You are on page 1of 9

Computers and Electronics in Agriculture 69 (2009) 24–32

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture


journal homepage: www.elsevier.com/locate/compag

Instantaneous quantitative and qualitative assessment of pear quality using near


infrared spectroscopy
Patricia Paz a, María-Teresa Sánchez a,∗, Dolores Pérez-Marín b,∗, José Emilio Guerrero b, Ana Garrido-Varo b

a
Department of Bromathology and Food Technology, University of Cordoba, Campus Rabanales, Charles Darwin Building, 14071 Cordoba, Spain
b
Department of Animal Production, University of Cordoba, Campus Rabanales, 14071 Cordoba, Spain

a r t i c l e i n f o a b s t r a c t

Article history: A comparative study was made of the performance of different spectrophotometers as part of a research
Received 12 December 2008 into the potential of NIR reflectance spectroscopy as a non-destructive method for predicting soluble
Received in revised form 21 April 2009 solid content, firmness and postharvest shelf-storage time in intact pears. Three commercially available
Accepted 15 June 2009
spectrophotometers were used: a scanning monochromator (SM) of 400–2500 nm, a combination of diode
array and scanning monochromator (DASM) of 350–2500 nm and a diode array (DA) of 900–1700 nm. A
Keywords:
total of 332 pears (Pyrus communis L.) cv. ‘Blanquilla’ and ‘Conference’ were used to develop calibration
NIR spectroscopy
models in different spectral regions, using various spectral signal pretreatments. Models to predict soluble
Intact pear
Quality
solid content yielded coefficients of determination (r2 ) ranging from 0.39 to 0.76, and a standard error of
Non-destructive analysis cross-validation (SECV) of between 0.59 and 1.49 ◦ Brix, depending on the instrument used. Models for
Soluble solid content predicting firmness yielded r2 values of between 0.45 and 0.79 and SECVs ranging from 5.33 to 7.36 N.
Firmness Subsequently, the ability of these instruments to classify fruit as a function of postharvest shelf-storage
Shelf-storage time time (0, 6 and 8 days) was tested using partial least squares discriminant analysis (PLS-DA), which yielded
percentages of correctly classified samples (ratio of correctly classified samples to total samples) ranging
from 81.1 to 94.4%, depending on the instrument used. The results obtained highlight the potential of
these NIR instruments for assessing internal quality indices and predicting postharvest storage duration
in pears, particularly when the instruments SM and DASM were used.
© 2009 Elsevier B.V. All rights reserved.

1. Introduction ness and SSC as a function of postharvest storage time could provide
valuable information for commercial decision-making, since fruit
The fresh fruit market is becoming increasingly demanding with sold to the consumer must meet given quality standards based on
regard to product quality; this, in turn, governs fruit prices. Today, these two parameters.
one of the main aims of the fruit industry is to provide the consumer At the same time, the consumer buying the product in the super-
with products meeting high internal quality standards, rather than market needs to be sure how long it will subsequently last. Shelf-life
fruit which looks mouthwatering but actually tastes insipid or has can be seen as a synthesis of all the physical/chemical characteris-
an undesirable texture. tics of the product, and its prediction is therefore of great interest
Parameters such as soluble solid content (SSC) and, especially, to retailers.
firmness are among the major quality attributes of pears (Nicolaï et The quality control of fruits and vegetables has traditionally been
al., 2008). They must therefore be taken into account when deter- based on destructive methods. As a result, only a small number of
mining the effect of postharvest storage time on product quality. control samples could be analysed per batch, giving rise to problems
Good-quality pears develop a buttery, juicy texture, following soft- caused by the individual variation between products in the same
ening related to changes in cell wall structure (Murayama et al., batch (Galvis-Sánchez et al., 2004).
2006). Firmness is thus the best indicator of degree of ripening, The fruit sector would benefit greatly from a non-destructive,
and must be measured at both the shipping and the retail stages rapid, precise, low-cost, non-contaminant and versatile method
(Thompson and Crisosto, 2002; Valero et al., 2007). Charting firm- for quality control. NIRS technology would appear to be ideal for
this purpose, in that it meets all these requirements. It also yields,
instantly, a single spectral analysis which potentially provides
∗ Corresponding authors. Tel.: +34 957 212576; fax: +34 957 212000. information on a whole range of parameters, including chemi-
E-mail addresses: teresa.sanchez@uco.es (M.-T. Sánchez), pa2pemad@uco.es cal attributes (e.g. soluble solid content), physical properties (e.g.
(D. Pérez-Marín). firmness) and a global index (shelf-life). This method is espe-

0168-1699/$ – see front matter © 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.compag.2009.06.008
P. Paz et al. / Computers and Electronics in Agriculture 69 (2009) 24–32 25

Table 1
Statistical analysis of the calibration and prediction sample sets, i.e. data range, mean and standard deviation (SD) and coefficient of variation (CV).

Parameter Item Blanquilla Conference Blanquilla + Conference

C1 C12 V C2 C22 V C3 C32 V

SSC (◦ Brix) Number 168 146 22 164 142 22 332 288 44


Minimum 10.70 10.70 11.90 11.10 11.35 11.10 10.70 10.70 11.40
Maximum 17.70 17.70 15.40 21.95 21.95 20.30 21.95 21.95 17.70
Mean 13.89 13.89 13.90 15.13 15.15 15.03 14.50 14.57 14.08
SD 1.20 1.23 1.00 2.10 2.05 2.43 1.81 1.86 1.42
CV (%) 8.65 8.88 7.18 13.86 13.53 16.17 12.50 12.73 10.12

