G Model
POSTEC-9318; No. of Pages 8
ARTICLE IN PRESS
Postharvest Biology and Technology xxx (2011) xxx–xxx
Contents lists available at ScienceDirect
Postharvest Biology and Technology
journal homepage: www.elsevier.com/locate/postharvbio
Assessment of sugar and starch in intact banana and mango fruit
by SWNIR spectroscopy
P.P. Subedi ∗ , K.B. Walsh
Plant Sciences Group, Building 7, Central Queensland University, Rockhampton 4702, Australia
a r t i c l e
i n f o
Article history:
Received 28 March 2011
Accepted 19 June 2011
Keywords:
Fruit quality
Maturity index
Near infrared spectroscopy
Non-invasive
Ripening
a b s t r a c t
The prediction accuracy of models based on visible-short wavelength near infrared spectra (VIS–SWNIR;
500–1050 nm) collected from intact fruit using a partial transmittance optical geometry was considered
for dry matter (DM) and total soluble solids (TSS) content of mesocarp tissue of banana (Musa acuminata,
cv. Robusta) and mango (Mangifera indica, cv. Keitt) fruit. The DM content was modelled well across all
2
stages of maturity for mango, with a cross validation correlation coefficient of determination (Rcv
) > 0.75
and root mean square error of cross-validation (RMSECV) of <0.70% DM. However, the performance of
the banana mesocarp DM model was relatively poor, presumably due to the thickness of the peel. For
2
mango, TSS content was modelled well only in ripened fruit (typical Rcv
> 0.75, RMSECV < 0.60%), and
was predicted poorly across ripening stages (Rp2 < 0.75). This result is consistent with an inability to
discriminate between starch and soluble sugars when using spectra of intact fruit. Better results were
achieved for banana pulp TSS, however, this outcome is interpreted as an indirect assessment, with
mesocarp TSS content highly correlated (R2 > 0.85) with skin colour (Hunter a and a/b) in the populations
assessed. VIS–SWNIR is recommended for assessment of the ripening stage of mango and banana fruit
and for assessment of DM in intact mango, but not banana fruit. The technique is also not recommended
for assessment of TSS content across ripening stages of banana or mango fruit.
© 2011 Elsevier B.V. All rights reserved.
1. Introduction
Mango and banana fruit accumulate carbohydrate reserves
dominated by starch during maturation. A minimum level of carbohydrate accumulation is associated with fruit maturation (the
point at which the fruit is physiologically capable of independent
ripening). During fruit ripening the starch reserve is converted to
soluble sugars, while dry matter (DM) content, as an index of total
carbohydrate content, remains relatively unchanged. Final eating
quality is linked to total soluble solids (TSS) at the fully ripe stage.
Potential eating quality (i.e. final TSS) can be indexed by DM content
at any stage after physiological maturity. The relative proportion of
TSS to starch or DM content is thus an index of ripening and can be
an index of fruit maturity. Banana fruit has a final TSS of between 15
and 25%, while mango fruit has a final TSS of 14 and 20%, depending
on variety and growing conditions.
Banana and mango growers in Australia typically determine
fruit maturity based on subjective assessment of fruit shape and/or
estimates of harvest time as time from flower initiation or fruit set.
However, these attributes are useful only with certain varieties, and
∗ Corresponding author. Tel.: +61 7 49232149; fax: +61 7 49306536.
E-mail address: [email protected] (P.P. Subedi).
under consistent growing conditions. A non-invasive, field portable
system capable of assessment of fruit DM thus has potential use in
gauging time of harvest.
For both banana and mango fruit, the progress of ripening is
commercially assessed through correlation with skin colour, or
by measurement of fruit firmness (Ramaswamy and Tung, 1989).
However, certain ripening conditions (e.g. high temperature) can
result in an uncoupling of the rate of pulp ripening and skin
colour change (Marriott, 1980). For mango, growing conditions (e.g.
high nitrogen conditions) can influence skin colour change during
ripening. A technique for non-invasive evaluation of mesocarp TSS
and starch content would be an useful tool for monitoring the mesocarp ripening process.
