ATR-FTIR and FORS Fingerprints for Authentication of Commercial Sunflower Oils and Quantification of Their Oleic Acid

Abstract

The composition of sunflower oil, rich in fatty acids, largely depends on the seed variety. Commercial sunflower oils are classified as low (SFO), medium (MOSFO), and high (HOSFO) oleic, distinguished by their oleic and linoleic acid content. Higher oleic acid levels enhance health benefits and oxidative stability. Due to their differing market values, ensuring the correct quality and authenticity of these oils is essential. Unsupervised chemometric methods have been applied to visualise the natural behaviour of sunflower oils, while supervised models have been used for authentication based on Attenuated Total Reflection Fourier Transform Infrared Spectroscopy (ATR-FTIR) fingerprints obtained from a benchtop spectrometer. Authentication of MOSFO is particularly challenging because of its wider oleic acid range (43.1–74.9%) and production via genetic modification or blending SFO/HOSFO. To address this, two multivariable PLS-R regression models were developed using ATR FT-IR and Fibre Optic Reflectance Spectroscopy (FORS) fingerprints, the latter obtained with a portable, cost-effective device. The results indicate that FORS could be used as a rapid quality control tool for on-site quantification. In contrast, ATR FT-IR is a more accurate tool for confirmation and quantification, achieving excellent results (Residual Predictive Deviation, RPD = 7.09 and Range Error Ratio, RER = 17.82).

1. Introduction

Sunflower oil is cultivated in numerous regions globally, with approximately two-thirds of worldwide production concentrated in southern and eastern Europe. Other significant production areas include Argentina and certain regions of North America [1,2]. This oil is obtained from the seeds of Helianthus annuus L. [3] and is one of the most sought-after vegetable oils around the world.

There are different types of sunflower oil on the market; these are distinguished by their chemical composition, which is rich in fatty acids (FAs) [4]. Conventional sunflower oil (SFO) has been found to contain a higher percentage of linoleic acid (45.4–74.0%) in comparison to oleic acid (up to 43.0%) [5,6]. Conversely, high oleic sunflower oil (HOSFO) is characterised by its elevated oleic acid content (>75.0%) [5,6] and is derived from genetically modified Helianthus annuus L. seeds (GM-seeds) [7,8]. The most common process for modifying these seeds is the inactivation of the FAD2-1 gene, which encodes the enzyme responsible for converting oleic acid into linoleic acid (oleate desaturase) [9]. Mutations in this gene reduce the activity of the enzyme, leading to an increased concentration of oleic acid [10]. Finally, medium oleic sunflower oil (MOSFO) can be obtained from both GM-seeds [7,8,9,11] and through industrial blends of SFO and HOSFO in different proportions, with a variable oleic acid content ranging from 43.1% to 74.9% [5,6].

These characteristics give rise to two main problems: firstly, oils labelled as MOSFO or HOSFO may not meet the minimum oleic acid values established (43.1% and 75.0%, respectively) [5,6] due to variations in composition after the oil has been obtained from GM-seeds. Secondly, oils produced by blending SFO and HOSFO may not meet the minimum oleic acid content required for labelling. This may be due to insufficiently homogeneous blending, incorrect measurements of the oleic acid content of the base oils, or the economic interests of producers.

These facts pose significant challenges for the authentication and quality control of sunflower oils, especially considering that varieties with high oleic acid content are increasingly being cultivated due to their superior performance [1]. The accurate quantification of oleic acid is imperative, as this analyte not only delineates the nutritional value of the oil but also serves as a pivotal marker for differentiating between diverse types of sunflower oil. Furthermore, an elevated oleic acid content confers several additional advantages, including enhanced oxidative stability, thereby improving the resistance of the oil to elevated temperatures during the frying process and prolonging its shelf life [11]. In addition to these benefits, oleic acid content provides significant health benefits [12].

To address these challenges, various chromatographic techniques have traditionally been used, including High Performance Liquid Chromatography (HPLC), Gas Chromatography (GC), and Thin Layer Chromatography (TLC). These techniques have official methods applied to the study of different fractions of vegetable oils, particularly olive oils [13,14,15]. They can be applied both using a quantitative approach and through the methodology of the instrumental fingerprinting [16,17]. However, these methods are not without their limitations, including high solvent consumption, the use of environmentally harmful reagents and solvents, extended analysis times, and the necessity for meticulous sample preparation. Consequently, there is an increasing focus on the development of more sustainable analytical methods, particularly those grounded in rapid and straightforward spectroscopic techniques, combined with chemometric tools, under the instrumental fingerprinting approach [18,19,20,21]. These methodologies offer the possibility of performing quality controls in situ, in contrast to conventional methods [22,23].