Firmness (N) Number 168 146 22 164 142 22 332 288 44


Minimum 2.37 2.37 4.47 3.16 3.16 3.97 2.37 2.37 3.19
Maximum 47.54 47.54 33.53 54.31 54.31 51.29 54.31 54.31 47.97
Mean 15.09 15.11 14.94 14.31 13.70 18.22 14.70 14.48 16.21
SD 9.71 9.92 8.36 13.90 13.35 16.89 11.95 11.47 14.81
CV (%) 64.34 65.65 55.96 97.17 97.43 92.71 81.30 79.19 91.36

cially valuable for predicting complex parameters such as shelf-life, variety or two: C1 (168 ‘Blanquilla’ pears for SSC and firmness), C2
which is governed by multifactorial changes (Camps et al., 2007) (164 ‘Conference’ pears for SSC and firmness) and C3 (332 pears for
and thus requires the information provided by the spectrum as a SSC and firmness).
whole. Once the two best instruments, signal pretreatments and spec-
Recent advances in NIR instrumentation include an increased tral ranges had been selected for each calibration set and analytical
use of diode-array equipment varying in wavelength resolution, parameter, new NIRS calibration models were developed and
detector type, and electronic stability. This has enabled NIRS to validated with an external set (not used in the calibration devel-
be used for on-site and on-line product testing, although much opment). After elimination of outlier spectra (i.e. wrong spectra or
research is still required before this can be achieved on industrial spectra belong to fruits with external damages), calibration and
level. validation samples were selected solely on the basis of spectral
Several published papers report on the use of NIRS technology data, following the method proposed by Shenk and Westerhaus
for the quality control of fruit and vegetable products, as discussed (1991), using the Center algorithm included in the WinISI II soft-
in a number of specific reviews (Nicolaï et al., 2007; Saranwong and ware package version 1.50 (Infrasoft International, Port Matilda, PA,
Kawano, 2007). The application of NIRS technology to the study of USA) (ISI, 2000). This algorithm performs a principal component
pears has focussed on the use of NIRS to predict internal damage in analysis, reducing the original spectral information (log 1/R values)
Chinese pears (Han et al., 2006), to identify varieties (Fu et al., 2007), to a small number of linearly independent variables, thus facilitat-
to predict physical/chemical parameters such as pectin constituents ing the calculation of spectral distances. These new variables were
in Japanese pears (Sirisomboon et al., 2007), sugar content, titrable used to calculate the centre of the spectral population and the dis-
acidity and pH (Ying and Liu, 2008) and soluble solid content and/or tance (expressed as the Mahalanobis distance) of each sample in
firmness (Fu et al., 2006; Liu et al., 2007; Tsai et al., 2007; Nicolaï the calibration set from that centre. Having ordered the sample set
et al., 2008). Few studies have sought to compare NIRS instruments by spectral distances (from smallest to greatest distance to the cen-
differing in terms of cost, optical design, and suitability for on-site tre), the 22 samples forming the validation set for the single-variety
use. No hitherto-published research has dealt with the prediction models were selected by taking one sample out of every seven in the
of shelf-life in pears, despite the commercial importance of this initial set. After this procedure, the new calibration and validation
parameter. sets thus comprised the samples shown in Table 1.
This study aimed to compare the performance of three commer-
cially available NIR instruments in providing models predicting SSC 2.3. Reference data
and firmness in pears, and subsequently to evaluate the ability of
NIRS technology to measure changes taking place in pears during Firmness was measured based on the resistance of the fruit flesh
postharvest storage, using discriminant analysis. to penetration by the puncture probe. The Magness–Taylor (M–T)
technique was used to measure firmness. Skin from opposite cheeks
2. Materials and methods of each fruit was removed with a knife for M–T measurements.
Firmness of all individual fruit was measured at two positions
2.1. Fruits around the equator, approximately 180◦ apart, and perpendicular to
the stem–calyx axis. Measurements were carried out with a Univer-
A total of 336 pears (Pyrus communis L.) cv. ‘Blanquilla’ and ‘Con- sal Testing Machine (Model 3343 Single Column, Instron, Norwood,
ference’, were received by our laboratories between March and MA, USA) with a 8 mm diameter probe mounted on a 1000 N load
April 2007; for each variety, 14 batches of 12 pears each were cell with a crosshead speed of 0.0016 m/s (100 mm/min) and pene-
received. Four ‘Conference’ pears were discarded for failing to meet trating the fruit to 10 mm. Flesh firmness was calculated as an aver-
established quality standards, leaving a total of 332 pears. age of two measurements per fruit and expressed in Newtons (N).
In order to determine pear storage time, batches were arranged Soluble solid content (SSC) were taken by removing two longitu-
on shelves at 20 ◦ C and 70% relative humidity. The product was dinal wedges (from stem end to calyx end) from each fruit, pressing
kept under these conditions throughout the trial period, with sam- through cheesecloth, and the SSC of the juice was measured with a
ples drawn for analysis at 6- and 8-day intervals. The raw material temperature-compensated refractometer (model ATC-1, Atago Co.,
sample (0 days’ storage) served as control. Tokyo, Japan).

2.2. Definition of calibration and validation sets 2.4. Spectra collection

Three different calibration sets were used to obtain NIRS predic- For collecting NIR spectra, three NIR-instruments – working in
tion models, depending on the use of sets that comprised an unique reflectance mode (log 1/R) and differing mainly in the wavelength
26 P. Paz et al. / Computers and Electronics in Agriculture 69 (2009) 24–32

Table 2
Basic technical features of three spectrophotometers: scanning monochromator (SM: FNS-6500), diode array and scanning monochromator (DASM: LabSpec® Pro A108310)
and diode array (DA: NIR-512).

