Three different pomegranate cultivars (cv. Wonderful, Acco, Herskawitz) were procured from Sonlia pack-house, Wellington, Western Cape region. A total of 180 fruit or 60 fruit per cultivar was used for this study. At the research laboratory, fruit without any physical defects was sorted and manually cut open for the edible aril portion at ambient conditions (21°C ± 65% RH). Cheesecloth was employed to separate kernels from the arils. Kernels were extracted from arils and then washed with distilled water to eliminate the residual aril sacs before being dried at a temperature of 60°C for 24 h in a hot air oven (PROLAB, South Africa). Pomegranate kernels were dried to a moisture content of 1.7 wt. % (dry basis). After drying the seed the final seed weight averaged 12 ± 2.5 g per fruit. Dried pomegranate kernels were then packed in a polyethylene bag and stored at −20°C until further processing.

In this study, pomegranate oil was extracted using the solvent extraction method as described by Ampem (2017). Dried kernels were grinded into a powder with a particle size of 0.25 mm using a Sunbeam coffee grinder (Model SCG-250, 60 g capacity, South Africa) in preparation for oil extraction (Eikani et al., 2012). Hexane solvent was used to extract oil from the kernel powder. Pomegranate kernel powder (30 g) was weighed into a glass flask and extracted twice, respectively, with 300 ml of hexane solvent at a time, reaching a total volume of 600 ml solvent solution for each sample. The mixture (600 ml) was sonicated in an ultrasonic bath (Model DC 400H, Haifa, Israel) which was operated at 40°C for 40 min. The oil filtrates from repeated extractions were pooled and recovered through distillation using a rotary evaporator (Heidolph Instruments GmbH & Co. KG, Germany). Thereafter, samples were placed within a vacuum oven at 60°C for 1.5 h to remove any remaining hexane solution (Parashar et al., 2009). A total of 6 ml oil was obtained from each fruit and transferred into a 9 ml glass tube and stored in a dark environment at room temperature until further analysis. A total of 45 oil samples composed of 15 samples each from three different cultivars (Acco, Wonderful, Herskawitz) were used for this study.

The Alpha-P ATR FT-IR spectrometer (Bruker Optics, Ettlingen, Germany) and the Multi-purpose analyser (MPA) spectrometer (MPA, Bruker Optics, Ettlingen, Germany) were used for spectral data acquisition. Samples were kept in 8 mm glass vials, and sample temperature was maintained at ±50°C using a heating block before spectra recording. This was to ensure that sample temperature was stable as studies have shown temperature to impact the intensity of the bands (Jiang et al., 2008; Cayuela and García, 2017; Özdemir et al., 2018). The temperature of 50°C was chosen through preliminary trials and consultation with Bruker Optics, South Africa. For the MPA spectrometer, the spectral data were acquired over the range of 12,500 to 4,000 cm⁻¹ (scanning resolution 4 cm⁻¹; scanner frequency 10 kHz; background with air, 128 scans). The spectral acquisition occurred almost immediately for the Multi-purpose analyser (MPA) spectrometer since the instrument does not have a temperature control system. For the Alpha-P ATR FT-IR spectrometer, sample spectral data were acquired over the range of 4,000–400 cm⁻¹. The Alpha-P spectrometer was equipped with a diamond crystal plate (area 2 mm²) that maintained the sample temperature at 50°C. The temperature was monitored using OPUS software and spectral acquisition would only occur when the diamond crystal plate and sample reached a temperature of 50°C. The average time taken to acquire spectral data for one sample was 120 s using the following instrument settings: 4 cm⁻¹ resolution scan, 10 kHz scanner frequency and 128 averaged scans per spectrum. The sample stage was cleaned in-between measurements with soft paper and undiluted methanol to avoid cross-contamination (Foudjo et al., 2013).

All chemical reagents were obtained from Sigma–Aldrich–Fluka Co. Ltd. (South Africa) unless otherwise stated.

The spectral acquisition occurred with OPUS software (version 7.0), while data processing and analysis were achieved with SIMCA and OPUS software. Qualitative analysis (modeling of cultivar difference) was carried out using PCA and OPLS-DA using SIMCA software, and quantitative analysis (developing calibration models) was carried out with PLSr using OPUS software. For this study, several preprocessing methods were evaluated, baseline correction spectra were subjected to several filtering techniques, which included Savitzky–Golay transformation (first derivative), multiplicative scattering correction (MSC), and standard normal variate (SNV) correction. Separate OPLS-DA models were built for both NIR and MIR spectral data, each pair of two successive stages by using a dummy variable with a value of 1 assigned to samples that belonged to a specific group and a value of 0 to samples that did not belong to that group.

