|

HORTSCIENCE 46(1):5763. 2011.

Assessment of Yellow-fleshed Kiwifruit

(Actinidia chinensis Hort16A) Quality
in Pre- and Post-harvest Conditions
Using a Portable Near-infrared
Spectrometer
Jinquan Feng1, Andrew V. McGlone, Mike Currie, Chris J. Clark,
and Bob R. Jordan
The New Zealand Institute for Plant & Food Research Limited, Private Bag
92169, Auckland 1142, New Zealand
Additional index words. dry matter, soluble solids, flesh hue, FANTEC FQA-NIRGUN, field
measurement, temperature, ambient light, model transfer
Abstract. Dry matter content (DM), soluble solids content (SSC), and flesh hue (FH) are
important quality attributes of yellow-fleshed kiwifruit (Actinidia chinensis Hort16A).
Rapid non-destructive measurement tools enable effective assessment of quality attributes in the field or in packhouses and offer the potential for selective harvesting and
differential postharvest treatments. This study investigates a few practical issues for
use of one such tool, a portable near-infrared spectrometer, the FANTEC FQANIRGUN. Experimental results indicated that the precision of the NIRGUN in predicting
the DM, SSC and FH of Hort16A kiwifruit, given as the root mean square error of
calibration (RMSEC), is 0.6%, 0.9% and 1.48, respectively, and is similar to that reported
for standard laboratory benchtop spectrometers. These levels of precision are good
enough to distinguish differences in fruit quality between fruit from different vines. However, this is only possible when measurements are carried out under similar conditions
either within a short period of time in the field when fruit temperature is stable or in
a temperature-controlled laboratory. The calibration developed on one instrument could
be transferred to another instrument, but bias correction is necessary after the
calibration transfer. Hence, the absolute accuracy of predictions from the NIRGUN is
limited by significant and variable bias offset issues, particularly for field measurements
when fruit are on the vine.
Hort16A (Actinidia chinensis Planch.
var. chinensis) is a yellow-fleshed kiwifruit
cultivar that is marketed globally as ZESPRI
GOLD kiwifruit. SSC at harvest is an indicator
of fruit maturity, and DM content reflects the
potential SSC at eating ripe, i.e., when starch
reserves in the berry have been converted to

Received for publication 17 June 2010. Accepted

for publication 13 Oct. 2010.
This study was funded by ZESPRI International
Limited.
We thank Alistair Mowat of ZESPRI Innovation
Ltd. and Shirley Miller of Plant & Food Research
for their efforts to initiate and support this project.
We also thank Daniel Mathieson of ZESPRI Group
Ltd. and Tim Torr of Apata Ltd. for arranging
access to the NIRGUNs; AgFirst BOP Ltd. and
Linda Boyd of Plant & Food Research for sharing
fruit and data; Yuko Maeda of FANTEC Research
Institute for technical advice; and Dr. Nihal DeSilva
of Plant & Food Research for assistance in statistical
data analysis. We also thank Alison Duffy and Philip
Martin of Plant & Food Research for assistance with
destructive measurements.
Use of trade names does not imply an endorsement
of the products named or criticism of similar ones
not named.
1
To whom reprint requests should be addressed;
e-mail Ringo.Feng@plantandfood.co.nz.

HORTSCIENCE VOL. 46(1) JANUARY 2011

sugars during fruit softening. FH is an important quality parameter for Hort16A, because
its yellow color is a primary factor in the
appeal of the fruit to the consumer (Martin
and Luxton, 2005). Flesh color changes from
green to yellow as the fruit matures. However,
there is still a considerable degree of variability present in the properties of the fruit at
harvest (Schaare and Fraser, 2000).
Fruit variability at harvest leads to problems in post-harvest handling. For example,
1 of 2 weeks of conditioning at 5 C is used
for fruit with FH values above 105. This
enables the flesh degreening process to occur
but can result in excessive softening of fruit
with FH below 103. A line of fruit may have
a mix of FH. Thus, identification of fruit with
lower FH (which are generally associated with
more advanced maturity) in the orchard,
coupled with staged selective harvesting of
the mature fruit, would lead to reduced variation in quality and overall improvement in
the quality of the harvested crop. In addition,
the ready identification of poorly performing
or slowly maturing vines would enable targeted remedial measures to be undertaken.
Near-infrared (NIR) spectroscopic tools
suitable for orchard use would enable fast,
non-destructive, and effective monitoring of
fruit maturity on-vine. The color and taste