Properties Instruments

SM: FNS-6500 DASM: LabSpec® Pro A108310 DA: NIR-512

Detector type Silicon, 400–1100 nm. Lead sulphide, Modular silicon array and two Peltier 512 element InGaAs. Linear array
1100–2500 nm cooled InGaAs detectors
Wavelength range (nm) 400–2500 350–2500 900–1700
Spectral bandpass 10 ± 1 nm in reflectance 3 nm at 700 nm. 10 nm at 3 nm
1400–2100 nm
Spectral data rate 1.8 scans/s 0.1 s 0.1–3 s
Dispersion Pre Post Post
Light source Full spectrum Full spectrum Full spectrum
Analysis mode Reflectance Reflectance Reflectance

range used and the measuring principle involved – were used. The partial least squares (MPLS) as regression method (Shenk and
main features of these instruments are shown in Table 2. Westerhaus, 1995a). Shenk and Westerhaus (1995a) indicated that
The FNS-6500 (FOSS NIRSystems, Silver Spring, MD, USA) scan- MPLS is often more stable and accurate than the standard PLS
ning monochromator (SM) was interfaced to a remote reflectance algorithm for agriculture applications. In MPLS, the NIR residuals
fiber optic probe (NR-6539-A) with a 43 × 43 mm window. In order at each wavelength, obtained after each factor is calculated, are
to shield the detector assembly, a dark perspex compartment standardized (divided by the standard deviations of the residuals
(340 mm × 238 mm × 222 mm) was used. Each fruit was placed in at a wavelength) before calculating the next factor. When devel-
the probe with the stem–calyx axis horizontal, so that the desired oping MPLS equations, cross-validation is recommended to select
fruit location was centred on, and in direct contact with, the probe. the optimal number of factors and to avoid overfitting (Shenk and
The optical absorption spectrum from 400 to 2500 nm in 2 nm steps Westerhaus, 1995a). For cross-validation, the calibration set was
was measured at two different locations. The first measurement partitioned in four groups; each group was then predicted using
was collected at a random location around the circumference, and a calibration developed on the other samples. Finally, validation
the next at approximately 180◦ rotation from the initial site. The errors were combined to obtain a standard error of cross-validation
two spectra were averaged to provide a mean spectrum for each (SECV).
pear. NIR spectra data were collected using the WinISI II software All multivariate regression equations were obtained using the
package version 1.50 (Infrasoft International, Port Matilda, PA, USA). standard normal variate and detrending methods for scatter correc-
The LabSpec® Pro A108310 (Analytical Spectral Device, ASD, Inc., tion (Barnes et al., 1989). Moreover, four derivative mathematical
Boulder, Colorado, USA), a diode array scanning monochroma- treatments were tested in the development of NIRS calibrations:
tor (DASM), spectrophotometer, was equipped with a fiber-optic 1,5,5,1; 2,5,5,1; 1,10,5,1 and 2,10,5,1 (the first digit is the order of the
contact probe, working in reflectance mode in the spectral range derivative, the second is the gap over which the derivative is calcu-
350–2500 nm, at 1 nm intervals. Pears were placed steadily upon lated, the third is the number of data points in a running average
the fruit holder, with the stem–calyx axis horizontal. On each pear, or smoothing, and the fourth is the second smoothing) (Shenk and
a reflectance spectrum was measured on two opposite, equatorial Westerhaus, 1995b).
positions and the averaged spectrum per fruit was used for analysis. For calibration of both quality indices a number of spectral
The NIR-512 (Ocean Optics, Inc., Dunedin, FL, USA) diode array regions were used. For the NIR-512 instrument all measurements
(DA), instrument was equipped with a 2 mm-diameter angled were made over 900–1700 nm. The other two instruments were
reflection probe, using a 30◦ window to remove specular reflec- each tried at 800–2200, 900–1700, and 1100–2200 nm.
tion effects, working in reflectance mode in the spectral range The statistics used to select the best equations were: standard
900–1700 nm. Spectra were captured using the same method as error of calibration (SEC), coefficient of determination for calibra-
for the LabSpec® Pro A108310. tion (R2 ), standard error of cross-validation (SECV), coefficient of
For the latter two instruments, the signal was captured using determination for cross-validation (r2 ), RPD or ratio of the standard
IndicoTM Pro (ASD, Inc., Boulder, Colorado, USA) and Spectra Suite deviation of the original data (SD) to SECV, and the CV (coefficient
packages (Ocean Optics, Inc., Dunedin, FL, USA), respectively, and of variation) or ratio of the SECV to the mean value of the reference
subsequently pretreated using the Unscrambler version 7.5 pro- data for the training set. This last statistic enables SECV to be stan-
gram (CAMO, ASA, Oslo, Norway). dardized, facilitating comparison of the results obtained with sets
of different means (Williams, 2001).
2.5. Chemometric data treatment The best-fitting equations, as selected by the above statisti-
cal criteria particularly taking into account the lowest SECV and
Spectra data were processing using the WinISI II software pack- the highest RPD, were subsequently evaluated by external valida-
age version 1.50 (Infrasoft International, Port Matilda, PA, USA) (ISI, tion, a procedure determining the predictive ability of an equation
2000). Before developing NIRS calibrations, the structure and spec- based on a sample set which has not been used in the calibration
tral variability of the sample population was determined using the procedures. This statistical process is based on the determination
Center algorithm included in the WinISI II software. This program of a known significant error, termed “bias”, and an unexplained
performs an initial principal component analysis (PCA) to calculate significant error, termed SEP(c) (standard error of performance,
the centre of the population and the distance of samples from that bias-corrected) (Windham et al., 1989).
centre in an n-dimensional space, using the Mahalanobis distance; Generally, for calibration groups comprising 100 or more sam-
samples with a statistical distance greater than 3 were considered ples, and validation groups containing nine or more samples,
outliers or anomalous spectra (Shenk and Westerhaus, 1991, 1995a). assuming a confidence level of 90%, the following control limits are
Quantitative calibrations were obtained for predicting soluble assumed (Shenk et al., 1989; Shenk and Westerhaus, 1996): limit
solid content (SSC) and firmness in intact pears, using modified control SEP(c) < 1.30 × SEC; limit control bias < ±0.60 × SEC.
P. Paz et al. / Computers and Electronics in Agriculture 69 (2009) 24–32 27