Among the most promising alternatives are Attenuated Total Reflection Fourier Transform Infrared Spectroscopy (ATR-FTIR) and Fibre Optic Reflectance Spectroscopy (FORS), which allow the instrumental fingerprints of oils to be obtained through rapid, non-destructive analysis. In the case of ATR-FTIR, spectral information is generated from the vibrations of the organic bonds present in the chemical composition of the oil [24]. In contrast, FORS spectra primarily reflect indirect absorptions associated with molecular vibrations, manifesting through overtones and combination bands [25,26]. The use of portable spectroscopic instruments, such as those based on FORS [27], Raman [21], or NIR (Near-Infrared) [22,23,25], renders them ideal tools for preliminary or screening evaluations of commercial samples. Conversely, benchtop instruments, particularly those grounded in ATR-FTIR, are more apt for comprehensive quality control, as they facilitate high-throughput analysis of food components with real-time measurements throughout all phases of production [28]. Although ATR-FTIR has previously been used to characterise vegetable oils, including sunflower oils [29], no studies have yet reported on the use of FORS for this purpose. However, the employment of these techniques for authenticating the three commercial types of sunflower oil (SFO, MOSFO, and HOSFO) has not yet been explored. In the present article, it has been demonstrated that both spectral fingerprints provide unique, characteristic, and unambiguous information about the chemical composition of each sample [30].

In this context, the main aim of this study was to develop an approach based on ATR-FTIR and FORS spectral instrumental fingerprints for the correct authentication of commercial sunflower oils and the quantification of oleic acid in commercial samples of the three types of sunflower oils covered by legislation [5,6]. This study makes a significant contribution to the field of quality control methodology, offering novel, rapid, reliable, non-destructive tools for the food and regulatory industries. Moreover, this approach does not require sample pre-treatment, the use of reagents or solvents, thereby reducing analysis costs and time.

2. Materials and Methods

2.1. Sample Bank

2.1.1. Commercial Sunflower Oils

A total of 85 commercial sunflower oils were collected from local stores and supermarkets in Spain, primarily in Andalusia (south of Spain). The sample bank comprised two groups: the first group consisted of 24 samples of refined commercial HOSFO, whereas the second group comprised 61 samples, including 42 samples of refined SFO and 19 samples of refined SFO blended with refined HOSFO to form MOSFO. Currently, there are no commercial samples of MOSFO available in Spain in which the class “medium” is declared on the label. The term that appears instead is “sunflower oil blended with high oleic sunflower oil”. A new subset was used, selecting randomly 19 samples from the original data bank, and they were analysed in accordance with the official International Olive Council (IOC) method for quantifying fatty acids by Gas Chromatography with Flame Ionisation Detection (GC-FID) [14].

2.1.2. Binary Blends (SFO/HOSFO and MOSFO/HOSFO)

From the subset previously analysed by GC-FID, 9 samples (comprising 4 SFOs, 4 HOSFOs, and 1 MOSFOs) were selected to prepare a new subset constituted by 100 binary blend samples. This blend subset contained a variety of concentrations ranging from 5% to 95% of HOSFO blended with SFO or MOSFO (see Table S1).

Binary blends were prepared by gravimetrically weighing the required amounts of each oil into 19 mL glass vials at 22 °C. Samples were homogenised by oscillatory agitation (900–1000 oscillations/min) for 10 min, followed by vortexing at 2500 rpm for 30 s. The blends were analysed using an ATR-FTIR benchtop instrument and stored in darkness under refrigeration until subsequent FORS analysis.

2.2. GC-FID for the Determination of Oleic Acid Content in Commercial Sunflower Oils

To determine the oleic acid content in the 19 commercial sunflower oil samples, the reference method [14] was applied. This method is based on the conversion of the FAs present in sunflower oils to their corresponding methyl esters (FAMEs).

2.2.1. Reagents and Solvents

Methanol (MeOH) (99.9%) and n-hexane (99%) were acquired by Panreac AppliChem (Darmstadt, Germany), while potassium hydroxide (KOH) (85%) was supplied by Panreac (Barcelona, Spain). A total of 2.8 g of KOH was dissolved in 25 mL of MeOH, with stirring, to prepare a 2 mol·L⁻¹ solution (KOH/MeOH). As an internal standard, tridecanoic acid methyl ester ≥ 97% (GC) (Sigma Aldrich, St. Louis, MO, USA) was used.

2.2.2. Chromatographic Conditions

Gas chromatographic analyses were performed using a GC Varian CP 3800 Gas Chromatography system (Varian Inc., Palo Alto, CA, USA) equipped with a flame ionisation detector (FID) and a CTC PAL Autosampler (CombiPal; CTC Analytics, Zwingen, Switzerland). The elution process was carried out utilising an Agilent HP-INNOWAX capillary column (30 m × 0.25 mm × 0.25 μm) using helium as the carrier gas at 1.5 mL min⁻¹. Samples (1 μL) were injected in split mode (1:10) with the injector maintained at 250 °C. The oven temperature programme started at 165 °C (8 min), increased to 210 °C at 1.5 °C/min, and was held for 2 min, resulting in a total run time of 40 min. These chromatographic conditions were adapted from the reference method [14].

