Assessment of internal quality of blueberries using hyperspectral images with selected wavelengths

Gabriel A. Leiva-Valenzuela
a

a,b,

Renfu Lu , José Miguel Aguilera

a

b

United States Department of Agriculture, Agricultural Research Service, Michigan State University, /224 Farrall Hall, Michigan State University, East Lansing, Michigan, United States of America 48824. b Department of Chemical and Bioprocess Engineering, Pontificia Universidad Católica de Chile, Avenida Vicuña Mackenna 4860, Macul, Santiago, Chile. gmleiva@uc.cl

ABSTRACT Hyperspectral imaging has been used in previous studies for assessing firmness and soluble solids content of fresh fruit. To apply this technique for automatic sorting it is necessary develop robust prediction models with fewer wavelengths. In this study, a hyperspectral imaging system was used to acquire reflectance images from 400 blueberries for the spectral region of 400-1,000 nm. Mean spectra were extracted from the regions of interest for the hyperspectral images of each blueberry. Calibration models for firmness index (FI) and soluble solids content (SSC) were developed using partial least squares regression for the entire reflectance or transmittance spectra as well as their combined data. Further, interval partial least squares (iPLS) regression with nine different intervals of 10 wavelengths was used to reduce the spectral dimensionality. Overall, the best Rp was 0.85 for SSC and 0.77 for FI. The prediction errors for iPLS, on average, increased by 7%, compared to PLS for the whole spectra. This preliminary research has demonstrated that it is feasible to implement hyperspectral imaging technique for sorting blueberries for SSC and possibly firmness, using appropriate wavelengths.

1

Introduction

In recent years, countries in the southern hemisphere have increased blueberry export to the northern hemisphere. Because of the long distance shipment, it is especially important to perform internal and external quality determination for individual fresh blueberries to ensure their quality upon arrival at the destination. Soluble solids content (SSC) and firmness (or softening) is important quality parameters that determine the flavor and shelf life of blueberries. Hence individual blueberries should be inspected for firmness to avoid spoilage and possible rejection by the consumer. Hyperspectral imaging technique has emerged as a new technique for quality and safety inspection of food and agricultural products (ElMasry and Sun, 2010). However, the need for processing a large number of spectral images has been the main obstacle in the application of hyperspectral imaging for high-speed online sorting and grading of agricultural products. Consequently, considerable studies were focused on the development of spectral data reduction algorithms. Interval partial least squares (iPLS) is one of the commonly used algorithms in selecting the most efficient wavelength regions for developing an optimized local PLS model with fewer variables (Nørgaard et al., 2005). Our previous study (Leiva-Valenzuela et al., 2013) evaluated hyperspectral imaging technique for predicting SSC and Firmness of blueberries. This work presents preliminary results about using hyperspectral imaging technique, coupled with PLS and iPLS for selection of wavelengths, to predict internal firmness and SSC of blueberries with either whole spectra or selected wavelengths in the spectral region of 400 – 1,000 nm. This manuscript presents preliminary results about hyperspectral imaging technique for predicting internal quality attributes of blueberries using either whole spectra or selected wavelengths in the visible and short-wave near-infrared region of 400 – 1,000 nm.

2 2.1

Material and Methods Image acquisition

Commercial “Rabbit edge” blueberries (Vaccinium corymbosum. Var Tifblue), originally produced in Chile, were visually inspected for appearance and surface defects. Only those fruit free of visual defects (such as
InsideFood Symposium, 9-12 April 2013, Leuven, Belgium 1|P a g e

