in a band-interleaved-by-line (BIL) format.

2.3. Image segmentation and extraction of spectral data

Image segmentation is one of the most crucial fundamental steps in precisely extracting spectral information from hyperspectral images. The main purpose of image segmentation is to identify regions of interest (ROI) of the tested sample from the background by creating a mask. In this study, image segmentation was mainly used to identify regions with the flour sample as ROI and remove the background of each image. For this study, the mask was created at 872 nm (a band with the greatest contrast between the flour sample and the background) with a threshold of 0.5. All the background portions in the mask were set to zero and the non-zero part of the mask was determined to be the main ROI. The determined main ROI was then used to extract the reflectance spectrum from the hyperspectral imaging. The average of all pixels was taken to obtain one spectrum for each image/sample.

2.4. Data analysis

All analyses performed in this study used the software MATLAB R2021b (The MathWorks, MA, USA) coupled with PLS-toolbox_90 (Eigenvector Research Inc., MA, USA).

2.4.1. Spectral pre-treatment

Pre-treatment is an important step in spectral analysis before building models to reduce possible scatterings in the raw spectral data and make the model more accurate. Different pre-treatment methods can generate different impacts on the raw dataset, so it is crucial to choose an appropriate pre-treatment method and avoid creating inaccurate models. In this study, the five pre-treatment used are first derivatives (FD), second derivatives (SD), standard normal variate (SNV), multiplicative scatter correction (MSC), and detrend (DT). These pre-treatments had been commonly used in spectral analysis. Among them, FD eliminates the linear baseline offsets, SD decreases certain peak overlaps, SNV and MSC are used to effectively remove the scatterings from the original data, and DT reduces the influences from the non-uniform particle size and baseline variations (Azzouz et al., 2003; Osborne, 2000).

2.4.2. Partial least squares regression (PLSR)

The partial least squares regression (PLSR) analysis is a regression method that relates two data matrixes to create a linear multivariate model (Wold et al., 2001). It is commonly used with spectral data to develop precise prediction models (Wold et al., 2001). In this study, PLSR was used to create a global calibration model to detect the levels of pure quinoa flour in adulterated samples. In PLSR, selecting an optimum number of latent variables for the analysis is crucial to develop an accurate model. Lowering the number of latent variables while keeping the RMSECV as small as possible can increase the accuracy of the PLSR model (Shariati-Rad and Hasani, 2010).The number of latent variables was determined by analyzing the root mean square errors of cross-validation (RMSECV) vs. latent variable number plot. PLSR models were created with both raw and pre-treated datasets and the model statistics were presented for comparison.

2.4.3. Evaluation of the PLSR model

The PLSR models were evaluated based on the value of the coefficient of determination from calibration (R²c), the coefficient of determination from prediction (R²p), the root mean square error for calibration (RMSEC), the root mean square error for prediction (RMSEP), the ratio of prediction to deviation (RPD), and the range error ratio (RER). In general, a better model was created when R²c and R²p get closer to 1 while having a small root mean square value. At the same time, having the value RMSEC and RMSEP relatively close to each other is another indication of a good model. A model can be classified as excellent if its corresponding RPD is over 4.1 and this model can be used for any application, including but not limited to quality and process control (Heil and Schmidhalter, 2021). Meanwhile, it is generally accepted that the higher the RER value is, the better the model (Ndlovu et al., 2021a; Williams and Norris, 2001).