POSTHARVEST BIOLOGY AND TECHNOLOGY

attributes of Hort16A are currently determined in the New Zealand kiwifruit industry
by destructive measurement of FH, DM, and
SSC. Laboratory-based NIR spectroscopy on
samples of Hort16A kiwifruit post-harvest
has demonstrated that it is possible to measure these attributes with prediction errors of
0.5%, 0.9%, and 1.1 for DM, SSC, and FH,
respectively (Clark et al., 2004). Portable NIR
instruments have been developed and used to
assess maturity and quality indices of many
fruits (Aoki et al., 2010; Costa et al., 2002;
Golding et al., 2006; Khuriyati et al., 2004;
Kusumiyati et al., 2008; Saranwong et al.,
2003a, 2003b; Zude et al., 2008). However,
the performance of these instruments under
different temperature and light conditions has
not been studied rigorously for in situ measurements of Hort16A kiwifruit in the field.
The study reported here evaluated the
performance of a commercially available
FQA-NIRGUN released by the FANTEC
Research Institute (Shizuoka, Japan) with
an attempt to determine possible influences
of temperature and light conditions. The
possibility of transferring predictive models
developed on one NIRGUN to another NIRGUN was also tested.
Materials and Methods
The instrument
The FQA NIRGUN (FANTEC Research
Institute) is a gun-shaped NIR spectrophotometer weighing less than 1 kg (Fig. 1) that
can be easily operated in both the field and
laboratory. This instrument was chosen because it was commercially available back in
2005 when this study started and it fulfilled
the needs for outdoor use with a small-size,
lightweight, rechargeable battery, weatherresistant, and easy data handling (Saranwong
et al., 2003a, 2003b). The head of the NIRGUN
contains apertures for the light source (a small
halogen lamp) and detector and is arranged so
that an interactance measurement is made on
the fruit. A spectral measurement, made by
pressing a trigger button in the handle, consists
of a series of light intensity measurements
made at 2-nm intervals between 588 nm and
1092 nm (256 data points per spectrum). Up to
three fruit quality attributes can be predicted
simultaneously from the spectrum using onboard models, and results are displayed immediately on the NIRGUNs LCD screen. With a
fully charged battery, the NIRGUN can make
2000 measurements when operating in the
standalone mode. The on-board memory is
able to store 5000 predicted values or spectra
and predicted values for 500 measurements,
which can be downloaded to a computer for
offline analysis.
Measurement procedure
The NIRGUN was placed on the fruit in
such a way that the light source shone onto
the blossom side of the fruit with the detector
facing the middle (Fig. 1). In this position, the
curvature of the fruit formed a good fit with
the geometry of the NIRGUN head. Care was
taken to make measurements on the flattest

57

Fig. 1. Measurement of Hort16A kiwifruit using

a FANTEC FQA-NIRGUN.

side of the fruit to maximize the light-sealing

capability of the external rubber seal around
the head. The NIRGUN was held on the fruit
with moderate pressure to ensure that the internal rubber ring that surrounds the detector
aperture was firmly against the fruit surface,
thus preventing source light directly entering
the detector without having first passed through
the fruit. This measurement procedure also
proved to be easy and practical for orchard
use while fruit were still on the vine.
An integration time of 30 ms was selected
based on preliminary measurement of a range
of typical Hort16A kiwifruit so that the spectral peak height occupied 50% of the full
range (as recommended by FANTEC). It takes
2 s to complete a measurement.
Referencing
Referencing is a process that monitors
instrument performance and ensures that it
continues to operate in a stable mode. A reference spectrum taken on the supplied white
reference tile (specific to each instrument) is
preinstalled in memory, and it is not necessary to do referencing after every measurement. However, referencing needs to be
redone after replacing the halogen lamp or
if the instrument has not been used for more
than 90 d. The integration time for referencing was fixed at 15 ms and is independent of
the integration time set for fruit measurement.
There is no information about the dark reference
in the instrument manual. However, the same
baseline and white reference spectrum appear
when the spectrum from fruit is displayed using the vendor-supplied software. We can
download absorbance and first and second derivative data, but not the raw data. It is possible that a dark reference is taken every time
when the white reference is being renewed.
Development of calibration models
Calibrations were generated using the proprietary software Ca_Maker supplied with
the NIRGUN. By default, calibration was performed using the second derivative of the absorbance spectra generated with what appears
to be a Norris Derivative scheme (Hopkins,
2001) with segment, gap, and smoothing sizes
of 14, 0, and 20 nm, respectively. Segment is

58

the wavelength range used to apply second

differential calculus on the spectrum. Gap is
the distance between the last wavelength in
one segment and the first wavelength in the
next segment. Smoothing size is the wavelength range used for smoothing (Ca_Maker
uses moving averages for smoothing). These
settings can be altered to suit particular applications using more advanced calibration software called Professional Ca_Maker available
from FANTEC by special request. The software used multiple linear regressions for calibration modeling, searching for the best
combination of up to five wavelengths over the
truncated spectral range (700 to 950 nm by
default but can be altered when using Professional Ca_Maker). Calibrations were judged
using the standard measures of RMSEC and
R2, the coefficient of determination.
Three different data subsets of Hort16A
fruit were collected for preliminary investigations, which were then combined to make
a full calibration data set covering the expected variation resulting from differences in
temperature and orchard location.
The first calibration subset consisted of
300 fruit harvested from the Te Puke Plant &
Food Research Orchard in the Bay of Plenty
(North Island), New Zealands major kiwifruit production region. The fruit were equilibrated to 20 C in a laboratory before NIR
spectra were collected. This was followed by
destructive assessments of DM, SSC, and FH
using standard industry techniques as outlined
in Schaare and Fraser (2000) and McGlone
et al. (2002). Briefly, FH was measured using
a Minolta Chroma Meter (CR-300, D65 light
source, calibrated with a white calibration
plate; Minolta, Japan) after the removal of
skin and flesh to a depth of 2 mm. Two color
measurements were made on each fruit at two
sides at 90 to each other along the equator of
the fruit. The two readings measured from
each fruit were averaged to give one value for
each fruit. Soluble solids content was measured
using a digital refractometer (Atago, Tokyo,
Japan; 0% to 20%) with juice squeezed from
the blossom and stem ends separately, and the
two readings were averaged to give one value
for each fruit. DM was measured using a slice
3 mm in thickness (including skin, flesh,
seeds, and core tissues) cut along the equator of
each fruit. The slice was weighed immediately
after cutting and then after dried to constant
weight (typically 24 h at 65 C).
Two NIRGUN measurements were taken
on opposite sides of each fruit, giving a total
of 600 spectra. All 600 spectra were used for
initial calibration with each destructive measurement being matched to the corresponding
pair of spectra from the same fruit.
The second calibration set consisted of
a subsample of 90 fruit in addition to the 300fruit data calibration subset described previously. These fruit were randomly assigned
to two different batches of 45 fruit each. The
first batch was warmed to 20 C for the first
NIRGUN measurement, then cooled to 15 C
and re-measured, and finally cooled to 8 C
and remeasured a third time. These repeated
measurements were made on the same two