Each hypercube consisted of 150 scanning lines X 688 spatial pixels X 520 wavelengths covering the spectral region of 362 1010 nm. a cylindrical reference disc made from white Teflon (6 mm in diameter and 6 mm in thickness) was mounted next to the frame of the sample holding unit and it was connected with a fiber optic line from the front of the lens in the light enclosure toward the disc. Michigan. Mean reflectance were computed by averaging over all pixels from the regions of interest for each segmented blueberry image.. Dark current images were acquired with the camera shutter being closed. the stem end and equator facing vertically toward the imaging system). USA) using the same imaging parameters as that for blueberries.A (Ariana and Lu. East Lansing. U. hyperspectral reflectance images were also acquired from a 20% diffuse reflectance standard (Spectralon® Reflectance Standard SRS-20-010.scars. 2010) and then segmented (Otsu.. 1. Each blueberry was placed on the sample holding unity and then imaged for each of the three orientations (i.S. For transmittance images correction.2 Image processing and prediction models Hyperspectral images were corrected (Eq. Fig. cuts. etc. They were stored in refrigerated air at 4 °C for various time periods from three to 14 days before experiments were performed. 2006) (Fig.e. Labsphere. 2. shrivel.) were selected for the experiment.30 nm/pixel and the spatial resolution of 0.167 mm/pixel. NH. For correction of the light source effect. with the calyx end. Schematic of a hyperspectral imaging system. Hyperspectral reflectance images were acquired for 400 blueberry samples using a prototype hyperspectral imaging system developed in the USDA/ARS postharvest engineering laboratory at Michigan State University.1). (1) P a g e |2 . The hypercubes had a spectral resolution of 1. North Sutton. 1979). Inc.1) by the reference image pixels by pixels (Ariana and Lu. Blueberries were removed from cold storage three hours before the experiment was begun to allow the fruit to reach room temperature (~22 ºC).

2). The models were evaluated using an independent set of samples (Fig. the calibration models were used to predict the independent samples. Interval PLS algorithm (iPLS) was used to select the best wavelengths. ID is the dark frame image value. The same procedure was implemented for reflectance and transmittance. Fig. IS is the sample image value. Calibration models for predicting FI and SSC were developed using PLS with ¾ of the samples. and IR is the reference image pixel value (Fig. 2a and 2b).2). Leave one out cross-validation was used to determine the number of latent variables for the calibration models.where RI is the relative reflectance or transmittance at pixel location (x.3 Image processing and prediction models Calibration models for predicting FI and SSC were developed using PLS for the best 20 principal components. 2. 2 Flow diagram for the assessment of internal quality.y). RMSEP = (2) 3|P a g e . referred to as calibration variance and imprecision of prediction model respectively. Thereafter. Three folds (or 75% of all samples) were used for calibration and the remaining fold was used for independent test or prediction. The models were evaluated using root mean squares errors for cross validation (RMSECV) and prediction (RMSEP) (See Eq.

4 with concordance with Table 1 could have two consequences: -Possible sorting 2 classes of firmness: soft and firm (RPD values close 1. Reflectance increased rapidly starting from 650 nm and reached a peak at ~800 nm and another peak at 900 nm. Hence. Firmness Index (FI) was determined from the slope of the forcedeformation curve obtained from the compression test of blueberries with a Texture Analyzer (Model TA. The rotated out samples were then used as the new test samples.. defined as the ratio between the sample standard deviation and RMSEP. This method builds a series of sub PLS models in adjacent but non-overlapping intervals with the same dimensions as that for the conventional PLS to minimize the RMSECV with cross validation (Nørgaard et al. 2000). For reference measurement. 3 Mean relative reflectance for 400 blueberry samples. Rc.9 % SSC) as criterion to remove tasteless fruit in the packing line. A simple visual inspection of results suggests differences in prediction models for SSC and Firmness.5). low relative reflectance (< 10%) and smaller variation (~20%) were observed among the samples. The above procedure was repeated four times.000 nm (Fig. RMSECV.2): sweetest and tasteless based in mean value of sample distribution (12. For each new run.9 and 2.. RMSEP and RPD) from the four runs were averaged to estimate the final performance of the models.e. it then decreased rapidly at the wavelengths of 900-1. . 3). To select the best wavelengths. RPD values measure the ability of a model for classification (Nicolaï et al.XT2i).. 3 Results and Discussion In the visible spectral region (up to 650 nm).In addition. correlation coefficients for calibration (Rc) and prediction (Rp) and residual predictive deviation (RPD). P a g e |4 . SSC actual and predicted values adjust better than for Firmness index. Rp. 2007). Fig.To predict SSC or sort berries in two sweetness classes (RPD values of 1. one fold of samples in the calibration set was rotated out and replaced with the test samples from the previous run. the results (i. These differences observed in Fig. using a digital refractometer. were also calculated. soluble solids content (SSC) was measured from the juice extracted from berry halves. interval PLS (iPLS) was used. Finally.