2.6. Discriminant models

Discriminant models were built, for each variety, to classify


pears in terms of postharvest shelf-storage time as an approach for
shelf-life estimation, using the PLS2 discriminant analysis (Naes et
al., 2002) included in the WINISI II version 1.50 software package.
PLS discriminant analysis uses a training set to develop a qualita-
tive prediction or calibration model which may subsequently be
applied for the classification of new unknown samples. This model
seeks to correlate spectral variations (X) with defined classes (Y),
attempting to maximize the covariance between the two types of
variable. In this type of approach, the Y variables used are not con-
tinuous, as they are in quantitative analysis, but rather categorical
“dummy” variables created by assigning different values to the dif-
ferent classes to be discriminated.
Three categories or classes were established for each cultivar:
day 0 or baseline and two storage intervals (6 and 8 days). For each of
the three storage times, 56 pears were taken for each variety, except
on day 8 of storage for the ‘Conference’ variety, when 52 pears were
analyzed, thus giving a total of 332 pears (168 ‘Blanquilla’ and 164
‘Conference’).
Two spectral regions were used for the development of the
classification models: 900–1700 nm for the three instruments
tested and 800–2200 nm for the FNS-6500 and LabSpec® Pro
A108310 instruments. Spectra were pre-treated with standard nor-
mal variate and detrending methods for scatter correction and
four derivative mathematical treatments (1,5,5,1; 2,5,5,1; 1,10,5,1;
2,10,5,1).
The following statistical parameters were used to evaluate the
predictive capacity of the models: the SECV and the percentage of
correctly classified samples.

3. Results and discussion

3.1. Calibration database composition

The main characteristics of the three calibration sets used, i.e. the
number of samples analyzed for each variety, together with mean,
Fig. 1. Changes in quality parameters during on-shelf storage. Vertical bars represent
range, standard deviation (SD) and variation coefficient (CV) for the
SD (n = 56 except for ‘Conference’ pears on day 8, n = 52).
reference values, are shown in Table 1.

Table 3
Calibration statistics for the equations obtained for the prediction of soluble solid content (SSC, ◦ Brix) for the different instruments and spectral ranges studied.

Instrument Set Spectral range (nm) Mathematical treatment Mean SD SEC R2 SECV r2 RPD CV (%)

FNS-6500 C1 800–2200 2,10,5,1 13.81 1.10 0.57 0.73 0.71 0.59 1.55 5.14
C1 900–1700 2,5,5,1 13.85 1.08 0.49 0.80 0.59 0.71 1.83 4.26b
C1 1100–2200 2,10,5,1 13.80 1.10 0.56 0.74 0.75 0.55 1.47 5.43
C2 800–2200 2,10,5,1 15.04 1.86 0.94 0.74 1.16 0.62 1.60 7.71
C2 900–1700 2,10,5,1 14.99 1.89 0.99 0.72 1.16 0.63 1.63 7.74a
C2 1100–2200 1,5,5,1 15.06 1.87 1.05 0.68 1.16 0.62 1.61 7.70
C3 800–2200 2,10,5,1 14.40 1.67 0.77 0.78 0.88 0.72 1.90 6.10
C3 900–1700 2,10,5,1 14.42 1.67 0.77 0.78 0.82 0.76 2.04 5.69b
C3 1100–2200 1,5,5,1 14.37 1.66 0.85 0.74 0.93 0.69 1.78 6.47

LabSpec® Pro A108310 C1 800–2200 1,10,5,1 13.90 1.14 0.54 0.78 0.74 0.58 1.54 5.32a
C1 900–1700 1,5,5,1 13.90 1.14 0.59 0.73 0.76 0.55 1.50 5.47
C1 1100–2200 1,5,5,1 13.90 1.13 0.61 0.71 0.82 0.48 1.38 5.90
C2 800–2200 1,10,5,1 14.90 1.99 1.09 0.70 1.21 0.63 1.64 8.12
C2 900–1700 1,10,5,1 14.90 1.99 1.02 0.74 1.13 0.68 1.76 7.58b
C2 1100–2200 1,10,5,1 14.85 1.92 1.04 0.70 1.20 0.61 1.60 8.08
C3 800–2200 1,5,5,1 14.33 1.59 0.77 0.77 1.00 0.61 1.59 6.98
C3 900–1700 1,10,5,1 14.43 1.62 0.74 0.79 0.89 0.70 1.82 6.17a
C3 1100–2200 1,10,5,1 14.39 1.70 1.05 0.61 1.17 0.52 1.45 8.13

NIR-512 C1 900–1700 1,5,5,1 13.92 1.20 0.77 0.59 0.94 0.39 1.28 6.75a
C2 900–1700 2,5,5,1 15.13 2.05 1.18 0.67 1.49 0.47 1.38 9.85a
C3 900–1700 1,5,5,1 14.41 1.61 1.09 0.54 1.21 0.44 1.33 8.40a

Mean: mean of the calibration set. SD: standard deviation. SEC: standard error of calibration. R2 : coefficient of determination of calibration. SECV: standard error of cross-
validation. R2 : coefficient of determination of cross-validation. RDP: SD/SECV. CV: coefficient of variation.
a
Best equation.
b
The best of the best equations for each calibration set.
28 P. Paz et al. / Computers and Electronics in Agriculture 69 (2009) 24–32

Table 4
Calibration statistics for the equations obtained for the prediction of firmness (N) for the different instruments and spectral ranges studied.