2.3. Instruments

2.3.1. ATR-FTIR Benchtop Instrument—Specifications and Methodology

The FT/IR-4X benchtop spectrophotometer (Jasco, Tokyo, Japan) was employed, equipped with a diamond ATR PRO 4X accessory (Jasco, Tokyo, Japan). Spectra were collected in the range 4000–400 cm⁻¹ (MIR region) at 8 cm⁻¹ resolution with 32 scans per spectrum. The infrared beam was incident on the ATR crystal at an angle of 45°, which determines the penetration depth of IR radiation into the liquid sample. No sample preparation was required, and samples were directly applied to the ATR crystal.

Background spectra were recorded prior to each analysis and automatically subtracted. Commercial samples were analysed on the same day, while binary blends were measured in triplicate on three separate days in random order to minimise preparation errors and potential experimental deviations; mean spectra were calculated for further analysis.

2.3.2. UV–VisNIR (FORS) Portable Device—Specifications

Reflectance spectra were recorded using a FieldSpec4STD portable spectroradiometer (Analytical Spectral Devices Inc., Boulder, CO, USA) over the 350–2500 nm range. Before commencing the measurement process, the instrument underwent a 30 min stabilisation period, followed by calibration using a 3.62″ Spectralon white reference with near 100% reflectance, incorporating a dark current correction (DCC), and recalibrated every 25 spectra. The contact probe A122307 (Analytical Spectral Devices Inc., Boulder, CO, USA) with a halogen light source and a measurement area of approximately 2 cm in diameter was employed. The spectral resolution was 3 nm at 700 nm and 10 nm between 1400 and 2100 nm, yielding a total of 2151 variables.

The measurements were conducted in a darkened environment with the open vials positioned vertically inside a plastic box lined internally with graphite paper to minimise possible reflection of the vial walls. Three replicates per sample were recorded, rotating the vial; averaged spectra were calculated for chemometric study.

2.4. Spectroscopic Analysis

2.4.1. ATR-FTIR—Data Pre-Processing and Spectral Elucidation

For each commercial oil and binary blend, an IR reflectance spectrum was acquired using instrument software (SpectraManager 2.5). Each spectrum had wavenumber (cm⁻¹) and reflectance (%) values. The spectral data were imported into MATLAB (version 9.3, Mathworks Inc., Natick, MA, USA), generating two data matrices representing instrumental ATR-FTIR fingerprints: (i) commercial sunflower oils matrix of dimensions 75 × 1869, which is conceived for resolved problems the authentication of different sunflower oil types, and (ii) ATR-FTIR blends oils matrix of dimensions 100 × 1869, which is formed for quantify the oleic acid.

For analysis of these spectroscopic fingerprints, 3 steps were carried out: (i) remove of irrelevant variables since too much noise against relevant information, (ii) a normalisation applying a Standard Normal Variate (SNV) filter for reduce the difference in the baseline, this effect is very notorious when the spectrums were recorded in different days with diverse backgrounds and, (iii) applying a Savitzky–Golay (SG) smoothing filter and the first derivative transformation (second order polynomial and 9-point window) to enhance signal resolution and correct for baseline drifts [31,32,33]. Finally, the two matrices of study (pure sunflower oils and blends) had dimensions of 75 × 375 and 100 × 375, respectively. The complete signal processing pipeline is illustrated in Figure S1.

The most informative variables are found in 3 regions (A, B, and C) corresponding to variables 355–418 (region A), 624–727 (region B), and 1195–1401 (region C), directly associated with the wavenumbers 1081.87–1023.36 cm⁻¹, 1600.63–1799.26 cm⁻¹, and 2701.78–3099.05 cm⁻¹, respectively. Region A contains information about tension vibrations in C-O ester groups (Peak 1 in Figure 1). Specifically, it is useful for differentiating subtle changes, as SFO and MOSFO have a band shifted slightly towards higher wave numbers with respect to HOSFO. Furthermore, HOSFO exhibits a band at approximately 1118 cm⁻¹, the presence of which is not discernible in other types [34]. Region B is associated with changes in vibration of the bonds C=C in 1650 cm⁻¹ [35] and tension vibration of the carbonyl group (C=O) of the ester of the fatty acids, represented by an intensity band of triglycerides, which has a maximum peak in 1743 cm⁻¹ (Peak 2 in Figure 1). Region C contains 3 informative bands, and at 2853 cm⁻¹, the asymmetric vibration of tension of the bond C-H (CH₂) [36], due to the alkyl rest of the triglycerides [33], has been found, which is also found in higher proportions in the edible vegetable oils (Peak 3 in Figure 1). At 2922 cm⁻¹, the same explanation, but now for the asymmetric vibration of tension of the bond C-H (CH₃) (Peak 4 in Figure 1). Finally, in 3003 cm⁻¹ for HOSFO and 3008 cm⁻¹ for SFO have found the tension vibration of the bond (C–H) (cis) (>C = CH-), a higher presence of unsaturation more displaced is this band, in fact, the MOSFOs count with band between the bands of HOSFO (3003 cm⁻¹ ) and SFO (3008 cm⁻¹) [37] (Peak 5 in Figure 1).