5 Selected wavelength intervals for predicting soluble solids content (SSC) and firmness index (FI). 5|P a g e . RPD ~ 2 suggest the feasibility of sorting blueberries into two SSC classes. Fig.. 4 PLS calibration and prediction of: a) soluble solids content and b) firmness index for the complete data set. a reduction of spectral dimensionality was considered using interval Partial Least Square (iPLS) (Nørgaard et al. Comparison results between complete data set and reduced models are shown in (Table 1).Fig. Therefore. 2000). The speed of image acquisition by a hyperspectral imaging system is not sufficient to meet the requirement for the sorting lines in the packinghouse. For SSC. 5). which could allow hyperspectral imaging technique to be implemented online as a multispectral imaging system with a few wavelengths (Fig.

2006. J..0 Moreover.03 0.E. R.89± 0. L.1 -14. Peirs. B. G. Stacey Stark.. 62-66.. Multivariate near-infrared and Raman spectroscopic quantifications of the crystallinity of lactose in whey permeate powder.797± 0.4 16 ± 3 0.849 ± 0. London. Gabriel A.935± 0.. Knudsen.A.08 1. Computers and Electronics in Agriculture 53..3 Dif (%) 0.M.. A threshold selection method from gray-level histograms. Beullens. L. N. Near-infrared hyperspectral reflectance imaging for detection of bruises on pickling cucumbers.010 0. J.P. Wagner. Engelsen.-W. J. (Ed. Leiva-Valenzuela thanks The National Commission for Science and Technology (CONICYT) of Chile for providing a fellowship.. Hahn.004 0.Table 1 Prediction results for firmness and SSC using the complete spectral data with PLS and reduced data set with iPLS.06 RPD 2. Nielsen.2 ± 0. 91-98. Nicolaï. 1261-1270.A. Otsu. K.020 0.02 1.. Miranda Sperry and Ms. International Dairy Journal 15..B. Most of the selected wavelengths for SSC and firmness prediction were from the NIR region.I..01 1. Journal of Food Engineering 115..)... Acknowledgements Mr. 60-70. Theron. P a g e |6 .020 1.04± 0. D.. Saeys.892± 0. Farhat.9 ± 0. Principles of Hyperspectral Imaging Technology. ElMasry.026 0. W.. I.69 ± 0. Engelsen. Leiva-Valenzuela. p. Munck. In: Sun. 4 Conclusions The visible spectral region is not effective for SSC or firmness predictions. Aguilera..T. selection showed that using the reduced data set increased the standard error of prediction for the FI and SSC in average 7%. 2005.85 ± 0. D.011 0. D. L. 1979.. 2000. graduate students in Biosystems Engineering at Michigan State University. M.5± 0.0 9±4 0. Sun.31 ± 0. 413-419.-W.. 2007. It is feasible to implement hyperspectral imaging technique using a few selected wavelengths for online sorting of blueberries for firmness and SSC. R. Lu... 2013..B.014 15 ±3 0. Saudland. Hyperspectral Imaging for food quality analysis and control.04 0. Lammertyn..771 ± 0. UK.1 0.036 0.28 ± 0. References Ariana. Nørgaard. Applied Spectroscopy 54.757 ± 0. Treatment SSC prediction Whole spectra iPLS reduced FI prediction Whole spectra iPLS reduced PC 19 ± 1 R c RMSECV 0.34 ± 0.02 0.900± 0. Prediction of firmness and soluble solids content of blueberries using hyperspectral reflectance imaging.34 ± 0. Bobelyn. K.P. 2010.B.02 R p RMSEP 0. Interval partial least-squares regression (iPLS): A comparative chemometric study with an example from near-infrared spectroscopy.5 ± 0. L. G. J... 477. E. Guyer.850± 0. Nondestructive measurement of fruit and vegetable quality by means of NIR Nørgaard. S. Man and Cybernetics 9.M. The authors also acknowledge helpful experimental aid from Ms. Elsevier Inc. D.. A... S. A. IEEE Transactions on Systems. Lu.