Estimation of Starch Hydrolysis in Sweet Potato ( Beni haruka ) Based on Storage Period Using Nondestructive Near-Infrared Spectrometry

: Sweet potatoes are a substantial source of nutrition and can be added to processed foods in the form of paste. The moisture and starch contents of these potatoes affect the physicochemical properties of sweet potato paste. In this study, the changes in the moisture, starch, and α -amylase content of sweet potatoes were measured for eight weeks after harvest. Using nondestructive near-infrared analyses and chemometric models, the moisture and starch contents were predicted. The partial least squares (PLS) method was used for prediction, while linear discriminant analysis (LDA) was used for discrimination. To increase the accuracy of the model, the near-infrared spectrum was preprocessed using the Savitzky–Golay derivative (S–G), standard normal variate (SNV), and multiplicative scattering correction methods. When applying PLS to the moisture content, the best calibration model accuracy was obtained using the S–G preprocessed spectrum. Furthermore, the best calibration model accuracy for starch content was obtained using the SNV preprocessed spectrum. The moisture and starch contents were categorized into ﬁve classes for LDA, with results indicating that the internal quality of sweet potatoes can be predicted and classiﬁed using chemometric models through nondestructive detection.


Introduction
Sweet potatoes originate from Central and South America. Total worldwide production currently exceeds 105 million metric tons annually, with Asia producing approximately 82% of this yield [1]. Carbohydrates are the primary components of sweet potatoes, but they also contain other nutrients such as large amounts of vitamin C and potassium, dietary fiber, yalapin, and seraffin [2,3]. Beni haruka is one of the most popular varieties of sweet potato in Asian countries, and can be prepared in a variety of styles, such as dried, roasted, boiled, steamed, baked, and fried [4]. The consumption of sweet potatoes has recently increased owing to their nutritional advantages and sweet taste. They are sold worldwide in the form of processed foods such as pastes, purees, jerky, chips, and fries [5,6]. Sweet potato paste, which is the most consumed sweet potato-based processed food owing to its ease of transportation and use [7], is obtained by steaming and baking sweet potatoes, followed by peeling, crushing, and removing their fibers using a sieve. More than 90% of the paste is composed of sweet potatoes, and the quality of the final product is determined by the moisture and starch content of the raw sweet potatoes used.
The quality of agricultural products can be classified as either external or internal based on a combination of various attributes, such as appearance, taste, physical qualities, and hardness [8]. External quality is evaluated based on visual appearance and hand

Sweet Potatoes
The sweet potatoes (Beni haruka) in our study were grown and harvested in Gimje-si, Jeollabuk-do, South Korea (35 • 53 27.6" N 126 • 54 45.7" E). They were cured at 34 • C under 100% relative humidity for 3 days, and stored at 12 • C under 85% relative humidity in a storage warehouse at the Gongdeok Nonghyup Agricultural Products Processing Plant (Gongdeok 7-gil, Gongdeok-myeon, Gimje-si, Jeollabuk-do, Korea). A total of 144 sweet potatoes were stored, and 18 sweet potatoes were delivered to the laboratory in each of weeks 1-8, involving 2 h of transportation. The NIR spectra and internal qualities (moisture, starch, and α-amylase activity) of these sweet potatoes were analyzed.

Measurement of Quality Characteristics during Storage 2.2.1. Moisture
After conducting nondestructive NIR spectroscopy, the sweet potatoes were cut into 1 mm thick slices, and the slices near the center were selected to verify the internal quality. The moisture content was determined using a modified Association of Official Analytical Chemists (AOAC) method [29]. A dish with a lid was left in an oven at 105 • C until it reached a constant weight, and was then cooled in a desiccator for 30 min. A sweet potato sample was added to the dish, the dish was covered with a lid, and the weight (W1) was measured. After placing the dish in the oven, the lid was opened slightly, and the sample was dried for 3 h. The sample in the cooled dish was stored in a desiccator for approximately 30 min, after which its mass was measured. The sample was then dried again for 1 to 2 h, and the same operation was repeated until a constant weight was attained (W2). The moisture content was calculated using the following equation: moisture (%) = [(W1-W2)/W1] × 100.