2.4.4. Important wavelength selection

There are many different methods to determine key wavelengths. However, each experimental dataset is different and complicated, so there is no single best method for all datasets (Pu et al., 2015). Some of the variable selection methods, including variable in importance projection (VIP), stepwise regression (SR), correlation coefficient (CC), principle component analysis (PCA), and successive projection algorithm (SPA) have been implemented for hyperspectral imaging datasets (Pu et al., 2015). In the present study, bootstrapping soft shrinkage (BOSS) was applied as a novel variable selection technique for hyperspectral imaging data. BOSS is a recently developed variable selection method that combines the fundamental principle of weighted bootstrap sampling (WBS) and model population analysis (MPA) (Deng et al., 2016). PLSR sub-models were first created by variables randomly selected through bootstrap sampling (BSS) processes, and weights were assigned to the variables by calculating regression coefficients and the value of root mean square error of cross-validation (RMSECV) of each sub-model (Deng et al., 2016). MPA was then implemented to extract information and modify the assigned weights of each variable (Deng et al., 2016). Soft shrinkage gave smaller weights to variables with less importance, and the iteration of the calculation stopped when the number of variables was equal to one (Deng et al., 2016). Ultimately, the sub-model with the lowest RMSECV calculated was chosen, and the most important variables were selected.

2.5. Image analysis

Each image created from the hyperspectral imaging data acquisition process contains a spectrum for every pixel. Applying a calculated model to each pixel of an image can result in a spatial prediction map. Only image corresponding to the key wavelengths selected was used for developing a prediction map. During the prediction map generation process, the hyperspectral images at selected key wavelengths were first unfolded, creating a two-dimensional (2-D) column vector. Then, the regression coefficient from the best PLSR model was multiplied by each pixel of the images. Subsequently, the matrix was folded back to colored images with the same dimension as the original images before unfolding. A linear color scale was used to illustrate the different adulteration rates in every pixel of the prediction map.

3. Results and discussion

3.1. Spectral characteristics

Fig. 1 shows the raw spectra plots for quinoa flour adulterated with each of the four adulterant type. Lighter color represents a higher adulteration rate (lower % quinoa), and vice versa. In each subplot, the color changes gradually, corresponding to the change in concentration. With an increase in the adulteration level, the four plots showed a different concavity in shape, especially in the visible range between wavelengths 450–550 nm, due to the increase in the concentration of the different adulterants containing different chemical components. Fig. 2 (a) shows the spectral plot for all the samples together. Each spectrum in the plot is colored by its corresponding adulteration rate. A darker color indicates a lower adulteration rate, and they tend to present in the middle of the graph, while the spectrum with a higher adulteration rate tends to present at either the lower or the upper region with the increasing concentration of different adulterant types, resulting in a diverse shape of the spectrum. With the increase in the impurity in the sample, the discrepancy of the plot became larger due to the increasing concentration of the different adulterant types. Overall, the spectrum shows consistency in the shape trend regardless of the adulteration rate for the sample. At the same time, it was observed that there were no distinctive peak points in the spectrum plot, so further wavelength selection method is needed to determine the key wavelengths containing the most important spectrum information.

Fig. 1.

Raw spectrum plots separated by the adulterant types. A darker color indicates lower adulteration rate (higher % of quinoa), and vice versa. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

Fig. 2.

Raw and preprocessed spectrum plots from VNIR hyperspectral imaging in the wavelength range 400–1000 nm. The color bar indicates the percentage of adulteration (percent of adulterant in the sample). From top to bottom and left to right, the figure titles are as follows: a. Raw; b. FD; c. SD; d. SNV; e. MSC; f. Detrend. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

The five pre-treatment methods used in the analysis for this study (FD, SD, SNV, MSC, and detrend) were applied to the spectrum and their corresponding plots were shown in Fig. 2(b–f), respectively. The shape of the spectrum plots exhibited significant variances after applying different pre-treatment methods compared to the raw spectral plot. Similar to the raw spectral plot, the spectrum for the less adulterated samples were presented in the middle with less variation in the reflectance values. Moreover, the spectra plot corresponding to 400–550 nm region presents a greater variance and a larger slope of change, while the spectrums demonstrated general stability at higher wavelengths. According to plots 2 b, c, d, and e, the noise in the spectra in the spectral range of 900–1000 nm (Fig. 2a) was reduced after pre-treatment with FD, SD, SNV, and MSC, respectively. These pre-treatments also removed the scattering effect from the raw spectra. However, pre-treatment with detrend did not improve the noise and scatter effect from the spectra (Fig. 2f). The variance in the pre-treatment results emphasized the influence of different pre-treatment methods and the importance of choosing the most appropriate pre-treatment methods for model development.

