Rapid Nondestructive Postharvest Potato Freshness and Cultivar Discrimination Assessment

Background: Quality and safety of potato is both cultivar and postharvest management dependent. The precise assessment of freshness and cultivar are complex tasks requiring time-consuming, expensive, and destructive techniques. Method: Potatoes from three commercial cultivars were stored for 5 months at 5 °C. Color and chlorophyll fluorescence were recorded, Red-Green-Blue (R-G-B), Red-Green-Near infrared (R-G-NIR) and Red-Blue-Near infrared (R-B-NIR) digital images, as well as hyperspectral images were acquired both on the external periderm of the tuber and in the inner flesh part. Partial least square regression (PLSR) and discriminant analysis, combined with feature selection techniques were implemented, in order to assess the potato freshness and to classify them into the respective genotypes. Results: The PLSR analysis of visible/near infrared (Vis/NIR) spectra reflectance most reliably predicted potato freshness, with a cross-validated regression coefficient equal to 0.981 and 0.947, as determined by external or internal measurements, respectively. Variance inflation factor, variable importance scores, and genetic algorithms identified specific wavelength regions that mostly affected the accuracy of the model in terms of strongest regression and lowest collinearity and root mean cross validation error. Conclusions: Vis/NIR spectra reflectance data from the skin of the potato tubers may be reliably used in the assessment of postharvest storage life, as well as in the cultivar discrimination process.


Introduction
Potato is considered an important source of human nutrition and is ranked high in the global food production and consumption. Apart from their distribution as fresh commodity in the retail market, potatoes constitute the raw material for the food processing industry (e.g., in chips or frozen food production) or even for the chemical industry, as a source of ethanol and starch [1]. In order to fulfill the demand all year round, a common practice is to store the tubers for several months and thus increase their marketability, as well. However, during a long period of storage, the dormancy of the tuber is interrupted and consequently storage may significantly affect the quality of the raw or the processed product in some of them. However, no study has been employed in order to estimate the freshness of whole fresh tubers at several stages of the postharvest storage by capturing the spectra reflectance on several potato cultivars simultaneously. Indeed, the capacity of a system to reliably assess a tuber's freshness, irrespective of the variety and at all steps of the potato supply chain, is crucial and this system should be rapid and functional and should also combine the ability to be applied during all stages of the postharvest life until the disposal at the retail market.
In the case of potatoes, there is a considerable variation in size, shape, and color of the cultivars available in the market which complicates even more the implementation of a robust, universal sorting technique. Therefore, it is important to point out that most of the published literature highlights the need to confirm the reliability of the non-destructive methods on a variety of cultivars [24]. This detail further highlights the need of classifying potatoes into distinct cultivars during sorting of a mixed bunch of tubers.
In this particular study, color along with chlorophyll fluorescence, R-G-B, R-G-NIR, and R-B-NIR imaging systems, as well as spectral reflectance were evaluated for their efficacy in monitoring the freshness of potatoes postharvest, as well as to classify the tubers into the corresponding genotypes, while combined with several chemometric approaches.

Plant Material
Potatoes from three commercial cultivars, Spunta, Banba, and Sylvana (Figure 1), were grown in 2018 following the usual practices, in terms of soil cultivation, fertilization, irrigation, and plant protection, in neighboring plots in a production area under protected geographical indications (PGI) under the same weather conditions and by the same grower. Spunta tubers have an oval to long shape, white to yellow skin color, light to yellow flesh color, and smooth to intermediate skin texture. Banba tubers have oval to long shape, yellow skin color, light yellow flesh color, and intermediate skin smoothness, while Sylvana has oval shaped tubers with white to light beige skin color, medium yellow flesh color, and medium skin smoothness (https://www.europotato.org; http://varieties.ahdb.org.uk) (Accessed: 11 March 2021). Transplanting and harvesting dates were the same for all the cultivars. Plants were uprooted by hand in several sites within each plot and around 500 tubers from each cultivar were thoroughly examined by eye before selecting the 90 more representative, in terms of shape and size, free of visible defects. The samples were collected at the same day of harvest, early in the morning, were transferred to the laboratory within 90 min, were washed under tap water to remove dust and foreign particles from the surface, wiped with a paper towel, and were dried before finally choosing half of them (45 tubers) from each cultivar and storing them for 5 months at 5 • C in continuous darkness, in order to retard sprouting and maintain other various quality attributes [2]. The dimensions of the cold room were 4.2 × 3.5 × 2.1 m 3 (length × width × height) and the average temperature and relative humidity were 5.1 ± 0.53 • C and 89.8 ± 5.53%, as recorded by HOBO Pro v2 Temp/RH data loggers (ONSET, Bourne, MA, USA). Forty-five tubers were selected per cultivar, free of external disorders and were subjected to various measurements with optical and remote sensing technologies at 3 different storage periods (0, 2.5, and 5 months). The measurements were taken both on the external periderm (outside) of each tuber, as well as in the flesh (inside) after halving the potatoes longitudinally.