Starch
After conducting the NIR measurements and sampling for moisture and α-amylase assays, the remaining portion of the sweet potato was baked in an oven at 180 • C for 90 min. The baked sweet potatoes were cut near the center to obtain a cross-sectional slice. RGB images of the sample were captured in a white plastic box with dimensions of 30 cm × 30 cm × 30 cm, using a camera placed approximately 30 cm above the baked sweet potato. The amount of light entering the box was controlled by a black curtain. The white balance was set using a neutral gray card. The acquired images were processed using ImageJ software (NIH, Bethesda, MD, USA). An RGB image showing only the baked sweet potato cross-section and excluding the background was selected, and converted into a 256-level grayscale image with black corresponding to 0 and white to 255. The images were translated into a histogram with the gray intensity level (GIL) ranging from 0 to 255 on the x-axis and pixel count on the y-axis. The starch content was determined by the GIL, where the starch corresponded to the lighter area where GIL>100 (A1) and the hydrolyzed starch part corresponded to darker areas where GIL<100 (A2). The starch content was calculated using the following equation: starch content (%) = [A1/(A1 + A2)] × 100.

α-Amylase Assay
Sweet potato samples with a thickness of 1 mm were homogenized, and enzyme extract was obtained by adding a McIlvaine buffer (pH 7.0) solution in amounts that were 10 times the volume of the sweet potato sample. The substrate solution contained 1% (w/v) soluble starch. Five milliliters of a preincubated substrate solution, 4 mL of McIlvaine buffer, and 1 mL of 0.1% CaCl 2 were incubated with 10 mL of the enzyme extract at 37 • C for 20 min. The reaction was terminated using ice to reduce the temperature to 0 • C for 10 min. The supernatant was collected after centrifugation at 10,000 rpm at 4 • C for 10 min. Next, 0.2 mL of the supernatant and 0.6 mL of dinitrosalicylic acid (DNS) solution were reacted at 100 • C for 5 min, and then cooled at room temperature. The absorbance was measured at 540 nm using a UV-Vis spectrophotometer (S-3100, SCINCO, Seoul, Korea) with a blank solution of 0.2 mL of McIlvaine buffer (pH 7.0) and 0.6 mL of DNS solution, comprising 1% dinitrosalicylic acid, 1% sodium hydroxide, 0.2% phenol, and 0.05% sodium sulfite [30].

Quality Evaluation Using Nondestructive Techniques
The NIR spectrum was obtained using halogen tungsten lamps (JCR12V-100WBAU, USHIO, Tokyo, Japan) and a detector (Ava Spec-UV/VIS/NIR, AVANTES, Apeldoorn, the Netherlands) controlled via software developed by Hansung Engineering (Cheonansi, Chungcheongnamdo, Korea). The spectral acquisition of uncut sweet potatoes was conducted within a spectral range of 500 to 2500 nm, using a sampling interval of 1.0 nm. The transmission mode was used, and transmittance (%T) was converted into absorbance value using the conversion equation absorbance = 2 − log(%T). The range of 760-1420 nm was used for the spectrum analysis, with the spectrum based on the average values of 16 scans collected three times for each sample. The measurement equipment included an NIR spectrophotometer, as shown in Figure 1A. The spectra were obtained and the chemometric model was applied via NIR penetration of the central portion of intact sweet potatoes during the transmission mode, which was then collected by the detector. It was confirmed that the NIR spectra could be obtained from sweet potatoes using an industrial device, and that the chemometric model could be applied ( Figure 1B).

Measurement of NIR Spectrum
The NIR spectrum was obtained using halogen tungsten lamps (JCR12V-100WBAU, USHIO, Tokyo, Japan) and a detector (Ava Spec-UV/VIS/NIR, AVANTES, Apeldoorn, the Netherlands) controlled via software developed by Hansung Engineering (Cheonansi, Chungcheongnamdo, Korea). The spectral acquisition of uncut sweet potatoes was conducted within a spectral range of 500 to 2500 nm, using a sampling interval of 1.0 nm. The transmission mode was used, and transmittance (%T) was converted into absorbance value using the conversion equation absorbance = 2 − log(%T). The range of 760-1420 nm was used for the spectrum analysis, with the spectrum based on the average values of 16 scans collected three times for each sample. The measurement equipment included an NIR spectrophotometer, as shown in Figure 1A. The spectra were obtained and the chemometric model was applied via NIR penetration of the central portion of intact sweet potatoes during the transmission mode, which was then collected by the detector. It was confirmed that the NIR spectra could be obtained from sweet potatoes using an industrial device, and that the chemometric model could be applied ( Figure 1B).