3.2. Calibration models using full wavelengths range

Many studies were conducted using spectroscopy or hyperspectral imaging to detect adulteration in flour and other agricultural commodities (An et al., 2022; Su and Sun, 2017; Verdú et al., 2016; Zheng et al., 2022). These studies focused on detecting a specific adulterant in authentic flour/foods. This is not convenient for adulteration detection in the food supply chain, which is very complex. Therefore, a global calibration model is needed to predict the authenticity level of adulterated foods, irrespective of adulterants. That is why, instead of predicting adulterants, the main focus of this study was to predict the level of quinoa flour in all possible adulterated samples. However, calibration models were also developed to predict the levels of adulteration.

PLSR calibration models were developed using raw and pre-treated spectra to predict the level of quinoa in each sample. A total of 114 samples were used to develop the calibration, and another 36 samples were used to validate the efficacy of calibration models. It is very critical to select an optimum number of LVs for developing a stable calibration model. In this study, the number of LV was selected based on the minimum value of RMSECV. However, the values of RMSEC and RMSEP and the difference between RMSEC and RMSEP were also considered while choosing an optimum LVs.

Table 1 shows the PLSR models created for the raw and pre-treated data (with FD, SD, SNV, MSC, and DT) using full spectral range. The performance of PLSR with raw and pre-treated spectra revealed that none of these pre-treatments improved the model predictability. Most of them had a deteriorative effect on model performance compared to the raw spectra. Many authors reported similar results for detecting adulteration using spectroscopic (Kamruzzaman et al., 2022; Mishra et al., 2021; Wang et al., 2022) or hyperspectral imaging techniques (He et al., 2019; Kamruzzaman et al., 2022). Therefore, PLSR model developed using raw spectra will be discussed in the following sections.

Table 1.

PLSR model for detecting % quinoa created with the raw and pre-treated data from VNIR hyperspectral images (with all four adulterants).

Pre-treatment	No. LVs	Calibration				Prediction
		R2c	RMSEC (%)	RPD	RER	R2p	RMSEP (%)	RPD	RER
None (Raw)	11	0.98	3.63	8.15	27.00	0.99	3.08	8.77	25.32
FD	11	0.98	3.82	7.74	25.66	0.94	6.46	4.21	12.08
SD	13	0.99	2.62	11.3	37.14	0.93	7.58	3.72	10.29
SNV	12	0.99	3.58	8.26	27.39	0.95	7.08	3.47	11.02
MSC	13	0.99	3.55	8.33	27.61	0.91	8.92	3.22	8.74
DT	11	0.98	3.74	7.91	26.20	0.95	6.38	4.32	12.23

From Table 1, it can be concluded that the PLSR models created using the raw spectra were very accurate as indicated by the high values in both R²c (0.98) and R²p (0.99) with relatively low values for RMSEC (3.63%) and RMSEP (3.08%) compared to the models with pre-treatments. The performance of calibration models was also evaluated based on RPD and RER, and the high value of RPD (8.77) and RER (25.32) indicated the accuracy of PLSR models.

This model can be considered an excellent calibration model for detecting quinoa levels in adulterated samples due to its high R² value, a relatively low value of RMSEC and RMSEP, small difference between RMSEC and RMSEP, and a high RPD and RER value. Usually, the prediction results should always be worse than the calibration if both calibration and prediction data are in the same range. For the best model, we obtained RMSEC of 3.63% and RMSEP of 3.08%. This is mainly due to the level of ranges used for both calibration and validation. In our case, we developed the calibration in the range of 2–100% but we used validation samples in the range of 14–92%. This is why the prediction error such as RMSEP was smaller than the calibration error (i.e., RMSEC). However, RER and RPD values for both calibrations were higher than the prediction set. These results indicated that the development model was stable, not overfitting, and the majority of variance resided in the measured values was reproduced in the prediction model. Therefore, this model will be used to compare the predictability of PLSR model after selecting important wavelengths.

