Research ArticleOpen Access

Quantification of potassium concentration with Vis–SWNIR spectroscopy in fresh lettuce

Yating Xiong

Graduate School of Science and Technology, Niigata University, 8050 Ikarashi 2-no-cho Nishi-ku, Niigata 950-2181, Japan

Search for more papers by this author

Shintaroh Ohashi

Faculty of Agriculture, Niigata University, 8050 Ikarashi 2-no-cho, Nishi-ku, Niigata 950-2181, Japan

E-mail Address: sohashi@agr.niigata-u.ac.jp

Corresponding author.

Search for more papers by this author

Kazuhiro Nakano

Graduate School of Science and Technology, Niigata University, 8050 Ikarashi 2-no-cho Nishi-ku, Niigata 950-2181, Japan

Search for more papers by this author

Weizhong Jiang

College of Water Resources & Civil Engineering, China Agricultural University, 17 Qinghua Donglu Beijing 100083, P. R. China

Search for more papers by this author

Kenichi Takizawa

Faculty of Tourism Management, Niigata University of Management, 2909-2 Kibougaoka, Kamo-shi Niigata 959-1321, Japan

Search for more papers by this author

Kazuyuki Iijima

Graduate School of Science and Technology, Niigata University, 8050 Ikarashi 2-no-cho Nishi-ku, Niigata 950-2181, Japan

Search for more papers by this author

, and

Phonkrit Maniwara

Postharvest Technology Research Center, Faculty of Agriculture, Chiang Mai University, 239 Huay Kaew Road, Muang District, Chiang Mai 50200, Thailand

Search for more papers by this author

https://doi.org/10.1142/S1793545820500297Cited by:6 (Source: Crossref)

Abstract

Chronic kidney disease (CKD) is becoming a major public health problem worldwide, and excessive potassium intake is a health threat to patients with CKD. In this study, visible–short-wave near-infrared (Vis–SWNIR) spectroscopy and chemometric algorithms were investigated as nondestructive methods for assessing the potassium concentration in fresh lettuce to benefit the CKD patients’ health. Interactance and transmittance measurements were performed and the competencies were compared based on the multivariate methods of partial least-square regression (PLS) and support vector machine regression (SVR). Meanwhile, several preprocessing methods [first- and second-order derivatives in combination with standard normal variate (SNV)] and wavelength selection method of competitive adaptive reweighted sampling (CARS) were applied to eliminate noise and highlight the spectral characteristics. The PLS models yielded better prediction than the SVR models with higher correlation coefficients ( $R^{2}$ $R^{2}$ ) and residual predictive deviation (RPD), and lower root-mean-square error of prediction (RMSEP). Excellent prediction of green leaves was obtained by the interactance measurement with $R^{2} = 0.93$ $R^{2} = 0.93$ , $RMSEP = 24.86$ $RMSEP = 24.86$ mg/100g, and $RPD = 3.69$ $RPD = 3.69$ ; while the transmittance spectra of petioles provided optimal prediction with $R^{2} = 0.92$ $R^{2} = 0.92$ , $RMSEP = 27.80$ $RMSEP = 27.80$ mg/100g, and RPD=3.34, respectively. Therefore, the results indicated that Vis–SWNIR spectroscopy is capable of intelligently detecting potassium concentration in fresh lettuce to benefit CKD patients around the world in maintaining and enhancing their health.

Keywords:

1. Introduction

Potassium is one of the three major nutrients required for normal plant growth,^1,2,3 and it is also an essential nutrient for human health. However, for those diagnosed with chronic kidney disease (CKD) or the hyperkalemia patients, large amounts of potassium intake can cause physical damage and even sudden death. The dietary potassium intake for patients suffering from CKD is very restrictive.⁴ In 2004, the NKF KDOQI (National Kidney Foundation’s Kidney Disease Outcomes Quality Initiative) published CKD-specific guidelines with a daily potassium intake recommendation of 0.4g/d for a population with CKD stages 1 and 2, and 2–4g/d for those with CKD stages 3 and 4. Owing to the recent rapid increase in kidney disease patients,⁵ helping patients with CKD manage their potassium intake from food reasonably and safely has become a critical issue globally.

Lettuce is a popular potassium-rich vegetable usually eaten raw in salad, but CKD patients must face a strong demand and an unmet need for fresh lettuce.⁶ At present, there are some studies on potassium deficiency cultivation to produce low-potassium lettuce, but it is reported that there are side effects including the reduction of yield, quality, and unacceptable taste.^7,8,9,10 Therefore, developing rapid and accurate detection and presentation of potassium concentration in lettuce is an effective method to manage potassium intake for the CKD patients worldwide.

Visible–short-wave near-infrared (Vis–SWNIR) spectroscopy involves energy transfer between light and matter. The spectral features of samples in the Vis–NIR spectral region are associated with the vibrational modes of the functional group in the target substance.¹¹ Vis–SWNIR spectroscopy has recently been applied for plant quality evaluation, especially for the trace elements determination, because of its prominent advantages of being nondestructive, having a rapid response, being simple to operate and requiring no complex pretreatments.¹² Based on former researches, the Vis–SWNIR spectroscopy has been proved to be useful for measuring the potassium concentration in plant’s leaf. Zhang and Li¹³ applied support vector machine regression (SVR) and reflect spectral data to estimate the phosphorus content present in cucumber leaves. Zhai¹⁴ estimated the contents of nitrogen (N), phosphorus (P), and potassium (K) of different plants based on SVR and partial least-square regression (PLS) method combined with the Vis–SWNIR reflectance measurement. Ozyigit and Bilgen¹⁵ determined the N, P, and K contents of rangeland plants using spectral reflectance value. Liu et al.¹⁶ studied the potassium content of tobacco leaves by establishing a PLS prediction model with a coefficient of determination ( $R^{2}$ $R^{2}$ ) of 0.909 and a root-mean-square error of prediction (RMSEP) of 0.119 in the range of near-infrared spectrum. Menesatti et al.¹⁷ explored the trace elements in orange tree leaves to detect the nutritional status of plants by visible–near-infrared spectroscopy. Rustioni et al.¹⁸ discriminated the mineral deficiency symptoms of grape leaves by Vis–SWNIR reflectance spectroscopy, and Mutanga et al.¹⁹ reported that phosphorus and potassium elements affected absorption in the visible region of the spectrum.

