In this work, argan oil, noted AO, was used to assess its adulteration by two commercial vegetable oils noted VO1 and VO2. VO1 is commercial sunflower oil and VO2 is commercial edible oil sold without any indication of its origin. Cheap vegetable oils (VO1 and VO2) were purchased from a market in Oujda (East of Morocco) and pure argan oil (handmade) was obtained from Agadir (south of Morocco).

The samples were divided into two groups. The first one contained samples of AO adulterated by VO1 (54 samples), and the second one contained AO adulterated by VO2 (55 samples). Cheap oils were heated at 200 °C for 30 minutes many times, until the oil had the same color as AO. VO1 and VO2 were heated for the reason that fraudsters use waste frying oil to adulterate argan, since this oil has the same color as argan oil and the detection of adulteration becomes more difficult. The process of adulteration was achieved as follows: AO was adulterated by VO1 from 0 to 31% and by VO2 from 0 to 32%. A total of 109 samples were prepared. The first sample in each group contained 100% AO and the second one contained 99.6% AO and 0.4% adulterant (VO1 or VO2). The procedure was repeated for all samples, increasing the number of drops of adulterant for each sample by one drop. Therefore, for each sample, a drop of pure AO oil was replaced by a drop of adulterant. The drops were added using a micropipette and masses were measured using a digital scale of very high sensitivity. Finally, the prepared samples were homogenized and stored in the dark at ambient temperature.

LIF is a physical phenomenon in which a molecule absorbs an amount of the energy from a laser beam. There is thus a transition from a ground state S₀ to an excited state S₁ with a change in the electron orbital. This excited state S₁ has a very short lifetime (a few nanoseconds). Changes in conformation and interactions with surrounding molecules change the molecule from the excited state S1 to low vibrational levels of S₁: this is the internal conversion. In the case of fluorescent molecules, the transition from the excited state S₁ to the ground state S₀ takes place with the release of a photon of lower energy. This phenomenon is laser induced fluorescence, which occur at wavelengths greater than the incident excitation wavelength (Mazouffre 2009). The LIF system used in this work was an assembled system. The spectral measurement of the LIF technique was obtained by irradiating the samples with an Nd: YAG laser beam (532 nm) set at 2 mW, with 3.3 V DC operating voltage and 0.4 A current. During irradiation, the fluorescence emitted by the samples was collected by an optical fiber (2 m long and 400 μm in diameter), equipped with an SMA type connector at each end. The axis of the optical fiber was positioned at an angle of 90° to the direction of the laser beam in order to collect as much fluorescent signal as possible and avoid the detection of photons from the incident beam. The beam was dispersed by a spectrophotometer Avantes, model USB2000, grating 600 lines, blazed at 750 nm, L2 lens, 100 µm slit, equipped with a 2048-pixel charge coupled detector CCD. It consisted of a linear array of silicon (Si) diode and was linked to a computer by a USB connection in order to visualize the fluorescent spectra of each sample. The spectrophotometer used in this work covered the spectral range 500-1000 nm. Note that the spectra were performed in three replicates, under the same conditions. Analyses were made using the average.

The obtained spectral data were converted into Microsoft Office Excel format for Matlab software analysis. The spectral range was reduced from 500-1000 nm to 540-750 nm. Then, to reduce noise and baseline shifts, the spectra were corrected using preprocessing, whose objectives were the attenuation of non-linearity between variables, the elimination of interference and reduction of random noise (Verboven et al., 2012).

Before multivariate data analysis, all LIF spectra were subjected to Savitzky-Golay smoothing (1 point, 2 orders) then the spectra obtained were subjected to a Multiplicative Scatter Correction (MSC) combined with a second derivative for PCA analysis and combined to Baseline for PLS modeling (Savitzky and Golay, 1964; Zheng et al., 2015). MSC is a diffusion correction method that was introduced in 1983 and developed in 1985 (Guidetti et al., 2012). It is based on the idea of correcting the level of dispersion of all sample spectra from an “ideal” spectrum, which is usually the average spectrum. The concept behind MSC is that artifacts or imperfections will be removed from the data matrix (spectra) prior to modeling. The MSC consists of two steps:

The first and second derivatives are used for baseline variation reduction and separation of overlapping bands, so hidden bands are enhanced. With regards to baseline preprocessing, most correction methods make the supposition that the observed spectrum is the combination of a useful signal and a signal of uncontrolled variation. Therefore, the correction consists of subtracting the background from the obtained signal.