Color
The color was determined at two diametrically opposite spots at the equatorial diameter of the potato tubers using a chroma meter (Minolta CR-400, Minolta, Osaka, Japan), equipped with an 8-mm measuring head and a C illuminant (6774 K). The meter was calibrated using the manufacturer's standard white plate. Color changes were quantified in the L*, a*, and b* color space and a*/b* was later calculated. Hue angle [(h° = 180 + tan −1 (b*/a*)] and chroma values [C* = (a* 2 + b* 2 ) 1/2 ] were calculated from a* and b* values [27]. In particular, hue angle and chroma correspond to the basic tint of a color (e.g., yellow and red) and saturation or vividness of the color, respectively.

R-G-B Digital Imaging
The image components of potato tubers were extracted after processing the images captured by a digital R-G-B camera, Nikon D3300 DSLR (Nikon Inc., Melville, NY, USA) that consisted of a 24.2 megapixel image sensor with 12-bit resolution. This type of cameras created images by combing red (R), green (G), and blue (B) channels. All images were taken in a dark room, after the tubers were placed 22 cm below the camera lens and were illuminated by 4 halogen lamps (70 W, 2800 K, 1200 lumen) placed at an angle of 45° above them. The captured images were processed with suitable software (Mutispec v4. 4,2015) and the R, G, and B values of a selected section around the center of each tuber were extracted. Two sections at two opposite areas around the equatorial diameter of each potato tuber covering >75% of the surface area were selected and an average for each tuber was calculated. The R-G-B images were calibrated using black and white references following the equation R = (I -Bl)/(W -Bl), where I is the raw R-G-B image and Bl and W are the signals for black and white references, respectively.

Color
The color was determined at two diametrically opposite spots at the equatorial diameter of the potato tubers using a chroma meter (Minolta CR-400, Minolta, Osaka, Japan), equipped with an 8-mm measuring head and a C illuminant (6774 K). The meter was calibrated using the manufacturer's standard white plate. Color changes were quantified in the L*, a*, and b* color space and a*/b* was later calculated. Hue angle [(h • = 180 + tan −1 (b*/a*)] and chroma values [C* = (a* 2 + b* 2 ) 1/2 ] were calculated from a* and b* values [27]. In particular, hue angle and chroma correspond to the basic tint of a color (e.g., yellow and red) and saturation or vividness of the color, respectively.

R-G-B Digital Imaging
The image components of potato tubers were extracted after processing the images captured by a digital R-G-B camera, Nikon D3300 DSLR (Nikon Inc., Melville, NY, USA) that consisted of a 24.2 megapixel image sensor with 12-bit resolution. This type of cameras created images by combing red (R), green (G), and blue (B) channels. All images were taken in a dark room, after the tubers were placed 22 cm below the camera lens and were illuminated by 4 halogen lamps (70 W, 2800 K, 1200 lumen) placed at an angle of 45 • above them. The captured images were processed with suitable software (Mutispec v4. 4,2015) and the R, G, and B values of a selected section around the center of each tuber were extracted. Two sections at two opposite areas around the equatorial diameter of each potato tuber covering >75% of the surface area were selected and an average for each tuber was calculated. The R-G-B images were calibrated using black and white references following the equation R = (I -Bl)/(W -Bl), where I is the raw R-G-B image and Bl and W are the signals for black and white references, respectively.

