Walnut samples were provided by a fruit producer’s association located in Alentejo, Fruteco (Borba, Portugal), in the 2018–2019 season. Samples for this study were obtained from five different varieties: “Chandler”, “Franquette”, “Howard”, “Lara”, and “Tulare”. Forty in-shell walnuts were taken for each variety (i.e., a total of 200 samples). They were carried to the laboratory where the walnut shells were removed. Walnut kernels were subjected to a visual inspection in order to identify samples with defects, such as insect damage, fungal growth, abnormal coloration, or detrimental disorders. Defective samples were discarded, and the remaining samples were stored at 4°C until spectra were acquired.

Walnut kernels were tempered and submitted to the acquisition of their NIR spectra. For that purpose, a NIR hyperspectral device was used. This device consists in a camera (Xenics Infrared Solutions, Inc., Leuven, Belgium), a spectrograph (Spectral Imaging Ltd., Oulu, Finland), two 70W tungsten iodine halogen lamps (Prilux®, Barcelona, Spain), a mirror scanner (Spectral Imaging Ltd., Oulu, Finland), and a computer system. Hyperspectral images covered the spectral range between 900 and 1,700nm with a spectral resolution of 3.25nm. The procedure performed for NIR spectra acquisition is based on Nogales-Bueno et al. (2020), with some modifications. Briefly, groups of 10 walnut kernels were disposed over a polyethylene tray and under the hyperspectral camera each time. Images were recorded using a time of exposition of 9ms and a frame rate of 50Hz. Two images were acquired for each 10-walnut group. Between both images, walnuts were rotated 180° aiming to acquire as much spectral information as possible. Hyperspectral device was controlled using the instrument acquisition software SpectralDAQ v. 3.62 (Spectral Imaging Ltd., Oulu, Finland).

Reflectance spectra were corrected following a two-point calibration. To this end, in each image acquisition session, two extra images were acquired: (1) a dark current spectrum, where lamps were turned off and the entry window of the mirror scanner was covered and (2) a white reference spectrum, where it was acquired the spectrum of a polytetrafluoroethylene tile (Labsphere Inc., North Sutton, United States) at operation conditions. Then, regions of interest (ROIs) were extracted from the images by means of a segmentation procedure. The segmentation algorithm can identify walnut kernels and background pixels and save only the reflectance spectra of the walnut. For the development of this algorithm, a number of walnut and background spectra were manually extracted from the calibrated images. A stepwise linear discriminant analysis (LDA) was carried out using NIR wavelengths as dependent variables and the membership of the spectra to kernel or background as categorical variable. Then, the LDA model was applied to all the calibrated images and individual walnut spectra were extracted. Noisy wavelengths were found at both extremes of the spectral range and only the subrange 950–1,650nm was kept. Finally, the spectral matrix was transformed to absorbance values and saved. Calibration and segmentation processes were carried out in the software Matlab (R2018a; TheMathWorks, Inc., MA, United States), ENVI 4.7 (ITT Corporation, White Plains, NY, United States), and SPSS 25.0 (SPSS, Inc., Chicago, IL, United States).

The entire procedure described above has been schematized in Figure 1.

Once the walnut kernel spectra were obtained, it was necessary to obtain the chemical values for the reference parameters in these walnuts. However, in order to reduce the number of samples required for the development of spectroscopic methods, a selection of samples was made following the method described by Nogales-Bueno et al. (2015) For that purpose, the spectral information of the samples was analyzed.

A principal component analysis (PCA) was applied to the spectral matrix. Previously, spectra were pre-treated with a standard normal variate (SNV; 2,5,5,1) spectral pre-treatment. Numbers in brackets, respectively represent the number of the derivative, the gap over which the derivative is calculated, the number of data points in a first running average or smoothing, and in a second smoothing (Shenk and Westerhaus, 1995). PCA is an unsupervised pattern recognition tool, which allows studying the latent structure of the matrix. The number of principal components (PC) was established, Mahalanobis distance (H) was measured, and samples were sorted considering their distance to the spectral center. Samples with H>3 were identified as spectral outliers and were removed from the spectral matrix. Furthermore, samples were grouped according to their neighborhood Mahalanobis distance (NH). Groups with NH<0.9 were created. Since each group can be considered spectrally homogenous, a calibration sample set was created by selecting one sample for each group. Moreover, by selecting another sample for each group (if possible), a validation set was created. These two sample sets (calibration and validation) were subsequently employed for the development of the spectroscopic methods and for their internal and external validation.

After sample selection, samples allocated into calibration and validation sets were weighted and minced in a mill (IKA® A11 basic) and then they were subjected to the determination of the reference parameters: fatty acid profile and total fat.

In addition of the LDA and PCA, described in the previous sections, other statistical tools were applied to the chemical and spectral data sets.

