A New Approach to Predict Acidity of Bayberry Juice by Using Vis/Near Infrared Spectroscopy

Abstract

A total of 76 bayberry juices were collected and their spectra features were got by using a vis/NIR spectroscopy. One mixed algorithm was used to predict the acidity (pH) of bayberry juice with partial least squares (PLS) and artificial neural network (ANN). PLS was used to find some sensitive spectra actives to acidity in juice, before doing this, the influence of various spectral pretreatments (standard normal variate, multiplicative scatter correction, S.Golay first derivative, wavelet package transform) were compared. The PLS approach with WPT preprocessing spectra was found to provide the best results, and the spectral reflectivity corresponding to them were regarded as the input neurons of ANN. Remnant values by subtracting standard values and validation values, were regarded as the output neurons of ANN. The calibration equation developed from them was used to predict the constituent values for the independent spectra of 30 samples. The results indicated that the observed results using PLS-ANN (r_p = 0.943) were better than those obtained by PLS (r_p =  0.932). At the same time, the sensitive wavelengths corresponding to the acidity of bayberry juices or some element at a certain band were proposed on the basis of regression coefficients by PLS. It indicates that using vis/NIRS technique to fast and nondestructive detection the acidity of bayberry juices was feasible.

Keywords:

• Vis/NIR spectroscopy
• PLS
• ANN
• PLS-ANN
• Bayberry juice
• Acidity (pH)

INTRODUCTION

Vis/NIR spectroscopy is the electro-magnetic wave whose frequency is between visual light and near infrared. Its wavelength is defined as from 380 to 2526 nm. Vis/NIR spectroscopy technology has been widely used in quantitative and qualitative analysis. In the all, it has become an important detecting technique in every field, especially in food science and agricultural science. Kawano et al.[1] studied the sugar content in peaches with an optical fibre in interactance mode. Slaughter[2] determined that visible and NIR-spectroscopy could be used to measure non-destructively the internal quality of peaches and nectarines as characterized by their SSC, sugar content, sorbitol content, and chlorophyll content. Wang et al.[3– 4] classified pecky rice kernels and fungal-damaged soybean seeds through using visible and near-infrared spectroscopy.

Bayberry (Myrica rubra Siebold and Zuccarini), or Yangmei in Chinese Mandarin, is indigenous to China. It has been cultivated more than 2000 years and the annual output is about 300,000 tons. Its flesh can be processed into sweets, jam, juice and wine, or canned in syrup. It is a very important fruit-product as juice or wine for export, and it has a high economical value. The acidity (pH) and sugar content are the important internal quality attribute to bayberry. Most the methods to measure these qualities are based on complex processing of samples, and involve a considerable amount of manual work. At present, research has been focused on the development of non-destructive measurements techniques for fruit and juice internal quality using vis/NIR spectroscopy techniques as told above. But the most important techniques for NIR are how to extract quantitative information from them. The objectives of the study were to evaluate the use of vis/NIR spectroscopy in measuring the acidity (pH) of bayberry juice, and try to find a better mathematic model to predict it by using vis/NIR.

MATERIALS AND METHODOLOGY

System Set-up and Experimental Material

Spectra were collected using a NIR scanning spectroradiometer ASD Handheld FieldSpec in reflectance mode. Measurements were made at ambient temperature (25°C) over the wavelength range 325–1075 nm at 1.5 nm intervals. The spectroradiometer plumb placed at a height of approximately 100 mm above the vessel center, a 150 W halogen lamp was used, and the angle between them was 30°. Figure 1 shows the collection system of bayberry juice. Reflectance data were stored as log (1/R) (R = reflectance) at 1nm intervals (751 spectra data points). Software of ASD View Spec Pro, Unscramble V9.2 (CAMO, PROCESS, AS, OSLO, Norway), and DPS (data procession system for practical statistics) were used in this article.

Figure 1 Collection system of bayberry juice.

Seventy six of bayberry juices were gained from the local market. There are 4 varieties of bayberry juice with 19 of each variety (They are Cixi bayberry juice, Xianju bayberry juice, Lishui bayberry juice, and linhai bayberry juice). The samples were placed in airtight glass vessel, stored in an ice filled cooler and transported to the laboratory to keep at cold temperature (4 ± 1°C). All samples were first allowed to equilibrate to room temperature (25°C) before NIR spectroscopy analysis. Each sample was individually numbered and placed in each vessel. In order to reduce the error of operation during the whole experiment, each vessel was glass service chose as the same diameter of 6.5 cm and height of 1.4 cm. For each bayberry juice the scan number were 10 at exactly the same position. The measurement of acidity was measured by a pH meter (SJ-4A, Exact instrument Co., Ltd., Shanghai, China) under the China standard.