Near infrared spectroscopy (based on the wavelength region
1100–2500 nm) is routinely used for carbohydrate analysis of
ground samples. Shortwave near infrared spectroscopy (SWNIRS;
700–1100 nm) is used for the assessment of DM content and/or TSS
content of intact, thin skinned fruit because of lower absorptivity
by water in this range, allowing assessment of a greater volume
of the sample. The effective optical sampling depth when using
reflectance geometry with intact fruit is typically 2–4 mm within
the 700–900 nm range (Lammertyn et al., 2000), and 10–30 mm
for the interactance geometry used in the current study (Walsh
et al., 2000). Despite some claims, application is limited to analysis
0925-5214/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.postharvbio.2011.06.014
Please cite this article in press as: Subedi, P.P., Walsh, K.B., Assessment of sugar and starch in intact banana and mango fruit by SWNIR
spectroscopy. Postharvest Biol. Technol. (2011), doi:10.1016/j.postharvbio.2011.06.014
G Model
POSTEC-9318; No. of Pages 8
2
ARTICLE IN PRESS
P.P. Subedi, K.B. Walsh / Postharvest Biology and Technology xxx (2011) xxx–xxx
of major constituents (such as DM, but not titratable acidity,
McGlone et al., 2003a,b) and to thin skinned fruit (Walsh et al.,
2004).
The use of partial least square (PLS) or multiple linear regression (MLR) models based on SWNIR spectra to assess DM content
of intact mango and TSS content of ripe intact mango has been
reported by a number of researchers (Guthrie and Walsh, 1997;
Saranwong et al., 2004; Subedi et al., 2006; Jha et al., 2010),
but the issue of prediction of TSS content of ripening mango
fruit (in which starch content is changing) has not been well
considered.
For intact banana, Zude (2003) reported excellent calibration
model statistics for the concentration of individual soluble sugars
in mesocarp tissue for a PLS model based on diffuse reflectance
spectroscopy of intact fruit using the wavelength window of
950–1700 nm (e.g. for glucose: cross-validation R2 = 0.96, root
mean square error of cross-validation, RMSECV = 1.8; for sucrose:
R2 = 0.94, RMSECV = 0.9 g/kg on a dry weight basis). It was concluded that “such accurate results indicated that NIR spectra
responded directly to pulp sugars, since a measurement based on
the sensor response to carbohydrates in the fruit skin would have
caused a low correlation coefficient for sucrose” (as skin and pulp
sucrose levels were poorly correlated). These results were very
encouraging; however, the degree of correlation between individual sugars, or between individual sugars and other attributes (e.g.
peel colour) was not reported. Thus it is not clear whether the sugars were assessed ‘directly’ in the models developed, or indirectly
through the relationship of an individual sugar to another attribute
(e.g. skin pigment or light scattering properties). Further, no testing of the model in terms of prediction of independent populations
was attempted. We postulated that this assessment might be over
optimistic, involving no prediction of independent test populations.
Further, the high R2 values achieved by Zude (2003) will reflect
the high range in sugar levels between unripe and ripe fruit, while
the reported RMSECV values (e.g. 2.7 g/kg for fructose) are similar to the standard deviation (SD) of these attributes at any given
ripening stage. In other words, the model may be predicting ripening stage rather than pulp sugar concentration per se. Given the
likely depth of optical sampling is restricted to the skin for these
reflectance spectra, a possible scenario is that the model represents
a measure of peel moisture content, which will vary with ripening
stage.
Apple and kiwifruit offers a parallel case to mango and banana,
as these fruit also contains starch at physiological maturity, converting to sugar during ripening. McGlone et al. (2003a,b) reported
accurate prediction of DM content of ‘Royal Gala’ apples using
SWNIR based models at any stage of fruit ripeness (correlation
coefficient of determination of prediction (Rp2 ) = 0.95; root mean
square error of prediction (RMSEP) < 0.32%; on SD > 1.3%). However, prediction of TSS was poor before full ripeness (i.e. in the
presence of starch; R2 = 0.79; RMSEP = 0.51%; on SD = 1.1%; cf. fully
ripe fruit; R2 = 0.94; RMSEP = 0.32%; on SD = 1.2%). It was noted
that prediction of DM presumably relies on spectral information
due to water and carbohydrate bands, while prediction of TSS in
unripe fruit presumably requires segregation of the carbohydrate
bands to distinguish the soluble and insoluble forms of carbohydrate. Thus SWNIR may be useful for determination of the total
carbohydrate content of intact fruit, but may be unsuitable for differentiating sugar from starch – i.e. the technique may be unsuited
to the assessment of TSS in fruit in which starch levels are changing, as in ripening mango or banana fruit. McGlone et al. (2007)
extended this consideration to kiwifruit, noting harvest time TSS
models were ‘indifferent to the spectral window’, and thus did not
represent a direct prediction of sugar, but rather were related to
maturity changes, such as a relationship between texture and light
scattering.
