Detection of Melamine in Soybean Meal Using Near-Infrared Microscopy Imaging with Pure Component Spectra as the Evaluation Criteria

Soybean meal was adulterated with melamine with the purpose of boosting the protein content for unlawful interests. In recent years, the near-infrared (NIR) spectroscopy technique has been widely used for guaranteeing food and feed security for its fast, nondestructive, and pollution-free characteristics. However, there are problems with using near-infrared (NIR) spectroscopy for detecting samples with low contaminant concentration because of instrument noise and sampling issues. In addition, methods based on NIR are indirect and depend on calibration models. NIR microscopy imaging offers the opportunity to investigate the chemical species present in food and feed at the microscale level (the minimum spot size is a few micrometers), thus avoiding the problem of the spectral features of contaminants being diluted by scanning.The aim of this work was to investigate the feasibility of usingNIRmicroscopy imaging to identifymelamine particles in soybeanmeal using only the pure component spectrum.The results presented indicate that using the classical least squares (CLS) algorithm with the nonnegative least squares (NNLS) algorithm, without needing first to develop a calibration model, could identify soybean meal that is both uncontaminated and contaminated with melamine particles at as low a level as 50mg kg.


Introduction
Soybean meal is one of the most important feed raw materials and one of the main ingredients in compound feed because it has a complete protein profile.In the past decade, the price of soybean meal has tripled (http://faostat3.fao.org/home/index.html).The price is dictated by the protein content: the higher the content, the higher the price.There have been recent cases of soybean meal being adulterated with melamine (1,3,5-triazine-2,4,6-triamine) in order to boost the protein content [1].There have also been cases where it was suspected that soybean meal had been contaminated with low melamine levels.The reference methods (wet chemistry) usually are time-consuming and expensive, cause damage to the sample, and need chemical reagent [2][3][4][5].So there is a real need for fast, nondestructive, and automatically controlled screening methods that will guarantee quality and security.
Near-infrared (NIR) spectroscopy is widely used in this context because it is a rapid, nondestructive, and nonpolluting method that requires minimum or no sample preparation [6].Applications vary from sample chemical composition to detecting adulteration or contaminant ingredients in raw materials and compound feed [7,8].Many studies have investigated the feasibility of using NIR to detect melamine [9][10][11][12][13][14], but there are still some problems here when it comes to testing samples with low melamine content because of the instrument noise and sampling design error [15,16].NIR can only acquire the spectra information but cannot provide the spatial information of the sample.NIR microscopy (NIRM) which combines NIR and digital images together is able to collect the spectrum of individual particles (the minimum spot size is a few micrometers) from samples [17].With the development of a high-precision - motion stage, NIRM imaging is obtained by successively measuring spectra while the sample is repositioned in the  and  spatial dimensions.Thus, the technique offers an opportunity to explore not only what kinds of chemical species are present at the microscale level, but also where they might be present [18].NIRM imaging is a type of NIR imaging and has better spatial resolution.In the past decade, NIR imaging has become a powerful analytical tool for detecting contaminants and defects in agrofood products [19][20][21][22][23][24][25].For feed products, several investigations have been conducted on the feasibility of using NIRM to detect, identify, and quantify processed animal byproducts [26][27][28][29][30] and on using NIR imaging for the complete screening of compound feeds [31].
The aim of this study was to investigate the feasibility of using NIRM imaging to identify melamine particles in soybean meal.The classical least squares (CLS) algorithm with the nonnegative least squares (NNLS) algorithm was used for the analysis of soybean meal adulterated with melamine.

Experimental Design.
A total of 20 soybean meal samples were collected from various feed manufacturers and were ground to pass through a 0.5 mm sieve.And 3 melamine samples were bought from three chemical reagent companies, with 99% of the particles being less than 0.25 mm (analytical reagent, C 3 H 6 N 6 content certified by the company ≥99.5%).For this study, three sets of samples were used.Set A was prepared artificially in order to develop a method without the need to create the calibration model.Set B was prepared using representative matrixes to validate the reliability and robustness of the method.Set C was prepared with low melamine content in order to investigate the sensitivity of the method.

Experimental Sample:
Set A. Samples were artificially prepared on Teflon Spectralon using soybean meal and melamine particles (Set A-1, Figure 1(a)): a melamine particle (m2#) was placed on top of a soybean meal particle (s2#); a melamine particle (m4#) was placed under a soybean meal particle (s4#); a melamine particle (m3#) and a soybean meal particle (s3#) were clustered together; a single melamine particle (m1#) was used alone; and a single soybean meal particle (s1#) was used alone.
If a melamine particle is placed under a soybean meal particle, the thickness of the latter would affect the detection of the melamine particle.In order to study how this thickness affected the method, other samples were artificially prepared with different soybean meal thicknesses (Set A-2, Figure 1(b)).S1, S2, S3, and S4 relate to soybean meal that was about 30, 50, 100, and 110 m thick, respectively.