spots on each fruit. The second batch was first

cooled to 8 C and measured before warming
to 15 C and finally to 20 C for repeat measurements. NIRGUN measurements were carried out once stable fruit temperatures had
been reached after 3-h equilibration at each
temperature (Feng et al., 2002). In all cases,
the NIRGUN was maintained at the same
temperature as the fruit themselves to replicate
conditions that would be expected for field
measurements. After the NIR measurements,
these fruit were then equilibrated to 20 C for
destructive assessments of DM, SSC, and FH.
The third calibration set consisted of 45-fruit
samples from six grower lines from three different production regions within New Zealand:
Nelson, Bay of Plenty, and Kerikeri. On arrival,
moisture and dirt on fruit surfaces were removed by hand wearing a cotton glove. Fruit
were then packed into single-layer kiwifruit
trays and held at 20 C to allow surfaces to dry
completely before NIRGUN measurements
were undertaken. This data set was designed
to test whether orchard location had a significant effect on NIR prediction and to allow the
calibration model to account for variation
between orchards.
The master calibration was developed using
a combined data set containing 990 observations, in which 180 observations were from Set
1, 360 observations from Set 2 (NIR measurement at 8, 15, and 20 C), and 450 observations
from Set 3. Some data from each of the three
subsets had to be eliminated from the combined master, Because the calibration software
supplied with the NIRGUN only allows a maximum of 1000 observations. The selection of
observations was arbitrary based on measurement condition and fruit variability to maximize the coverage.
Performance of the master calibration
in on-vine fruit measurement
The master calibration based on the combined master calibration data set was loaded
onto the NIRGUN and used to measure fruit on
vines in the field. Ten fruit per vine were
selected from each of nine kiwifruit vines in
the Plant & Food Research Te Puke orchard.
The measurements were made on a sunny fall
day during the main harvest season in 2005
(210 d after full bloom). The temperature,
relative humidity, and radiation during the day
were recorded at a weather station in the orchard.
Each fruit was measured three times during the
day, while it remained on the vine, using the
NIRGUN. Fruit were subsequently harvested
after sunset and the NIR measurements were
repeated again in a laboratory at 20 C before
destructive measurement of DM, SSC, and FH.
Performance of the master calibration
in a subsequent year using another
NIRGUN
Encouraging results from experiments in
2005 led to the purchase of six FANTEC
FQA-NIRGUN units by New Zealand packhouses in 2006. In 2007, the performance of
the predictive models established in 2005
was tested under a range of temperature and
light conditions with one of these new units.
HORTSCIENCE VOL. 46(1) JANUARY 2011

A random sample of 60 fruit was harvested

from girdled vines and 40 fruit from control
vines during the main harvest season (4 May
2007). Fruit from each treatment were randomly assigned into four groups of 25 fruit
(15 fruit from girdled vines and 10 fruit from
control vines) to test the effect of temperature
and light conditions.
Measurement at different combinations
of fruit and instrument temperatures. Fruit of
the first group were placed in storage at 10 C,
whereas those of the second group were placed
in an air-conditioned laboratory at 20 C.
NIRGUN measurements were carried out twice.
The first measurement was made when the instrument temperature was 20 C; the
NIRGUN (at 20 C) was taken to where the
fruit were kept and the measurements were
completed within 3 min. The second measurement was made when the instrument temperature was at 10 C; again the NIRGUN
(at 10 C) was taken to where the fruit were
kept and the measurements were completed
within 3 min. On each measurement occasion,
a single spectrum was taken on orthogonal
sides of the fruit at (i.e., the flat side and curved
side) and predicted values were averaged to
give a single reading for each fruit. Fruit kept at
10 C were equilibrated to 20 C before destructive measurement of DM, SSC, and FH.
This produced a matrix of two instrument
temperatures and two fruit temperatures to
investigate the effect of temperature on measured values.
Measurement at different ambient light
conditions. Fruit from the remaining two 25fruit groups were stored at 0 C for 2 d and then
equilibrated for 5 h at 20 C before NIR spectra
were obtained. NIRGUN measurements were
taken twice before destructive measurement
of DM, SSC, and FH, once under weak ambient
light (scattered light from windows when all
the lights in the laboratory were turned off)
and again under strong ambient light (800 mm
under two 58-W Philips daylight fluorescent
tubes; Philips, Holland). This was to test if
ambient light intensity had a significant effect
on predicted values.