The objective of the present study was to extend the consideration of McGlone et al. (2003a,b, 2007) on the ability to distinguish
soluble and insoluble forms of carbohydrate using SWNIRS, with
attention given to ripening mango and banana fruit. These two fruit
crops were chosen as relatively thin and thick skinned fruit, respectively; with skin thickness an issue in the application of an optical
technique such as SWNIRS.
2. Materials and methods
2.1. Fruit samples and reference methods
Mango (Mangifera indica cv. Keitt) fruit (n = 140 fruit) were harvested at the hard green stage, sprayed with 0.2% (w/v) ethephon
480 (A&C Rural Pty. Ltd., Queensland, Australia) within 12 h of harvest, and allowed to ripen at 25 ◦ C. At intervals of approximately
1 d, a randomly selected set of fruit (n = 40) were sampled for DM,
TSS and firmness, with the remainder allowed to continue ripening
at 25 ◦ C until fruit reached a senescent stage (over ripe). Banana
(Musa acuminata, cv. Robusta) fruit were sourced from three farms
in the Atherton Tablelands region of North Queensland. Fruit were
destructively sampled (n = 51 fingers at each of four maturity stages,
viz. green, ripening, ripe and over ripe) following spectral acquisition.
A destructive measurement of fruit firmness was made using
an Effegi penetrometer employed with a HortPlus (HKP Software
Ltd., Hasting, New Zealand) mechanical drive stand. Penetrometer
readings were made on the cut surface of transverse slices (1 cm
thick) of banana fruit, and on the cheek of whole mangoes, after
removal of the skin using a fruit peeler. The force (N) required to
plunge a 6 mm diameter probe 8 mm into the fruit tissue at a speed
of 20 mm min−1 was recorded. Measurements were made at the
position of spectral acquisition.
Dry matter was determined by weight loss of a mesocarp sample of 5–10 g fresh weight (FW) following 48 h drying in a forced
air oven (65 ± 2 ◦ C). To determine mango TSS, juice was extracted
from a 27 mm diameter, 10 mm deep core of the fruit mesocarp
using a garlic press. For banana, TSS content of both pulp and peel
was determined. A 30 g FW sample of tissue was blended with 90 ml
of distilled water using a stomacher (Lab-Blender 80, A.J. Steward,
UAC House, London, UK). Juice TSS was determined using a temperature compensated refractometer (Bellingham and Stanley RFM
320, Longfield Road, Kent, UK), corrected for sample dilution.
Both peel and internal flesh colour of banana and internal
flesh colour of mango was measured with a Chromameter CR-400
(Minolta, Osaka, Japan) as Hunter ‘L’, ‘a’ and ‘b’ values. Hunter ‘L’ is a
measure of luminosity, while ‘a’ and ‘b’ values index the red–green
and yellow–blue space, respectively. In the case of mango, each side
of each fruit was cut longitudinally midway between the seed and
skin, and then recut to take a 1 cm thick slice. A colorimeter reading
was taken of the skin and the cut flesh of each fruit sample, from
the location of spectra acquisition.