These previous studies had concentrated mostly on the regulation and control of nutrition elements that satisfied plant growth, and the spectral data were generally collected only by the reflection methods. However, there were few studies that focused on potassium concentration evaluation in post-harvest fresh lettuce by using Vis–SWNIR, but the feasibility of detecting the concentration of potassium in the cells by the transmittance and interactance methods was not discussed.

Successfully predicting the concentration of potassium in fresh lettuce will benefit CKD patients around the world for maintaining and enhancing their health. Therefore, this study aims at investigating the possibility of potassium concentration determination in fresh lettuce by Vis–SWNIR spectroscopy with the interactance and transmittance measurements and comparing the prediction models of two chemometric algorithms (PLS and SVR) to improve the prediction accuracy. In addition, this study also focuses on the best combination of valuable information wavelengths and data processing algorithms for different detection sites (green leaves and petioles) of fresh lettuce.

2. Materials and Methods

2.1. Lettuce material and chemical method

A total of 183 fresh lettuce leaves (20 bunches) were used for the experiments that were purchased from different local markets on different days (approximately five samples each time). In order to obtain enough variability in the range of the samples, two varieties of lettuce were collected. One was named the red-tip leaf lettuce (Lactuca sativa var. crispa), which is popular with consumers and often used for salads. The other was low-potassium lettuce (Frillice), a low-potassium variety especially cultivated for chronic kidney patients.

Because the distribution of potassium is unequal in the leaves, and the structures of the upper and lower layers of a single leaf are different, it is necessary to analyze the presence of potassium in the green leaf and petiole, respectively. In addition, based on the growth condition of lettuce, the potassium concentrations in the outer leaf and the inner layer are also different. Therefore, in this study, the leaves are numbered sequentially from the outer layers to the inner layers, and then each leaf is divided into green leaf and petiole along the boundary to analyze the potassium concentration, respectively. The samples are stored at 25^∘C for 3h before measurements to reach the equilibrium temperature with laboratory environment, and the measurements, including spectral collection and quality analysis of same samples, are carried out on the same day to ensure that the lettuces are fresh and potassium statuses in leaves remain stable. The reference potassium concentration data of fresh lettuce are measured by the use of chemical destruction method with a handheld fast detector (B-731; Horiba, Ltd., Kyoto, Japan). The Horiba B-731 is a handheld, battery-operated analyzer that uses ion-selective electrode technology to derive the effective K $^{+}$ $^{+}$ ions concentration; the measurement range is 39–3900ppm, with an accuracy of $\pm 10$ $\pm 10$ %. Each device contains a well with two sensors, the ion-selective electrode and reference electrode. The lettuce juice from fully ground lettuce leaves is dropped into the sample cell and then the two electrodes are covered for analysis.²⁰ In this study, the potassium concentration value is expressed as the mass ratio unit (mg/100g). And the chemical measurements are performed immediately after the Vis–SWNIR measurements.

2.2. Spectra collection

According to the different measurement setups for obtaining Vis–SWNIR spectra, there are transmittance, reflectance, and interactance modes. Reflectance mode measures the light that is reflected or scattered from the surface of samples. In contrast, the transmittance-measurement mode can detect the internal properties of the sample with the light that goes through the samples,²¹ and interactance-measurement mode is a method that can be used to obtain information on both the inside and surface of the samples.^22,23,24 In this study, the feasibility of collecting spectral data by the use of transmittance and interactance methods is discussed.

Figure 1 shows the schematic diagram of Vis–SWNIR spectroscopy system with the transmittance-measurement [Fig. 1(a)] and interactance-measurement [Fig. 1(b)] modes. In order to perform the interactance measurement, the light source (MHAA-100W; Moritex Co., Ltd., Japan) is provided by a portable optical fiber, and the detector is bundled in the middle of the circular optical fiber, which can obtain the spectra that contain the surface and internal diffused reflection information. During spectra collection, the lettuce leaves are placed directly horizontally on the fiber optic probe. In contrast, in the transmittance-measurement mode, the Vis–SWNIR light region (HR-k2150N; Hiroshi Industry Co., Ltd., Japan) is from two 12-V/100-W tungsten halogen lamps (MCR 12-150M). The light source and fiber optic detector probe are placed against each other and the samples are put in a holder between them, and the fiber optic probe is placed close to the sample surface to reduce light leakage. All interactance and transmittance light intensity spectra are obtained using a spectrophotometer (Handy Lambda II; Spectra Co-op Co., Ltd., Japan) within the range of 500–1000nm in wavelength increments (interval wavelengths) of 3.3nm. A white ceramic plate with a thickness of 1mm is used as a standard reference to get rid of the characteristics of the light source. Spectral measurements are performed and averaged from three random locations of green leaves and petioles for both the interactance and transmittance measurements by using Wave Viewer software (Spectra Co-op Co., Ltd., Japan).

2.3. Multivariate data statistics

2.3.1. Spectral data preprocessing

In this study, spectral data are obtained by averaging 10 spectra from each detected location and explained in terms of the logarithm of reciprocal absorbance [ $log (1 ∕ T)$ $log (1 ∕ T)$ ]. After the averaging, a data matrix composed in the form of an array of 183 rows (i.e., samples) $\times$ $\times$ 152 variables (spectral absorbance values) is divided into calibration and prediction subsets by way of the Kennard–Stone duplex algorithm²⁵ with a ratio of 2:1. Thus, 122 lettuce samples are selected as the calibration set to build multi-cultivar models, and the remaining 61 samples are used as the prediction set to evaluate the prediction model effect. The potassium concentration values of the calibration set exhibited a wider range than those of the prediction set for each cultivar. These features are beneficial for developing a stable and robust model.