2.4.2. FORS—Data Pre-Processing and Spectral Elucidation

For each binary blend, a FORS spectrum was recorded using the instrument software. Each spectrum comprised two data vectors: wavelength (nm), spanning from 350 to 2500 nm, and reflectance, expressed as a percentage between 0 and 0.2%. The spectral data were imported into MATLAB (version 9.3, MathWorks Inc., Natick, MA, USA), generating a 100 × 2151 matrix representing the FORS spectral fingerprints of the binary blends. The objective of this analysis was to quantify the oleic acid content in commercial sunflower oils.

The spectroscopic data underwent a three-step processing procedure: (i) normalisation using an SNV filter to minimise baseline variations; (ii) application of a SG smoothing and first-derivative transformation (second-order polynomial with a 9-point window) to enhance spectral resolution and correct baseline drifts [30,38]; and (iii) Mean Centre for data standardisation. The complete signal processing process is shown in Figure S2.

Despite the initial hypothesis that the UV-A (350–400 nm) and visible (VIS) (450–700 nm) regions would not provide relevant information for these samples, these regions were retained to account for possible variability in the spectral behaviour among oils. Within the range of 800–1650 nm, the main spectral distinction was attributed to differences in reflectivity. The higher reflectance (lower absorption) regions are associated with overtone vibrations of C-H and O-H stretching bonds present in various molecular components of the oils [27]. The most informative regions were identified as three spectral zones (A, B and C), corresponding to wavelengths of 700–1200 nm (Region A), 1201–2000 nm (Region B), and 2201–2500 nm (Region C), respectively, in the NIR region.

This region of the FORS spectrum covers the first, second, and third harmonics between 700 and 1900 nm [39]. The initial area of interest, illustrated in Figure 2 as Region A (930–980 nm), is associated with the second overtone of the O-H bond stretching. Beyond 1900 nm and up to 2500 nm, combination bands derived from one or more overtones become predominant. In edible vegetable oils, characteristic absorption peaks are observed around 1208, 1391, 1416, 1724, 1760, 2308, and 2348 nm (Regions B and C in Figure 2) [40]. In Region B, the band near 1215 nm corresponds to the second overtone of the C-H stretching vibration and is attributed to the presence of oleic acid, particularly in sunflower oils [37]. The absorptions observed within the 1391–1416 nm range are attributed to combined stretching and deformation modes of the same bond [41], while the band centred at 1450 nm corresponds to the first overtone of the O-H bond stretching. In a similar manner, the peaks observed at 1724 and 1760 nm are attributed to the first overtone of C-H vibrations present within methyl, methylene, and ethylene groups. In Region C (2000–2500 nm), the absorptions at 2308 and 2348 nm are attributed to combination bands associated with C-H stretching in fatty acids [42].

2.5. Chemometric Methodology

For the development of the different chemometric models [43], different pre-processing methods (see Tables S2–S4) were tested, the most adequate being the one that includes SNV + SG (1st derivative, 2nd-order polynomial, and 9-point window) for ATR-FTIR fingerprints. In the case of FORS, an additional Mean-Centring step was necessary.

2.5.1. Authentication of All Types of Commercial Sunflower Oils

The interest in using the commercial sunflower oil dataset was to distinguish between the three existing types. Initially, an exploratory visualisation was performed using unsupervised methods. Subsequently, a range of supervised chemometric models were developed, including soft (Soft Independent Modelling of Class Analogies, SIMCA) and hard (Partial Least Squares Discriminant Analysis, PLS-DA, and Support Vector Machine, SVM) models. During the development of discrimination/classification models, MOSFO samples were excluded and subsequently used as a prediction set. For the SFO and HOSFO samples, 75% of the data were used for model training and the remaining 25% for external validation, selected according to the Kennard–Stone (KS) algorithm [44] (see Figure 3).