Processing of the Optical Data

SNV is a mathematical transformation of log (1/R) spectra. It removes the multiplicative interferences of scatter, particle size, and the change of light distance.[5] MSC is used to correct for additive and multiplicative effects in the spectra. This method is useful when applying linear regression methods as it eliminated scattering effects.[6] S. Golay 1st-Der,[7] which is a moving window averaging method, consisting of a preliminary averaging over a fixed number of points (segment option), and the gap was set as three. All of the above three preprocessing methods were disposed in Unscramble V9.2. WPT[8] is a novel signal processing technique, which has dual localization in both frequency and time. It is an important extension of wavelet transform. In this paper, the wavelet function symlets 6 was adopted to decompose wavelet packet into five layers. It was carried out through using Matlab 7.0. To avoid low signal-noise ratio, only the wavelengths ranging from 400 to 1000 nm were used in this investigation. Figure 2 is the four original representative absorbance spectral data from 325 nm to 1075 nm (one of each variety).

Figure 2 Four original representative absorbance spectral data from 325 nm–1075 nm.

Partial Least Square (PLS)

Partial least square analysis, a method of NIR spectrum analysis, is the most widely used choices among many technologies.[9–11] Because it takes the advantage of the correlation relationship that already exists between the spectral data and the constituent concentrations. Besides, the model built by PLS is steadier, especially to the study on small samples and large variables. However, PLS is based on linear models and unsatisfactory results may be obtained when non-linearity behaviors are present. To model these non-linearities, additional factors have to be introduced into the linear PLS model. Nowadays these non-linearities are commonly modeled by using neural network.[12]

Artificial Neural Network (ANN)

The most popular type of neural network for use in analytical applications is artificial neural network (ANN)[13–15] with the back-propagation learning algorithm. It is a one-way multilayer feed-forward network. Back–propagation uses a learning process to minimize the global error of the system by modifying node weights. The weight increment or decrement is achieved by using the gradient descent rule. The network is trained by initially selecting the weights at random and then presenting all training data repeatedly. The weights are adjusted after every trial using external information specifying the correct result until the weights converge, and the errors are reduced to acceptable values. The delta rule was used as the learning rule, which specifies how connection weights are changed during the learning process. The sigmoid transfer function was chosen for the non–linear function that transfers the internally generated sum for each node to a potential output node.

RESULTS AND DISCUSSION

Near-infrared spectra are difficult to interpret directly because of the overlap of weak overtones and combinations of fundamental vibration bands. As a result, multivariate calibration is required for quantitative analysis of sample constituents by NIRS. Various calibration methods have been used to relate near-infrared spectra with measured properties of materials. Principal components regression (PCR), partial least squares (PLS), multiple linear regression (MLR), Fourier regression and artificial neural networks are the most used multivariate calibration techniques for NIRS.[16– 17] PLS was usually considered for a large number of applications in fruit and juice analysis and was widely used in multivariate calibration. Because it takes the advantage of the correlation relationship that already exists between the spectral data and the constituent concentrations. However, PLS is based on linear models and unsatisfactory results may be obtained when non-linearities are present. Artificial neural network (ANN) as a new technique has been widely used for its prediction and forecasting abilities especially in complex settings.

Multivariate Calibration by PLS-ANN for Acidity

In this article, a mixed algorithm was employed for building the nonlinear model of NIR and the acidity of bayberry juice. The mixed algorithm was combined with PLS method and ANN. The model based on the mixed algorithm was divided into two parts: linear part and nonlinear part, and the corresponding model of each part were built, respectively. In the linear part a total of 46 juice samples were used to build a calibration model by PLS. The calibration models for acidity with pre-processed spectra were developed using PLS regression with cross validation by the predicted error sum of squares (PRESS) function in order to avoid overfitting of the models. To evaluate the results, the standard error of calibration (SEC) and correlation coefficient (r) for the calibration set, and standard error of cross-validation (SECV) and r for validation set were considered. Good models should have lower SEC and SECV, higher r but smaller differences between SEC and SECV.

All the parameters of different calibration and validation sets were showed in Table 1. The results indicated that the pre-processing method affected the performance of models. Compared the four different preprocessing method in wavelength range of 400–1000 nm, WPT was the best, with higher correlation coefficient and lowest SEC and SECV values also smallest differences between SEC and SECV.