Most of the studies for adulteration detection using hyperspectral imaging were carried out using benchtop systems. All benchtop systems are laboratory-based, large, and heavy, and it is not possible to take the device to the food supply chain for screening foods. Therefore, developing portable/handheld devices for sensing food fraud in the food supply chain is more crucial. This model is accurate enough to use for future applications and the performance data was better than some other spectroscopic research on the adulteration of quinoa flour. The PLSR model developed by Xue et al. (2021) had the R²p value reported to be in the range between 0.939 and 0.967 for detecting adulteration levels for quinoa flour with maize and soybean flours using the front-face synchronous fluorescence spectroscopy (Xue et al., 2021). The R²p value for the best PLSR model established with NIR spectroscopy on detecting quinoa flour with wheat flour was stated by Wang et al. (2022) to be 0.94. With a higher R²p value (0.98) for the PLSR model developed in this study, VNIR hyperspectral imaging is believed to be the preferred method with better accuracy in predicting adulterations of quinoa flour. The PLSR models for adulteration detection (% adulterant) in the mixtures were also developed, and the statistical results were presented in the supplementary information (Table S1).

We also developed the PLSR model excluding samples adulterated with rice and validated the model with all validation samples adulterated with rice, wheat, soybean and corn. Table 2 shows the PLSR models created by including samples with only samples adulterated by wheat, soybean, and corn (excluding the samples adulterated with rice) in the calibration processes, and the models were validated with the original 36 validation sample sets with all four adulterants. This model shows great performance in prediction with the statistical parameters better than the original model in Table 1 with all four adulterants. The best model was found when the data were pre-treated with FD. This model has an R²p of 0.99, an RMSEP of 2.30%, a prediction RPD of 12.0, and a prediction RER of 33.9. These results proved a strong universality of the hyperspectral imaging model to be applied to detect quinoa flour in adulterated samples. Meanwhile, with an increase in the accuracy of the pre-treated data compared to the raw dataset, this model also proved the importance of applying the correct pre-treatments in certain situations. PLSR models created excluding the samples adulterated with rice in the calibration process for detecting % adulterants can be found in the supplementary information (Table S2).

Table 2.

PLSR model for detecting % quinoa excluding all samples adulterated with rice in calibration.

Adulterants (No. of Samples)	Pre-treatment	No. LVs	Calibration			Prediction
			R2c	RMSEC(%)	RER	R2p	RMSEP(%)	RPD	RER
wheat, soybean, and corn (90)	None (Raw)	7	0.99	3.48	27.6	0.99	3.16	8.81	24.7
	FD	8	0.99	2.74	35.0	0.99	2.30	12.0	33.9
	SD	8	0.99	2.42	39.7	0.98	4.13	6.88	18.9
	SNV	5	0.97	4.77	20.1	0.96	5.70	4.82	13.7
	MSC	5	0.97	4.84	19.8	0.96	5.66	4.85	13.8
	DT	7	0.99	2.75	34.9	0.92	8.52	3.37	9.20

3.3. Selection of important wavelengths to design a multispectral imaging

The major bottleneck of hyperspectral imaging is the speed of the system due to the size of the images. It takes a long time to process, display, and decide from hyperspectral data analysis, hindering real-time application. However, hyperspectral imaging is a good tool for selecting some important bands to design a multispectral imaging system without compromising the full accuracy of a hyperspectral imaging system. In fact, hyperspectral imaging can be used as a precursor to select some important bands to design a multispectral imaging system. On the other side, not all spectral bands in the hyperspectral images contain important information. Some bands have redundant, irrelevant, and useless information. Therefore, selecting some important information and discarding redundant, irrelevant, and useless information is very effective. In this study, a total 13 important wavelengths (409, 420, 423, 490, 508, 534, 602, 664, 820, 899, 954, 982, and 997 nm) were selected using BOSS to predict adulteration in quinoa flour using a low-cost multispectral imaging system.

