If You Can't Use This Word Document and You'd Like a PDF Cover Sheet Please Contact Sandy s1

Send your completed paper to Sandy Rutter at by 13 April 2007 to be included in the ASABE Online Technical Library.

If you can't use this Word document and you'd like a PDF cover sheet please contact Sandy.

Please have Word's AutoFormat features turned OFF and do not include live hyperlinks. Your paper should be no longer than 12 pages. For general information on writing style, please see http://www.asabe.org/pubs/authguide.html.

This page is for online indexing purposes and should not be included in your printed version.

Author(s)

First Name / Middle Name / Surname / Role / Email
Xiaping / Fu /
Yibin / (or initial) / Ying / ASABE Member /
Huirong / Xu /
Ying / Zhou /

Affiliation

Organization / Address / Country
CollegeofBiosystemsEngineeringandFoodScience,ZhejiangUniversity / 268 Kaixuan St. Hangzhou, 310029, / China

Publication Information

Pub ID / Pub Date
073057 / 2007 ASABE Annual Meeting Paper

The authors are solely responsible for the content of this technical presentation. The technical presentation does not necessarily reflect the official position of the American Society of Agricultural and Biological Engineers (ASABE), and its printing and distribution does not constitute an endorsement of views which may be expressed. Technical presentations are not subject to the formal peer review process by ASABE editorial committees; therefore, they are not to be presented as refereed publications. Citation of this work should state that it is from an ASABE meeting paper. EXAMPLE: Author's Last Name, Initials. 2007. Title of Presentation. ASABE Paper No. 07xxxx. St. Joseph, Mich.: ASABE. For information about securing permission to reprint or reproduce a technical presentation, please contact ASABE at or 269-429-0300 (2950 Niles Road, St. Joseph, MI 49085-9659 USA).

An ASABE Meeting Presentation

Paper Number: 073057

Principal Components-Artificial Neural Networks for Predicting SSC and Firmness of Fruits based on Near Infrared Spectroscopy

Xiaping Fu, PH.D Candidate

CollegeofBiosystemsEngineeringandFoodScience,ZhejiangUniversity,

268 Kaixuan St. Hangzhou, 310029, China,

Yibin Ying*, Professor

CollegeofBiosystemsEngineeringandFoodScience,ZhejiangUniversity,

268 Kaixuan St. Hangzhou,310029, China,

Huirong Xu, PH.D Candidate

CollegeofBiosystemsEngineeringandFoodScience,ZhejiangUniversity,

268 Kaixuan St. Hangzhou,310029, China,

Ying Zhou, Graduate student

CollegeofBiosystemsEngineeringandFoodScience,ZhejiangUniversity,

268 Kaixuan St. Hangzhou,310029, China,

Written for presentation at the

2007 ASABE Annual International Meeting

Sponsored by ASABE

Minneapolis Convention Center

Minneapolis, Minnesota

17 - 20 June 2007

Mention any other presentations of this paper here, or delete this line.

Abstract. The use of near infrared (NIR) spectroscopy was proved to be a useful tool for components analysis of many materials. Principal component analysis (PCA) and artificial neural networks using back-propagation algorithm (BP-ANN) were used to establish nonlinear model for the prediction of soluble solid content (SSC) and firmness of peach and loquat fruits from NIR spectral data. The first ten principal components extracted from original spectra and spectra after multiplicative scattering correction (MSC) were used as input nodes of BP-ANN. TANSIG and LOGSIG transfer functions and two to nine neurons were considered for the hidden layer of the network. For peaches, the best results were R train=0.940 and R test= 0.900 for SSC; R train=0.701 and R test =0.453 for firmness. For loquats, the best results were R train=0.962 and R test= 0.893 for SSC; R train=0.812 and R test =0.624 for firmness. The results of this study show that combination of PCA and BP-ANN is feasible for predicting fruit quality based on NIRS. For further researches, factors such as the number of neurons for input layer, the number of hidden layers, other learning algorithms and so on could be studied to improve modeling performance and predicting accuracy.

Keywords. NIRS, PCA, BP-ANN, SSC, firmness

(The ASABE disclaimer is on a footer on this page, and will show in Print Preview or Page Layout view.)

