*3.4. Regression Models*

PLS regression models were built with FT-NIR and Vis/NIR spectra to quantify moisture, oil content, soluble solids, TPC, and antioxidant activity of olive drupes. In order to make data more evenly distributed, TPC and DPPH• results were transformed in the inverse and the logarithmic values, respectively. The best models based on determination coefficients and errors are reported in Table 1 for each spectral acquisition system. Predicted vs. measured plots of the models are reported in the Supplementary Figure S1. In general, performances of the three acquisition systems were similar in calibration and cross-validation, while in prediction FT-NIR spectra gave better results, maybe due to the wider NIR range and the low complexity of the models resulting in a higher stability.


**Table 1.** Figures of merit of the best PLS regression models for olive chemical parameter prediction based on spectroscopic data.

TPC: total phenolic content; GA: gallic acid equivalent; inhib.: inhibition; DPPH•: radical 2,2 diphenyl-1-picrylhydrazyl; LVs: latent variables; R2 cal: calibration coefficient of determination; R2 cv: cross-validation coefficient of determination; R2 pred: prediction coefficient of determination; RMSEC, RMSECV, and RMSEP: root mean square errors of calibration, cross-validation, and prediction, respectively; SNV: standard normal variate; d1: first derivative.

> With respect to moisture content prediction, the three acquisition systems exhibited promising and similar prediction outcomes. The determination coefficients ranged from 0.77 to 0.92, with reasonably low values of errors (from 2.67 to 4.75%). However, the model calculated with the Vis/NIR spectra transformed in d1 showed a higher number of LVs (16 vs. 8 and 7 for the sphere and the probe, respectively), maybe due to the higher resolution of the spectra and the limited NIR range considered.

> Oil content was better predicted by FT-NIR spectra, pre-treated with SNV and d1. Prediction coefficients of determination were higher than those of the portable acquisition system (0.77 and 0.78 vs. 0.64), with lower RMSEP values (2.92 and 2.86% vs. 3.74%) and LVs (9 and 5 vs. 16). The outcomes of calibration and cross-validation coefficients of determination for the FT-NIR sphere and probe (0.77–0.93) were comparable to those reported in the literature (0.78–0.84) for a smaller number of samples (183) [18].

> Considering soluble solids, the regression model reliability appeared even more promising for FT-NIR spectrometer than for the portable device. Both the FT-NIR sphere and fiber-optic spectra pre-treated with a combination of SNV and d1 resulted in satisfactory determination coefficients in prediction (0.70 and 0.74, respectively) and low RMSEP (2.39 and 2.23 ◦Bx, respectively). The precision of the models was comparable to those observed for other fruits as, to the best of our knowledge, there is no study on NIR prediction of soluble solids in intact olives. For instance, quantitative determination of soluble solid content for quality prediction of intact strawberries using a handheld micro-electromechanical NIR system, resulted in R<sup>2</sup> pred of 0.37–0.47 and RMSEP of 1.02–0.87% [44]. With the spectra in the Vis/NIR range, the coefficient of determination in prediction decreased to 0.58, with a RMSEP of 3.02 ◦Bx.

> Similar model performances in calibration and cross-validation were obtained for 1/TPC for all the spectral acquisition systems (R<sup>2</sup> range, 0.76–0.89), whereas in prediction FT-NIR spectra, gave better results (R2 pred = 0.77–0.76) than Vis/NIR spectra (R<sup>2</sup> pred = 0.69). FT-NIR models are better than those reported in the literature for a filter-based NIR spectrometer [35]. The authors attributed the unsatisfactory output of their model (R<sup>2</sup> cal, 0.72; SEP, 13.35 g oleuropein/kgdm) to the exclusion of 8600–6900 cm−<sup>1</sup> range from the spectral bands, which was instead here considered. Our models were more promising also when compared to grape TPC prediction models developed using a portable NIR-

AOTF [45]; the authors observed determination coefficient values of 0.77 and 0.62 in calibration and cross-validation, respectively.

To the best of our knowledge, there is no other published paper in which the antioxidant activity of olive drupes is tentatively determined using rapid spectroscopic techniques. Therefore, our models seem fair especially when the FT-NIR probe was used, which generated comparatively highest R<sup>2</sup> pred and lowest RMSEP among the three spectral acquisition systems. The dynamic nature of this in vitro antioxidant activity makes its adaptation to spectroscopic techniques somewhat difficult. A more accurate NIR prediction of DPPH• radical scavenging activity was recorded in bean flours (R2 cal, 0.94–0.99; R2 val, 0.85–0.97) [46]. On the contrary, for a more bioactive horticultural product like *Hibiscus sabdariffa*, calibration and prediction determination coefficients are reported in the literature in the ranges 0.82–0.87 and 0.75–0.86, respectively, depending on spectra pretreatments [47].

From the inspection of the weighted regression coefficients of PLS models, for both the FT-NIR sphere and the probe the relevance of 7500–6100 and 5400–4500 cm−<sup>1</sup> regions for moisture and soluble solid prediction was confirmed. Moreover, the PLS model developed for oil content prediction were highly influenced by the 5800–5650 and 4350–4250 cm−<sup>1</sup> regions, attributed to the oil content of the drupes [40]. The same regions showed high weighted regression coefficients for TPC and DPPH• models, which were also characterized by high relevance of the 8700–8300 cm−<sup>1</sup> region, previously related to TPC of olives [19].