Instrument Set Spectral range (nm) Mathematical treatment Mean SD SEC R2 SECV r2 RPD CV (%)

FNS-6500 C1 800–2200 2,10,5,1 13.89 7.91 4.79 0.63 5.50 0.52 1.44 39.60a
C1 900–1700 2,10,5,1 14.03 8.32 5.33 0.59 5.76 0.52 1.44 41.05
C1 1100–2200 1,5,5,1 13.56 7.60 5.17 0.54 5.66 0.45 1.34 41.74
C2 800–2200 2,10,5,1 13.45 13.16 4.55 0.88 6.35 0.77 2.07 47.21a
C2 900–1700 1,5,5,1 12.22 11.79 6.21 0.72 6.63 0.68 1.78 54.26
C2 1100–2200 2,10,5,1 13.35 13.20 5.70 0.81 7.35 0.69 1.80 55.06
C3 800–2200 2,10,5,1 13.93 11.22 5.37 0.77 6.78 0.63 1.65 48.67a
C3 900–1700 2,5,5,1 13.49 10.60 5.58 0.72 6.53 0.62 1.62 48.41
C3 1100–2200 1,10,5,1 13.50 10.58 6.37 0.64 6.92 0.57 1.53 51.26

LabSpec® Pro A108310 C1 800–2200 1,5,5,1 14.46 9.72 4.45 0.79 5.99 0.62 1.62 41.42
C1 900–1700 1,5,5,1 14.20 9.08 3.49 0.85 5.33 0.65 1.70 37.54b
C1 1100–2200 1,5,5,1 14.61 9.57 4.27 0.80 5.93 0.61 1.61 40.59
C2 800–2200 1,5,5,1 14.55 14.17 3.27 0.95 6.45 0.79 2.20 44.33
C2 900–1700 1,10,5,1 15.38 14.90 4.66 0.90 6.81 0.79 2.19 44.28b
C2 1100–2200 1,5,5,1 14.66 14.44 4.65 0.90 7.36 0.74 1.96 50.20
C3 800–2200 1,5,5,1 14.67 12.12 4.50 0.86 6.75 0.69 1.80 46.01
C3 900–1700 1,10,5,1 15.16 12.71 5.39 0.82 6.80 0.72 1.87 44.85b
C3 1100–2200 1,10,5,1 14.40 11.86 5.36 0.80 6.89 0.67 1.72 47.85

NIR-512 C1 900–1700 2,5,5,1 14.74 9.34 3.88 0.83 5.67 0.63 1.65 38.28a
C2 900–1700 1,5,5,1 13.77 13.63 5.84 0.82 7.22 0.73 1.89 52.43a
C3 900–1700 1,10,5,1 13.72 10.98 6.09 0.69 6.85 0.61 1.60 49.93a

Mean: mean of the calibration set. SD: standard deviation. SEC: standard error of calibration. R2 : coefficient of determination of calibration. SECV: standard error of cross-
validation. r2 : coefficient of determination of cross-validation. RDP: SD/SECV. CV: coefficient of variation.
a
Best equation.
b
The best of the best equations for each calibration set.

It can be observed that set C2 (‘Conference’), displayed the great- important. This is interesting since the price of the NIR equipments
est variation for both parameters (for SSC, CV = 13.86% versus 8.65% is highly correlated with its spectral range. Specially, for the diode-
and 12.50% for sets C1 and C3, respectively; for firmness, CV = 97.17% array instruments working at wavelengths above 1700 nm is still
versus 64.34 and 81.30% for sets C1 and C3, respectively), even economically unviable.
though C2 contained a single variety whilst C3 contained both vari- The results displayed in Table 3 show differences in SSC
eties. prediction accuracy between all three instruments. Thus, the
Changes and evolution in SSC and firmness for ‘Blanquilla’ and FNS-6500 displayed the greatest predictive capacity (r2 = 0.76;
‘Conference’ pears during shelf-storage are shown in Fig. 1. It can SECV = 0.82 ◦ Brix; RPD = 2.04), with results similar to those pro-
be reported that no evolution is observed in SSC during the shelf- vided by the LabSpec® Pro A108310 (r2 = 0.70; SECV = 0.89 ◦ Brix;
storage time. Nevertheless, firmness declined markedly throughout RPD = 1.82), whilst the NIR 512 displayed the lowest predictive
postharvest storage. This decrease is observed for the two varieties capacity (r2 = 0.44; SECV = 1.21 ◦ Brix; RPD = 1.33). This may be due
of pear from day 0 to day 6, and also continues for the day 8. to the small probe diameter (2 mm) on the NIR-512, giving a smaller
At the start of measurements (day 0) mean firmness values, area of analysis.
around 25 N for ‘Blanquilla’ pears and around 31 N for ‘Conference’ Other authors (Tsai et al., 2007; Nicolaï et al., 2008; Ying and
pears, could be considered as “ready to buy” for the first variety and Liu, 2008) working with single-variety sets of Japanese pears, Euro-
“ripe” for ‘Conference’ (Gorny et al., 2000). During on shelf-storage, pean ‘Conference’ pears and Chinese pears obtained calibrations for
for ‘Blanquilla’ pears, firmness values fell to 11 N at day 6, lying in determining SSC with higher predictive capacity, with r2 ranging
the range considered as edible firmness or optimum (Murayama et from 0.60 to 0.85 and SEP values of between 0.40 and 0.62 ◦ Brix.
al., 1998) whilst at day 8 the firmness fell to 8 N. For ‘Conference’
pears, the loss of firmness was very pronounced, and on day 6, the 3.3. Firmness
reached firmness was lower than the optimum target, with values
around 5 N, being considered as too soft. The statistics for the best-fitting equations to determine firm-
ness in intact pears, for each spectral range assayed and for the
3.2. Soluble solid content three NIRS instruments and calibration sets used, are displayed in
Table 4.
The statistics for the best-fitting equations to predict SSC, Using the FNS-6500, the best models were obtained with
obtained with each calibration set tested and for the three NIR D2 log(1/R) and a spectral window of 800–2200 nm, although for
instruments studied are presented in Table 3. set C3 the differences between the statistics obtained at 800–2200
For the FNS-6500, the best calibration results were obtained and 900–1700 nm were minimal. Using the LabSpec® Pro A108310,
using D2 log(1/R) for the three calibration sets analyzed, whilst the best calibration equations were achieved with D1 log(1/R) and
using the LabSpec® Pro A108310, the best results were obtained a spectral window of 900–1700 nm. For the NIR-512, the best equa-
using D1 log(1/R) for all three sets. Using the NIR-512, the best tions were obtained with D2 log(1/R) in the ‘Blanquilla’ set, and with
results for the ‘Blanquilla’ set and the mixed-variety set were D1 log(1/R) in both the ‘Conference’ set and the mixed-variety set.
obtained with D1 log(1/R), whilst the best equations for the ‘Con- Slight differences in accuracy were noted for the firmness mod-
ference’ set were obtained using D2 log(1/R). els between all three instruments. The best calibration statistics
Regarding the spectral range, it seems that in general the best for the FNS-6500 were r2 = 0.77, SECV = 6.35 N, RPD = 2.07; for the
results were obtained using the region 900–1700 nm, except for LabSpec® Pro A108310 the best values were r2 = 0.79, SECV = 6.81 N,
set C1 on the LabSpec® Pro A108310. Nevertheless, in this case the RPD = 2.19; and for the NIR-512 r2 = 0.73, SECV = 7.22 N, RPD = 1.89. In
improvement of use the large range (800–2200 nm) is not very all three instruments, the best calibration statistics were recorded
P. Paz et al. / Computers and Electronics in Agriculture 69 (2009) 24–32 29