delivered only moderately performing models

for DM and FH and poor models for SSC,
possibly because of the fairly narrow range of
parameter distributions involved (Table 1).
The initial calibration was then applied to
the other two calibration data sets to investi-

gate the effect of fruit temperature and orchard

location on NIR measurements.
A strong temperature effect was readily
apparent in the second calibration data set,
particularly for SSC predictions. NIR-predicted values for DM and SSC increased as

Table 1. Statistics of the initial calibration based on 600 spectra measured at 20 C on 300 Hort16A
kiwifruit harvested from one orchard (initial calibration experiment) and the master calibration based
on 990 spectra measured at different temperatures on fruit harvested from different orchards
(combined data set from three calibration experiments).
Initial calibration
Master calibration
Mean SD
RMSEC
R2
Mean SD
Attribute
RMSEC
R2
Dry matter content (%)
0.67
0.79
16.0 1.50
0.60
0.85
16.4 1.52
Soluble solids content (%)
0.91
0.53
8.1 1.33
0.90
0.90
10.5 2.88
Flesh hue ()
1.38
0.77
104.6 2.93
1.39
0.86
105.1 3.67
RMSEC = root mean square error of calibration, R2 = coefficient of determination.

Fig. 2. Effect of temperature on NIRGUN measurements of Hort16A kiwifruit dry matter content (DM),
soluble solids content (SSC), and flesh hue using the initial calibration established based on the first
calibration data set measured at 20 C on 300 Hort16A kiwifruit harvested from one orchard (see
Tables 1 and 2 for details). Fruit were measured as they changed from low to high temperature. The
NIRGUN was maintained at the same temperature as the fruit themselves. Similar results for the
reverse change have been omitted for clarity.

Data analysis
The relationship between NIRGUN predictions and destructively measured values
was analyzed using regression and general
linear model procedures (REG, MENS, and
GLM) with the software package SAS Version 8.2 (SAS Institute Inc., Cary, NC). Predictions were judged on the basis of the
standard statistics of root mean square error
of prediction (RMSEP), bias (the average
difference between actual and predicted
values),
the SE of prediction (SEP, equivalent
q

to RMSEP2  BIAS 2 ), and R2. The two

predicted values from repeated measurements on each fruit were averaged to give
a single predicted value before these analyses
were undertaken.
Results
Calibration models. The initial calibration subset (N = 600 spectra from 300 fruit)
HORTSCIENCE VOL. 46(1) JANUARY 2011

Fig. 3. Second derivative (D2) spectra of the same 90 Hort16A kiwifruit measured at three different
temperatures (both the instrument and fruit were equilibrated to a specified temperature to replicate
conditions that would be expected for field measurements). The difference between D2 values
measured at 20 and 8 C (D220 C D28 C) was plotted at 10 times scale for clarity.

59

temperature decreased. On the other hand,

NIR-predicted values for FH decreased as temperature decreased (Fig. 2). In case of DM and
SSC, there was almost no bias between destructively measured and NIRGUN-predicted
values at 20 C for either parameter. However,
there was a significant bias for 20 C FH determination. NIR-predicted FH values at 20 C
were 3 C lower than destructively measured
values. This might have been caused by a
systematic error in the destructively measured
FH values. Destructively measured FH value
could be affected by thickness of the removed
skin and inaccurate calibration of the Minolta
colorimeter. Fortunately, the relative difference between fruit was not changed, i.e., bias
of NIRGUN-predicted values did not appear to
be dependent on measured values.
Comparison between NIR spectra taken at
three temperature conditions indicated that
temperature has a significant influence in the
second derivative spectra, particularly at the
chlorophyll absorption region of 680 nm
and the water absorption region 950 nm
(Fig. 3).
Bias resulting from orchard location was
not as marked as the temperature effect (Fig.
4).
The final master calibration results (Table
1) based on combined data from all three
subsets were encouragingly good with high
R2 values (0.85, 0.9, and 0.86 for DM, SSC,
and FH, respectively) and reasonable RMSEC
values (0.60, 0.9, and 1.39 for DM, SSC, and
FH, respectively) compared with the statistics
produced for Hort16A fruit using a benchtop
NIR unit operating under standardized conditions (Clark et al., 2004; Schaare and Fraser,
2000).
The wavelengths and coefficients to calculate fruit attributes from second derivative
spectra are listed in Table 2. The coefficients
and wavelengths used in the master calibration are different from those used in the initial
calibration. In general, the wavelengths used
in the master calibration had smaller temperature impact factors (difference between second derivative spectra measured at 20 and
8 C on the same 90 fruit expressed as a percentage relative to the average spectra at the
two temperatures). For example, temperature
impact factors were on average over 50%
smaller for the master calibrations and always the initial calibrations had the wavelengths with the single largest temperature
impact factor (Table 2). Therefore, it is perhaps not surprising that the master calibration
is less sensitive to temperature changes.
Performance of the master calibration
with on-vine fruit measurement in the field.
Environmental conditions in the orchard
showed the expected diurnal variation in
temperature, relative humidity, and radiation
flux (Fig. 5). The NIRGUN predictions for
the 90 fruit from nine vines were well
correlated with destructively measured quality attributes (R2 0.7 to 0.9). However, the
predictions also showed a similar diurnal
variation with bias values varying with time
and generally peaking at the 1400 HR measurement (Table 3).