Modified partial least square (MPLS) regressions were applied for developing calibration methods for the screening of total fat and fatty acid profile. For this purpose, different spectral pre-treatments were applied to the spectra allocated in the calibration set. SNV, multiplicative scattering correction (MSC), detrend, first and second derivatives, and none pre-treatments were tested and only the model with the best results was saved for each reference parameter. MPLS divides the calibration set into a number of subsets in order to perform cross-validation to establish the number of PLS factors and to reduce the possibility of overfitting (Shenk and Westerhaus, 1995). Chemical outliers were identified and removed by the evaluation of the T-statistic and setting the threshold in 2.5units. To do so, residual error is obtained by comparison between the predicted values and the chemical values. Finally, the SE of cross-validation (SECV) is evaluated and expressed as percentage.

Then, the goodness of each MPLS model was tested. To this end, models were applied to the samples allocated into the validation set, and the results were compared to the chemical values previously determined. In this way, a SE of prediction (SEP) in external validation was obtained for each reference parameter. Finally, models were also applied to those samples not selected for calibration or validation sets. Thus, a predicted fatty acid profile was obtained for these samples. Then, these predicted profiles were averaged for each walnut variety and compared to those obtained from the gas chromatography analysis, for samples allocated into the calibration and validation sets. MPLS models were developed and tested with Win ISI® (v1.50; Infrasoft International, LLC, Port. Matilda, PA, United States).

The comparison between the average values for the reference parameters of the different walnut varieties was carried out by means of ANOVA and post hoc Tukey test. Total fat, individual fatty acids, SFA, MUFA, and PUFA were used as dependent variables, and walnut variety was used as independent variable or factor. The statistically significant level was considered at α=0.05. ANOVA were performed using Statistica v.8.0 software (StatSoft Inc., OK, United States, 2007).

Walnut kernels were visually inspected and 35 of the initial 200 samples were identified as defective samples. They were discarded and not taken into account for the rest of this study. The remaining 165 walnut kernels (31 “Chandler”, 29 “Franquette”, 32 “Howard”, 35 “Lara”, and 38 “Tulare” samples) were subjected to the hyperspectral analysis. The segmentation procedure identified three wavelengths (1,233, 1,481 and 1,593nm) for the detection of the ROIs (Figure 1).

Next, the sample selection procedure was carried out. PCA did not identify any spectral outlier, and H and NH distances were measured in a space composed by 12 PCs. These 12 PCs explained the 99% of the spectral variability. Forty-three spectrally equivalent groups were found in this space and one sample for each group was selected and allocated into the calibration set. Therefore, this sample set contained most of the spectral information from the original sample set. Then, another sample of each group was selected, if possible, and allocated into the validation set. In consequence, the 165 initial samples were distributed as follows: calibration set (43 samples), validation set (20 samples), and non-selected set (102 samples).

Only walnut kernels allocated into the calibration and validation sets were subjected to chemical analysis. These sets of samples comprise 12 samples of “Chandler”, 8 samples of “Franquette”, 4 samples of “Howard”, 23 samples of “Lara”, and 16 samples of “Tulare” varieties. Differences in the number of samples for each variety indicate which varieties are more spectrally homogeneous (“Howard” or “Franquette”) and which are more spectrally heterogeneous (“Lara” or “Tulare”).

Table 1 shows the mean and SEs of the total fat and fatty acids (percent) in each walnut variety. Up to 16 individual fatty acids could be identified and quantified with the chromatographic analysis. Among them, four were SFA (palmitic acid or C16:0, margaric acid or C17:0, stearic acid or C18:0, and heneicosanoic acid or C21:0); eight were MUFA (palmitoleic acid or C16:1 ω7c, hypogeic acid or C16:1 ω9c, cis-10-heptadecenoic acid or C17:1 ω7c, civetic acid or C17:1 ω9c, elaidic acid or C18:1 ω9t, oleic acid or C18:1 ω9c, and vaccenic acid or C18:1 ω7c and gondoic acid or C20:1 ω9); and four were PUFA (linolelaidic acid or C18:2 ω6ct, linoleic acid or C18:2 ω6c, γ-linolenic acid or C18:3 ω6, and α-linolenic acid or C18:3 ω3).