To demonstrate that the prediction of the different selected quality parameters is from the actual IR spectra and not due to possible correlations with the other measured parameters, the reference data was subjected to Pearson’s correlation test using Statistica software (Statistica 16.0, StatSoft Inc., Tulsa, OK, United States).

For this study, reference data for the different parameters were normally distributed around the mean (Table 1). According to Lu et al. (2006), the accuracy and validation of calibration models normally depend on large variation in the present within the sample set in the physical and biochemical reference data. However, reports have indicated superior model accuracies using NIR spectroscopy when data with a large sample variation within the calibration and validation set is being considered (Magwaza et al., 2013, 2014). Table 2 presents the standard deviation, minimum-to-maximum range, and CV% statistics of most of the parameters. Most parameters had high CV% values of up to 43% for both calibration and validation data sets covering a wide range of values, aside from the refractive index. Pearson correlation was applied to investigate the interrelationships between selected reference data of pomegranate oil. From the result, it can be deduced that the prediction of these quality parameters (phenolics, carotenoids), and their concentrations should not correlate with one another. Correlation tests indicate that no correlation was observed between chemical indices such as phenolic and carotenoid content (0.227). Similarly, phenolic and carotenoid content showed no correlation with peroxide value (−0.009, 0.334) refractive index (0.260, 0.176) or oil yellowness index (0.172, 0.215). These results suggest that the prediction of the different studied parameters is actually from the IR spectra.

NIR spectroscopy is a powerful non-destructive technique used for the detection of various compounds, the NIR spectrum provides information on the vibrational absorption of hydroxyl (O–H), amido (N–H), and C–H bonds. The average for both NIR and MIR spectra of pomegranate oil is presented in Figure 1. Pomegranate is highly abundant in punicic acid (C18:3-9c), linoleic acid (C18:2), and oleic acid (C18:1). Band assignment was done according to the literature (Foudjo et al., 2013; Inarejos-García et al., 2013; Özdemir et al., 2018). In the NIR region, bands around 8,451 cm⁻¹ arise from second overtones of C–H stretching vibrations, while those at 7,502 and 7,498 cm⁻¹ are due to the combination bands of C–H in fatty acids. The bands at 5,774 and 5,450 cm⁻¹ can be ascribed, according to literature, from the first overtone of C–H stretching vibrations of methyl, methylene, and ethylene groups (Sinelli et al., 2011; Özdemir et al., 2018). Small bands at 4,659 and 4,597 cm⁻¹ are associated with combination bands of C–H and C–O stretching vibration. Several bands dominate the MIR spectra at 2,918, 2,556, 1,837, 1,463, 1,377, 1,238, 1,163, 1,114, 1,099, and 721 cm⁻¹. The absorbance band at 3,013 cm⁻¹ has been associated with the stretching of the functional group =C-H (cis-) found in unsaturated fatty acids such as punicic acid (Guillén and Cabo, 1997). Absorbance at 2,924 and 2,852 cm⁻¹ are due to bands from asymmetric CH₂ stretching vibration of acyl chains and methylene chains in fatty acids (punicic acid, linoleic acid, and oleic acid; Guillén and Cabo, 1997; Sun, 2009). The major band at 1,743 cm⁻¹ arises from C=O stretching vibrations of ν(C=O) ester in fatty acids (Guillén and Cabo, 1997, Sun, 2009). The band at 1,238, 1,163 and 1,114 cm⁻¹ has been associated with C-O or CH₂ stretching or bending vibration out-of-plane of functional groups from fatty acids (Guillén and Cabo, 1997; Rohman and Che, 2011). The band at 721 cm⁻¹ corresponds to the CH₂ rocking mode (Yang et al., 2005; Sinelli et al., 2010). The spectral profile for pomegranate oil is comparable to those reported for other oil samples like avocado oil (Foudjo et al., 2013), virgin olive oil (Dupuy et al., 2010; Sinelli et al., 2010), rapeseed oil blend (Ma et al., 2014), and palm oil (Mba et al., 2014).

The best FT-NIRs and FT-MIRs models were developed using 17 points, first derivative, second derivative, and straight-line subtraction, respectively. The model for each parameter was selected based on the evaluation of statistical parameters that gave higher R², high RPD values, lowest RMSEV and RMSEP, and lowest number of latent variables. The overall performance of the developed models for all quality parameters is represented in Table 2. Scatter plots of FT-NIR and FT-MIR spectroscopy for predicted data plotted against measured reference data are presented in Figure 4. Models developed in the NIR and MIR spectral regions had a major influence on the regression statistics. All three cultivars were combined to create models with high robustness and variability.