R-G-NIR Digital Imaging
A digital commercial camera, Canon Powershot SX260 HS (Canon USA Inc., Huntington, NY, USA), including a 12-megapixel backlit complementary metal oxide semiconductor-CMOS sensor, was used in order to capture images that later had the red (R), green (G), and near infrared (NIR) values extracted. This type of camera used a blue-blocking filter and created a NIR-green-red digital image, which was processed by a suitable software (Mutispec v4. 4,2015) and the R, G, and NIR values of the selected sections were extracted. The light conditions and tuber distance were similar as described above. Two sections at two opposite areas around the equatorial diameter of each potato tuber covering > 75% of the area were selected and an average for each tuber was calculated. The calibration was performed similarly to RGB images, but in this case, the equation was R = (NIR -Bl)/(W -Bl), where NIR is the raw NIR image, and Bl and W are the signals for black and white references, respectively.

R-B-NIR Digital Imaging
A digital commercial camera, Canon Powershot SX260 HS (Canon USA Inc., NY, USA), including a 12-megapixel backlit CMOS sensor, was used in order to capture images that later had the red (R), blue (B), and near infrared (NIR) values extracted. This type of camera used a green-blocking filter and created a NIR-blue-red digital image (Hunt et al., 2012), which was processed by a suitable software (Mutispec v4. 4,2015) and the R, B, and NIR values of the selected sections were extracted. The light conditions and tuber distance were similar as described above. Two sections at two opposite areas around the equatorial diameter of each potato tuber covering > 75% of the area were selected and an average for each tuber was calculated. The calibration was performed similarly to RGB images, but in this case the equation was R = (NIR -Bl)/(W -Bl), where NIR is the raw NIR image, and Bl and W are the signals for black and white references, respectively.

Chlorophyll Fluorescence
Chlorophyll fluorescence was recorded with Fluorpen FP100-MAX, PAM fluorometer (Photon Systems Instruments, Drásov, Czech Republic). Following the manufacturer's protocols, after 30 min of dark adaptation, the Kautsky curve was recorded and analyzed according the OJIP-test using the FluorPen 1.0.0.6 software (Photon Systems Instruments). Two measurements were taken per tuber at two opposite spots at the equatorial diameter of each potato tuber and an average of the two was calculated. Based on OJIP-test for each couple of measurements, twenty-six different parameters were further calculated [13].

Vis/NIR Spectroscopy
A hyperspectral camera with a SPECIM V10E ImSpector spectrograph (Specim OY, Oulu, Finland) was used in order to capture the reflectance spectra of the skin or the flesh in potato tubers. Three scans were acquired per tuber at the equatorial diameter of each potato tuber and an average of the three spectra was calculated. This sensor has a standard input slit with a fixed length and produces a reflectance spectrum of each point on a narrow linear stripe on the target surface. The field-of-view is determined by the length of the slit and the chosen lens, under controlled laboratory conditions, this spectrograph is capable of capturing the reflectance of tuber tissue at the 400-1000 nm region of the electromagnetic spectrum, with a spectral resolution of 2.8 nm. The light conditions and tuber distance were similar as described above. A white plate (spectralon diffuse reflectance standard) was used as the reference reflectance spectra, capturing almost 100% reflectance in the 400-1000 nm range, under the specific laboratory lighting conditions.