Pretreatment of NIR Spectrum
To improve the accuracy of the model, the NIR spectra were preprocessed. The Savitzky-Golay derivative (S-G), standard normal variate (SNV), and multiplicative scattering correction (MSC) methods were employed as pretreatment methods using an Unscrambler (Camo Analytics, Oslo, Norway) [31]. When using the S-G method, a polynomial derivative was applied. In this study, a first-order derivation was used with a secondorder polynomial and 11 smoothing points. The spectra were smoothed by reducing the signal-to-noise ratio of the existing finite difference before calculating the derivative [32].

Chemometrics Pretreatment of NIR Spectrum
To improve the accuracy of the model, the NIR spectra were preprocessed. The Savitzky-Golay derivative (S-G), standard normal variate (SNV), and multiplicative scattering correction (MSC) methods were employed as pretreatment methods using an Unscrambler (Camo Analytics, Oslo, Norway) [31]. When using the S-G method, a polynomial derivative was applied. In this study, a first-order derivation was used with a second-order polynomial and 11 smoothing points. The spectra were smoothed by reducing the signalto-noise ratio of the existing finite difference before calculating the derivative [32]. The SNV and MSC performed Savizky-Golay smoothing to remove noise before pretreatment, and were used to correct the spectral data for scatter effects [33].

Prediction of Post-Harvest Quality Using NIR
To apply the prediction model for internal quality, a calibration model was established using PLS and an Unscrambler ( Figure 2). Outliers were evaluated using Q residuals and Hotelling's T 2 values. Twelve samples with high Q residuals and high Hotelling's T 2 values were removed. Sweet potatoes were randomly categorized into two groups: a calibration set (100 samples) for modeling and a prediction set (32 samples) for validating the model. PLS is a popular linear regression algorithm typically used to construct models for value prediction [34], and is widely used in spectral calibration analyses by creating latent variables (LVs) that correspond to the projection of independent and dependent variables. The relationship between the LVs and their target attributes is established mathematically. The cross-validation set was calculated from the calibration set to indicate the error of the proposed calibration models [35]. The prediction set was calculated from the prediction data to evaluate the prediction ability of the PLS models. The accuracy of the prediction model was evaluated based on the fitting correlation coefficients for the calibration (R c 2 ), cross-validation (R cv 2 ), and prediction datasets (R p 2 ), along with their corresponding root mean square errors (RMSEC, RMSECV, RMSEP). The RMSEP/RMSECV ratio and the residual prediction deviation (RPD) were adopted to evaluate the robustness of the model. The RPD value is the ratio of the standard error of prediction to the standard deviation; if the RMSEP/RMSECV value of the model was less than 1.2, it was assumed that the performance of the model was robust and accurate [36,37]. An RPD value of 2 or more indicates good prediction capability, a value in the range of 1.5-2 indicates intermediate prediction capability, and < 1.5 indicates poor prediction capability [38]. The squared correlation coefficient (R 2 ) can be used to evaluate the goodness of fit between the actual and predicted values. The closer R 2 is to 1, the better the model fits. The SNV and MSC performed Savizky-Golay smoothing to remove noise before pretreatment, and were used to correct the spectral data for scatter effects [33].