60

Fig. 4. Effect of orchard location on NIRGUN measurement of dry matter content (DM), soluble solids
content (SSC), and flesh hue (FH) of Hort16A kiwifruit using the initial calibration established based
on the first calibration data set (see Tables 1 and 2 for details). The measurement was carried out at
20 C with the same temperature for both fruit and the NIRGUN. Data from two orchards are presented,
whereas those from the other four orchards have been omitted for clarity. Solid lines represent 1:1 lines.

Table 2. Model coefficients of the initial calibration based on 600 spectra measured at 20 C on 300
Hort16A kiwifruit harvested from one orchard (initial calibration experiment) and the master
calibration based on 990 spectra measured at different temperatures on fruit harvested from different
orchards (combined data set from three calibration experiments).z
Attribute
Dry matter content (%)

Soluble solids content (%)

Initial calibration
Wavelength (nm)
Coefficient
Intercept
20.6
774.1
820 (33.3)y
836 (118.9)
222.5
864 (21.1)
2242.4
900 (5.7)
1033.5
932 (2184.7)
343.5
Intercept
700 (0.8)
824 (87.5)
860 (21.7)
888 (14.6)
916 (2.3)

15.4
85.3
850.3
1290.7
1621.5
589.0

Master calibration
Wavelength (nm)
Coefficient
Intercept
32.3
720 (2.4)
85.7
752 (2.9)
160.8
828 (827.8)
424.0
872 (20.1)
2124.8
896 (7.8)
1129.6
Intercept
700 (0.8)
816 (18.1)
868 (20.6)
892 (10.6)
916 (2.3)

25.0
144.8
852.0
1826.2
1758.7
867.0

Flesh hue ()

Intercept
57.4
Intercept
75.1
708 (1.6)
550.4
732 (7.2)
1046.3
768 (67.7)
761.4
752 (2.9)
2107.9
860 (21.7)
2511.0
872 (20.1)
2766.9
888 (14.6)
2738.8
888 (14.6)
2170.1
916 (2.3)
1087.4
948 (8.1)
247.6
z
Both calibrations were based on the second derivative of the absorbance spectra.
y
The number in brackets after each wavelength is a temperature impact factor (TIF) calculated as a
percentage change in second derivative (D2) spectra based on data from 90 fruit measured at 20 C and
8 C (TIF = 100(D220 C D28 C)/[(D220 C + D28 C)/2].

For the purpose of detecting relative

differences between vines, means of DM,
SSC, and FH measurements of the nine vines
predicted from the NIRGUN ranked almost
identically to the means based on destructive
measurement regardless of diurnal time (Figs.
6, 7, and 8). Analysis of variance between fruit
from different vines based on values predicted
from the NIRGUN resulted in similar separation of means compared with that based on
destructive measurement because the least
significant differences between means were
in similar scales.
Model performance in a subsequent year
at different temperature conditions. The master calibration developed in 2005 based on
combined data sets from three calibration

experiments (Table 1) performed well in 2007

when the calibration was loaded onto another
NIRGUN. NIRGUN measurement at 20 C
explained over 95% of the variance in measured DM, over 82% of the variance in
measured SSC, and over 85% of the variance
in measured FH. The predictions at 10 C
were similar to that at 20 C except for SSC,
in which predicted values explained slightly
less (72%) than the variance in measured
SSC. SEPs were 0.31% to 0.40% DM, 0.10%
to 0.18% SSC, and 0.39 to 0.86 hue (Fig. 9).
Significant bias existed for all the predictions. The bias for DM was significantly
affected by instrument temperature but was
unaffected by fruit temperature. The prediction bias was 0.59% DM when the NIRGUN
HORTSCIENCE VOL. 46(1) JANUARY 2011

Fig. 5. Temperature, relative humidity, and radiation from the Hort16A kiwifruit orchard on the day that
field measurements using the NIRGUN were carried out.
Table 3. Effect of diurnal course on predicting dry matter content, soluble solids content, and flesh hue of
Hort16A kiwifruit (n = 90) using a FANTEC FQA-NIRGUN in the field with the master calibration
established based on combined data sets (see Tables 1 and 2 for details).
SEPy
Biasx
Slopew
R2v
Time/Place
RMSEPz
1000 HR/field
0.59
0.44
0.39
0.79*
0.83
1.26
0.39
1.20
0.78*
0.87
1400 HR/field
0.76
0.36
0.66
0.82*
0.89
1800 HR/field
2000 HR/laboratory
0.39
0.39
0.03
0.82*
0.87
1.28
1.13
0.61
0.57*
0.72
1000 HR/field
1.42
1.07
0.94
0.59*
0.75
1400 HR/field
1.33
1.08
0.77
0.62*
0.73
1800 HR/field
2000 HR/laboratory
1.12
1.08
0.26
0.61*
0.73
3.64
1.61
3.26
1.06
0.77
1000 HR/field
3.62
1.53
3.28
1.05
0.79
1400 HR/field
1800 HR/field
3.80
1.50
3.49
1.08
0.80
2.09
1.42
1.53
1.06
0.82
2000 HR/laboratory
z
Root mean square error of prediction.
y
SE of prediction.
x
Average difference between actual and predicted values (measured value-predicted value).
w
Slopes of the regression lines with asterisk are significantly different from 1 at P < 0.05 significance level.
v
Coefficient of determination.