The refractive index is an intrinsic property of oil measured based on light penetration through an oil sample (Aydeniz et al., 2014; Khoddami et al., 2014). Oil refractive index has been reported to be directly proportional to the degree of unsaturation of fatty acids and inversely related to its viscosity and can, therefore, be used to quantify the double bonds of fatty acids (Aydeniz et al., 2014; Khoddami et al., 2014). The statistical indicators for model fitness showed that both NIR and MIR spectra yielded relatively accurate PLS models for refractive index. PLS model development in the FT-MIRs provided slightly better prediction statistics (R² = 0.912, RMSEP = 0.0002, and RPD = 3.43) compared to the FT-NIRs (R² = 0.863, RMSEP = 0.0003, and RPD = 3.44; Table 2). The wavenumber range used during the development of the PLSr model for the refractive index was between 3,996 and 1,118 cm⁻¹, which is within the range reported by Yang and Irudayaraj (2001) for olive oil. The RPD value suggests the model developed can provide a satisfactory prediction for refractive index, while the low bias (<0.0002) of the developed model suggests that the model was stable and non-sensitive to factors such as cultivar.

For quality control in edible fats and oils, oxidation is one of the main parameters used for oil products. Oxidation of fats and oils produces either primary (peroxides) or secondary oxidation products. The PLSr models developed for peroxide value has shown that the FT-MIRs provided better prediction statistics (R² = 0.919, RMSEP = 1.05 meq, and RPD = 3.54) compared to the FT-NIRs (R² = 0.833, RMSEP = 1.78 meq, and RPD = 2.80; Table 2). Similar values were observed for RMSEV and RMSEP, suggesting that the developed models were not overfitted. Furthermore, the developed models had low bias values (0.036–0.86) indicating robust fitting and stability. This indicates that the models were not sensitive to external factors such as different cultivars. The RPD value (3.54) for the developed model suggests that satisfactory predictions can be made with FT-MIRs.

FT-MIR spectroscopy has been used to evaluate the peroxide value of coconut oil, where the authors reported a high coefficient of determination (R² = 0.982) and low RMSEP values (0.4978 meq; Marina et al., 2013). The wavenumber range reported in this study for peroxide value is similar to those reported for the development of models for various oil products (Liang et al., 2013; Marina et al., 2015; Zahir et al., 2017). For total carotenoid content, the NIR region of 7,498–940 and 6,102–5,774 cm⁻¹ provided better prediction statistics (R² = 0.843, RMSEP = 0.019 g β-carotene/kg) compared to FT-MIRs (R² = 0.632, RMSEP = 0.007 g β-carotene/kg), with RPD value of 2.28 suggesting that the model is fit for quantitative predictions. Similar prediction results for total carotenoid content were reported by Schulz et al. (1998) in essential oils, within the spectral region of 10,100 and 5,150 cm⁻¹. For total phenolic content, FT-MIRs in the region of 3,996 and 758 cm⁻¹, have been shown to provide rough predictions (RPD = 1.57), while those developed in the NIR region were shown to be unreliable (RPD = 1.26). Contrary to our results, Trapani et al. (2016) reported that the NIR spectral region of 12,500 to 4,000 cm⁻¹ provided relatively good prediction statistics (R² = 0.71, RMSEP = 0.08 mg/kg dm) for total phenolic content in olive oil. Model development for the yellowness index gave relatively poor prediction statistics for both FT-NIRs and FT-MIRs (Table 2). Low RPD values and high bias characterized these models, suggesting that the developed models were unreliable, and overestimation may have occurred for these quality attributes. The developed calibration models were only performed using internal cross-validation and thus only applicable to the three selected cultivars. It is well known that the real challenge with calibration models is that their predictive performance almost always reduces when tested on unknown sources such as fruit maturity, seasonality, and growing regions. Thus future research should include more variability (growing regions and seasonality) to improve the model’s robustness.

For this study, FT-MIRs were shown to be better suited for both qualitative and quantitative applications. The regression models developed within the MIR spectral region performed better than those developed within NIR spectral region. This can be because the mid-infrared spectrum contains wavenumbers for fundamental rotational molecular vibration, which is highly sensitive to specific chemical compositions. In contrast, the near-infrared spectrum is associated mainly with overtone and combination bands of fundamental transition, making it less reproducible and specific. Another advantage of ATR FT-MIR spectroscopy is temperature control via the ATR crystal, which reduces potential variation by maintaining constant sample temperature (Smyth and Cozzolino, 2011). However, FT-NIRS is more applicable to practical usage for online or inline implementation or the development of portable devices due to their relatively inexpensive instrumentation costs, more robust components, and it is easier to manufacture rugged instruments, involving no moving parts.