Prediction of Post-harvest Quality Using NIR
To apply the prediction model for internal quality, a calibration model was established using PLS and an Unscrambler ( Figure 2). Outliers were evaluated using Q residuals and Hotelling's T 2 values. Twelve samples with high Q residuals and high Hotelling's T 2 values were removed. Sweet potatoes were randomly categorized into two groups: a calibration set (100 samples) for modeling and a prediction set (32 samples) for validating the model. PLS is a popular linear regression algorithm typically used to construct models for value prediction [34], and is widely used in spectral calibration analyses by creating latent variables (LVs) that correspond to the projection of independent and dependent variables. The relationship between the LVs and their target attributes is established mathematically. The cross-validation set was calculated from the calibration set to indicate the error of the proposed calibration models [35]. The prediction set was calculated from the prediction data to evaluate the prediction ability of the PLS models. The accuracy of the prediction model was evaluated based on the fitting correlation coefficients for the calibration (Rc 2 ), cross-validation (Rcv 2 ), and prediction datasets (Rp 2 ), along with their corresponding root mean square errors (RMSEC, RMSECV, RMSEP). The RMSEP/RMSECV ratio and the residual prediction deviation (RPD) were adopted to evaluate the robustness of the model. The RPD value is the ratio of the standard error of prediction to the standard deviation; if the RMSEP/RMSECV value of the model was less than 1.2, it was assumed that the performance of the model was robust and accurate [36,37]. An RPD value of 2 or more indicates good prediction capability, a value in the range of 1.5-2 indicates intermediate prediction capability, and < 1.5 indicates poor prediction capability [38]. The squared correlation coefficient (R 2 ) can be used to evaluate the goodness of fit between the actual and predicted values. The closer R 2 is to 1, the better the model fits.   Discrimination of Sweet Potatoes Using NIR LDA was performed using NIR data from the calibration set of PLS. The moisture and starch contents were categorized into five classes within the range, and then analyzed using LDA for classification using MATLAB code (2019a, MathWorks, Natick, MA, USA). LDA classified the raw and preprocessed NIR spectra by reducing the number of dimensions and maximizing the ratios between and within classes [39]. Figure 3 and Table S1 show the changes in the internal quality of sweet potatoes during storage according to the moisture, starch, and α-amylase contents. The general range of moisture content was 59.44−64.83 wt%, which was maintained at an approximate constant through regulation of the humidity in the storage warehouse ( Figure 3A). The starch content decreased as the storage period increased ( Figure 3B). Figure 4 shows the cross-section of the baked sweet potatoes according to the degree of starch decomposition. As shown in Figure 3C, α-amylase activity increased during the first 3 weeks of the storage period, and decreased thereafter. In addition, α-amylase activity increased significantly during weeks 2 and 3 of the storage period. During a storage period of 3 weeks, the highest activity was measured at 7839 units/g. The α-amylase activity decreased significantly in weeks 7 and 8 during the storage period. After the amylase enzyme activity increased, the starch concentration rapidly decreased. This is a general observation of starch hydrolysis in sweet potatoes. A previous study reported the measured changes in the α-amylase activity of five sweet potato varieties grown in Uganda, and found that the α-amylase activity varied depending on whether the sweet potato flesh was cream, white, orange, or pale orange [40]. However, their results indicated that all varieties exhibited the highest levels of α-amylase activity at 3 weeks, comparable to our present study. Hagenimana et al. measured the distribution of α-amylase in the outer and inner portions of sweet potatoes using immunological detection and activity measurements [41]. Their immunological detection showed the presence of α-amylase in the outer portions of sweet potatoes, for example, laticifer, cambium, and parenchyma tissues, but not in the inner portion. The α-amylase activity was measured by mechanically separating the outer and inner portions. Furthermore, α-amylase activity was higher in the outer portion than in the inner portion. Sarikaya et al. measured the abilities of α-amylase and β-amylase from Bacillus amyloliquefaciens and B. cereus in degrading raw starch granules [42]. The two types of α-amylase showed different decomposition patterns, and the degradation ability of α-amylase was higher than that of β-amylase. Previous studies have shown that α-amylase is released from the outer portion of sweet potatoes after harvesting. Nabubuya et al. measured the changes in glucose and total starch content of cured sweet potatoes over a storage period of 8 weeks [12]. Their experimental results indicated that as the storage period increased, the glucose content also increased and total starch content decreased. In addition, a decrease in starch content was confirmed, which was considered to be responsible for the reduced paste viscosities of sweet potato flour.

Spectral Characteristics of NIR Acquisition
The raw spectrum is shown in Figure 5A, with the transmission curves expressed as the absorbance. The spectrum comprises two main peaks located at approximately 970 and 1270 nm. The spectra preprocessed through S-G, SNV, and MSC are shown in Figure 5B-D. It should be noted that the NIR spectra can be affected by water in agricultural products. Luck et al. reported that the stretching and bending of O-H occurred at 970, 1190, 1450, and 1940 nm, as indicated by the NIR spectra [43]. Water can form bonds with various organic molecules, creating a shift in the absorption wavelength during bond formation [44]. Moreover, the NIR spectra are mainly observed as a broad band caused by the overlapping absorption, corresponding to the combinations and overtones of the vibrations of the C-H,  [45]. In this study, it was observed that the NIR spectra were affected by moisture content. The NIR spectrum measurements after the dehydration of sweet potatoes indicated that the raw material contained approximately 30% water; although the 970 nm peak was weakened, the 1270 nm peak remained relatively strong ( Figure 6). It was confirmed that the 970 and 1270 nm peaks were the second overtones of the O-H and C-H stretching, respectively, and the intensity of the peak was altered due to hydrogen bonding with water. Therefore, the O-H stretching of water contributed to the 970 nm peak, based on the prediction model of starch in Section 3.2.2, and the 1270 nm peak is related to the C-H stretching of starch.