2.2. Spectroscopy
Spectra of mango and banana fruit were collected over the wavelength range 300–1150 nm, using an interactance optical geometry,
a silicon diode detector and a 100 W tungsten halogen light source
configured in the interactance optical configuration reported by
Greensill and Walsh (2000) (0◦ angle between illumination and
detected light rays, with detection probe viewing a shadow cast
by the probe onto the fruit). This unit was a prototype form of that
manufactured for on-line fruit grading applications, viz. ‘Insight’,
Colour Vision Systems, Bacchus Marsh, Australia, and for hand-
Please cite this article in press as: Subedi, P.P., Walsh, K.B., Assessment of sugar and starch in intact banana and mango fruit by SWNIR
spectroscopy. Postharvest Biol. Technol. (2011), doi:10.1016/j.postharvbio.2011.06.014
G Model
ARTICLE IN PRESS
POSTEC-9318; No. of Pages 8
P.P. Subedi, K.B. Walsh / Postharvest Biology and Technology xxx (2011) xxx–xxx
held application, viz. ‘Nirvana’, Integrated Spectronics, Sydney,
Australia.
Integration time was varied with sample type (typically
25–35 ms), with signal level maintained below saturation but above
66% of detector saturation value. Referencing was undertaken
within each experimental run, using a Teflon tile and a black stop.
Spectra were collected from two sides of each intact mango fruit,
at the equator of the fruit, and from two positions along each intact
banana fruit. In a separate exercise, spectra were first acquired of
intact green, part yellow and yellow banana fruit, and then of the
same area of peel removed from the fruit, against a Teflon tile (1 cm
thick) and against a dark background.
Raw count data were converted into absorbance data using dark
and white reference spectra. Second derivative of absorbance was
calculated using a Savitzky Golay second order fit over 25 nm (i.e.
9 data points).
2.3. Chemometrics
The chemometrics software package, The Unscrambler v. 7.6 and
9.1 (Camo, Oslo, Norway), was used for partial least squares regression (PLSR) calibration model development (prediction of a given
attribute from spectral data). Model performance was evaluated in
2 , and in terms of prediction of samples not
terms of RMSECV and Rcv
included in the calibration set (RMSEP and Rp2 ). Cross-validation was
performed using 20 segments with random selection of samples.
The number of PLS factors used in model development was
determined by the comparison of root mean square error of calibration (RMSEC) and RMSECV, with The Unscrambler chemometric
software suggested number of factors adopted. The optimal wavelength range for the PLSR model was established using a MatLab
based script by creating models varying in wavelength start and
stop wavelength at 3 nm increments, creating a ‘map’ of model performance based on optimal number of factors, RMSEC and RMSECV
by start and end wavelength, as described by Guthrie et al. (2005).
Spectra with a statistic (Shenk and Westerhaus, 1993) of (Mahalanobis distance)2 /f > 6, where f is the number of factors in the
model, were defined as outliers and excluded from the calibration
model.
3
3. Results and discussion
3.1. Mango
As expected, mango DM content, as an index of total carbohydrate content, was constant following harvest, over the ripening
period (at around 19%), while TSS increased (from 5 to 16%)
(Tables 1 and 2). Interactance spectra (720–920 nm) of intact
mango were used for PLSR models for fruit DM and TSS (following
Subedi et al., 2006). Dry matter content was modelled reasonably well at any given stage of ripening of mango fruit and in
2 typicombined data sets, across ripening stages (Table 1) (Rcv
cally > 0.8; RMSECV < 0.6%). Dry matter was also predicted well
across ripening stages (i.e. prediction of population at a ripening
stage not included in the development of the model; Rp2 > 0.85,
Table 1).
Total soluble solids were modelled well using spectra of a popu2 < 0.85, RMSECV = 0.49%),
lation of ripened mango fruit (day 8, Rcv
2 < 0.75; Table 2). This result was partly
but not for unripe fruit (Rcv
explained by the low range (and SD) of TSS in unripe fruit (day
2 was depressed. However, for populations
0 and 1), such that Rcv
harvested on days 3, 6, 7 and 8, when mean TSS of the fruit was
2 (and
increasing but the SD of TSS was relatively constant, Rcv
the ratio of SD to RMSECV) increased with each ripening stage
(Table 2). This result is consistent with a negative impact of changing starch levels on the performance of the TSS model. Thus it is not
surprising that a TSS model (720–920 nm) created on later ripening stages performed poorly in prediction of early ripening stages
(Rp2 = 0.01), and improved in prediction of late ripening stages, in
which starch to sugar conversion was presumably complete (e.g.
the day 7 data set, predicted using a model containing the day 8
data set; Rp2 = 0.8).