Statistical Analyses
In order to assess the postharvest storage period of fresh potatoes, namely the freshness of raw tubers, various partial least square regression analysis models were tested using color, chlorophyll fluorescence components, R-G-B, R-G-NIR, and R-B-NIR imaging indexes, as well as spectra reflectance data both externally on the peel or internally in the flesh and multiple linear regressions (MLR) were applied. MLR is the linear modeling of the relationship between a dependent and more than one independent variables [28], but it often encounters multicollinearity issues among the variables and has to be used with caution, in order to avoid model overfitting. Therefore, only models with variance inflation factor (VIF), an indicator of the severity of multicollinearity, with values lower than 4 were selected in the present study. Partial least squares regression (PLSR) analyses, an advanced modern regression technique was further employed using only the identified important variables, in order to cross-validate the models. The number of latent variables were selected based on the lowest root mean square cross validation error (RMSECV) using the 'leave-one-sample-out' method.
For the discrimination of the three potato cultivars, the partial least squares discriminant analysis (PLSDA) was used. PLSDA is a linear classification method which combines regression algorithms and discrimination techniques, which minimize the cross-validation error [29]. The performance of PLS was estimated according to the non-error rate (NER), an index of total correctly classified samples, according to the equations (Table 2).
Moreover, advanced feature extraction methods were adapted, in order to detect specific wavelength regions that exhibit the most significant effect in the regression or discriminant analysis models, such as the genetic algorithm (GA). All data analyses were carried out using Microsoft Excel 2016, SPSS v. 25, MATLAB (Version R2017, The Mathworks Inc., Natick, MA, USA) and PLS Toolbox 8.6 (Eigenvector Research Inc., Manson, WA, USA).