Meanwhile, the first-order derivative $[D_{1} log (1 ∕ T)]$ $[D_{1} log (1 ∕ T)]$ , second-order derivative $[D_{2} log (1 ∕ T)]$ $[D_{2} log (1 ∕ T)]$ , and standard normal variate (SNV) are combined as spectral data preprocessing methods to improve the signal-to-noise ratio of the spectra, aiming to reduce the influence of uncertain noises, which might be produced from the instrument, spectral transformation, manual operations, etc. Derivatives are easy to be performed and often used to reduce the scatter effects for continuous spectra.²⁶ The SNV method is based on the assumption that each data point of every spectrum conforms to the normal distribution. The calculation method involves subtracting the average value of all data points in the spectrum from each data point, and then dividing by the standard deviation.²⁷

Since the spectra with hundreds of variables contain lots of information besides potassium concentration, the competitive adaptive reweighted sampling (CARS) is used to select key variables to improve the model performance by removing uninformed variables.²⁸ In CARS, the absolute values of regression coefficients of the PLS model are used as a criterion for evaluating the importance of each variable. Monte Carlo (MC) sampling is used to select the modeling samples, and in each sampling run, exponentially decreasing function (EDF) and adaptive reweighted sampling (ARS) are utilized to select the key wavelengths. Finally, variables subset corresponding to the smallest root-mean-square error of cross-validation (RMSECV) is selected.^29,30

The data processing methods, PLS models, and SVR algorithm are implemented using the Unscramble X10.3 software (CAMO Software, Inc., Oslo, Norway). The CARS is implemented by the use of the MATLAB software (The MathWorks, Natick, MA, USA) and the algorithm codes for MATLAB are available for free at https://code.google.com/p/carspls/.

2.3.2. Regression methods

PLS is a common linear relation method used to make regression models because of its simplicity, rapid speed, good performance, and easy accessibility.^31,32 SVR is a promising method proposed by Vapnik in 1998.³³ SVR is processed based on the statistical learning theory to avoid overfitting and multidimensional problem,³⁴ and is commonly used as a good nonlinear regression method for quantitative analysis.²⁴ In this study, SVR is applied to build a nonlinear models for comparison of prediction performance with the linear PLS models.

The performances of two algorithms combined with the mentioned preprocessing techniques are compared by the use of statistical parameters, such as the coefficient of determination ( $R^{2}$ $R^{2}$ ) between the predicted values and actual values of each observation, the root-mean-square errors of calibration and prediction (RMSEC and RMSEP), and the ratios of performance to deviation, known as the residual predictive deviation (RPD) values.³⁵ The calculations of RMSEP (the definition is analogous for RMSEC), bias (predictive fault), and RPD are defined in the following equations. In PLS analysis, the number of variables is optimized into factors; including too many factors in the PLS model may lead to overfitting, whereas too few factors may result in underfitting.³⁶ The optimal number of factors corresponds to a compromise allowing a model to present both the relative lowest RMSEP value and highest $R^{2}$ $R^{2}$ value.

Generally, the closer the $R^{2}$ $R^{2}$ value is to 1, the better is the model correlation; the smaller the RMSEP and the larger the RPD, the higher the model prediction accuracy and the stronger the resolving ability. When $RPD > 3$ $RPD > 3$ , the model accuracy is supposed to be excellent^37,38,39 :

RMSEP=√∑ni=1(yi−ŷi)2n,<math display="block" altimg="eq-00033.gif"><mstyle><mtext mathvariant="normal">RMSEP</mtext></mstyle><mo>=</mo><msqrt><mrow><mfrac><mrow><msubsup><mrow><mo>∑</mo></mrow><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mrow><mi>n</mi></mrow></msubsup><msup><mrow><mo stretchy="false">(</mo><msub><mrow><mi>y</mi></mrow><mrow><mi>i</mi></mrow></msub><mo>−</mo><msub><mrow><mi>ŷ</mi></mrow><mrow><mi>i</mi></mrow></msub><mo stretchy="false">)</mo></mrow><mrow><mn>2</mn></mrow></msup></mrow><mrow><mi>n</mi></mrow></mfrac></mrow></msqrt><mo>,</mo></math>(1)

RPD=SDRMSEP,<math display="block" altimg="eq-00034.gif"><mstyle><mtext mathvariant="normal">RPD</mtext></mstyle><mo>=</mo><mfrac><mrow><mstyle><mtext mathvariant="normal">SD</mtext></mstyle></mrow><mrow><mstyle><mtext mathvariant="normal">RMSEP</mtext></mstyle></mrow></mfrac><mo>,</mo></math>(2)

Bias=∑ni=1(yi−ŷi)n,<math display="block" altimg="eq-00035.gif"><mstyle><mtext mathvariant="normal">Bias</mtext></mstyle><mo>=</mo><mfrac><mrow><msubsup><mrow><mo>∑</mo></mrow><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mrow><mi>n</mi></mrow></msubsup><mo stretchy="false">(</mo><msub><mrow><mi>y</mi></mrow><mrow><mi>i</mi></mrow></msub><mo>−</mo><msub><mrow><mi>ŷ</mi></mrow><mrow><mi>i</mi></mrow></msub><mo stretchy="false">)</mo></mrow><mrow><mi>n</mi></mrow></mfrac><mo>,</mo></math>(3)

where

$n$ is the number of samples in the calibration set or prediction set,

$ŷ_{i}$ is the experimentally measured reference result of the sample

$i$ , and

$y_{i}$ is the estimated result of the model for the corresponding prediction sample

$i$ . SD is the standard deviation of chemical measured values in the prediction set.

3. Results and Discussion

3.1. Potassium concentration in fresh lettuce

This study collected the Vis–SWNIR spectra of red-tip leaf lettuce and low-potassium lettuce with different potassium concentrations. To obtain robust prediction models, the variations in potassium concentration at different locations of the lettuce are considered as important factors; the prediction models will be analyzed based on green leaves and petioles separately.

Table 1 shows the results of potassium concentration in 183 pieces of fresh lettuce from the red-tip leaf lettuce and low-potassium lettuce. Concerning the red-tip leaf lettuce, the range of potassium concentration is 130–379mg/100g, much higher than that in the low-potassium lettuce of 38–180mg/100g. The mean values of potassium concentration in green leaves and petioles of the red-tip leaf lettuce are 242.32mg/100g and 271.84mg/100g, respectively. Meanwhile, in low-potassium lettuce, the mean values of potassium concentration are 67.16mg/100g and 88.15mg/100g, respectively. The results revealed significant differences in potassium concentration in different parts of the leaves, and it is obvious that the potassium concentration in petiole part is higher than that in the green leaf in both the types of lettuce. The coefficient of variation (CV) value is computed for each part of the two varieties of fresh lettuce. As shown in Table 1, petioles of low-potassium lettuce had the highest CV, while those of the red-tip leaf lettuce displayed the lowest among the sample sets tested. The CV value of each green leaf and petiole in low-potassium lettuce is higher than that of the red-tip leaf lettuce. Figure 2 shows the trend of potassium concentrations in each lettuce leaf, which exhibits the difference in potassium concentrations of petioles and green leaves visually. Furthermore, after grouping as shown in Table 2, in calibration sets, the potassium concentration ranges of green leaves and petioles are 38–352mg/100g and 45–379mg/100g, respectively, the highest and lowest values are assigned to the calibration set and the mean values of calibration sets and prediction sets are similar. It indicated that the distribution of samples is representative in the calibration sets to ensure the prediction model’s robustness, reliability, and reproducibility.^40,41