Attribute
Dry matter content
(%)
Mean = 16.4%
SD = 1.1%
Soluble solids content
(%)
Mean = 10.5%
SD = 2.0%
Flesh hue ()
Mean = 99.3
SD =2.8

Fig. 8. Comparison between Hort16A kiwifruit

average flesh hue (FH) measured destructively
and those measurements predicted from NIRGUN assessments in the field or laboratory at
different times during a sunny day. The master
calibration was established based on the combined data set (see Methods) measured using
the same NIRGUN. Each point represents the
mean of 10 fruit from a vine; solid line
represents a perfect match. Vertical bar represents least significant difference (LSD) to separate destructively measured means. Horizontal
bars represent LSDs to separate NIRGUN predicted means for the specified time.

was at 10 C compared with the bias of 0.31%

DM when the NIRGUN was at 20 C. Conversely, the biases for SSC and FH were significantly affected by fruit temperature but not
instrument temperature (Fig. 9). The magnitudes of temperature effects shown in Figure 9
were much smaller than those shown in Figure
2, indicating the superior performance of the
master calibration model based on the combined calibration data set compared with the
first calibration model based on a single data
set (Tables 1 and 2).
Model performance in a subsequent year
at different ambient light conditions. Predicted values for DM, SSC, and FH under
strong ambient light were very close to those
predicted under weak ambient light (Fig. 10).
Figure 10 also indicated that repeated NIRGUN
measurements on the same fruit gave very
consistent results (R2 > 0.95).
Discussion

Fig. 6. Comparison between average Hort16A

kiwifruit dry matter contents (DM) measured
destructively and those predicted from NIRGUN
assessments in the field or laboratory at different
times during a sunny day. The master calibration
was established based on the combined data
set (see Methods) measured using the same
NIRGUN. Each point represents the mean of 10
fruit from a vine; solid line represents a perfect
match. Vertical bar represents least significant difference (LSD) to separate destructively measured
means. Horizontal bars represent LSDs to separate
NIRGUN predicted means for the specified time.

HORTSCIENCE VOL. 46(1) JANUARY 2011

Fig. 7. Comparison between average Hort16A

kiwifruit soluble solids contents (SSC) measured
destructively and those predicted from NIRGUN
assessments in the field or laboratory at different
times during a sunny day. The master calibration
was established based on the combined data
set (see Methods) measured using the same
NIRGUN. Each point represents the mean of 10
fruit from a vine; solid line represents a perfect
match. Vertical bar represents least significant
difference (LSD) to separate destructively measured
means. Horizontal bars represent LSDs to separate
NIRGUN predicted means for the specified time.

On-vine fruit measurement. On-vine orchard experiments demonstrated that DM,

SSC, and FH can be predicted with the same
level of precision in the field as in the
laboratory. The analysis of variance revealed
very similar rankings of means for the nine
vines used in the experiment irrespective of
whether NIRGUN predictions or destructive
measurements were used. However, there
was a significant prediction bias associated
with many of the NIRGUN results, particularly with what looked like diurnal variation.
This means that the accuracy of the NIRGUN
prediction in the orchard depends on the time
of day at which the measurement is made.
Despite that variation, differences between
vines were clearly measurable and consistent
in a relative sense at any one time and would
enable the ready differentiation of vines in
terms of fruit maturity/quality.

61

Fig. 9. Effect of instrument and fruit temperature on prediction of Hort16A kiwifruit dry matter, soluble
solids content, and flesh hue in 2007 using the master calibration established in 2005 based on the
combined calibration data set measured using another NIRGUN.

Fig. 10. Effect of lighting on prediction of Hort16A kiwifruit dry matter (DM), soluble solids content
(SSC), and flesh hue in 2007 using the master calibration established in 2005 based on the combined
data set measured using another NIRGUN. Weak ambient light means near-infrared (NIR)
measurement was made under scattered light from windows when all the lights in the laboratory
were turned off. Strong ambient light means NIR measurement was made under two 58-W Philips
Daylight fluorescent tubes (800 mm distance between fruit and the lights).