The plots of PLS regression coefficients for DM and TSS models were smooth (Fig. 1), indicative that the models were not
over-fitted. The SWNIRS spectra of carbohydrates should possess
features relevant to O–H and C–H vibration; however, interpretation of a regression coefficient plot in terms of such spectral features
is difficult, in that the PLS model may weight area of the peak
shoulders.
Table 1
Mango: dry matter model calibration and prediction statistics. Partial least squares regression models were based on interactance spectra over the wavelength range
720–920 nm of intact mango fruit. Populations were drawn from a set of ripening fruit, with fruit hard green at day 0 and fully ripe by day 8. The wavelength region
(720–920 nm) was chosen using an optimisation procedure.
Calibration
Population
# Samples
# Factors
Mean
SD
2
Rcv
RMSECV
Day 0
Day 1
Day 3
Day 6
Day 7
Day 8
Day 0–8
Day 0–6 + 8
Day 0–1, 4–8
Day 1–8
40
40
40
40
40
40
240
200
200
200
5
4
5
5
4
5
6
5
6
5
19.31
19.52
19.01
18.73
17.72
19.33
19.36
19.30
19.01
19.36
0.87
0.80
1.61
1.06
1.59
1.32
1.28
1.22
1.61
1.31
0.81
0.72
0.94
0.81
0.81
0.92
0.88
0.76
0.76
0.88
0.39
0.39
0.38
0.43
0.68
0.38
0.52
0.56
0.59
0.63
Prediction
Calibration sets
Prediction sets
Rp2
RMSEP
SDR
Bias
Day 0–6 + 8
Day 0–1, 4–8
Day 1–8
Day 7
Day 3
Day 0
0.88
0.94
0.88
0.85
0.78
0.90
1.87
2.06
0.97
0.43
0.07
0.80
2
Rcv
= correlation coefficient of determination of cross-validation.
Rp2 = correlation coefficient of determination of prediction.
RMSECV = root mean square error of cross-validation.
SD = standard deviation.
RMSEP = root mean square error of prediction.
Please cite this article in press as: Subedi, P.P., Walsh, K.B., Assessment of sugar and starch in intact banana and mango fruit by SWNIR
spectroscopy. Postharvest Biol. Technol. (2011), doi:10.1016/j.postharvbio.2011.06.014
G Model
ARTICLE IN PRESS
POSTEC-9318; No. of Pages 8
P.P. Subedi, K.B. Walsh / Postharvest Biology and Technology xxx (2011) xxx–xxx
4
Table 2
Mango: total soluble solids model calibration and prediction statistics. Partial least squares regression models are based on interactance spectra over the wavelength range
720–920 nm of intact mango fruit. Populations were drawn from a set of ripening fruit, with fruit hard green at day 0 and fully ripe by day 8. The wavelength region
(720–920 nm) was chosen using an optimisation procedure, and was the same as used for the dry matter model.
Calibration
Population
# Samples
# Factors
Mean
SD
2
Rcv
RMSECV
Day 0
Day 1
Day 3
Day 6
Day 7
Day 8
Day 0–8
Day 0–6 + 8
Day 0–1, 6–8
Day 1–8
40
40
40
40
40
40
240
200
200
200
7
5
8
6
4
8
6
6
4
6
5.10
5.40
6.40
12.10
15.30
15.80
10.40
9.60
6.40
10.90
0.36
0.43
1.24
1.21
1.22
1.27
4.50
4.35
1.25
4.43
0.62
0.55
0.62
0.66
0.77
0.85
0.96
0.96
0.96
0.96
0.22
0.29
0.75
0.60
0.57
0.49
0.89
0.91
0.80
0.87
Prediction
Calibration sets
Prediction sets
Rp2
RMSEP
SDR
Bias
Day 1–8
Day 0–6 + 8
Day 0–1, 6–8
Day 0
Day 7
Day 3
0.01
0.79
0.33
1.19
0.69
1.14
0.30
1.77
1.09
1.10
0.36
0.78
2
Rcv
= correlation coefficient of determination of cross-validation.
Rp2 = correlation coefficient of determination of prediction.
RMSECV = root mean square error of cross-validation.
SDR = standard deviation ratio (SD/RMSEP).
SD = standard deviation.
RMSEP = root mean square error of prediction.