Table 5
Main calibration, validation statistics and control limits for predicting soluble solid content (◦ Brix) and firmness (N) in intact pears.

Parameter Instrument Set Calibration Prediction Control limits


2 2 2
R SEC r SECV r SEP SEP(c) Bias SEP(c) Bias

SSC (◦ Brix) FNS-6500 B 0.70 0.61 0.52 0.80 0.78 0.58 0.59 −0.08 0.79 ±0.37
C 0.50 1.37 0.45 1.42 0.62 1.08 1.08 0.20 1.78 ±0.82
B+C 0.78 0.79 0.73 0.87 0.74 0.83 0.83 0.08 1.03 ±0.47
LabSpec® Pro A108310 B 0.76 0.57 0.55 0.79 0.45 0.65 0.66 −0.11 0.74 ±0.34
C 0.77 0.91 0.66 1.13 0.75 1.10 1.12 0.07 1.18 ±0.55
B+C 0.81 0.76 0.73 0.90 0.72 0.74 0.75 0.03 0.99 ±0.46

Firmness (N) FNS-6500 B 0.49 6.55 0.42 7.02 0.48 5.64 5.78 −0.07 8.52 ±3.93
C 0.87 4.69 0.77 6.17 0.81 5.75 5.88 0.22 6.10 ±2.81
B+C 0.78 4.93 0.69 5.83 0.64 6.29 6.35 −0.34 6.41 ±2.96
LabSpec® Pro A108310 B 0.85 3.67 0.56 6.31 0.46 6.43 6.55 0.66 4.77 ±2.20
C 0.88 4.75 0.79 6.41 0.61 6.74 6.95 −0.21 6.18 ±2.85
B+C 0.82 5.01 0.72 6.28 0.80 6.01 6.00 −1.01 6.51 ±3.01

R2 : coefficient of determination of calibration. SEC: standard error of calibration. r2 : coefficient of determination of cross-validation. SECV: standard error of cross-validation.
r2 : coefficient of determination of prediction. SEP: standard error of prediction. SEP(c) : standard error of prediction bias-corrected. B: Blanquilla. C: Conference. B + C:
Blanquilla + Conference.

Fig. 2. Reference measured data versus NIRS predicted data for the validation set for SSC (◦ Brix) and firmness (N) using the best models obtained for the plurivarietal calibration
set.

for the single-variety ‘Conference’ set (C2), probably due to greater order to have an independent assessment of the equation accuracy
variability in firmness in this set (see Table 1). (Windham et al., 1989).
Generally speaking, calibrations displayed acceptable predictive The validation of the best calibration models obtained with
capacity for the physical firmness parameter; indeed, results were each calibration set were performed using sets comprising 22 pear
better than those reported by Nicolaï et al. (2008), who considered samples for the single-variety sets and 44 for the mixed-variety
the results obtained by NIRS unsatisfactory for modelling firmness set (Table 1). The validation set was obtained as described under
predictions. Section 2, and the remaining samples were used to generate new
calibration models (C12 for ‘Blanquilla’, C22 for ‘Conference’, C32 for
3.4. External validation ‘Blanquilla’ + ‘Conference’), using the optimum pretreatment and
spectral range selected with the results displayed in Tables 3 and 4.
The two instruments providing the best results (FNS-6500 and The cross-validation and validation statistics for the prediction
LabSpec® Pro A108310) were used for external validation, using of SSC and firmness, together with the control limits are presented
a sample set not included in the calibration (see Table 1), in in Table 5. It can be appreciated that the values obtained for SECV
30 P. Paz et al. / Computers and Electronics in Agriculture 69 (2009) 24–32

Fig. 3. D1 log(1/R) spectra for intact pear cv. ‘Blanquilla’ and ‘Conference’ during postharvest shelf-storage.