After chemically obtaining the total fat and fatty acid profile for the calibration and validation sets, these reference data were assigned, respectively to the correspondent spectral samples. Then, MPLS models for predicting these parameters were developed. MPLS regressions were carried out employing the spectral data of the calibration samples as independent (X) variables and the values of total fat, SFA, MUFA, PUFA, and the individual fatty acids as dependent (Y) variables. Table 2 describes the main descriptors for the MPLS models developed for total fat, SFA, MUFA, and PUFA: the best spectral pre-treatment, the number of samples (N) used to obtain the calibrations after deleting the chemical outliers (T), the number of PLS factors, the range of applicability of the models, and a number of statistical descriptors, which allow describing the goodness of these models. All models present an adequate coefficient of determination (RSQ) and SECV. This error is more noticeable if it is expressed as a percentage of the mean value of the range. In all cases, obtained errors were less than 10%. Afterward, using the spectra and chemical values allocated in the validation set, the developed models were tested and the SEP errors in external validation were obtained (Table 2). The good results obtained for fatty acid families suggested that good MPLS models might be obtained for individual fatty acids. Then, MPLS models were developed for the individual fatty acids identified in the gas chromatography system. Table 3 summarizes the results obtained.

Finally, the values of total fat, SFA, MUFA, PUFA, and individual fatty acids were predicted for the samples not used in the development or validation of the models, i.e., for the samples allocated into the non-selected set (Table 4). Regarding individual fatty acids, only those whose MPLS models had adequate RSQ, SECV, and SEP values were predicted.

Total fat ranged from 65.85% for “Chandler” to 68.86% for “Franquette” varieties. However, there were not found significant differences among varieties. According to the bibliography, walnuts contain an average value of 65% fat by weight (Hayes et al., 2016; Fatima et al., 2018). However, higher and lower values are also described. Muradoglu et al. (2010) reported average values of 50% for total fat content in walnuts, contrarily to Pereira et al. (2008) who reported values between 69 and 72% for total fat content by weight. These differences can be attributed to the influence of the variety, harvesting year and different environmental conditions, which may influence the chemical composition of the fruits. In fact, the oil content, the fatty acids, vitamin E, and tocopherols have been found by others to vary significantly among different walnut cultivars and environmental condition (Amaral et al., 2005).

In this study, the fatty acids compositions of walnut varieties ranged from 8.73 to 10.17% for SFA; from 13.54 to 14.44% for MUFA, and from 73.03 to 77.48% for PUFA. These results are consistent with the fact that walnut is the consumable tree nut with the lowest percentage of SFA and the highest percentage of PUFA. Regarding to individual fatty acids, the 69 and 25% of the SFA were, respectively palmitic and stearic acids and 90% of the MUFA was oleic acid. Also 81 and 19% of the PUFA correspond to linoleic and α-linolenic acid, respectively. Similar results can be widely found for walnuts in the literature (Greve et al., 1992; Amaral et al., 2003; Li et al., 2007; Pereira et al., 2008; Poggetti et al., 2018; Red BEDCA, 2019; U.S. Department of Agriculture, 2019).

Considering differences between walnut varieties, “Franquette” has significantly higher amounts of SFA and MUFA and a lower concentration of PUFA. Similar trends in the fatty acid profile were reported between “Lara” and “Franquette” by other researchers (Amaral et al., 2003; Pereira et al., 2008), between “Franquette” and “Howard” (Poggetti et al., 2018; Cittadini et al., 2020) and also between “Tulare” and “Franquette” (Cittadini et al., 2020). Moreover, similar profiles for “Chandler” and “Howard” varieties were also reported in Greve et al. (1992) and in Cittadini et al. (2020).

The higher amount of SFA in “Franquette” can be easily linked to a greater oxidative stability of walnut oil (Rabrenovic et al., 2011). Furthermore, the higher concentration of MUFA and lower concentration of PUFA leads to a higher MUFA/PUFA ratio in this variety than in the others. In fact, MUFA/PUFA ratio has also been linked to oxidative stability in tree nuts (Sena-Moreno et al., 2016). The higher the MUFA/PUFA ratio, the greater the oxidative stability. Indeed, MUFA/PUFA ratio for “Franquette” variety is 0.23 while for the remaining varieties is 0.18–0.19. Therefore, “Franquette” is more stable than the other varieties, and can be considered a good option for walnut oil production. Nevertheless, it is necessary to take into account the stability problems of walnut oil when compared with other tree nuts, mainly due to the much higher concentration of PUFA in walnut (Savage et al., 1999). In addition, this greater oxidative stability of “Franquette” suggests a better aptitude for the sale of walnut kernel, whereas the remaining varieties are more suitable for in-shell sale. In fact, shells protect the kernel from oxidative agents (Hosseini et al., 2014). On the other hand, a higher concentration of PUFA is linked to greater health profits (Ros et al., 2004; Kris-Etherton, 2014), which suggests that “Franquette” walnuts could have a slightly lower health benefits than other varieties.