2.5.2. Quantification of Oleic Acid in Commercial Sunflower Oils

Two multivariate PLS-R (Partial Least Squares Regression) models were developed, both based on 100 binary blends comprising 80 SFO/HOSFO and 20 MOSFO/HOSFO. The aim was to quantify the oleic acid content in commercial samples using two measurement techniques (ATR-FTIR and FORS). Both models were constructed using the same methodology: (i) exploratory analyses were performed to identify and eliminate anomalous samples at an early stage, and (ii) models were calibrated and cross-validated using 75% of the binary blends, and externally validating them with the remaining 25% (21 and 24 samples for ATR-FTIR and FORS, respectively), this data split was performed by the KS algorithm [44]. Finally, once both models had been established, their predictive capacity was evaluated using 10 commercial samples of the three types of sunflower oil. The oleic acid concentration of these 10 samples had previously been determined by the Tello Laboratory (Table S5), which is accredited for the application of the official method [14]. The laboratory belongs to the Tentamus Group, which is recognised worldwide as a leader in agri-food analysis, as illustrated in Figure 4.

The present study utilised error metrics that have been widely recognised and employed in previous research to ascertain the efficacy of chemometric multivariate regression models in determining the actual oleic acid content in commercial sunflower oil samples. The Standard Error of Prediction (SEP) parameter is regarded as adequate when it exhibits relatively low values. In a similar manner, the criterion adopted is that a Residual Predictive Deviation (RPD) greater than 3 indicates excellent performance, while values greater than 2 are considered acceptable [45,46]. Finally, for the Range Error Ratio (RER) parameter, values greater than 4 are considered appropriate for screening applications, values greater than 10 are considered adequate for quality control, and values greater than 15 indicate high quantitative capacity [47].

3. Results and Discussion

3.1. Unsupervised Pattern Recognition Methods of ATR-FTIR Sunflower Oil Fingerprints

3.1.1. Hierarchical Cluster Analysis (HCA)

The reduced ATR-FTIR data matrix underwent an HCA following the Ward’s method and the Mahalanobis distance, such as the linkage criterion and the distance measure between pairs of observations, respectively. The number of clusters was determined using Dlinkage = 2/3 of Dmax as the internal criterion. As shown in Figure 5, sunflower oils naturally cluster according to their primary raw material, oleic acid content, and the oleic/linoleic ratio. Thus, two main clusters were identified: Group I comprised oils with higher oleic acid content, including all HOSFOs and one MOSFO with a priori high content of oleic acid, whereas Group II encompassed the remaining MOSFOs and all SFOs.

3.1.2. Principal Component Analysis (PCA)

The data matrix of pure sunflower oil was subjected to PCA, a process that yielded three principal components (PCs). Due to was observed that from the fourth PC onwards, the cross-validation (CV) error exhibited glow up. Collectively, these components accounted for 99.95% of the total variance observed in the dataset. Figure 6 displays the natural behaviour in the commercial sunflower oils in the PCA plot (PC1 vs. PC2). Therefore, samples are grouped according to their oleic acid content, consistent with the results of the HCA dendrogram (Figure 5). PC1 does not distinguish between the different types of sunflower oil. For PC2, HOSFOs obtained positive scores (Group I), whereas the low oleic sunflower oils received negative scores (Group II). Moreover, some of the medium oleic sunflower oils were grouped with the conventional sunflower oils, while a small subset of samples with positive PC2 scores appeared slightly separated from Group I.

3.2. Supervised Chemometric Models of ATR-FTIR Sunflower Oil Fingerprints

3.2.1. SIMCA Model for Authentication of All Sunflower Oil Types

Three PCs were selected for the target class (HOSFO), and four PCs for the non-target classes (not HOSFO), explaining 99.99% of the cumulative variance in both cases. Figure 7a,b show the Cooman’s classification plots for the analysed samples. Region A corresponds to the HOSFO class, region B identifies samples that are neither HOSFO nor not HOSFO (really not SFO), region C includes samples classified as both HOSFO and non-HOSFO, and region D comprises samples identified as not HOSFO (certainly SFO). In Figure 7a, almost all HOSFO samples, with one exception, are correctly located within their class (region A). The remaining oils are appropriately assigned to their respective not HOSFO class (region D), except for one sample that deviates slightly towards an atypical region (region B). This sample corresponds to an SFO with the lowest amount of oleic acid and the highest linoleic acid content.

Once the prediction set has been entered into the established model (see Figure 7b), there are different trends within the MOSFO group (green triangles). One of the samples evidently approaches region A, as was also the case in the unsupervised analyses (appearing to be HOSFO). Most of the group was in region B (neither HOSFO nor SFO); thus, these samples are defined as true MOSFOs. Finally, a few samples were found to resemble conventional sunflower oils. It is plausible that these samples have a lower oleic/linoleic acid ratio, attributable to a lower oleic acid content than the minimal levels stipulated by the Spanish legislation and by the Food and Agriculture Organization of the United Nations (FAO) [5,6] (Region C).