Spectral Characteristics of NIR Acquisition
The raw spectrum is shown in Figure 5A, with the transmission curves expressed the absorbance. The spectrum comprises two main peaks located at approximately and 1270 nm. The spectra preprocessed through S-G, SNV, and MSC are shown in Fig  5B-D. It should be noted that the NIR spectra can be affected by water in agricultu products. Luck et al. reported that the stretching and bending of O-H occurred at 9 1190, 1450, and 1940 nm, as indicated by the NIR spectra [43]. Water can form bonds w various organic molecules, creating a shift in the absorption wavelength during bond f mation [44]. Moreover, the NIR spectra are mainly observed as a broad band caused the overlapping absorption, corresponding to the combinations and overtones of the brations of the C-H, N-H, and O-H chemical bonds [45]. In this study, it was observ that the NIR spectra were affected by moisture content. The NIR spectrum measureme after the dehydration of sweet potatoes indicated that the raw material contained appr imately 30% water; although the 970 nm peak was weakened, the 1270 nm peak remain relatively strong ( Figure 6). It was confirmed that the 970 and 1270 nm peaks were second overtones of the O-H and C-H stretching, respectively, and the intensity of peak was altered due to hydrogen bonding with water. Therefore, the O-H stretching water contributed to the 970 nm peak, based on the prediction model of starch in Sect 3.2.2, and the 1270 nm peak is related to the C-H stretching of starch.   Table 1 presents the calibration, cross-validation, and prediction of PLS with raw, S-G, SNV, and MSC preprocessing spectra for the moisture and starch contents of sweet potatoes. The regression plots and regression coefficient plots of the PLS are shown in Figures S1-S8. The model performance for the moisture content was highest over the S-G preprocessed spectrum for calibration (RC 2 = 0.7878, RMSEC = 1.1147), and the MSC preprocessed spectrum for prediction (RP 2 = 0.8065, RMSEP = 1.2480). For the starch content, the calibration model performance was highest over the SNV preprocessed spectrum (RC 2 = 0.8320, RMSEC = 10.5383), and the highest prediction model performance was over the MSC preprocessed spectrum (RP 2 = 0.7811, RMSEP = 11.4115). The raw spectrum is affected by noise and scatter effects, but inaccurate information of the raw spectrum can be removed by preprocessing using S-G, SNV, and MSC. Noise is caused by the electronics and mechanical vibration of the instrument, and the scatter effect causes the diffuse reflectance. Therefore, the PLS model performances of the S-G, SNV, and MSC preprocessed spectra were improved compared to the raw spectrum. The RMSEP/RMSECV values of all measured models ranged from 0.7710 to 1.0424. Furthermore, the model accuracy was measured based on the RPD values. The RPD values of all measured models ranged from 1.0211 to 2.1749, indicating moderate to high prediction accuracy of the models. Furthermore, a PLS regression model analysis was performed for α-amylase. However, both the calibration and prediction models indicated R 2 values of 0.2 or less, and were consequently not included in the experimental results. Farhadi et al. predicted the starch, sugar, and moisture contents, which were considered qualitative variables for potatoes, using PLS and Vis/NIR spectroscopy [46]. Rady and Guyer predicted and classified the glucose and sucrose contents of entire and sliced potatoes using NIR reflectance [23]. Guoquan et al. predicted the physiochemical quality (amylose content, total starch content, protein content, phosphorus content, solubility, crystallinity, and hot paste viscosity) of sweet potato starch using NIR reflectance spectroscopy and the PLS model [47]. The measurement results showed that the R 2 value of the prediction model was 0.614-0.917, implying a high correlation between the NIR spectrum and the physiochemical quality of sweet potato starch. Thus, it was confirmed that sweet potato starch affects the NIR spectrum. As a result, it was assumed in this study that starch content could be predicted using the PLS model.  Table 1 presents the calibration, cross-validation, and prediction of PLS with raw, S-G, SNV, and MSC preprocessing spectra for the moisture and starch contents of sweet potatoes. The regression plots and regression coefficient plots of the PLS are shown in Figures S1-S8. The model performance for the moisture content was highest over the S-G preprocessed spectrum for calibration (R C 2 = 0.7878, RMSEC = 1.1147), and the MSC preprocessed spectrum for prediction (R P 2 = 0.8065, RMSEP = 1.2480). For the starch content, the calibration model performance was highest over the SNV preprocessed spectrum (R C 2 = 0.8320, RMSEC = 10.5383), and the highest prediction model performance was over the MSC preprocessed spectrum (R P 2 = 0.7811, RMSEP = 11.4115). The raw spectrum is affected by noise and scatter effects, but inaccurate information of the raw spectrum can be removed by preprocessing using S-G, SNV, and MSC. Noise is caused by the electronics and mechanical vibration of the instrument, and the scatter effect causes the diffuse reflectance. Therefore, the PLS model performances of the S-G, SNV, and MSC preprocessed spectra were improved compared to the raw spectrum. The RMSEP/RMSECV values of all measured models ranged from 0.7710 to 1.0424. Furthermore, the model accuracy was measured based on the RPD values. The RPD values of all measured models ranged from 1.0211 to 2.1749, indicating moderate to high prediction accuracy of the models. Furthermore, a PLS regression model analysis was performed for α-amylase. However, both the calibration and prediction models indicated R 2 values of 0.2 or less, and were consequently not included in the experimental results. Farhadi et al. predicted the starch, sugar, and moisture contents, which were considered qualitative variables for potatoes, using PLS and Vis/NIR spectroscopy [46]. Rady and Guyer predicted and classified the glucose and sucrose contents of entire and sliced potatoes using NIR reflectance [23]. Guoquan et al. predicted the physiochemical quality (amylose content, total starch content, protein content, phosphorus content, solubility, crystallinity, and hot paste viscosity) of sweet potato starch using NIR reflectance spectroscopy and the PLS model [47]. The measurement results showed that the R 2 value of the prediction model was 0.614-0.917, implying a high correlation between the NIR spectrum and the physiochemical quality of sweet potato starch. Thus, it was confirmed that sweet potato starch affects the NIR spectrum. As a result, it was assumed in this study that starch content could be predicted using the PLS model.