**Table 1. Potassium concentrations in fresh lettuce.**
	Red-tip leaf lettuce		Low-potassium lettuce (Frillice)
Potassium concentration	Green leaves	Petioles	Green leaves	Petioles
Range (mg/100 g)	130–352	180–379	38–110	45–180
Mean value (mg/100 g)	242.32	271.84	67.16	88.15
Std. deviation	50.18	47.36	15.07	27.36
Coefficient of variation (%)	20.71	17.42	22.44	31.04

**Table 2. Potassium concentration values in calibration sets and prediction sets.**
	Green leaves		Petioles
Items	Calibration set	Prediction set	Calibration set	Prediction set
No. of samples	122	61	122	61
Range (mg/100 g)	38–352	42–336	45–379	49–359
Mean value (mg/100 g)	179.66	177.31	210.02	195.98
Std. deviation	94.89	92.47	97.71	97.62

Fig. 2. Potassium concentrations by the use of chemical method for green leaves and petioles.

3.2. Spectral data processing

Concerning interactance measurement, light entered the lettuce leaves, then internal unabsorbed light and surface diffusely reflected light got reflected into the detector. Meanwhile, in the transmittance measurement, when the light passed through the samples, the unabsorbed light got transmitted to the detector.

Figure 3(a) shows the average absorbance spectra of green leaves and petioles based on the interactance and transmittance measurements; it can be found that the shapes of the absorption spectra are basically the same in the two methods. Because of the different light source intensity in each experiment, the light absorption value is shifted to some extent. It is also obvious that average spectra showed similar strong absorption peaks at around 670 nm and 970 nm. According to previous studies, these ranges are related to the absorption of chlorophyll and water, respectively, and the change of potassium concentration leads to the chlorophyll concentrations in plants.^42,43 Some works also have shown that potassium deficiency has a significant effect on spectral absorption in the red and near-infrared regions.^44,45 Figure 3(b) shows the mean spectrum processed by the first- and second-order derivatives combined with SNV. After preprocessing, the difference between the spectra is more significant and the spectral information of peaks have been enhanced. From Tables 3 and 4, it can be found that compared with the models of raw spectral data, better results of both PLS and SVM models are found when the pretreated spectra method is applied. However, because of the complicated nature of Vis–NIR spectra, it is necessary to apply an effective wavelength selection algorithm to extract the characteristic variables related to potassium concentration to reduce noise and simplify the calibration model.

Fig. 3. (a) Average original absorbance and (b) processed spectra for green leaves and petioles.

3.3. Wavelength selection

In this study, applying CARS to select variables is a critical step for data analysis. Although potassium cannot directly affect the absorption of the Vis–NIR spectrum, it can cause indirect effects by affecting the compounds in the leaves.⁴⁶

Based on interactance measurement, the number of selected wavelengths showed a rapid decrease from 152 to 24 and 26 in green leaves and petioles, respectively. Meanwhile, the numbers of selected wavelengths are 21 and 29 by transmittance measurement, respectively. The complexity of the model is simplified and numerous variables are discarded. According to the PLS modeling results (Table 3), 24 and 29 wavelength prediction models obtained better performance on green leaves and petioles, respectively. Figure 4 presents the sensitive wavelengths for interactance and transmittance measurements from each observation, that are 552, 555, 582, 589, 596, 609, 629, 652, 686, 693, 699, 716, 729, 749, 773, 806, 809, 816, 839, 852, 888, 895, 964, and 973 nm of green leaves; and 545, 549, 566, 602, 609, 639, 646, 663, 689, 703, 709, 713, 716, 723, 726, 736, 746, 759, 766, 776, 792, 802, 822, 865, 875, 905, 921, 954, and 990 nm of petioles. Since potassium is soluble in water and exists in plants in the form of ions, the absorption of water plays an important role in the measurement of potassium concentration; the wavelength range of 730–900nm is theoretically related to the third overtone of C–H and O–H bonds vibration.⁴⁷ What is more, 650–700-nm range is related to strong absorption of light by the pigments (chlorophyll, carotenoids, etc.).^48,49 By observing the results of PLS models (Table 3), CARS is found to be a good candidate. The selected wavelengths by CARS clearly revealed that sensitive variables of Vis–SWNIR regions significantly influenced the regression modeling, which effectively provided interesting information on organic compounds related to potassium and water. The information contributing toward the predictions are retained by CARS, which confirmed that CARS is a suitable method to discard irrelevant variables for the determination of potassium concentration in fresh lettuce.

**Table 3. Calibration and prediction results of PLS method with the interactance and transmittance measurements.**
				Calibration		Prediction
Sample	Mode	Model	Number of wavelengths	$R^{2}$	RMSEC (mg/100g)	$R^{2}$	RMSEP (mg/100g)	Factors	RPD
Green leaves	Interactance mode	Raw PLS	152	0.86	33.43	0.84	36.78	9	2.49
		$D_{1} log (1 ∕ T)$ –SNV PLS		0.91	26.82	0.88	30.08	10	3.04
		$D_{2} log (1 ∕ T)$ –SNV PLS		0.91	27.27	0.84	35.95	8	2.55
		$D_{1} log (1 ∕ T)$ –SNV–CARS PLS	24	0.94	23.82	0.93	24.86	11	3.69
		$D_{2} log (1 ∕ T)$ –SNV–CARS PLS		0.92	26.39	0.90	28.74	12	3.12
	Transmittance mode	Raw PLS	152	0.84	38.31	0.77	43.93	10	2.09
		$D_{1} log (1 ∕ T)$ –SNV PLS		0.90	28.64	0.87	33.31	9	2.75
		$D_{2} log (1 ∕ T)$ –SNV PLS		0.91	28.28	0.82	40.61	10	2.26
		$D_{1} log (1 ∕ T)$ –SNV–CARS PLS	21	0.92	26.26	0.89	30.86	11	2.92
		$D_{2} log (1 ∕ T)$ –SNV–CARS PLS		0.94	23.92	0.88	32.86	9	2.74
Petioles	Interactance mode	Raw PLS	152	0.86	36.34	0.82	40.68	10	2.28
		$D_{1} log (1 ∕ T)$ –SNV PLS		0.88	33.62	0.86	35.65	7	2.60
		$D_{2} log (1 ∕ T)$ –SNV PLS		0.90	30.94	0.84	36.55	8	2.53
		$D_{1} log (1 ∕ T)$ –SNV–CARS PLS	26	0.91	28.69	0.90	31.17	11	3.02
		$D_{2} log (1 ∕ T)$ –SNV–CARS PLS		0.93	26.90	0.89	30.93	12	3.05
	Transmittance mode	Raw PLS	152	0.84	38.88	0.77	45.29	11	2.05
		$D_{1} log (1 ∕ T)$ –SNV PLS		0.89	32.11	0.82	40.95	10	2.27
		$D_{2} log (1 ∕ T)$ –SNV PLS		0.92	27.72	0.81	40.09	9	2.32
		$D_{1} log (1 ∕ T)$ –SNV–CARS PLS	29	0.94	23.82	0.91	29.27	9	3.19
		$D_{2} log (1 ∕ T)$ –SNV–CARS PLS		0.95	22.22	0.92	27.80	9	3.34