Postharvest Storage Assessment
Given that the postharvest storage of potatoes cannot be easily estimated by judging sensorial indexes, such as visual color or firmness, it is of high importance to assess the freshness of a potato tuber, especially at low temperature using an objective remote sensing method. A prolonged storage of potatoes at low temperature is associated with chlorophyll and total glycoalkaloids increase, ascorbic acid decrease, and starch conversion into reducing sugars [30] that are involved both in quality deterioration such as brown discoloration during food preparation at high temperatures, as well as in safety issues, such as the synthesis of asparagine, an amino acid that is considered as a precursor of acrylamide synthesis, which in turn has been linked with health risks [6]. In our study, external color measurements proved to be an unreliable index of potato freshness, as long as cross-validated regression coefficient (Rcv) was less than 0.458, using the pooled data of all cultivars for the whole 5-month storage period at 5 • C (Table 1). Interestingly though, when data were tested in each individual cultivar, it turned out that external color changes were more pronounced in Spunta (Rcv = 0.728), than in Banba or Sylvana (Rcv = 0.406-0.456), implying that color in Spunta tubers is changing during storage, even in total darkness conditions. On the other hand, internal color changes were induced in all three cultivars ( Table 1). The regression coefficients of color changes during 5 months of storage in each cultivar was 0.901, 0.803, and 0.883 for Spunta, Banba, and Sylvana, respectively, and indeed the regression coefficient Rcv = 0.859 of the cross-validated data set irrespectively of the cultivar can be considered as a reliable index of postharvest storage in potato, which in turn depicts color changes in the internal flesh tissue of the tubers. A model combining digital imaging indices (NIR -R/NIR + R), (NIR -B/NIR + B) and chlorophyll fluorescence parameters (Fi and Fv) resulted in a high correlation with postharvest storage period (R ALL CVs = 0.885), which is also confirmed even by the models generated for each of the three individual cultivars (R SPUNTA = 0.914, R BANBA = 0.829, and R SYLVANA = 0.922). Fi expresses the fluorescence at the I-step (30 ms) of OJIP transient and Fv (=Fm-Fo) is an index of maximal variable fluorescence [31]. Fm expresses the maximal fluorescence, when all photosystem (PS) II reaction centers (RCs) are closed and Fo is the minimal fluorescence, when all PS II RCs are open (at t = 0). Therefore, the multiple regression model can be of high importance in assessing postharvest freshness of potato, not only due to the strong regression with storage time, but also given that each of the components that participate in that have a significant effect (p < 0.05) and at the same time exhibit a VIF lower than 4, which was set as a threshold, in order to identify the best algorithm. In addition, multiple regressions were statistically tested in order to select the combination of specific parameters that: (i) provide a high regression coefficient with storage time, (ii) exhibit a significant effect in the generated model (p < 0.05), and (iii) present a collinearity (VIF) score below 4, in order to be considered as reliable.
A non-contact system based on analyzing images captured above potatoes has already been found to detect external defects on whole tubers (skin cutting, shatter bruise, common scab, greening, cracks, etc.) using color images with classification rates higher than even 95% [26,32]. The basis of this method is the image analysis of the sample and comparison with a standard and later a decision making related to the acceptance or rejection of the samples [25]. Simultaneously, the fluorescence indexes have been successfully applied on potatoes, in order to assess postharvest changes during storage [1], and it has been shown that even though chlorophyll is not produced, the tubers still possess a photosynthetic apparatus whose physiological state can be measured by fluorometry. According to Maxwell and Johnson [33], after light illumination of plant tissue, a small percentage of energy is emitted during photosynthesis in the form of fluorescence, upon the return of an exited electron back to its ground state.
The partial least square regression analysis of spectra reflectance at the Vis/NIR region of the electromagnetic spectrum (400-1000 nm) that was captured at three distinct areas on the potato tubers, externally or internally, turned out to the best reliable indicator of the potato freshness (Table 1). In particular, the model that was generated based on all pooled data, irrespectively of the cultivar, had a cross-validated regression coefficient (Rcv) equal to 0.981 and 0.947, as determined by external and internal measurements, respectively, not significantly lower than when only each individual cultivar was tested (0.975-0.980 and 0.930-0.963, respectively) ( Table 1). Data normalization followed by the Savitzky-Golay first derivatization algorithm combined with mean centering were identified as the most appropriate preprocessing techniques to be applied on the spectral data among several others tested ( Figure 2) and partial least squares regression analysis was later performed using six latent variables based on the lowest root mean square error of cross-validation (RMSECV).
However, the efficient use of a sorting system, despite its' accuracy, is dependent on the rate of productivity (t/h), which is a crucial factor in determining its functionality and marketability [32]. In line with the above, spectroscopy can be used as a tool to estimate optimal spectra based on specific criteria that can be commercially implemented in an online sorting system at appropriate speed for commercial use. Therefore, the genetic algorithm was selected as a feature extraction method in order to identify specific regions of interest that mostly affect the accuracy of the model and indeed it was concluded that the spectra captured externally at the regions of 517-575, 604-633, 664-721, 780-809, 839-896, and 926-1000 nm and internally at 458-487, 546-633, 664-692, 721-750, 780-838, and 868-935 nm (Table 1) apparently provide the most useful information regarding the reliable assessment of potato postharvest storage life at 5 • C (Figure 3), without significantly reducing the efficacy of the model generated either from the pooled data or from each cultivar's reflectance measurements, while minimizing the number of latent variables to as low as 4 ( Figure 4). Table 1. Regression coefficients (R) of cross-validated partial least square analysis between storage period (0, 2.5, and 5 months) and color, digital imaging, and chlorophyll fluorescence as well as Vis/NIR spectroscopy data of the whole 400-1000 nm and the genetic algorithm (GA) selected regions. Data were either pooled from all three varieties (Spunta, Banba, or Sylvana) or individually within each one of them.