and SEP were very similar, confirming the suitability of the cross region, variations in absorption bands were attributable to differ-
validation method for estimating prediction statistics. According ences in skin and flesh colour between shelf-storage times; this
with the validation procedure outlined by Shenk et al. (1989), for was particularly true of bands at 570–590 and 630–690 nm, whilst
the two parameters studied, both the SEP(c) and the bias were in the NIR region, areas of variation were observed at around 950,
below confidence limits, thus ensuring accurate predictions, except 1450 and 1900 nm, corresponding to water-absorption bands, thus
for the SEP(c) for firmness in the two single-variety sets using the allowing shelf-storage times to be distinguished on the basis of
LabSpec® Pro A108310. Graphically, the NIRS predicted data versus differences in free water content due to changes occurring during
the reference values for the best models developed using the pluri- postharvest storage. Sugar-related peaks were observed at around
varietal calibration set (i.e. ‘Blanquilla’ + ‘Conference’) with the two 1100, 1390 and 2000 nm.
instruments selected, can be observed in Fig. 2. This work assesses the NIRS prediction of shelf-storage time
in pears as an approach for estimating the shelf-life of the prod-
3.5. Prediction of postharvest shelf-storage time uct, following the request of the company collaborator, which is
based on the demand of the supermarket chain that distributes
In order to determine the influence of variety during posthar- their fruits. This chain used to remove the product from the shelf
vest shelf-storage time, prediction models were developed using after 6 days, and they were interested in having an instantaneous
training sets for the two cultivars both separately and together. and non-destructive tool that would permit them to confirm the
Fig. 3 provides the typical D1 log(1/R) spectra for intact ‘Blan- shelf-storage time, to guarantee the shelf-life of the product to the
quilla’ and ‘Conference’ pears during shelf-storage. In the visible consumer.
P. Paz et al. / Computers and Electronics in Agriculture 69 (2009) 24–32 31

E: 900–1700 nm
The best classification results obtained for predicting pear shelf-
storage time (all instruments and training sets), using PSL2-DA, are
shown in Table 6. A high rate of classification of pear samples as a

D: 1,5,5,1
A: 92.1%
NIR-512

B: 0.28
function of storage time was noted: over 80% of samples were cor-

99.2%

70.0%
91.8%
C: 15
rectly classified in all cases. The three instruments proved suitable
for this application, although the LabSpec® Pro A108310 yielded the

LabSpec® Pro A108310


highest overall percentage of correctly classified samples—over 77%
for all sets, both for global and for partial classification.

E: 900–1700 nm
Although classification accuracy seemed to be good for the three
shelf-storage times analyzed, the highest percentage of correctly

D: 1,5,5,1
A: 93.7%
B: 0.27 classified samples was obtained on the first day of study, for all

96.2%
77.5%
97.4%
C: 15 instruments and sample sets. Models proved to be less able to
distinguish between samples by days 6 and 8 of storage, due to phys-
E: 900–1700 nm
iological changes taking place in the fruit, particularly the decrease
in firmness, as illustrated in Fig. 1.
Conference

D: 2,10,5,1
FNS-6500

A: 83.0%

Generally speaking, the best discriminant models were obtained


B: 0.33

46.9%
99.0%
76.2%
using D1 log(1/R) except for the ‘Blanquilla’ set on the NIR-512
C: 8

and the ‘Conference’ set on the FNS-6500, where the best dis-
criminant models were obtained using D2 log(1/R). In all cases, the
E: 900–1700 nm

best results were recorded in the wavelength range 900–1700 nm.


A: percentage of samples correctly classified training samples after cross validation. B: model SECV. C: number of factors. D: math treatment. E: spectral range.

This indicates that higher wavelengths do not significantly enhance


D: 2,5,5,1
A: 91.0%
NIR-512

B: 0.27

pear classification on the basis of changes taking place throughout


94.6%
92.7%
77.3%
C: 14

postharvest shelf-storage (20 ◦ C, 70 RH). Cheaper equipment can


therefore safely be used for this type of application.
LabSpec® Pro A108310

No published studies have addressed the use of NIR-


spectroscopy models for classifying pears as a function of
E: 900–1700 nm

postharvest shelf-storage time. The percentage of correctly classi-


fied samples in this study was higher to that reported by Camps et al.
D: 1,5,5,1
A: 94.0%
B: 0.25

(2007) in apples, that with mixed-variety sets obtained a 75–83% of


92.9%
80.0%
100%
C: 12

correctly classified fruits stored according to storage time; increas-


ing, as in the present work, until 98% when a single-variety set was
E: 900–1700 nm

used.
D: 1,10,5,1
Blanquilla

FNS-6500
Cultivars

A: 85.0%

4. Conclusions
B: 0.30

56.4%
83.1%
97.2%
C: 13

The results show that NIRS technology is an useful non-


destructive sensor for the assessment of intact pears by internal
E: 900–1700 nm

quality – soluble solids content, firmness and postharvest stor-


age time – allowing increased sampling and control of the whole
D: 1,5,5,1
A: 87.6%
NIR-512

B: 0.28

production. The results obtained suggest that the SM and DASM


64.3%
84.6%
97.3%
C: 15

provide similar results for physicochemical composition, being bet-


ter the SM for shelf-life prediction. The DA instrument provided less
LabSpec® Pro A108310

precision and accuracy in all cases.


E: 900–1700 nm

References
Percentage of intact pears correctly classified by storage interval. PLS-DA.

D: 1,5,5,1
A: 94.4%
B: 0.25

Barnes, R.J., Dhanoa, M.S., Lister, S.J., 1989. Standard normal variate transformation
98.7%

80.9%
95.1%
C: 14

and de-trending of near infrared diffuse reflectance spectra. Appl. Spectrosc. 43,
772–777.
Camps, C., Guillermin, P., Mauget, J.C., Bertrand, D., 2007. Discrimination of storage
E: 900–1700 nm

duration of apples stored in a cooled room and shelf-life by visible-near infrared


Global model

spectroscopy. J. Near Infrared Spectrosc. 15, 169–177.