The bias might not be a problem, particularly if small subsets of fruit could be
efficiently collected and destructively measured. This subset of measurements could
then be used to calculate an appropriate
adjustment to make the mean NIRGUN predictions match the destructive measurements.
Whereas such a strategy could work well for
SSC and FH, in which destructive measurements can be made rapidly, it would seem

62

unlikely that it could work as effectively for

DM, in which fruit samples must be dried for
7 h or longer. Thus, the presence of bias in
DM prediction represents the most serious
limitation to the use of the NIRGUN if the
accuracy of the predictions, rather than
merely the precision, is important.
We consider there are two possible factors
causing the bias. Either the instrument is not
particularly stable against ambient environ-

mental variations or the fruit themselves are

not. Certainly the instrument proved very
stable at each of the three set laboratory
temperatures to which it was equilibrated.
The study in 2007 at a combination of two
instrument temperatures and two fruit temperatures indicated that instrument temperature could have a significant influence in DM
prediction, although the absolute difference
between DM predicted with the instrument at
10 or 20 C was minor (0.27% DM, Fig. 9A
B). Instrument temperature did not appear to
have any significant influence on predicted
SSC and FH (Fig. 9CF). This suggests that
the instrument is at least reasonably stable
against variation in temperature for parameters of interest to the kiwifruit industry. The
study in 2007 also indicated that the effect of
fruit temperature on predicted values was
greatly reduced when using the predictive
models established in 2005 based on a combined data set collected at a range of temperature and fruit conditions. The bias of DM
prediction for 10 C fruit was not significantly different from that for 20 C fruit. The
bias difference between 10 and 20 C fruit
was statistically significant for predicted SSC
and FH, but the difference was only 0.2%
SSC and 1.4 hue.
The dependence of prediction bias on
diurnal time suggests that bias variation is
related to environmental factors affecting the
fruit such as sunlight flux and temperature.
Changes in sunlight flux during the day were
not considered likely as a major cause of bias
offsets, because all the measurements were
made in a mature orchard with heavy and
established canopies with little chance of
sampled fruit being exposed to direct sunlight. This assumption was further confirmed
by the identical prediction results from measurements on the same fruit at different
ambient light conditions (Fig. 10). Direct
temperature fluctuations also seemed unlikely to be the major factor causing prediction bias, because the calibration data
set already included data collected at different temperatures and the bias difference
between fruit at 10 and 20 C in 2007 (Fig.
9) was much smaller than the difference
between diurnal times (Figs. 6, 7, and 8).
We can speculate further that another
cause for bias could be fruit water loss during
the day. Kiwifruit vines may experience water stress around noon and early afternoon.
The high temperature and low relative humidity observed during this time (Fig. 5) may
have caused excessive water loss from the
fruit. Although fruit water loss is unlikely to
result in a 1.2% increase in DM or a 0.9%
increase in SSC on a whole-fruit basis, such
increases in DM and SSC could be possible
within the localized skin tissue regions from
where NIRGUN measurements were made.
Predicted FH values in the field measurement
were over 1.7 higher than that predicted in
laboratory conditions at 2000 HR and 3
higher than the values measured destructively. The difference among the three field
measurements was relatively small (less than
0.3). The large prediction bias of the field
HORTSCIENCE VOL. 46(1) JANUARY 2011

measurements may be related to excitation of

chlorophyll during the day and the fact that
the destructive color measurement is taken at
a depth of 2 mm or more, whereas the NIR
includes information from more superficial
regions. Certainly, the superficial color of
Hort16A kiwifruit is much greener than the
interior coloration and can have a substantive
influence on color measurement (B. Jordan
and M. Loeffen, personal communication).
Moisture on fruit surfaces may also be contributing. The thickness of fruit tissue, which
is effectively measured using the NIRGUN,
might be reduced as the fruit surface dries out.
Comment on flesh color measurement.
Analysis of pigments across different species
and genotypes of genus Actinidia indicated
that the yellow flesh color of Hort16A
kiwifruit is mainly the result of the absence
of chlorophylls in the fruit rather than a significant increase in the carotenoid concentration (McGhie and Ainge 2002; Montefiori
et al., 2009). Therefore, it was expected that
chlorophyll absorption region 680 nm
would be important for developing calibration for flesh color. However, extending
wavelength range down to 650 nm did not
result in better calibration for FH compared
with the calibration based on the default
wavelength range (data not shown). Therefore,
the master calibration was based on the default
wavelength range of 700 to 950 nm (general
users do not have access to the Professional
Ca_Maker for wavelength extension).
The master calibration for FH used five
wavelengths ranging from 732 to 948 nm
(Table 2), indicating the correlation between
FH and carbohydrates. Indeed, FH was found
to be negatively correlated with dry matter
content (R2 = 0.56 based on the combined
calibration data set to develop the master
calibration). This may be attributed to the
reflecting and scattering effect of starch
granules that account for up to 50% of the
dry matter during the harvest season.
Model transfer between instruments in
different years. The calibration models developed in 2005 using one NIRGUN were
able to predict DM, SSC, and FH of fruit
harvested in 2007 using another NIRGUN
with SEP of 0.27% to 0.39% DM, 0.60% to
0.77% SSC, and 0.68 to 1.00 hue, respectively (Fig. 9). Predicted values from the
NIRGUN at 20 C correlated with measured
values with R2 of 0.95 for DM, 0.82 for SSC,
and 0.85 for FH (Fig. 9). The prediction bias
observed in 2007 (Fig. 9) was not substantively
different from that observed in 2005 under laboratory conditions (2000 HRh/laboratory results in Table 3). These model transfer results
are similar to those achieved with a benchtop
NIR instrument (Clark and McGlone, unpublished data). This is very encouraging because