3.2.2. PLS-DA Model for Authentication of All Sunflower Oil Types

The development of the PLS-DA model utilised three latent variables (LVs), accounting for 99.92% of the cumulative variance. The same criterion applied for the selection of principal components (PCs) in the PCA was followed. CV was performed using both the Leave-One-Out (LOO-CV) and Venetian Blinds (VB-CV) methods, with one sample per blind and ten data splits.

As demonstrated in Figure 8a,b, a threshold of 0.8 is established, so that samples with scores above this value are classified as belonging to the “High oleic” class. The intermediate or doubtful region that falls between 0.8 and 0.4 is defined by the inability of samples to be unequivocally assigned to either the HOSFO or SFO classes. Conversely, samples with scores below 0.4 are regarded as exhibiting a lower oleic acid content compared to linoleic acid, an inherent characteristic of conventional sunflower oils. The training set allows for complete discrimination between the classes, with no inconclusive samples. Conversely, the external validation set counted two doubtful samples, each of a different type of sunflower oil, as illustrated in Figure 8a and as depicted in Figure 7a (SIMCA model). Finally, in Figure 8b, when MOSFOs are introduced to predict their behaviour, three clear trends can be distinguished again: one red triangle (HOSFO), eight green triangles (MOSFO), and six blue triangles (SFO).

Consequently, the model successfully distinguished higher-quality sunflower oils from the rest of the samples and is capable of visualising different trends along the MOSFO group. It was to be expected that the model would show greater variability for the MOSFOs due to the range (43.1–74.9%) of oleic acid permitted under Spanish legislation and international normative [5,6].

3.2.3. SVM Model for Authentication of All Sunflower Oil Types

The SVM model used the Kernel algorithm with a radial basis function (RBF) and was optimised with the cost = 100 and gamma = 0.032 parameters, employing 25 support vectors. For the calibration and external validation blocks (Figure 9a) and for the calibration and prediction blocks (Figure 9b).

All samples in the calibration set were correctly classified into their corresponding class. A single sample of SFO from the external validation set was found to be in the region of indecision, showing behaviour like that observed in the SIMCA and PLS-DA models. Regarding the prediction set, the results coincided with those obtained using PLS-DA, although a slight increase was observed in the distances between the true MOSFO and the MOSFO predicted as SFO.

In general, the three chemometric models developed demonstrate that the ATR-FTIR fingerprint allows highly effective discrimination between high oleic sunflower oil and conventional sunflower oil. However, these models identified a sample labelled as SFO that does not conform to the definition of SFO, as its oleic acid content exceeds the limits established [5,6].

Conversely, the SIMCA and PLS-DA models detected a sample labelled as HOSFO, which exhibited a lower oleic acid content compared to its class, thereby placing it in a region of indecision. Even the SVM model, despite “forcing” its classification within the HOSFO class, clearly reflects that this sample differs from the rest of its group.

When medium oleic sunflower oils are incorporated into the models, their greater compositional variability becomes evident, with three distinct trends being observed: one sample with a composition close to HOSFO, the majority truly representative of MOSFO, and 6 of 15 samples whose chemical composition is more similar to that of SFO than to that of MOSFO.

In order to address this issue and comply more strictly with current legislation and normative, multivariate quantification models for oleic acid in sunflower oils were developed. The aim of this study is to elucidate, clarify, and accurately distinguish those samples that gave rise to uncertainty during the development and prediction of the models.

3.3. Quantification of Oleic Acid in Commercial Sunflower Oils by PLS-R Model

Two multivariate regression models (PLS-R) were developed:

(i)

Model 1: Built with ATR-FTIR. MIR fingerprints.

(ii)

Model 2: Built with FORS. NIR fingerprints.

For the development of these models, two criteria for sample exclusion are established at all stages: Criteria for exclusion of outliers: >Residual value ± 5 ó > Residual studentized value ± 3, as described previously in Figure 4.

3.3.1. PLS-R Model 1 Built with ATR-FTIR: MIR Fingerprints

The ATR-FTIR blends matrix, which has dimensions of 100 × 1869, undergoes the same pre-processing stages as the pure oils matrix and ends up with dimensions of 100 × 375.

(1)

Tuning 1: Outlier detection (Exploratory PLS-R)

The analysis employed four latent variables (LVs), accounting for 99.68% of the cumulative variance. The exploratory analysis presented Root Mean Square Error for calibration (RMSEC) and cross-validation (RMSECV) of 3.61 and 3.95, respectively. As illustrated in Figure S3a, the regression line fit (slope = 0.946), the coefficients of determination (R²_Cal = 0.95 and R²_CV = 0.93), and the bias values (~0) for the calibration and CV sets are demonstrated. Moreover, the plot shows the relationship between gravimetrically determined oleic acid content (x-axis) and model-predicted content (y-axis). Figure S3b shows the studentised residual plot, where no sample was considered an outlier according to this criterion. However, in Figure 10, 14 atypical samples (X_{ATR FR-IR} = 14) were eliminated for the construction of the PLS-R Model 1.