BOSS identified important wavelengths, which were mostly (409, 420, 423, 490, 508, 534, 602, 664 nm) in the visible region (400–700 nm) and short wave NIR region (820, 899, 954, 982, and 997 nm). These absorption bands in the visible region are mainly associated with the presence of pigments which imparts the color of the samples (Kurek and Sokolova, 2020). The NIR region represents the overtones of three chemical bonds, such as O–H stretching (water or carbohydrate), C–H stretching (fat or starch) and N–H stretching (protein). The absorption bands at 820 and 954 nm are related to the fat content of the samples (Bogomolov and Melenteva, 2013; Kamboj et al., 2017; Kawamura et al., 2007), while the absorption at 899 nm is related to both protein and fat (Bogomolov and Melenteva, 2013). The absorption at 982 nm is associated with water (Kawamura et al., 2007; Omar et al., 2012; Rady and Adedeji, 2018), and the absorption at 997 nm is highly correlated with the presence of carbohydrates such as starch and sugar (Jie et al., 2004; Kamboj et al., 2017).

BOSS selected appropriate wavelengths from the full spectra. Fig. 3 shows the regression coefficients (RCs) plot of PLSR models using the full spectral range. RCs quantify the weight contribution of each variable to the regression model. The high absolute value of RCs indicate a greater contribution to the regression model and the low value of RCs indicate no or minimum contribution to the regression model. It is clear from the RCs plot of PLSR model with full spectral range that important wavelengths with a high absolute value of RCs were selected using BOSS. Therefore, important wavelength selection using BOSS was effective for adulteration detection in quinoa flour.

Fig. 3.

Regression coefficients vs. wavelengths plot and important wavelengths selected using BOSS.

3.4. Calibration models using selected important wavelengths

After applying BOSS as the variable selection methods, the number of variables was reduced from 202 to 13. PLSR model was again generated using these reduced spectral bands. Table 3 shows calibration and validation performance using only 13 important wavelengths. The results indicated that the PLSR model using 13 important wavelengths was better than the model developed using the full spectral range of 202 bands. This new PLSR model had good performance with an R²c of 0.99, an RMSEC of 3.41%, an RPD of 8.67 and an RER of 28.75. When this model was applied to validation samples, it produced an R2p of 0.99, an RMSEP of 2.93%, an RPD of 9.18, and an RER of 26.60. These results are not surprising. When the redundant/interference/less informative variables were excluded from the calibration dataset, most co-linearity problems among variables were alleviated, which resulted in a better model in terms of accuracy and robustness. Many authors also presented better results using only a few important wavelengths than the whole spectral range (Li et al., 2021a, Li et al., 2021b; Sun et al., 2021; Wang et al., 2022; Zhang et al., 2022). Generally, 2 x RMSEP is considered a 95% confidence interval in the prediction model. In this study, the 2 x RMSEP was 5.86% for PLSR model using 13 important wavelengths. This means that the PLSR model using 13 important wavelengths will be able to predict samples very well if the level of quinoa is above 5.86%. Below this percentage, the model cannot accurately predict the level of quinoa flour in adulterated samples.

Table 3.

PLSR model for detecting % quinoa with full and selected wavelengths.