The wavelength range used for LIF analysis was reduced to 540 - 750 nm to keep only the part that contains relevant information and to eliminate noise. PCA was exploited to get main information from spectra and reduce the number of variables. Then PLS algorithm was applied on LIF spectra to establish a model that can predict the percentage of adulteration. PLS calibration gives optimum results compared to many other multivariate calibration methods. An important aspect of PLSR is that it collects the relevant spectral information in a few linear combinations of the spectral measurements. These combinations or components can be used to facilitate interpretation of the relationship between concentrations and spectra as well as the relationships among the spectral variables themselves.

The samples were divided into calibration/validation and prediction datasets. The optimum number of latent variables was obtained using the full cross validation method. The prediction performance of each model was evaluated throughout the root mean square error (which represents the standard deviation of the residuals), the prediction standard errors, and the coefficient of determination (R²) of both calibration and validation data sets. In general, as low as the RMSEC/P and SEC/P values can be, and R² as close as possible to 1, the better the model’s predictions will be. Equations corresponding to each parameter are:

Figure 1 shows an overlay of LIF spectra of AO, VO1 and VO2 oils. The spectrum of VO2 oil showed emission at around 650 - 700 nm, This emission was due to the fluorescence of chlorophyll compounds. The AO spectrum showed a fluorescent band around 550 - 625 nm, although this emission was practically absent from the VO1 and VO2 spectra. Figure 2 represents LIF spectra of the same samples after heating VO1 and VO2. The spectra showed the same shape with different intensity. In this figure, the absence of the peak in the red region 650 - 700 nm could be observed, which means the absence of chlorophyll compound emission. It can be concluded from Figures 1 and 2 that heating produced the degradation of the Chlorophyll.

Figure 3 shows seven spectra in the spectral range of 544 - 750 nm. The first spectrum correspond to AO’s fluorescence and the other spectra represent the fluorescence of the adulterated samples, showing the tendency of the spectra according to the concentrations of adulterant. It can be seen from Figure 3 that the spectra are similar to each other and they are a bit noisy. A vertical baseline shift and intensity difference between the spectra can also be seen. Actually, when an oil sample is adulterated, whether the amount of adulterant is small or large, the whole spectrum of the pure sample is affected. These deformations are corrected using chemometric pretreatments. In order to reduce variability in the spectra and improve the signal-to-noise ratio, the spectra were fitted using smoothing preprocessing and then MSC combined to the second derivative (Savistky - Golay) was applied as a suitable preprocessing for PCA analysis and MSC combined to the baseline for PLS analysis. After the spectra were pretreated, chemometric methods were used to extract hidden information.

After preprocessing the spectra using smoothing combined to MSC and the second derivative, PCA was performed to explore the similarities and differences in samples, to extract relevant information and reduce the number of variables. Figure 4 presents the score plot of PCA applied to LIF spectra. A total 109 samples were measured in this work. The first two principal components accounted for approximately 73% of the total variability in the data. PC1 explained 72% of the total variance and 1% of the information was explained according to PC2. The interpretation of the results was made via loading plots; they describe how much each variable contributes to a particular principal component. Large loadings (positive or negative) indicate that a particular variable has a strong relationship to a particular principal component. There was a strong relationship between noise in the loadings and the optimal number of components. The concepts of underfitting, overfitting and optimal prediction ability have an important connection to the shape of loading plots, which can be seen when plotting loadings versus wavelength number for all the components computed. For the first few components corresponding to the largest eigenvalues, the shapes of these plots looked like spectra, but as we went towards the components with smaller eigenvalue, we introduced more and more noise (overfitting) and this is sometimes clearly seen as ripples and irregularities in the loading plots. There was a tendency, that close to the optimal number of components, the noise started to become clearly visible in these plots.