(2)

Tuning 2: Establishment of the PLS-R Model 1

The PLS-R Model 1 was developed using the remaining 86 samples, and the selection of four VLs resulted in an explanation of 99.99% of the cumulative variance. As illustrated in Figure 4, the division of data using the KS algorithm resulted in the separation of 65 and 21 samples for calibration and external validation blocks, respectively.

The RMSEs obtained were 2.33, 2.48, and 2.06, and the R² values were 0.98, 0.96, and 0.97 for the calibration, CV, and external validation sets, respectively. Figure S4a provides a visual representation of the regression lines, with slopes of 0.976 (fit), 0.979 (calibration), and 0.963 (external validation), along with R² values of 0.98 and bias (~0) in all blocks. As demonstrated in Figure S4b, the studentised residuals reveal an absence of outliers under this criterion. However, Figure 11 revealed the presence of four anomalous samples (Y_{ATR FR-IR} = 4) in the calibration set, which were excluded from the final model.

The evaluation of the prediction of externally validated samples is achieved by the utilisation of the SEP, RPD, and RER metrics [47], as illustrated in

Table 1. Error metrics calculated for the evaluation of the chemometric models.

Model	SEP	RPD	RER	Step in the Development of the Models
PLS R Model 1	2.11	7.37	23.34	External validation
PLS R Model 2	2.24	6.51	22.67	External validation
ATR-FTIR (MODEL 1)	3.07	7.09	17.82	Prediction
FORS (MODEL 2)	4.73	4.60	11.58	Prediction

.

3.3.2. PLS-R Model 2 Development with FORS: NIR Fingerprints

The FORS-blend matrix, which has dimensions of 100 × 2151, was subjected to the pre-processing steps described in Figure S2 (applying SNV, SG filter, and Mean Centre).

(1)

Tuning 1: Outlier detection (Exploratory PLS-R)

For this exploratory PLS-R, 8 LVs were used, accounting for 95.36% of the cumulative variance. This exploratory analysis yielded error metrics RMSEC and RMSECV of 1.99 and 3.99, respectively. Figure S5a shows the regression fit (slope = 0.984), the coefficients of determination (R²_Cal = 0.98 and R²_CV = 0.93), and the bias (~0) for both the calibration and CV blocks. The figure also displays the multivariate regression line plot. Figure S5b and Figure S5c present the studentised residuals and residual plots, respectively, where analogous samples were excluded—two in this case (X_FORS = 2).

(2)

Tuning 2: Establishment of the PLS-R Model 2

The PLS-R model was constructed using the remaining 98 samples and established with 6 LVs, explaining 95.00% of the cumulative variance. Data split carried out using the KS algorithm as PLS-R Model 1 (see Figure 4).

The RMSEs obtained were 2.37, 3.03, and 2.58, with R² values of 0.98, 0.96, and 0.97 for the calibration, CV, and prediction (external validation) sets, respectively. In all cases, the bias values were found to be less than 1 and close to 0 (see Figure S6a). As shown in Figure S6b, the studentised residuals do not exclude any samples, while the residuals detect 3 outliers in the calibration set and 2 additional ones in the external validation set (Y_FORS = 5), as can be seen in Figure S6c.

Finally, the predictive capability of the externally validated samples was evaluated by means of the SEP, RPD, and RER metrics [47], whose values are presented in Table 1.

3.3.3. Evaluation of the Predictive Capability of the PLS-R Models

Once the model was developed, the prediction stage was performed using 10 samples of commercial sunflower oils (SFO, MOSFO, and HOSFO), in accordance with the scheme illustrated in Figure 4.

To predict the 10 unknown samples using both models, four VLs were utilised in each case. These accounted for 99.99% and 87.08% of the cumulative variance in the PLS-R Model 1 and PLS-R Model 2, respectively. It was ascertained that the RMSEs, R², and bias values remained within the designated ranges in all blocks. The parameters, in conjunction with the adjustments of the multivariate regression lines, are presented in Figure S7a,b. The SEP, RER, and RPD metrics, which evaluate the predictive power of PLS-R Models 1 and 2, are shown in Table 1. The results of the determination of oleic acid content obtained using the developed PLS-R chemometric models are shown in Table S2.

In the case of PLS-R Model 1 (ATR-FTIR Model), the prediction of commercial sunflower oil samples was found to be highly successful, as evidenced by the low SEP value (3.07), the RPD value that significantly exceeded the threshold of 3 (7.09), and the RER value that was notably higher than 15 (17.82). Consequently, the model is deemed appropriate for the accurate quantification of unknown samples.