As for the model developed with the Vis/NIR spectra, the inspection of the weighted regression coefficients revealed that both visible and NIR range influenced the prediction of moisture, oil content, and soluble solids. In particular, the range 880–970 nm showed the highest influence in the models for moisture and oil content prediction, whereas the maximum recorded weight for soluble solids corresponded to 970 nm. Moving to TPC and DPPH• prediction, it has been noticed that the highest values of the weighted regression coefficients were related to the visible range (550–700 nm), maybe linked to the olive color modification occurring during ripening, due to compounds like chlorophylls, carotenoids, anthocyanins, and polyphenols. Actually, other authors demonstrated that during olive ripening a rise in some bands of the visible range occurs (i.e., 600–650 and 550–625 nm), due to the presence of anthocyanin and other pigments related to reddish as well as green and yellow color [18].

#### *3.5. Regression Model Comparison*

The effectiveness of the prediction ability was at first established comparing the intermediate precisions (SEL) of the regression models with those of the reference methods (Table 2). The SEL values for the different NIR systems were generally higher than those obtained for the reference analyses, except for the SEL of the oil content predicted by the FT-NIR probe measurements. Indeed, the SEL values of NIR systems are more affected by the drupe heterogeneity, since spectra are collected on entire olives without the sample preparation phase of the chemical analyses, which is carried out by grinding and homogenizing the olive pulp.

The SELref values were also compared with the prediction performances of the models in terms of SEP. As expected, SEP values were always higher than those of SELref, because they include not only the sampling and analysis errors, but also the spectroscopy and model errors. The SEP obtained for the FT-NIR probe models were the lowest and the closest to the corresponding SELref values. If the SEP is <2SELref, the prediction performance of the model should be considered as good [48]. This was the case of models developed from FT-NIR probe spectra for moisture, oil content, and 1/TPC prediction.

Furthermore, the *t*-test for paired samples demonstrated that the biases for the models developed with the three spectral acquisition systems were comparable, i.e., the null hypothesis was confirmed (*p* values between 0.1 and 0.8; data not shown). On the other hand, the comparison between the standard deviations of the models [33] returned some differences as reported in the last three columns of Table 2. For moisture, the FT-NIR probe

model resulted significantly different from those based on sphere and portable device spectra, due to a better performance resulting in a lower RMSEP (Table 1). All the other comparisons resulted in similar performance of the FT-NIR sphere and probe models, whereas the portable device models resulted significantly different because of the worse performance in terms of R2 pred, RMSEP, and SEP.

**Table 2.** Comparison of regression models calculated for olive chemical parameter prediction based on three different FT-NIR and Vis-NIR acquisition systems.


TPC: total phenolic content; GA: gallic acid equivalent; DPPH•: radical 2,2 diphenyl-1-picrylhydrazyl; inhib.: inhibition; SELref: standard error of laboratory for reference analyses; SELNIR: standard error of laboratory for NIR systems; SEP: standard error of prediction; n.s.: not significantly different standard deviation values (*p* > 0.05); \*: statistically different standard deviation values (*p* ≤ 0.05).

#### **4. Conclusions**

The benefits of different NIRS acquisition systems as green technology for quality characterization of intact olive drupes were explored. Generally, the calculated PLS models were remarkably encouraging in terms of determination coefficients and errors, both in internal validation and prediction. The model comparison highlighted a general better performance of both the FT-NIR sphere and probe acquisition systems with respect to the handheld device. However, the Vis/NIR device, being portable and relatively cheaper, is worthy of further investigations, because its use could be in any case very useful for preliminary quick quality assessment of olive drupes directly in the field. On the contrary, an on-line or in-line application of the FT-NIR optical probe in the olive mill should be promoted in order to quickly classify the drupes for a better quality design of the olive oil and a more sustainable management of the production chain.

**Supplementary Materials:** The following are available online at https://www.mdpi.com/article/ 10.3390/foods10051042/s1, Figure S1: Regression lines obtained for the prediction of entire olive chemical parameters with models developed by FT-NIR integrating sphere, FT-NIR fiber-optic probe, and portable Vis/NIR device. Table S1: Variation ranges of the chemical parameters for the different olive cultivars.

**Author Contributions:** Conceptualization, S.G., V.G., E.C. and C.A.; methodology, S.G., V.G., A.T., G.S., P.C., A.D.B., C.A.; formal analysis, S.G., O.S.J.; investigation, S.G., A.T., G.S., P.C., A.D.B.; resources, G.S., P.C., A.D.B., F.F.; data curation, S.G., O.S.J., C.A.; writing—original draft preparation, S.G., O.S.J., C.A.; writing—review & editing: V.G., A.T., G.S., P.C., A.D.B., F.F.; supervision, E.C., C.A.; funding acquisition, E.C. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by AGER 2 Project, grant no. 2016-0105.

**Acknowledgments:** The authors wish to thank prof. Francesco Caponio for the supervision of the whole research project.

**Conflicts of Interest:** The authors declare no conflict of interest. The funder had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.