After applying PCA to the pretreated spectra, samples were significantly classified. It can be seen from Figure 4 that PCA can distinguish samples from each other successfully. Adulterated argan oils are different after passing through the PCA process, and it can be seen from the figure that objects were dispersed according to a two-dimensional space (PC1 and PC2) and divided into two sets according to the type of adulterants. The lower right group in the score graph consists of AO samples adulterated with VO2, noted group B and the higher left group consists of AO samples adulterated with VO1, noted group A. By projecting samples according to PC1, it can be easily noticed that the first component contained information which was responsible for the separation between samples; a low separation existed by projecting samples according to PC2. Group B is located in the negative part of PC2 and group A is located in the positive part of PC2. It can be concluded that, by this model, it was possible to determine the type of adulterant for an unknown sample by taking its spectrum and injecting it into this model.

PLS was exploited as a multivariate calibration technique. It constructs a mathematical model based on the features of PCA and multiple regression to find a linear mathematical relationship between two datasets, X (spectra) and Y (level of adulteration) (Morsy and Sun, 2013).

The spectral data were arranged in a 2D matrix (X), the rows of this matrix represent the samples (109 samples) and the columns contain the number of variables. One column vector (Y) containing the concentration of each adulterant was added to this matrix and the data were then analyzed. To make sure that the obtained model was neither over nor underfit, a cross validation using the leave-one-out method was considered. The linearity of the regression model was evaluated by fitting the reference adulteration value against the predicted ones.

The above PCA model successfully identified the type of cheap edible oil added to AO. To further predict the contents of VO1 and VO2 in the blended oil samples, PLS regression was performed. This method has been successfully used in several studies to predict the percentage of adulterants in oils. PLS regression was applied on raw and pretreated spectra and the one that gave the best result was kept. In this work, LIF data were preprocessed by taking a smoothing (1 point) combined to MSC and Baseline. After removing the outliers, the set of samples was randomly divided into three sample sets samples, in which two sets were used as the experimental group, calibration/validation sets, containing almost 85% of the total set of samples (70% calibration, 15 % validation) and one set as the testing group containing the 15% remaining.

The efficiency of the calibration model can be determined generally from three statistical parameters: the correlation (R), the standard error of calibration (SEC) and the mean square error of calibration (RMSEC). a R value greater than 0.9 indicates a good response for the parameter studied and less than 0.7 indicates a poor response. When a new spectrum is inserted into the matrix, it is compared to the spectral basis, the closest spectrum is then used to make a PLS regression.

The graphic display of the calibration/validation and prediction produced using the PLS model with the best performances is shown in Figure 5. The calibration results of 109 samples using PLSR produced excellent overall models with a correlation greater than 0.99 and low values for RMSEC and SEC. The same was true for the predicted model. It gave excellent statistical values with a correlation of 0.92, RMSEP and SEP of about 2%. The results indicate that the method proposed in this work is feasible for the detection and quantification of argan oil adulteration by cheap vegetable oils. Table 1 shows all the statistical results of the calibration and prediction models. A similar study was published using near infrared spectroscopy to evaluate the adulteration of argan oil by cheap oils (Farres et al., 2019). When comparing the results obtained in this work to those obtained in the published study, it can be easily seen that the R value is higher and error values are low. Therefore, it can be concluded that LIF spectroscopy was revealed to be most suitable technique to study the adulteration of argan oil, especially since both techniques were carried out on the same samples. Table 2 groups together the results obtained using near infrared spectroscopy (published work) and laser-induced fluorescence spectroscopy.

To prove the efficiency of these models, they were tested on two non-synthetic samples and the results are shown in Figure 6. Two non-synthetically adulterated samples were inserted into the PCA model, and as can be seen from the figure, one of samples was close to group A, and the second one was close to group B. This mean that the first sample was adulterated by VO1 and the second one with VO2.