Fig. 4. Results of wavelength selection by the use of CARS.

3.4. Comparation of PLS and SVR predictions

Different PLS models are built with the previously mentioned preprocessing and evaluated by the coefficient of determination ( $R^{2}$ ), RMSEP, and RPD. As shown in Table 3, for green leaves, the interactance measurement produced a better performance than transmittance measurement because of the higher values of RPD and $R^{2}$ and a lower RMSEP. Compared to others, the first-order derivative processing method combined with SNV and CARS got the highest RPD and $R^{2}$ and the lowest RMSEP, which proved to be the best preprocessing method. As for the petioles, the best choice is second-order derivative preprocessing based on the transmittance measurement. What is more, the optimal prediction results of PLS models are obtained from the spectral data of green leaves, which are $R^{2} = 0.93$ , $RMSEP = 24.86$ , and $RPD = 3.69$ .

In comparison, the evaluations of SVR models are shown in Table 4. By judging the prediction performance based on the results, it is indicated that the interactance measurement is still the best method for green leaves, same as in PLS prediction, and transmittance measurement is still suitable for petioles. The preprocessing spectra of the second-order derivative combined with SNV and CARS got the best performance for both green leaves and petioles, instead of the first-order derivative in PLS models. By comparison of the evaluated values, it is found that although the RPD values of prediction results of the optimal SVR models in green leaves and petioles are both over 2.5, the overall performance in predicting the potassium concentration is worse than that of PLS models, where the RPD values are over 3 in both green leaves and petioles. Meanwhile, the $R^{2}$ and RMSEP also showed worst performance in SVR prediction than that in PLS models, respectively. The reason is supposed to be the nonlinear principle of SVR, which increases the complexity of the prediction model, leading to that the nonlinear SVR algorithm is unsuitable for potassium concentration prediction in fresh lettuce; it shows that there is a more linear relationship between potassium concentration and visible–near-infrared spectrum information. Additionally, it is also proposed to be affected by the complex composition characteristics of fresh lettuce leaves. Similar results reported that the modeling of PLS based on Vis–NIR spectroscopy achieved promising prediction performances in numerous quality parameters, such as moisture content of cassava rhizome ground,⁵⁰ fructose and glucose contents of honey,⁵¹ inorganic phosphorus in soybeans,⁵² macronutrients, and the energy content of snack products.²⁶ In this study, separate investigation of the potassium status in green leaf and petiole prediction models is innovative, confirming the potential of the combination of PLS algorithm, suitable processing, and CARS wavelength selection method in the determination of potassium concentration.

**Table 4. Calibration and prediction results of SVR method with the interactance and transmittance measurements.**
						Calibration		Prediction
Sample	Mode	Model	Number of wavelengths	$C$	Gamma	$R^{2}$	RMSEC (mg/100g)	$R^{2}$	RMSEP (mg/100g)	RPD
Green leaves	Interactance mode	Raw SVR	152	0.1	0.01	0.78	49.89	0.72	50.34	1.62
		$D_{1} log (1 ∕ T)$ –SNV SVR		10	0.1	0.83	38.15	0.81	41.53	1.97
		$D_{2} log (1 ∕ T)$ –SNV SVR		100	1	0.88	31.89	0.85	35.51	2.30
		$D_{1} log (1 ∕ T)$ –SNV–CARS SVR	24	1	10	0.85	36.54	0.84	38.58	2.12
		$D_{2} log (1 ∕ T)$ –SNV–CARS SVR		1	1	0.91	29.04	0.89	32.01	2.58
	Transmittance mode	Raw SVR	152	10	0.1	0.78	52.68	0.74	55.91	1.46
		$D_{1} log (1 ∕ T)$ –SNV SVR		100	0.01	0.82	40.45	0.78	44.33	1.85
		$D_{2} log (1 ∕ T)$ –SNV SVR		0.1	1	0.88	33.51	0.84	38.07	2.15
		$D_{1} log (1 ∕ T)$ –SNV–CARS SVR	21	10	10	0.80	42.04	0.77	45.16	1.76
		$D_{2} log (1 ∕ T)$ –SNV–CARS SVR		1	1	0.90	31.54	0.87	35.70	2.27
Petioles	Interactance mode	Raw SVR	152	0.1	0.1	0.79	44.62	0.74	50.39	1.62
		$D_{1} log (1 ∕ T)$ –SNV SVR		1	1	0.83	39.89	0.78	44.67	1.83
		$D_{2} log (1 ∕ T)$ –SNV SVR		1	10	0.90	29.96	0.86	36.59	2.24
		$D_{1} log (1 ∕ T)$ –SNV–CARS SVR	26	10	1	0.85	37.70	0.81	43.11	1.95
		$D_{2} log (1 ∕ T)$ –SNV–CARS SVR		10	1	0.92	28.90	0.88	34.30	2.55
	Transmittance mode	Raw SVR	152	1	0.1	0.76	50.62	0.71	56.16	1.46
		$D_{1} log (1 ∕ T)$ –SNV SVR		0.1	0.01	0.82	41.75	0.77	47.29	1.73
		$D_{2} log (1 ∕ T)$ –SNV SVR		1	0.1	0.86	34.17	0.83	40.82	2
		$D_{1} log (1 ∕ T)$ –SNV–CARS SVR	29	10	10	0.82	42.67	0.78	46.69	1.71
		$D_{2} log (1 ∕ T)$ –SNV–CARS SVR		100	10	0.93	26.81	0.90	31.92	2.69