Standard error of prediction error in external validation is the best indicator of the goodness of the models. SEP ranged from 2.12% for PUFA to 13.08% for MUFA. Also for total fat, the result obtained for SEP was 3.01% which is an expected value since fat is the main component of these samples and the reference method for its determination is relatively simple. However, SEP values obtained for the fatty acid families are remarkable since the reference method used, gas chromatography, has a relatively high complexity. These SEP errors indicate that near infrared hyperspectral imaging can be a reliable tool for the control of these parameters in walnut kernels in a simple, non-destructive and environmentally friendly way. However, it should be taken into consideration that to obtain definitive models it would be necessary to employ more sample sets in order to consider seasonal and regional variations.

Regarding the MPLS models developed for the individual fatty acids identified in the gas chromatography system (Table 3), the first interesting outcome is the number of PLS factors selected by the software in the development of the different MPLS models. Attending the PLS factors, two groups of models can be defined: a first group with few PLS factors (1–4) and a second group with numerous PLS factors (5–7). In MPLS regression, each PLS factor is a compromise, trying to explain spectral variation while trying to correlate highly with the Y variable without overfitting (Shenk and Westerhaus, 1991; Westerhaus, 2014). Therefore, a model with few PLS factors might indicate that this compromise was hardly achievable. In this model, PLS factors mainly describe X variability without describing Y. The low concentration of the predicted fatty acids in the samples might be the cause of this. The gas chromatograph could detect these compounds, but, their concentration is really near to the minimum detectable concentration of the system. Noise or small errors in the integration procedure can easily disrupt the results obtained. Finally, the lack of PLS factors make hard for these models to achieve a good prediction. It can be observed how these models show low RSQ value and, in some of them, high errors in cross-validation (SECV) and in external validation (SEP).

However, the second group of MPLS models with numerous PLS factors shown better results. These reference parameters are marked in bold in Table 3. Most of these models were developed for fatty acids with a higher concentration than the models described above. In these cases, RSQ>0.6 were obtained, sometimes reaching 0.8. SECV ranged from 5.3 to 23.0%. When external validation was carried out, quite good results were obtained for palmitic, margaric, oleic, linolelaidic, linoleic, and α-linolenic acids. Therefore, near infrared hyperspectral imaging also proved to be a promising tool for the screening of some individual fatty acids in walnut kernels.

Similar results to those obtained in the present study were obtained for total fat in Yi et al. (2017). However, these authors used a traditional NIR spectroscope for measuring powder walnut. No more studies have been found for walnuts. However, similar errors (between 1 and 17%) are described in other studies developed in olive oil, peanuts, and almonds for total fat and SFA, MUFA, PUFA, and some individual fatty acids (Milinovic et al., 2019; Teixeira and Sousa, 2019).

The application of the MPLS models for the prediction of total fat, SFA, MUFA, PUFA, and individual fatty acids in the non-selected sample set was carried out with three different objectives: (1) to verify the general trends obtained from the chemical analyses previously performed for the calibration and validation sets, (2) to check the cultivar trends previously observed in the chemical analysis, and (3) to check the feasibility of the sample selection procedure. Thus:

In summary, total fat and fatty acid profiles of five different walnut varieties (“Chandler”, “Franquette”, “Howard”, “Lara”, and “Tulare”) produced in Alentejo region have been obtained. Although no significant differences in total fat have been found between these varieties, the assessment of the identity of walnut oil produced in this region and with these varieties will be of major importance to protect the local walnut production. Significantly higher amounts of SFA and MUFA and a lower concentration of PUFA were found for “Franquette” than for the remaining varieties. These results may translate into greater oxidative stability and slightly lower health benefits of “Franquette” walnuts and oil comparing with other varieties. Therefore, “Franquette” variety is more suitable for oil production or for kernel sale whereas the remaining varieties are more suitable for in-shell sale.

In addition, MPLS models have been developed from the near infrared hyperspectral data and from the chemical data (total fat, SFA, MUFA, PUFA, and individual fatty acids) of walnuts. In general, good results were obtained in cross and external validations. Hyperspectral imaging is a promising tool for the screening of total fat, SFA, MUFA, PUFA, and some fatty acids. Using a set of samples additional to the validation set, the trends found in the fatty acid profiles of the different varieties, the applicability of the MPLS models, and the feasibility of the sample selection procedure carried out in this study have been verified.

Since gas chromatography is a relatively complex technique, requiring a large amount of different resources (time, chemical reagents, capital, etc.), the MPLS models obtained in this study acquire greater importance. In this study, first steps have been taken to develop robust, reliable, non-destructive, non-polluting, and rapid methods to control families of fatty acids and individual compounds in walnuts. More sets of samples from different walnut varieties, regions or seasons need to be collected and MPLS models need to be improved with these data. This would result in more reliable and robust methods. In addition, other studies using near-infrared hyperspectral imaging, such as control of tocopherols, triglycerides, etc. or detection of faulty samples, can be developed to help improve the walnut and walnut oil industry.