Experimental Sample: Set B.
For Set B, using 20 soybean meal and 3 melamine samples, 20 sample mixtures contaminated with 5, 10, 25, and 50 g kg −1 of melamine, respectively, with five replicate samples for each melamine content level, were prepared with a mixer (REAX 20/8; Heidolph, Schwabach, Germany) in the laboratory (Table 1).In order to achieve a homogeneous distribution of melamine in the soybean meal, a stepwise dilution procedure was applied to ensure that in each dilution step the ratio of the two materials to be mixed did not exceed a factor of 3 [32].1, 4 sample mixtures contaminated with 1,000, 500, 100, and 50 mg kg −1 of melamine, respectively, were prepared in the laboratory using 4 soybean meal and 3 melamine samples.

Data Acquisition.
All the samples (Set A, Set B, Set C, and 20 pure soybean meal and 3 pure melamine samples) were analyzed using a NIRM imaging system (Spotlight400, Perkin Elmer), with 16 spectra being acquired simultaneously from the line detector.The detector in this instrument is a mercury cadmium telluride (HgCdTe or MCT).The spatial resolution of scanning is 25 × 25 m and the spectral resolution is 32 cm −1 .Ratio spectra ( =  sample / teflon ) were collected using a Teflon Spectralon (Spectralon5 Targets, Labsphere, Inc., North Sutton, New Hampshire) as the reflectance standard and then converted into absorbance () by  = log 10 (1/).Each spectrum was the average of four scans across the wavenumber range of 7,808-4,000 cm −1 .One image (an area of 8.75 × 8.75 mm coupled with 350 × 350 pixels; 122,500 spectra) was scanned for each of Set B and Set C (samples with 1,000, 500, and 100 mg kg −1 melamine), respectively.In Set C, four images of samples with 50 mg kg −1 melamine were scanned, because it was more difficult to identify melamine particles at low concentrations.
One image (an area of 1.25 × 1.25 mm coupled with 50 × 50 pixels; 2,500 spectra) was scanned for each of the 20 pure soybean meal and 3 pure melamine samples.

Data Analysis
2.3.1.Preprocessing.The main purpose of preprocessing was to remove spectral and spatial artifacts such as rough surfaces, optic effects, and detector noise.The NIRM imaging data cube and the pure component spectra were both preprocessed by applying a first derivative using the Savitzky-Golay algorithm with a five-point filter width and a degree 2 polynomial [33].The noisy part at the end of the spectra was removed by reducing the spectral range to between 7,300 and 4,100 cm −1 .

Chemometric Tools.
The classical least squares (CLS) algorithm was used to extract melamine distribution maps from whole wavelengths [34,35].This algorithm is a suitable method involving minimizing the sum of squared residuals in order to predict concentrations using reference spectra only.It was based on the assumption that the absorbance spectra from a pixel in NIRM imaging can be viewed as the weighted sum of the absorbance of each pure component spectrum constituting the sample, as well as the experimental noise.Initially, the data cube ( ×  × ) was unfolded into a two-dimensional matrix  = ( × ).Matrix  was then decomposed as follows: where   represents the pure component signals;  is the relative concentration matrix; and  is the error matrix.
was estimated by the pseudoinverse  = (  ) −1 using the nonnegative least squares (NNLS) algorithm.The concentration of melamine and soybean meal, respectively, was then calculated as follows [36]: where  mel (g kg −1 ) is the melamine concentration of sample;   is the melamine concentration of each pixel in the NIRM image predicted by CLS;   is the soybean meal concentration of each pixel in the NIRM image predicted by CLS;  is the number of pixels in one NIRM image; and  soy (g kg −1 ) is the soybean meal concentration of the sample.
All   or   values formed a chemical image in which higher pixel intensity reflected higher target concentration.Matlab5 (The MathWorks, http://www.mathworks.com/) and the PLS Toolbox (Eigenvector Research, http://www.eigenvector.com/)were used to perform this analysis.