The PLS regression coefficients for DM (in hard green stage fruit)
and TSS (ripe stage fruit) were very similar (Fig. 1). It is therefore
expected that a model based on TSS of ripe stage fruit will fail to
differentiate starch and sugar, when applied in prediction of unripe
fruit.
chlorophyll peak, Fig. 2). However, somewhat unexpectedly, the
best calibration models (lowest RMSECV) for peel Hunter ‘a’ value
were based on a wavelength region that included SWNIR as well as
visible wavelengths (530–931 nm; data not shown) although the
model weighted the 500–770 nm region. Presumably, the model
3.2. Banana
As expected, there was no significant change in banana DM
content (at around 25%) over the ripening period, while TSS content increased (from 2 to 19%) (Table 3). Peel Hunter ‘a’ value
also increased (Table 3), while pulp firmness decreased (data not
shown). The PLS models on these attributes were reasonable when
2 > 0.85) (Table 3).
based on all populations (Rcv
The change in skin colour with fruit ripening is obviously visually evident to a human observer. This change is also clearly
evident in the visible region of the interactance spectra (e.g. 680 nm
Fig. 1. Mango: regression coefficients for a partial least square model of dry matter
based on spectra of hard green stage fruit and for a model of total soluble solids
based on spectra of ripe stage fruit.
Fig. 2. Banana: mean absorbance (A) and second derivative of absorbance spectra
(B) of green, ripening, ripe and over-ripe fruit (representative samples chosen from
each stage) using interactance (partial transmittance) optics.
Please cite this article in press as: Subedi, P.P., Walsh, K.B., Assessment of sugar and starch in intact banana and mango fruit by SWNIR
spectroscopy. Postharvest Biol. Technol. (2011), doi:10.1016/j.postharvbio.2011.06.014
G Model
ARTICLE IN PRESS
POSTEC-9318; No. of Pages 8
P.P. Subedi, K.B. Walsh / Postharvest Biology and Technology xxx (2011) xxx–xxx
5
Table 3
Banana: calibration model performance statistics for the attributes of peel Hunter a, mesocarp dry matter (%DM) and total soluble solids (%TSS) for fruit in green to ripening
stages. The wavelength region was chosen using an optimisation procedure (see Figs. 3 and 4) at 500–931 nm for peel Hunter a, 800–1045 nm for DM of mesocarp tissue and
536–847 nm for TSS of mesocarp tissue.
Ripeness stage
Peel Hunter (a)
Green
Ripening
Ripe
Over ripe
All stages
Mesocarp (%DM)
Green
Ripening
Ripe
Over ripe
All stages
Mesocarp (%TSS)
Green
Ripening
Ripe
Over ripe
All stages
# Samples
Mean
SD
# Factors
2
Rcv
RMSECV
102
102
102
102
408
−17.54
−15.74
2.03
4.51
−6.10
1.24
2.54
1.55
1.34
9.85
2
10
6
4
2
0.31
0.83
0.77
0.80
0.98
0.99
1.38
0.98
0.81
2.19
102
102
102
102
408
25.65
25.45
24.36
22.50
24.60
0.94
1.21
0.64
0.79
1.53
9
7
10
8
9
0.79
0.80
0.75
0.72
0.88
0.57
0.71
0.41
0.51
0.73
102
102
102
102
408
1.78
3.90
15.61
19.20
11.07
0.32
2.53
1.66
0.74
7.79
9
6
6
6
7
0.47
0.96
0.87
0.67
0.97
0.19
0.69
0.80
0.50
1.77
2
Rcv
= correlation coefficient of determination of cross-validation.
RMSECV = root mean square error of cross-validation.
SD = standard deviation.
Fig. 3. Banana: difference spectra (absorbance spectra collected using interactance
optics of banana peel removed from the fruit and placed on a black background,
relative to spectra of the intact fruit).
for Hunter ‘a’ must represent both a direct assessment of peel
colour, and an assessment of attributes which change in concert
with colour during fruit ripening.