Introduction

Nowadays, non-destructive, simple and credible methods for fruit internal quality measurements are required by both fruit planters and sellers. Near infrared spectroscopy (NIRS) technique has gained wide application in food quality analysis mainly due to its suitability for acquiring the spectra of different sample states at low cost without pretreatment, which lead to its non-destructive and rapid characters (Sáiz-Abajo et al., 2004). However, NIR spectra typically consist of broad, weak, non-specific, extensively overlapped bands (Fidêncio et al., 2002). These characteristics hindered expansion of the NIR technique until multivariate calibration methods became widely available and accepted (Blanco et al., 2000).

There are numerous multivariate calibration methods for quantitative analysis. The most widely used linear approaches were multiple linear regression (MLR), principal component regression (PCR) and partial least-squares regression (PLSR), which assume a linear relationship between the measured parameters of the samples and spectra characteristics (Xu et al., 2004). These methods were applied both for analyzing the concentrations of specific chemicals and sample properties such as color, firmness or viscosity, which produce a spectral response.

For the reason of the presence of substantial non-linearity, which arise from scattered light or intrinsic non-linearity in the absorption bands, methods for correcting non-linearity were applied, such as mathematical processing of spectra or alternative non-linear calibration methods. Mathematical processing such as spectrum derivation, the standard normal variate (SNV) and multiplicative scattering correction (MSC) can minimize or avoid some of the non-linearity (Chu et al., 2004). Another way to solve this problem is non-linear calibration methods, artificial neural networks (ANNs) for instance. The wide range of applications of ANNs is due largely to their ability to deal with complex functions, enabling the modeling of non-linear relationships. (Cirovic, 1997).

ANNs imitate the structure and functioning of the human nervous system, to build parallel, distributed and adaptive information-processing systems, able to display a degree of intelligent behavior (Perez-Marin et al., 2007). ANNs are typically organized in layers where these layers are made up of a number of interconnected nodes which contain an activation function. Input vectors are presented to the network via the input layer which communicates to one or more “hidden layers” where the actual processing is done via a system of weighted “connections”. Most ANNs contain some form of “learning rule” which modifies the weights of the connections according to input patterns that it is presented with (Ramadan, 2004). Feed-forward networks based on back-propagation algorithm (BP-ANN) are among the mostly used networks. These networks transfer the information on the input layer to one or several hidden layers; subsequently, the information held by the neurons in the hidden layers is combined via non-linear functions - frequently of the sigmoidal type - in order to obtain the output data, i.e. the target parameter(s). This type of algorithm is highly suitable as it lends itself readily to supervised learning, i.e. to learn from data with known responses and to use the acquired knowledge to predict the answers for other problems (Blanco et al., 1999). It has wide application in NIR spectra modeling for many research objectives, such as pharmacy (Dou et al., 2005, 2006; Deng et al., 2005), soil (Ramadan, 2004), beer (Inon et al., 2006), chemical compounds (Dorn et al., 2003),dairy (Shao et al., 2007), grains (Liu et al., 2004; Lin et al., 2004; Ji et al., 1999) and so on. Most of these research concluded that ANNs approach obtained a better result than those of other linear approaches. And some of the researchers pointed out that it is very essential for the networks to use pressed variables from pretreated spectral data as input (Dou et al., 2006; Shao et al., 2007).

In this study, principal components analysis (PCA) and BP-ANN were applied on fruit NIR spectra for determining soluble solid content (SSC) and firmness. The goal of this study was to predict fruit internal quality using non-linear BP-ANN model based on spectra pretreatments (including derivative and MSC) and data reduction.

Materials and Methods

Fruit Samples

White peaches were purchased from Jinhua County, Zhejiang Province in June, 2006. Loquats were picked from Tangxi, Zhejiang province in May, 2006. Samples were transported to the laboratory immediately and kept at cold storage (about 4 °C). 24 hours prior to measurements, samples were taken out from refrigerator and maintained at 25 °C and 60% RH to equilibrate to room condition. Samples with surface defects were removed from the samples set. Finally, a total of 120 peach samples and 100 loquat samples were used in this study. For each kind of fruit, three fourths of the samples were used for the training the network and the remaining one fourth were used for testing. For each sample, spectral acquisition, firmness and soluble solid content measurements were carried out in two hours.