Results and Discussion
3.1.Soybean Meal and Melamine Component Spectra.The melamine reference spectrum was obtained by averaging all the pixel spectra of 3 pure melamine images, and the soybean meal reference spectrum was obtained by averaging all the pixel spectra of 20 pure soybean meal images (Figure 2).Melamine is a triazine heterocyclic organic compound, composed of nitrogen heterocyclic rings and -NH 2 .The NIR region offers a special advantage in the measurement of the primary amine NH 2 group because of a unique combination band [37].Melamine has three strong characteristic peaks in the 6,900-6,450 cm −1 range, especially near 6,805 cm −1 , which is the N-H combination band (]N-H asymmetric and ]N-H symmetric combination) from primary amides.Figure 2 showed that melamine has a distinctive spectral feature compared with soybean meal, facilitating the identification of melamine particles present in mixtures.2 showed the minimum   of 20 soybean meal samples which were calculated by CLS.Meanwhile, the mean value of the minimum values of   is 0.8822 and the standard deviation is 0.0484.In consideration of the conciseness and effectiveness, the liminal value of soybean meal's   was 0.8. 3 showed the detection results of Set A samples using CLS algorithm.Figure 3(a) was a soybean meal image produced by CLS.Four soybean meal particles (s1#, s2#, s3#, and s4#) could be correctly identified.Figure 3(b) showed a melamine image produced by CLS.CLS could detect the presence of melamine m1#, m2#, and m3#, but not m4# which was placed under a 150 m thick soybean meal particle.In order to investigate the effect of soybean meal thickness on the detection of melamine under the soybean meal particle using the CLS algorithm, an image of sample Set A-2 was obtained and analyzed using CLS.When the soybean meal was 30 m and 50 m thick, CLS could detect the presence of melamine, but it could not do so if the melamine was under soybean meal that was thicker than 100 m.In addition, the signal of the melamine under soybean meal that was 50 m thick was weaker than the signal when the thickness was 30 m, illustrating the difficulty of detecting the presence of melamine when melamine particles were embedded in soybean meal particles.Sample Set B containing 20 contaminated samples were analyzed using the CLS algorithm and the results were shown in Table 3.The concentrations of melamine calculated by CLS were 0.84±0.30,3.51±0.75,8.17±1.21,and 32.23±4.33g kg −1 for the 5, 10, 25, and 50 g kg −1 melamine content samples, respectively.The melamine in the 20 mixture samples was successfully detected.Although the melamine concentration was underestimated by CLS, there was a good linear relationship between the predicted and real values (predicted values = 0.57 × real values,  2 = 0.90).Particle size, surface roughness, and density of the sample spread on a Teflon Spectralon surface in a single layer were important factors in quantifying melamine correctly.It is also likely that the melamine concentration was underestimated because of the difficulty in detecting melamine if the particles were under or tightly embedded in the soybean meal particles, as shown in sample Set A.

Melamine Detection Using CLS. Figure
The pixel spectra corresponding to the   maximum of melamine for the 20 sample mixtures are shown in Figure 4.These melamine spectral characteristics indicated reliable extraction by CLS.In addition, the results showed that NIRM imaging data combined with the CLS algorithm could provide a visual melamine distribution map (Figure 5) for the analyst, which could not be done by NIR because the spatial information is lost.Sample Set C was used to investigate the method's sensitivity.  < 0.8 and GH > 3 were combined together for melamine detection in sample Set C. The GH is the standardized Mahalanobis distance between each spectrum and average spectrum, and GH > 3 is used to identify the outliers [38,39].In this work, 480 representative spectra were picked out from the 20 pure soybean meal images.And the GH values between the target spectrum and the 480 soybean spectra were calculated to identify whether the target spectrum is soybean spectrum or not.
As shown in Figure 6, all the 1000, 500, and 100 mg kg −1 of melamine images had spectra out of both GH and   liminal values.Two of four 50 mg kg −1 melamine images had spectra out of both GH and   liminal values.Using GH and   liminal values to analyze 20 pure soybean meal samples, there were also some abnormal spectra in Soy 6.The abnormal spectra detected by   < 0.8 and GH > 3 of Set C and Soy 6 were shown in Figure 7.All the spectra separated from Set C had the melamine spectra characteristics which could be recognized by visual inspection.Meanwhile the spectra separated from Soy 6 were not similar with melamine spectrum or soybean meal average spectrum; this part of the spectra was more like fullfat soybean which could be seen in the article published by Shen et al. [39].

Figure 2 :Figure 3 :
Figure 2: Average spectrum of soybean meal and melamine particles: raw spectra (a) and first derivative spectra (b).

Figure 5 :Figure 6 :Figure 7 :
Figure 5: The melamine distribution map predicted by CLS applied to the NIRM imaging data.Only one sample was taken as an example of each concentration.(a), (b), (c), and (d) represent 5, 10, 25, and 50 g kg −1 of melamine, respectively.

Table 1 :
Description of the samples Set B and Set C.

Table 2 :
The minimum   of 20 soybean meal samples calculated by CLS.

Table 3 :
Discriminant results of the analysis of the Set B samples using CLS.