HORTSCIENCE VOL. 46(1) JANUARY 2011

it indicates that the calibration models can be

transferred to other instruments for measurement in subsequent years. Usually, the calibration model would need to be updated to
cope with the variability between fruit produced from different years. Such a training
process requires large data sets that cover a
wider range of variability in both fruit and
measurement conditions. The calibration software currently available with the NIRGUN can
only hold 1000 observations and there is a
maximum of five wavelengths that can be
selected for calibration. This limits the ability
of operators to use robust calibration models.
For the NIRGUN to provide robust measurements with a wide variety of orchards and environmental conditions, improvements to the
calibration software are needed so that these
limitations can be overcome.
Calibration transfer between the NIRGUNs changed the regression slope as well
as the bias (Fig. 9). This means that corrections cannot be made simply by just applying
a fixed offset. This problem could be resolved
using sophisticated calibration procedures for
more accurate prediction. However, the model
and instrument as they stand now would be
useful to detect relative differences between
batches of Hort16A kiwifruit where measurements are made under similar conditions,
either in the field when fruit temperature is
stable or in a temperature-controlled laboratory situation.
Conclusion
The precision of the NIRGUN in predicting DM, SSC, and hue in Hort16A kiwifruit
is similar to the level reported for standard
laboratory benchtop instruments. The NIRGUN measurements can be used to distinguish fruit from different vines. For this
purpose, then, the predictive models developed using one NIRGUN could be transferred to another unit for measurements in
a subsequent year. However, this is only
possible in a relative sense in which all fruit
measurements are made under the same
measurement conditions, either within a short
period of time in the field or in a laboratory
under set conditions. This is because the
accuracy of NIRGUN predictions is affected
by significant and variable bias offsets, especially for field measurements made while
kiwifruit are on the vine. Bias corrections
are possible and feasible for some quality
attributes by taking small subsets of fruit for
destructive measurement.
Literature Cited
Aoki, H., T. Tanabe, and T. Akinaga. 2010. Pre and
postharvest data collecting system using the

compact NIR analyzer. Acta Hort. 864:299

304.
Clark, C.J., V.A. McGlone, H.N. De Silva, M.A.
Manning, J. Burdon, and A.D. Mowat. 2004.
Prediction of storage disorders of kiwifruit
(Actinidia chinensis) based on visible-NIR
spectral characteristics at harvest. Postharvest
Biol. Technol. 32:147158.
Costa, G., M. Noferini, G. Fiori, and O. Miserocchi. 2002. Determination of indications
of harvest and fruit quality in kiwifruit. Rivista
di Frutticoltura e di Ortofloricoltura 64:2023.
Feng, J., K.M. Maguire, and B.R. MacKay. 2002.
Effects of package and equilibration time on
physiochemical attributes of Hayward kiwifruit. Acta Hort. 599:149155.
Golding, J.B., S. Satyan, C. Liebenberg, K. Walsh,
and W.B. McGlasson. 2006. Application of
portable NIR for measuring soluble solids concentrations in peaches. Acta Hort. 713:461464.
Hopkins, D.W. 2001. What is a Norris derivative?
NIR News 12:35.
Khuriyati, N., T. Matsuoka, and S. Kawano. 2004.
Precise near infrared spectral acquisition of
intact tomatoes in interactance mode. J. Near
Infrared Spectrosc. 12:391396.
Kusumiyati, T., M. Akinaga, Tanaka, and S.
Kawasaki. 2008. On-tree and after-harvesting
evaluation of firmness, color and lycopene
content of tomato fruit using portable NIR
spectroscopy. J. Food Agr. Environ. 6:327
332.
Martin, R.A. and P. Luxton. 2005. The successful
commercialization of ZESPRI gold kiwifruit.
Acta Hort. 694:3540.
McGhie, T.K. and G.D. Ainge. 2002. Color in fruit
of the genus Actinidia: Carotenoid and chlorophyll compositions. J. Agr. Food Chem. 50:117
121.
McGlone, V.A., R.B. Jordan, R. Seelye, and P.J.
Martinsen. 2002. Comparing density and NIR
methods for measurement of kiwifruit dry
matter and soluble solids content. Postharvest
Biol. Technol. 6:191198.
Montefiori, M., T.K. McGhie, I.C. Hallett, and G.
Costa. 2009. Changes in pigments and plastid
ultrastructure during ripening of green-fleshed
and yellow-fleshed kiwifruit. Sci. Hort. 119:377
387.
Saranwong, S., J. Sornsrivichai, and S. Kawano.
2003a. Performance of a portable NIR instrument for Brix value determination of intact
mango fruit. J. Near Infrared Spectrosc. 11:175
181.
Saranwong, S., J. Sornsrivichai, and S. Kawano.
2003b. On-tree evaluation of harvesting quality of mango fruit using a hand-held NIR
instrument. J. Near Infrared Spectrosc. 11:283
294.
Schaare, P.N. and D.G. Fraser. 2000. Comparison
of reflectance, interactance and transmission
modes of visible-near infrared spectroscopy for
measuring internal properties of kiwifruit (Actinidia chinensis). Postharvest Biol. Technol. 20:
175184.
Zude, M., M. Pflanz, C. Kaprielian, and B.L.
Aivazian. 2008. NIRS as a tool for precision
horticulture in the citrus industry. Biosystems
Eng. 99:455459.

63