Table 1 Calibration and validation results of partial least square for analyzing acidity of bayberry juice

Method	Calibration			Validation
	SEC	RMSEC	r	SECV	RMSECV	r
SNV	0.1204	0.1193	0.9660	0.1675	0.1661	0.9334
MSC	0.1611	0.1597	0.9383	0.1904	0.1887	0.9128
1st-Der	0.1482	0.1469	0.9480	0.2028	0.2010	0.9021
WPT	0.1206	0.1195	0.9659	0.1664	0.1650	0.9343
SNV: standard normal variate; MSC: multiplicative scatter correction; 1 st-Der: S. Golay first derivative; WPT: wavelet package transforms; SEC: standard error of calibration; RMSEC: root mean square error of calibration; r: correlation coefficient; SECV: standard error of cross-validation; RMSECV: root mean square error of cross-validation.

At the same time, PLS method was used to find some sensitive spectra actives to acidity in juice where 25 wavebands were obtained for its higher regression coefficients, and then regarded the spectral reflectivity corresponding to them as the input neurons of ANN. Calculate the remnant values by subtracting standard values and validation values, and regarded them as the output neurons of ANN. So in the nonlinear part, a remnant ANN was built which regard sensitive spectral reflectivity as the input neurons and remnant values as the output neurons. A three-layer ANN was built. The transfer function of each layer was sigmoid function. The node of input layer was 25. The node of hidden neuron layer was 6. The node of output neuron layer was 1. The goal error was set as 0.001. The speed of learning was set as 0.2. The time of training was set as 1000.

Prediction Results of PLS-ANN Regression Analysis for Acidity

In the prediction, using PLS models with four different preprocessing methods were used to predict 30 unknown juice samples, and got the best prediction values of them. It showed that PLS model with WPT preprocessing is an optimal method to preprocess data of the spectra, and then regard spectral reflectivity corresponding to the 25 sensitive wavebands of these 30 samples as the input neurons of ANN, and predict the remnant values. The final predicted acidity values of 30 juice samples were got through adding predicted values (by calibration model of PLS) and remnant values (by remnant ANN model). The results indicated that the observed results using PLS-ANN were better than those obtained by PLS. The r_p for PLS-ANN prediction is 0.943, RMSEP = 0.145, bias = −0.101; However, the r_p for PLS prediction is 0.932, RMSEP  =  0.168, bias = −0.117. The model based on PLS-ANN shows superior performance for determination the acidity of bayberry juice.

Sensitive Wavelengths Analysis Based on Regression Coefficients

The sensitive wavelengths reflecting the characteristics of spectra for acidity was obtained based on regression coefficients. From Figure 4, we can find that wavelengths of 685∼695 nm, 910∼925 nm might be of particular importance for the acidity calibration. While the wavelengths between 700 to 950 nm are possible that it results from a 3rd overtone stretch of CH and 2nd and 3rd overtone of OH in bayberry juice which was referred by Slobodan and Yukihiro in their article about detailed band assignment for the short-wave NIR region useful for various biological fluids.[18] However, it is likely that the wavelengths 685∼695 nm for acidity mainly attributed to the color of bayberry juice because there is non-existent of organic acids in this region of the spectrum. So in our research, to acidity, 910∼925 nm were better. This found was similar to the earlier literature, such as He found wavelengths near 900 nm were sensitive wavelengths corresponding to organic acid of oranges.[19]

Figure 3 Prediction results for the unknown 30 samples from the PLS-ANN models.

Figure 4 Regression coefficients with corresponding wavelengths for acidity.

CONCLUSIONS

The results indicate that it is possible to use this non-destructive technique for measuring the acidity of bayberry juice. Through multivariate calibration technique of PLS a 46 juice samples calibration model was established. And at the same time, 25 sensitive spectra actives to acidity in juice were found, the spectral reflectivity corresponding to them were regarded as the input of ANN. Remnant values which were got by subtracting standard values and validation values were regarded as the output neurons of ANN. The predicted values of the unknown 30 juices were got through adding predicted values (by calibration model of PLS) and remnant values (by remnant ANN model). The model based on PLS-ANN show superior performance for determination the acidity of bayberry juice, that is, the correlation coefficient r = 0.943, RMSEP  =  0.145, bias  =  −0.101. At the same time, the sensitive wavelengths corresponding to the acidity of bayberry juices or some element at a certain band were proposed on the basis of regression coefficients by PLS. For acidity, wavelengths 910∼925 nm might be of particular importance. It indicates that using vis/NIRS technique to fast and nondestructive detection the acidity of bayberry juices was feasible.

ACKNOWLEDGMENTS

This study was supported by the Teaching and Research Award Program for Outstanding Young Teachers in Higher Education Institutions of MOE, PRC, Natural Science Foundation of China (Project No: 30270773), Specialized Research Fund for the Doctoral Program of Higher Education (Project No: 20040335034) and Natural Science Foundation of Zhejiang Province, China (Project No: RC02067).