In summary, compared with SVR models, the PLS models performed the best with the lowest RMSEP and highest $R^{2}$ and RPD of green leaves and petioles. Based on 24 spectral wavelengths used for 11 PLS factors, the $R^{2}$ of 0.93, RMSEP of 24.86 mg/100 g, and RPD of 3.69 are achieved for green leaves by the use of $D_{1}$ log(1/ $T)$ –SNV–CARS processing. For petioles, the optimized PLS prediction model obtained 29 wavelengths, nine PLS factors, and the $R^{2}$ of 0.92, RMSEP of 27.80 mg/100 g, and RPD of 3.34 by the use of $D_{2}$ log(1/ $T)$ –SNV–CARS method. The RPD values of green leaves and petioles are higher than 3, which showed an excellent prediction accuracy. Moreover, optimal sample presentation methods are stable to predict potassium concentration in fresh lettuce. Due to the different thicknesses and physical shapes of green leaves and petioles, interactance measurement is suitable for green leaves, while transmittance measurement is the best choice for petioles.

Figure 5 shows the fitting between predicted values and reference values by the PLS of green leaves and petioles, and there is no systematic error present in the predictions; the points are randomly distributed along the diagonal line. The above results show that the PLS model has an excellent prediction effect that can be used to measure the potassium concentration in fresh lettuce.

Fig. 5. Correlation diagrams between predicted values and reference values of the PLS models.

4. Conclusions

In this study, the potentials of interactance and transmittance measurements were compared based on PLS and SVR models in evaluating the potassium concentration in fresh lettuce. And the different data preprocessing methods were compared. The results confirmed that Vis–SWNIR spectroscopy can be useful for developing highly accurate prediction models of potassium concentration in fresh lettuce. The PLS models with $D_{1} log (1 ∕ T)$ were found to provide the best prediction accuracy for green leaves ( $R^{2} = 0.93$ , $RMSEP = 24.86$ mg/100g, and $RPD = 3.69$ ) by the use of interactance measurement, and the prediction model with $D_{2} log (1 ∕ T)$ got the best performance for petioles ( $R^{2} = 0.92$ , $RMSEP = 27.80$ mg/100g, and $RPD = 3.34$ ) with transmittance measurement. The results suggest the following: First, potassium is not uniformly present in the lettuce leaves and a higher potassium concentration exists in petioles than green leaves. Second, interactance measurement is a better method for green leaves and transmittance measurement is more useful for petioles. Third, spectral preprocessing and wavelength selection methods can be performed well to mathematically treat the extracted spectra and reduce the input variables. Finally, the linear PLS models fit better with the prediction of potassium concentration in fresh lettuce and got a high prediction accuracy (RPD values of over 3), compared with the nonlinear SVR models.

Vis–SWNIR spectroscopy coupled with the linear or nonlinear model is an intelligent method to evaluate the potassium concentration in fresh lettuce nondestructively; it is of great significance for maintaining and enhancing the health in CKD patients. Furthermore, in this study, the best sample spectrum collection modes for green leaves and petioles are different, which increased the difficulty in practical applications of Vis–SWNIR spectroscopy technology. In the future, it is necessary to gradually optimize the model and achieve consistency measurements.

Conflict of Interest

The authors have no conflicts of interest relevant to this article.

Acknowledgments

This work was supported by the Japan Society for the Promotion of Science (JSPS) (Grant No. 19K06312).