No. of Variables	Pre-treatment	No. LVs	Calibration				Prediction
			R2c	RMSEC (%)	RPD	RER	R2p	RMSEP (%)	RPD	RER
202	Raw	11	0.98	3.63	8.15	27.00	0.99	3.08	8.77	25.32
13	Raw	12	0.99	3.41	8.67	28.75	0.99	2.93	9.18	26.60

Hyperspectral imaging is an indirect method based on calibration, which is commonly developed using a sample set incorporating all variances usually encountered during routine sampling. The properties of agricultural crops may vary over time due to changes in growing conditions, variety, locations, harvesting, processing, storage and other external factors. As a result, collecting data for calibration that includes all variations in future predictions is sometimes challenging; consequently, calibration will only work for a specific period. After that, calibration performance will be less accurate. Therefore, periodic maintenance and updates of calibration model are necessary to ensure the continued stability and reliability of calibration.

3.5. Visualization maps

Each pixel in the hyperspectral image has a spectrum. That means it is possible to predict attributes in each pixel in the images. But it is not possible to determine attribute values in pixel levels in the laboratory. However, applying the calibration model to each pixel spectra is possible for pixel level detection. This study used the calibration model at selected wavelengths to create visualization maps to predict adulteration in each pixel. This visualization map was created by applying PLSR model at selected wavelengths in each pixel in the images and it was done by using a dot product between each pixel of the selected image bands and the regression coefficients obtained from the PLSR model. Fig. 4 shows the visualization map (also called prediction map) for each pixel in the image to predict % of quinoa flour in a binary mixture of (Quinoa + wheat, quinoa + corn, quinoa + soybean, quinoa + rice). In the visualization maps, pixels with similar spectral characteristics produced similar predicted values and it was shown with a linear color scale. The visualization maps indicate how adulteration level varies from pixel to pixel within the same sample or from sample to sample. Using RGB images or human eyes, it is challenging to see the adulteration of quinoa flour in each binary mixture; however, hyperspectral imaging revealed pixel-wise hidden information about the samples. Although the detection of adulteration in flour is a challenging task, the results indicated that VNIR hyperspectral imaging could be used as a rapid tool for predicting adulteration in flour. The spatial distribution of particular attributes in the image is the most promising application of hyperspectral imaging. However, if the calibration model is inaccurate, a misleading distribution map will be obtained.

Fig. 4.

Visualization map for the prediction of % quinoa.

4. Conclusions

Food fraud is a persistent global problem due to fraudulent practices and complex food supply chain. Therefore, fraud detection is challenging for researchers, industries, and regulatory bodies. This is the first reported study to use portable hyperspectral imaging for detecting adulteration in quinoa flour for the food supply chain. A global model was created to detect the percentage of quinoa in adulterated samples. The results indicated that portable VNIR hyperspectral imaging combined with appropriate chemometrics could become a promising tool for detecting adulteration in the food supply chain. Only 13 important wavelengths are enough to design and develop a low-cost multispectral imaging system. The PLSR model with only 13 wavelengths showed strong performance for prediction in unknown samples (R²p = 0.99, RMSEP = 2.93%, RPD = 9.18, RER = 26.60). This model was applied to each pixel in the images to show the visualization map that varies from sample to sample or from pixel to pixel within the same sample, which is an added benefit of hyperspectral imaging that other techniques cannot achieve. Although calibration results are promising, more research is needed to develop a calibration to predict a relatively small amount (<5%) of adulterant in quinoa flour.

CRediT authorship contribution statement

Qianyi Wu: Conceptualization, Methodology, Formal analysis, Investigation, Writing – original draft, Writing – review & editing. Magdi A.A. Mousa: Writing – review & editing, Funding acquisition. Adel D. Al-Qurashi: Writing – review & editing, Funding acquisition. Omer H.M. Ibrahim: Writing – review & editing, Funding acquisition. Kamal A.M. Abo-Elyousr: Writing – review & editing, Funding acquisition. Kent Rausch: Writing – review & editing, Funding acquisition. Ahmed M.K. Abdel Aal: Writing – review & editing, Funding acquisition. Mohammed Kamruzzaman: Conceptualization, Methodology, Resources, Writing – review & editing, Supervision, Funding acquisition.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following is the Supplementary data to this article.

Data availability

Data will be made available on request.