Apparatus and software

NIR diffuse reflectance spectra of intact the fruit samples were acquired using a commercial FT-NIR spectrometer (Thermo Electron Corp., USA) with a bifurcated optical fiber cable. The fiber optic probe was enclosed in a 16 mm diameter stainless steel cylindrical tube. Both light source beams and receptor beams were enclosed in it randomly. The spectrometer consisted of a wide-band light source (50W quartz halogen lamp), an interferometer and an InGaAs detector with a spectral range of 800-2500 nm. Considering the physiological properties of fruit samples and the performance of the spectrometer, the parameters for spectra acquisition were set up as follows: resolution of 16 cm-1 and mirror velocity of 0.9494 cm s-1. A standard Teflon block was used as reference.

The spectrometer is connected to a computer, and specific software OMNIC6.1a (Thermo Electron Corp., USA) was available for spectra acquisition, pretreatments and saving. The diffuse reflectance spectra of samples were saved as log(1/R). Each spectrum was the average of 64 successive scans. For each sample, several replicate were taken and the mean spectrum was calculated.

SSC and firmness measurements

According to the China standard for SSC measurement in fruit (China standard, 1990a), SSC was measured by a digital refractometer (ATAGO PR-101, Tokyo, Japan).Firmness were conducted using a standard 6 mm MT probe mounted in an Instron universal testing machine (Instron 5543, USA) with a loading rate of 20 mm/min, interfaced to a computer to obtain continuous force-deformation curves. SSC and firmness measurements were taken from the corresponding locations where NIR diffuse reflectance spectra were acquired on each fruit.

Data Analysis

Spectra from the OMNIC software were exported in suitable format and analyzed using the TQ Analyst software (Thermo Electron Corp., USA) and Matlab software. Nowadays, multiple developed neural network structures are available. In this study, a feed-forward type neural network structure with error back-propagation learning rule was selected based on graphical user interface (GUI) toolbox in Matlab software.

Input data for the network can be either the original/preprocessed spectra or the result of any data reduction procedure based on the original/preprocessed spectra (Blanco et al., 2000). Spectra preprocessing of derivation and multiplicative scattering correction were applied on the original spectra. Because the number of wavelengths of the spectral data were too large (including 1102 points in each spectrum) and the power of calculation for training process of the neural network, it is impossible to use the original variables directly as input layer in artificial neural networks. For establishing and optimizing neural network models, the original independent variables (wavelengths) were compressed by principle component analysis in order to reduce the dimensionality of the spectral data. The new variables (calculated principle components (PCs)) can describe the original data matrix by retaining as much information as possible (Mouazen et al., 2006; Otsuka, 2004). The scores of such PCs were used as input of the ANN.

The network leading to the highest correlation coefficient and the lowest error was adopted as optimal. The calculation of root mean square error (RMSE) of training indicated how well the model fits the calibration data. And the parameter of RMSE of testing was reported as a indication of the model’s prediction capability.

Table1 Statistics of SSC and firmness for fruit in training and testing data sets.

variety / parameters / Training / Testing
SSC (°Brix) / Firmness (N) / SSC (°Brix) / Firmness (N)
peach / No. of samples / 90 / 30
Mean / 11.11 / 23.08 / 11.08 / 23.26
S.D. / 1.58 / 4.38 / 1.59 / 4.30
Max. / 14.43 / 34.10 / 14.05 / 33.11
Min. / 7.30 / 8.93 / 7.87 / 14.05
loquat / No. of samples / 75 / 25
Mean / 14.04 / 17.88 / 14.15 / 18.09
S.D. / 2.64 / 4.37 / 2.54 / 4.54
Max. / 19.80 / 31.58 / 19.1 / 29.74
Min. / 5.70 / 9.41 / 9.00 / 10.95

Results and Discussion

SSC and firmness statistics

Statistics of SSC and firmness for fruit in both training and testing data sets measured by standard destructive methods are summarized in Table1. As table1 shown, it’s clear that the SSC and firmness of samples in both training set and testing set are appropriately distributed. The mean value and standard deviation (S.D.) of both SSC and firmness in the two sets are very close to each other.