References

1. H. Marschner, Function of Mineral Nutrients: Micronutrients, 2nd Edition, Academic Press, London (1995). Google Scholar
2. A. Lebaudy, A. Vavasseur, E. Hosy, I. Dreyer, N. Leonhardt, J. B. Thibaud, H. Sentenac, “Plant adaptation to fluctuating environment and biomass production are strongly dependent on guard cell potassium channels,” Proc. Natl. Acad. Sci. USA 105, 5271–5276 (2008). Crossref, Web of Science, Google Scholar
3. K. Nabiollahy, F. Khormali, K. Bazargan, S. H. Ayoubi, “Forms of K as a function of clay mineralogy and soil development,” Clay Miner. 41(3), 739–749 (2006). Crossref, Web of Science, Google Scholar
4. S. J. S. Bajwa, I. S. Kwatra, “Nutritional needs and dietary modifications in patients on dialysis and chronic kidney disease,” J. Med. Nutr. Nutraceut. 2, 46–51 (2013). Crossref, Google Scholar
5. Q.-L. Zhang, D. Rothenbacher, “Prevalence of chronic kidney disease in population-based studies: systematic review,” BMC Public Health 8, 117 (2008). Crossref, Web of Science, Google Scholar
6. T. Yoshida, K. Sakuma, H. Kumagai, “Nutritional and taste characteristics of low-potassium lettuce developed for patients with chronic kidney diseases,” Hong Kong J. Nephrol. 16, 42–45 (2014). Crossref, Google Scholar
7. C. W. Bednarz, D. M. Oosterhuis, “Physiological changes associated with potassium deficiency in cotton,” J. Plant Nutr. 22, 303–313 (1999). Crossref, Web of Science, Google Scholar
8. I. Cakmak, C. Hengeler, H. Marschner, “Changes in phloem export of sucrose in leaves in response to phosphorus, potassium and magnesium deficiency in bean plants,” J. Exp. Bot. 45, 1251–1257 (1994). Crossref, Web of Science, Google Scholar
9. S. H. Jin, J. Q. Huang, X. Q. Li, B. S. Zheng, J. S. Wu, Z. J. Wang, M. Chen, “Effects of potassium supply on limitations of photosynthesis by mesophyll diffusion conductance in Carya cathayensis,” Tree Physiol. 31, 1142–1151 (2011). Crossref, Web of Science, Google Scholar
10. D. Schwarz, G. B. Öztekin, Y. Tüzel, B. Brückner, A. Krumbein, “Rootstocks can enhance tomato growth and quality characteristics at low potassium supply,” Sci. Hortic. 149, 70–79 (2013). Crossref, Web of Science, Google Scholar
11. N. Asgari, S. Ayoubi, J. A. M. Dematte, A. C. Dotto, “Carbonates and organic matter in soils characterized by reflected energy from 350–25000 nm wavelength,” J. Mt. Sci. 17, 1636–1651 (2020). Crossref, Web of Science, Google Scholar
12. R. Ciccoritti, M. Paliotta, T. Amoriello, K. Carbone, “FT-NIR spectroscopy and multivariate classification strategies for the postharvest quality of green-fleshed kiwifruit varieties,” Sci. Hortic. 257, 108622 (2019). Crossref, Web of Science, Google Scholar
13. X. J. Zhang, M. Z. Li, “Analysis and estimation of the phosphorus content in cucumber leaf in greenhouse by spectroscopy,” Guang Pu Xue Yu Guang Pu Fen Xi (Spectrosc Spectr. Anal.) 28, 2404–2408 (2008). Google Scholar
14. Y. F. Zhai, L. J. Cui, X. Zhou, Y. Gao, T. Fei, W. X. Gao, “Estimation of nitrogen, phosphorus, and potassium contents in the leaves of different plants using laboratory based visible and near-infrared reflectance spectroscopy: Comparison of partial least-square regression and support vector machine regression methods,” Int. J. Remote Sens. 34, 2502–2518 (2013). Crossref, Web of Science, Google Scholar
15. Y. Ozyigit, M. Bilgen, “Use of spectral reflectance values for determining nitrogen, phosphorus, and potassium contents of rangeland plants,” J. Agric. Sci. Technol. 15, 1537–1545 (2013). Web of Science, Google Scholar
16. D. S. Liu, L. S. Jin, Z. H. Yang, X. R. Han, G. L. Lin, “Rapid determination of potassium content in flue-cured tobacco leaves by near infrared reflectance spectroscopy,” Chin. J. Soil Sci. 41, 417–419 (2010). Google Scholar
17. P. Menesatti, F. Antonucci, F. Pallottino, G. Roccuzzo, M. Allegra, F. Stagno, F. Intrigliolo, “Estimation of plant nutritional status by Vis–NIR spectrophotometric analysis on orange leaves [Citrus sinensis (L) Osbeck cv Tarocco],” Biosyst. Eng. 105, 448–454 (2010). Crossref, Web of Science, Google Scholar
18. L. Rustioni, D. Grossi, L. Brancadoro, O. Failla, “Iron, magnesium, nitrogen and potassium deficiency symptom discrimination by reflectance spectroscopy in grapevine leaves,” Sci. Hortic. 241, 152–159 (2018). Crossref, Web of Science, Google Scholar
19. O. Mutanga, A. K. Skidmore, H. H. T. Prins, “Predicting in situ pasture quality in the Kruger National Park, South Africa, using continuum-removed absorption features,” Remote Sens. Environ. 89, 393–408 (2004). Crossref, Web of Science, Google Scholar
20. L. B. Baker, C. T. Ungaro, K. A. Barnes, R. P. Nuccio, A. J. Reimel, J. R. Stofan, “Validity and reliability of a field technique for sweat Na $^{+}$ and K $^{+}$ analysis during exercise in a hot-humid environment,” Physiol. Rep. 2(5), e12007 (2014). Crossref, Google Scholar
21. X. Fu, Y. Ying, H. Lu, H. Xu, “Comparison of diffuse reflectance and transmission mode of visible-near infrared spectroscopy for detecting brown heart of pear,” J. Food Eng. 83, 317–323 (2007). Crossref, Web of Science, Google Scholar
22. I. Kavdir, R. Lu, D. Ariana, M. Ngouajio, “Visible and near-infrared spectroscopy for nondestructive quality assessment of pickling cucumbers,” Postharvest Biol. Technol. 44, 165–174 (2007). Crossref, Web of Science, Google Scholar
23. J. Wang, K. Nakano, S. Ohashi, K. Takizawa, J. G. He, “Comparison of different modes of visible and near-infrared spectroscopy for detecting internal insect infestation in jujubes,” J. Food Eng. 101, 78–84 (2010). Crossref, Web of Science, Google Scholar
24. P. P. Subedi, K. B. Walsh, “Assessment of avocado fruit dry matter content using portable near infrared spectroscopy: Method and instrumentation optimization,” Postharvest Biol. Technol. 161, 111078 (2020). Crossref, Web of Science, Google Scholar
25. R. W. Kennard, L. A. Stone, “Computer aided design of experiments,” Technometrics 11, 137–148 (1969). Crossref, Web of Science, Google Scholar
26. E. Benes, A. Gere, M. Fodor, “Predicting macronutrients and energy content of snack products using FT-NIR analysis and chemometric techniques,” J. Food Eng. 280, 109954 (2020). Crossref, Web of Science, Google Scholar
27. X. L. Chu, H. F. Yuan, W. Z. Lu, “Progress and application of spectral data pretreatment and wavelength selection methods in NIR analytical technique,” Prog. Chem. - Beijing 16, 528–542 (2004). Web of Science, Google Scholar
28. A. Wang, L. Xie, “Technology using near infrared spectroscopic and multivariate analysis to determine the soluble solids content of citrus fruit,” J. Food Eng. 143, 17–24 (2014). Crossref, Web of Science, Google Scholar