Technique
Tuber

Cultivar Discrimination
The need for developing a proper rapid technology able to properly classify the potato genotypes is further amplified by the demand to prevent a deceptive practice in the retail marketing of potatoes, which is the distribution of a mixed bunch of cultivars containing morphologically similar tubers, which are not, however, of the same quality standards. However, this practice is drifting the consumers' preference, purchase, and consumption to inferior quality potatoes inversely proportional to the increased purchase cost. As a result, popular areas, reputable in the production of optimal quality potatoes, such as of protected designations of origin (PDO) or protected geographical indications (PGI) are subjected to substantial profit loss. Consequently, the high value of traditional cultivars that are grown in the above areas is capitalized by retailers that distribute potatoes of similar external traits but of different cultivars with inferior quality grown in miscellaneous geographical sites.
Color of the peel or the flesh resulted only in 66% accuracy (non-error rate) in both calibrated or cross-validated classification models, which implies that there were not significant differences of the peel or the inner tissue color among these three cultivars (Spunta, Banba, and Sylvana) and, therefore, color measurement cannot be considered as a reliable discrimination method. Neither did the normalized parameters of the digital images (R-G-B, R-G-NIR, or R-B-NIR) provide valuable information in separating the three potato cultivars with accuracy scores lower than 51%, 44%, and 34%, respectively, when captured above the peel of the tuber, and lower than 43%, 34%, and 47%, respectively, when captured in the flesh pith (data nor shown). Better discrimination models were generated when chlorophyll fluorescence parameters were tested with Fi, Fv, and Phi_Eo exhibiting the highest VIP scores (a measure of a variable's importance in the PLSDA model) (data not shown). However, when digital imaging indexes were combined with the above chlorophyll fluorescence parameters, several models with higher classification accuracy were generated. Therefore, a combination of (NIR -G)/(NIR + G), (NIR -B)/(NIR + B), Fi, and Fv measured on the external area of the tubers was able to separate the three cultivars by 76% in the cross-validated model and by only 61% in the internal flesh, confirming once more the assumption that the flesh components are more or less similar in Spunta, Banba, and Sylvana (Table 2).
In our study, the best discrimination of the three potato cultivars, was achieved using the spectral reflectance measurements upon proper preprocessing of the raw data ( Figure 6) of the skin of the tubers (99% and 93% of the calibration and the cross-validation sample sets), almost similar to the destructive spectral acquisition on the flesh after halving the tubers (96% and 95%, respectively) ( Table 2).
Similarly, chemometric treatment of the Fourier Transformed Infrared Spectroscopy (FTIT) spectra allowed the discrimination of eight different potato cultivars with a good accuracy [34]. Being a rapid, and/or noninvasive method, the spectroscopy technique is suitable in on-line applications combining less time consumption, greater robustness, higher reproducibility, and cost effective performance compared to human labor or other destructive laboratory methods used in quality assessment [24,35]. Therefore, spectra reflectance data from the skin of the potato tubers may be reliably used in the assessment of postharvest storage life, as well as in the cultivar discrimination process. The scientific value and industrial benefit of this study lies on the detection of specific regions of the Vis/NIR electromagnetic spectrum or even distinct single wavelengths that can be incorporated in models and operate in portable devices or station based screening equipment. Thus, potatoes may be classified into separate cultivars and decisions regarding postharvest management and retail distribution be made in relation to their specific requirements and consequently even affect the quality of the processed product in the potato chip industry. Table 2. Calibration and cross-validation classification of correctly discriminated potatoes into the three cultivars (Spunta, Banba, or Sylvana) using the partial least squares discriminant analysis (PLSDA) classifier from color, digital imaging, and chlorophyll fluorescence as well as Vis/NIR spectroscopy data of the whole 400-1000 nm region, as well as from the selected regions upon the genetic algorithm implementation.