In the case of PLS-R Model 2 (FORS Model), the performance in predicting samples decreased considerably, especially for those corresponding to the SFO type. However, despite the higher error rate in comparison to the other PLS-R models, it still exceeds the criterion of RPD > 3 (4.60), thus being considered an excellent model. Furthermore, with an RER greater than 10, it fulfils its objective as a FORS methodology combined with PLS-R, making it suitable for screening and quality control purposes in various scenarios. Due to its portability, it is suitable for application both in the laboratory and in situ (factory, distributor, and final point of sale).

3.3.4. Quantification of Oleic Acid in MOSFOs and Inconclusive Samples of the Study of Authentication Using ATR FTIR Fingerprints

Once PLS-R models had been established based on the binary blends, their predictive capacity was evaluated using 10 reference samples, whose oleic acid content had been previously determined by GC-FID, following the official method [14].

Subsequently, PLS-R Model 1 is used to predict the oleic acid content of an SFO sample and an HOSFO sample that were in ambiguous regions within the discrimination/classification models developed using SIMCA, PLS-DA, and SVM, as well as the set of 15 MOSFO samples.

Table 2. Oleic acid content prediction in commercial sunflower oils using PLS-R Model 1.

Code	Predicted Oleic Acid Content (%) Using PLS-R Model 1 (ATR-FTIR)	Sunflower Oil Type
SFO-24	56.02	MOSFO
MOSFO-39	56.40	MOSFO
MOSFO-40	39.19	SFO
MOSFO-41	72.29	MOSFO
MOSFO-42	39.19	SFO
MOSFO-43	53.23	MOSFO
MOSFO-44	52.40	MOSFO
MOSFO-45	41.03	SFO
MOSFO-46	42.51	SFO
MOSFO-47	53.49	MOSFO
MOSFO-48	56.91	MOSFO
MOSFO-49	53.86	MOSFO
MOSFO-50	43.83	MOSFO
MOSFO-51	50.88	MOSFO
MOSFO-52	39.45	SFO
MOSFO-53	50.88	MOSFO
HOSFO-54	67.32	MOSFO

The sample codes are assigned based on the commercial label, and the type of sunflower oil is determined using the PLS-R 1 model (ATR-FTIR). - Yellow corresponds to conventional sunflower oil samples (<43.1% oleic acid). - Blue corresponds to sunflower oil samples with medium oleic acid content (43.1–74.9% oleic acid). - Pink corresponds to sunflower oil samples with high oleic acid content (>75.0% oleic acid).

presents the oleic acid content prediction values that were obtained using the PLS-R Model 1.

The findings derived from the PLS-R Model 1 corroborate the observations recorded in the discrimination/classification models. Specifically, a sample labelled as conventional sunflower oil was identified as actually corresponding to medium oleic sunflower oil, and similarly, an oil classified as high oleic showed similar behaviour since really a MOSFO sample.

In the medium oleic sunflower oil group, the PLS-DA and SVM models also revealed the presence of six samples that corresponded to SFO, and in the SIMCA model, there were four samples. However, PLS-DA and SVM models incorrectly detected sample MOSFO-50 as medium oleic, since PLS-R Model 1 positively detected it, although the error-relative difference from the SFO class was less than 1.7%. Compared with the SIMCA model, this model incorrectly detected the sample coded as MOSFO-46, such as the MOSFO sample, since the relative error is 1.37%.

4. Conclusions

Given the variability of the MOSFO group and the particular interest in oleic acid, which not only defines the nutritional value of the oil but also constitutes a key marker for differentiating between different types of sunflower oil. Instrumental fingerprints from a robust benchtop equipment based on ATR-FTIR and portable FORS were obtained and pre-processed to propose quantification methods of oleic acid in three commercial types of sunflower oil (SFO, MOSFO and HOSFO).

For ATR-FTIR fingerprints used as a raw data matrix, unsupervised pattern recognition methods, such as PCA and HCA, allowed for the observation of the natural behaviour attending to commercial sunflower oil type. Under this basis, supervised models (SIMCA, PLS-DA, and SVM) were developed to discriminate/classify the commercial types. In addition, as a final step of the procedures, an oleic acid quantification model has been developed by means of PLS-R. On the other hand, considering equivalent chemical information obtained with a FORS portable analyser was also used to develop oleic acid quantification models.

The implementation of both methodologies demonstrated their efficacy in the recording of instrumental fingerprints of commercial oils and binary blends prepared at the laboratory level. It was determined that the portable FORS-based equipment is ideal for rapid screening and in situ quality control of oleic acid, while ATR-FTIR-based benchtop equipment is a suitable tool for comprehensive and demanding quantification, achieving outstanding results (RPD = 7.09 and RER = 17.82).

Finally, it is shown that, despite SFO characteristics, its higher linoleic acid content can, in some samples analysed, have a higher oleic acid content than expected. Furthermore, not all MOSFOs exhibit the same behaviour.