Discrimination of Sweet Potatoes by NIR
LDA was performed to classify the moisture and starch content of sweet potatoes into a qualitative discriminant model (Figure 7). For LDA, the S-G and SNV methods were used to preprocess the NIR spectra for the moisture and starch content, respectively; these were selected as the most robust prediction models. The LDA of the moisture content based on the raw NIR spectrum indicated content classification, but with ambiguous boundaries. After the S-G method was used to preprocess the raw NIR spectrum, it was confirmed through LDA that the content was accurately classified. Likewise, the results of starch content based on the raw NIR spectrum indicated content classification with ambiguous boundaries. The boundary between the 71-80% and 0-10% classes was accurate; however, it was subsequently concluded that these two categories can perform adequate classification. After SNV preprocessing, the LDA results showed that the classification boundaries of the 0-10%, 11-30%, and 51-70% categories were close, although the boundaries of the others were clearly classified. Ding et al. dried and pulverized purple sweet potatoes, and the resulting powder was mixed with white sweet potato flour at specific ratios [48]. The samples were evaluated using NIR spectroscopy, and the mixing ratios of the samples were classified using LDA. Carvalho et al. performed LDA with raw NIR spectra and preprocessed NIR spectra using MSC and S-G methods to distinguish the geographical origin of sugarcane [49]. The accuracy of LDA was higher for the preprocessed spectra compared to that of the raw spectrum. These observations agree with the results of this study.

Conclusions
For the industrial use of sweet potatoes, a suitable nondestructive NIR method was developed. The internal quality, including moisture content, starch content, and α-amylase activity, was measured during the storage period after harvesting. It was confirmed that, except for α-amylase activity, the internal quality can be predicted using PLS. The selection of suitable sweet potatoes for paste preparation was accomplished using LDA after acquiring the NIR spectra. This study confirms that the NIR spectra of sweet potatoes can be obtained inline. The NIR apparatus is already manufactured to be inline ready, and the technology for prediction and classification via PLS or LDA models can be applied inline after further study in the near future. Table S1: Changes in internal quality in terms of moisture, starch, and α-amylase content of sweet potatoes during storage; Figure S1: (A) Regression plot and (B) regression coefficient plot of partial least squares (PLS) modeling about calibration and prediction set for raw spectrum and moisture content of sweet potatoes; Figure S2

Conclusions
For the industrial use of sweet potatoes, a suitable nondestructive NIR method was developed. The internal quality, including moisture content, starch content, and α-amylase activity, was measured during the storage period after harvesting. It was confirmed that, except for α-amylase activity, the internal quality can be predicted using PLS. The selection of suitable sweet potatoes for paste preparation was accomplished using LDA after acquiring the NIR spectra. This study confirms that the NIR spectra of sweet potatoes can be obtained inline. The NIR apparatus is already manufactured to be inline ready, and the technology for prediction and classification via PLS or LDA models can be applied inline after further study in the near future.

Data Availability Statement:
The data presented in this study are available on request from the corresponding author.