29. H. D. Li, Y. Z. Liang, Q. S. Xu, D. S. Cao, “Key wavelengths screening using competitive adaptive reweighted sampling method for multivariate calibration,” Anal. Chim. Acta 648, 77–84 (2009). Crossref, Web of Science, Google Scholar
30. X. P. Ma, J. F. Pang, R. Dong, T. Chen, Y. X. Shu, Y. K. Li, “Rapid prediction of multiple wine quality parameters using infrared spectroscopy coupling with chemometric methods,” J. Food Compos. Anal. 91, 103509 (2020). Crossref, Web of Science, Google Scholar
31. H. Witjes, M. van den Brink, W. J. Melssen, L. M. C. Buydens, “Automatic correction of peak shifts in Raman spectra before PLS regression,” Chemom. Intell. Lab. Syst. 52, 105–116 (2000). Crossref, Web of Science, Google Scholar
32. M. van den Brink, J.-F. Hansen, P. de Peinder, A. M. van Herk, A. L. German, “Measurement of partial conversions during the solution copolymerisation of styrene and butyl acrylate using on-line Raman spectroscopy,” J. Appl. Polym. Sci. 79, 426–436 (2001). Crossref, Web of Science, Google Scholar
33. V. N. Vapnik, Statistical Learning Theory, John Wiley and Sons, Inc., New York (1998). Google Scholar
34. A. Besalatpour, M. A. Hajabbasi, S. Ayoubi, A. Gharipour, A. Y. Jazi, “Prediction of soil physical properties by optimized support vector machines,” Int. Agrophys. 26(2), 109–115 (2012). Crossref, Web of Science, Google Scholar
35. P. Maniwara, K. Nakano, S. Ohashi, D. Boonyakiat, P. Seehanam, P. Theanjumpol, P. Poonlarp, “Evaluation of NIRS as nondestructive test to evaluate quality traits of purple passion fruit,” Sci. Hortic. 257, 108712 (2019). Crossref, Web of Science, Google Scholar
36. A. A. Gowen, G. Downey, C. Esquerre, C. P. O’Donnell, “Preventing over-fitting in PLS calibration models of near-infrared (NIR) spectroscopy data using regression coefficients,” J. Chemom. 25, 375–381 (2011). Crossref, Web of Science, Google Scholar
37. J. Gao, C. Q. Wang, “Regression model for Apples’ near infrared spectroscopy based on LS-SVM,” Comput. Meas. Control 19, 176–179 (2011). Google Scholar
38. H. Zheng, H. F. Lu, “A least-squares support vector machine (LS-SVM) based on fractal analysis and CIELab parameters for the detection of browning degree on mango (Mangifera indica L.),” Comput. Electron. Agric. 83, 47–51 (2012). Crossref, Web of Science, Google Scholar
39. M. Ghaedi, A. M. Ghaedia, M. Hossainpour, A. Ansari, M. H. Habibi, A. R. Asghari, “Least square support vector (LS-SVM) method for modeling of methylene blue dye adsorption using copper oxide loaded on activated carbon: Kinetic and isotherm study,” J. Ind. Eng. Chem. 20, 1641–1649 (2014). Crossref, Web of Science, Google Scholar
40. M.-T. Sánchez, M.-J. De la Haba, I. Serrano, D. Pérez-Marín, “Application of NIRS for nondestructive measurement of quality parameters in intact oranges during on tree ripening and at harvest,” Food Anal. Methods 6, 826–837 (2013). Crossref, Web of Science, Google Scholar
41. M. Li, W. Lv, R. Zhao, H. Guo, J. Liu, D. Han, “Non-destructive assessment of quality parameters in ‘Friar’ plums during low temperature storage using visible/near infrared spectroscopy,” Food Control 73, 1334–1341 (2017). Crossref, Web of Science, Google Scholar
42. P. C. Williams, “Implementation of near-infrared technology, ” Near-Infrared Technology in the Agricultural and Food Industries, 2nd Edition, P. C. Williams, K. H. Norris, Eds., pp. 145–171, American Association of Cereal Chemists, St. Paul, Minnesota (2001). Google Scholar
43. D. Zhao, D. M. Oosterhuis, C. W. Bednarz, “Influence of potassium deficiency on photosynthesis, chlorophyll content, and chloroplast ultrastructure of cotton plants,” Photosynthetica 39, 103–109 (2001). Crossref, Web of Science, Google Scholar
44. M. T. Gómez-Casero, F. López-Granados, J. M. Peña-Barragán, M. Jurado-Expósito, L. G. Torres, R. Fernández-Escobar, “Assessing nitrogen and potassium deficiencies in olive orchards through discriminant analysis of hyperspectral data,” J. Am. Soc. Hortic. Sci. 132, 611–618 (2007). Crossref, Web of Science, Google Scholar
45. R. O. Pacumbaba, Jr., C. A. Beyl, “Changes in hyperspectral reflectance signatures of lettuce leaves in response to macronutrient deficiencies,” Adv. Space Res. 48, 32–42 (2011). Crossref, Web of Science, Google Scholar
46. D. I. Givens, E. R. Deaville, “The current and future role of near infrared reflectance spectroscopy in animal nutrition: A review,” Aust. J. Agric. Res. 50, 1131–1145 (1999). Crossref, Google Scholar
47. P. Williams, K. Norris (Eds.), Near-Infrared Technology in the Agricultural and Food Industries, American Association of Cereal Chemists, St. Paul, Minnesota (2001). Google Scholar
48. P. Stchur, D. Cleveland, J. Zhou, R. G. Michel, “A review of recent applications of near infrared spectroscopy, and of the characteristics of a novel PbS CCD array-based near-infrared spectrometer,” Appl. Spectrosc. Rev. 37, 383–428 (2002). Crossref, Web of Science, Google Scholar
49. B. Jamshidi, S. Minaei, E. Mohajerani, H. Ghassemian, “Reflectance Vis/NIR spectroscopy for nondestructive taste characterization of Valencia oranges,” Comput. Electron. Agric. 85, 64–69 (2012). Crossref, Web of Science, Google Scholar
50. N. Nakawajana, J. Posom, “Comparison of analytical ability of PLS and SVM algorithm in estimation of moisture content, higher heating value, and lower heating value of Cassava Rhizome ground using FTNIR spectroscopy,” IOP Conf. Ser., Earth Environ. Sci. 301, 012032 (2019). Crossref, Google Scholar
51. S. F. Li, X. Zhang, J. J. Li, Y. Shan, Z. Huang, “Non-destructive detecting fructose and glucose content of honey with Raman spectroscopy,” Trans. Chin. Soc. Agric. Eng. 30(6), 249–255 (2004). Google Scholar
52. S. R. Delwiche, L. O. Pordesimo, A. M. Scaboo, V. R. Pantalone, “Measurement of inorganic phosphorus in soybeans with near-infrared spectroscopy,” J. Agric. Food Chem. 54(19), 6951–6956 (2006). Crossref, Web of Science, Google Scholar

Vol. 13, No. 06

Metrics

Downloaded 1,038 times

History

Received 26 May 2020

Accepted 9 September 2020

Published: 28 October 2020

Information

This is an Open Access article. It is distributed under the terms of the Creative Commons Attribution 4.0 (CC-BY) License. Further distribution of this work is permitted, provided the original work is properly cited.

Keywords

PDF download

System Upgrade on Tue, May 28th, 2024 at 2am (EDT)

Quantification of potassium concentration with Vis–SWNIR spectroscopy in fresh lettuce

Abstract

1. Introduction

2. Materials and Methods