Calibration
Cross-Validation   Classification of potato tubers into 3 commercial potato cultivars (Spunta, Banba, and Sylvana) that were stored for 0, 2.5, and 5 months at 5 °C based on the scores of LV1, LV2, and LV3 after the partial least squares discriminant analyses on the data from specific wavelength regions that were extracted with the genetic algorithm as a feature extraction technique of the spectra that were captured externally on the skin of the tubers.
Similarly, chemometric treatment of the Fourier Transformed Infrared Spectroscopy (FTIT) spectra allowed the discrimination of eight different potato cultivars with a good accuracy [34]. Being a rapid, and/or noninvasive method, the spectroscopy technique is suitable in on-line applications combining less time consumption, greater robustness, higher reproducibility, and cost effective performance compared to human labor or other destructive laboratory methods used in quality assessment [24,35]. Therefore, spectra reflectance data from the skin of the potato tubers may be reliably used in the assessment of postharvest storage life, as well as in the cultivar discrimination process. The scientific value and industrial benefit of this study lies on the detection of specific regions of the Vis/NIR electromagnetic spectrum or even distinct single wavelengths that can be incorporated in models and operate in portable devices or station based screening equipment. Thus, potatoes may be classified into separate cultivars and decisions regarding postharvest management and retail distribution be made in relation to their specific requirements and consequently even affect the quality of the processed product in the potato chip industry. Figure 6. Spectra reflectance data upon pre-processing (normalization, Savitzky-Golay first derivative and mean centering) that were externally captured from 3 commercial potato cultivars (Spunta, Banba, and Sylvana) during storage for 0, 2.5, or 5 months at 5 • C externally on the skin (A) or internally in the flesh (B) of the tubers. In each cultivar, the data were pooled from the samples stored for 0, 2.5, or 5 months at 5 • C.
Appl. Sci. 2021, 11, x FOR PEER REVIEW 13 of 15 Figure 6. Spectra reflectance data upon pre-processing (normalization, Savitzky-Golay first derivative and mean centering) that were externally captured from 3 commercial potato cultivars (Spunta, Banba, and Sylvana) during storage for 0, 2.5, or 5 months at 5 °C externally on the skin (A) or internally in the flesh (B) of the tubers. In each cultivar, the data were pooled from the samples stored for 0, 2.5, or 5 months at 5 °C.

Figure 7.
Classification of potato tubers into 3 commercial potato cultivars (Spunta, Banba, and Sylvana) that were stored for 0, 2.5, and 5 months at 5 °C based on the scores of LV1, LV2, and LV3 after the partial least squares discriminant analyses on the data from specific wavelength regions that were extracted with the genetic algorithm as a feature extraction technique of the spectra that were captured externally on the skin of the tubers.
Similarly, chemometric treatment of the Fourier Transformed Infrared Spectroscopy (FTIT) spectra allowed the discrimination of eight different potato cultivars with a good accuracy [34]. Being a rapid, and/or noninvasive method, the spectroscopy technique is suitable in on-line applications combining less time consumption, greater robustness, higher reproducibility, and cost effective performance compared to human labor or other destructive laboratory methods used in quality assessment [24,35]. Therefore, spectra reflectance data from the skin of the potato tubers may be reliably used in the assessment of postharvest storage life, as well as in the cultivar discrimination process. The scientific value and industrial benefit of this study lies on the detection of specific regions of the Vis/NIR electromagnetic spectrum or even distinct single wavelengths that can be incorporated in models and operate in portable devices or station based screening equipment. Thus, potatoes may be classified into separate cultivars and decisions regarding postharvest management and retail distribution be made in relation to their specific requirements and consequently even affect the quality of the processed product in the potato chip industry. Figure 7. Classification of potato tubers into 3 commercial potato cultivars (Spunta, Banba, and Sylvana) that were stored for 0, 2.5, and 5 months at 5 • C based on the scores of LV1, LV2, and LV3 after the partial least squares discriminant analyses on the data from specific wavelength regions that were extracted with the genetic algorithm as a feature extraction technique of the spectra that were captured externally on the skin of the tubers.

Conclusions
The partial least square regression analysis of spectra reflectance at the Vis/NIR region of the electromagnetic spectrum (400-1000 nm) that was captured at three distinct areas on the potato tubers, externally or internally, produced the most reliable models in assessing potato freshness, irrespectively of the cultivar, with a cross-validated regression coefficient (Rcv) equal to 0.981 and 0.947, as determined by external or internal measurements, respectively. Feature selection algorithms (variance inflation factor, variable importance scores, and genetic algorithms) identified specific regions of interest that mostly affected the accuracy of the model, in order to provide the most accurate multilinear model in terms of highest probability, strongest regression, and lowest collinearity and root mean cross validation error. Similarly, the best discrimination of the three potato cultivars, was achieved with the spectra reflectance measurements on the skin of the tubers (93% of the cross-validation sample sets), while R-G-NIR and R-B-NIR image components combined with selected chlorophyll fluorescence parameters (Fi, Fv, and Phi_Eo) performed better than color measurements in both potato freshness assessment or cultivar classification in all non-destructive trials, irrespective of the cultivar tested.