D: 1,10,5,1

Han, D., Tu, R., Lu, C., Liu, C., Wen, Z., 2006. Nondestructive detection of brown core
FNS-6500

A: 81.1%

in the Chines pear ‘Yali’ by transmission visible-NIR spectroscopy. Food Control


B: 0.33

98.6%

52.4%
70.9%
C: 15

17, 604–608.
Fu, X.P., Ying, Y.B., Liu, Y.D., Lu, H.S., 2006. Detection of pears firmness using near
infrared diffuse reflectance spectroscopy. Spectrosc. Spect. Anal. 26, 1038–1041.
Qualitative groups: storage time (days)

Fu, X.P., Zhou, Y., Ying, Y., Lu, H., Xu, H., 2007. Discrimination of pear varieties using
three classification methods based on near-infrared spectroscopy. Trans. ASABE
50, 1355–1361.
Galvis-Sánchez, A.C., Fonseca, S.C., Morais, A.M.M.B., Malcata, F.X., 2004. Effects of
preharvest, harvest and postharvest factors on quality of pear (cv. ‘Rocha’) stored
under controlled atmosphere conditions. J. Food Eng. 64, 161–172.
Gorny, J.R., Cifuentes, R.A., Hess-Pierce, B., Kader, A.A., 2000. Quality changes in
fresh-cut pear slices as affected by cultivar, ripeness stage, fruit size, and storage
regime. J. Food Sci. 65, 541–544.
ISI, 2000. The Complete Software Solution Using a Single Screen for Routine Anal-
ysis, Robust Calibrations, and Networking. Manual FOSS NIRSystems/TECATOR.
Infrasoft International, Sylver Spring, MD.
Table 6

Liu, Y., Chen, X., Sun, X.D., Ying, Y.B., 2007. Non destructive measurement of pear
Day 6
Day 8
Day 0

internal quality indices by visible and near-infrared spectrometric techniques.


N. Z. J. Agric. Res. 50, 1051–1057.
32 P. Paz et al. / Computers and Electronics in Agriculture 69 (2009) 24–32

Murayama, H., Takahashi, T., Honda, R., Fukushima, T., 1998. Cell wall changes in pear Shenk, J.S., Westerhaus, M.O., Abrams, S.M., 1989. Protocol for NIR calibrations: mon-
fruit softening on and off the tree. Postharvest Biol. Technol. 14, 143–149. itoring analysis results and recalibration. In: Martens, G.C., Shenk, J.S., Barton Jr.,
Murayama, H., Sekine, D., Yamauchi, Y., Gao, M., Mitsuhashi, W., Toyomasu, T., 2006. F.E. (Eds.), Near Infrared Spectroscopy (NIRS): Analysis of Forage Quality Agri-
Effect of girdling above the abscission zone of fruit on ‘Bartlett’ pear ripening on culture Handbook No. 643. US Government Printing Office, Washington, DC, pp.
the tree. J. Exp. Bot. 57, 3679–3686. 104–110.
Naes, T., Isaksson, T., Fearn, T., Davis, A., 2002. A User-Friendly Guide to Multivariate Sirisomboon, P., Tanaka, M., Jujita, S., Kojima, T., 2007. Evaluation of pectin con-
Calibration and Classification. NIR Publications, Chichester, UK. stituents of Japanese pear by near infrared spectroscopy. J. Food Eng. 78, 701–
Nicolaï, B.M., Beullensa, K., Bobelyna, E., Peirsa, A., Saeysa, W., Theronb, K.I., Lam- 707.
mertyn, J., 2007. Non destructive measurement of fruit and vegetable quality by Thompson, J.F., Crisosto, C., 2002. Handling at destination markets. In: Kader, A.A.
means of NIR spectroscopy: a review. Postharvest Biol. Technol. 46, 99–118. (Ed.), Postharvest Technology of Horticultural Crops, 3rd ed. DANR Publication,
Nicolaï, B.M., Verlinden, B.E., Desmet, M., Saevels, S., Saeys, W., Theron, K., Cubeddu, pp. 271–277.
R., Pifferi, A., Torricelli, A., 2008. Time-resolved and continuous wave NIR Tsai, C.Y., Chen, H.J., Hsieh, J.F., Sheng, C.T., 2007. Fabrication of a near infrared online
reflectance spectroscopy to predict soluble solids content and firmness of pear. inspection system for pear fruit. Int. Agric. Eng. J. 16, 57–70.
Postharvest Biol. Technol. 47, 68–74. Valero, C., Crisosto, C.H., Slaughter, D., 2007. Relationship between non-destructive
Saranwong, S., Kawano, S., 2007. Applications to agricultural and marine products: firmness measurements and commercially important ripening fruit stages for
fruits and vegetables. In: Ozaki, Y., McClure, W.F., Christy, A.A. (Eds.), Near- peaches, nectarines and plums. Postharvest Biol. Technol. 44, 248–253.
Infrared Spectroscopy in Food Science and Technology. John Wiley & Sons, Inc., Williams, P.C., 2001. Implementation of near-infrared technology. In: Williams, P.C.,
New Jersey, pp. 219–242. Norris, K.H. (Eds.), Near-Infrared Technology in the Agricultural and Food Indus-
Shenk, J.S., Westerhaus, M.O., 1991. Population definition sample selection and cali- tries. AACC, Inc., St. Paul, Minnesota, pp. 145–171.
bration procedures for near infrared spectra and modified partial least squares Windham, W.R., Mertens, D.R., Barton Jr., F.E., 1989. Protocol for NIRS calibration
regression. Crop Sci. 31, 469–474. sample selection and equation development and validation. In: Martens, G.C.,
Shenk, J.S., Westerhaus, M.O., 1995a. Analysis of Agriculture and Food Products by Shenk, J.S., Barton III, F.E. (Eds.), Near Infrared Spectroscopy (NIRS): Analysis of
Near Infrared Reflectance Spectroscopy. Monograph NIRSystems. Forage Quality Agriculture Handbook No. 643. US Government Printing Office,
Shenk, J.S., Westerhaus, M.O., 1995b. Routine Operation, Calibration, Development Washington, DC, pp. 96–103.
and Network System Management Manual. NIRSystem Inc., Silver Spring, MD. Ying, Y., Liu, Y., 2008. Non destructive measurement of internal quality in pear using
Shenk, J.S., Westerhaus, M.O., 1996. Calibration the ISI way. In: Davies, A.M.C., genetic algorithms and FT-NIR spectroscopy. J. Food Eng. 84, 206–213.
Williams, P.C. (Eds.), Near Infrared Spectroscopy: The Future Waves. NIR Pub-
lications, Chichester, UK, pp. 198–202.

You might also like