Banana fruit has a relatively thick skin (approximately 3–4 mm),
which will degrade the ability of optical systems to assess attributes
of the edible mesocarp of intact fruit. Further, mesocarp tissue carbohydrate content indices of DM and TSS are not well correlated
against DM and TSS, respectively, in the peel (Table 4). Nonetheless, mesocarp DM was relatively well modelled using interactance
2 typically > 0.75;
spectra of intact fruit at all ripening stages (Rcv
Table 3). This result is consistent with the notions that the interactance method employed gathered a significant level of information
(light) from mesocarp tissue, as suggested by Zude (2003).
To test this assertion, interactance spectra were acquired of
whole fruit, and of the removed skin against a scattering background (a Teflon tile) and against a light trap (black mat). The
apparent absorbance (true absorbance plus scattering) of removed
skin with a black backing was higher than the apparent absorbance
of intact ripe fruit, at least in the 700–930 nm range (Fig. 3). The
difference between skin with dark background and intact fruit in
the 700–900 nm range was approximately 0.05 Abs unit. This result
is consistent with the interpretation that a measurable amount of
light does penetrate through the skin, scatter within pulp tissues,
and re-pass through the skin to emerge in the field of view of the
detector. The absorbance of the removed skin with a scattering
background (Teflon tile) was higher than that of intact fruit, but not
as high as for the black ground (data not shown). Thus, the amount
of scattering was apparently greater for the mesocarp of the intact
fruit than the Teflon tile (data not shown). Less difference between
skin with dark background and intact fruit at wavelengths less than
700 and greater than 930 nm is consistent with high absorption due
to pigments (to 1.5 units in green fruit, Fig. 2) and water, respectively, such that less light transits the peel.
However, while the above exercises indicate that interactance
spectra of intact fruit contained information about the pulp, such
spectra would predominately hold information about the peel.
Further, in fruit in which starch reserves are converted to soluble sugars during the ripening process, there are typically many
attributes changing simultaneously (e.g. skin colour, flesh firmness,
pigmentation, firmness, and light scattering properties). For the
banana fruit monitored, mesocarp TSS content was correlated to
skin colour (Hunter ‘a’ and Hunter a/b) (R2 > 0.85) and mesocarp
firmness (penetrometer reading) (R2 = 0.97, across all stages). It is
therefore possible that a SWNIR PLS model ostensibly either on
mesocarp firmness or TSS may be assessing a correlated attribute
such as peel colour.
Indeed, the best PLSR calibration models (lowest RMSECV) for
mesocarp firmness were based on a visible wavelength region
(530–733 nm; data not shown). Thus the PLSR model for ‘firmness’
probably represents an assessment of chlorophyll content (note
Fpen – Hunter a/b R2 = 0.90, Table 4).
The best calibration models (lowest RMSECV, low RMSEC, low
number of PLS factors) for banana mesocarp TSS also included a visible wavelength region (536–847 nm; Fig. 4), with the PLS model
regression coefficients heavily weighted in the 550–700 nm region
(Fig. 4D). Such models presumably achieve an indirect assessment
of mesocarp TSS content through association with peel pigmentation. However, while the cross validation results were encouraging,
prediction of independent sets (varying in stage of ripeness) was
poor (Table 5). Such a model is likely to perform even more poorly
in prediction of fruit ripened at different temperatures, where a
change in the relative rates of peel colour change (plus possibly
other peel attributes) and mesocarp starch to sugar conversion is
expected. Restricting the PLS model on TSS to the spectroscopically relevant wavelengths of 536–847 nm resulted in a model of
Please cite this article in press as: Subedi, P.P., Walsh, K.B., Assessment of sugar and starch in intact banana and mango fruit by SWNIR
spectroscopy. Postharvest Biol. Technol. (2011), doi:10.1016/j.postharvbio.2011.06.014
G Model
ARTICLE IN PRESS
POSTEC-9318; No. of Pages 8
P.P. Subedi, K.B. Walsh / Postharvest Biology and Technology xxx (2011) xxx–xxx
6
Table 4
Banana: (i) population attribute statistics (mean and standard deviation) on pulp and peel dry matter (DM) and total soluble solids (TSS), penetrometer (Fpen ), and colorimeter
reading (Hunter L, a, b, a/b), and (ii) correlation coefficient of determination (R2 ) between attributes [R2 values above 0.75 are given in bold].
Attribute
L
a
b
a/b
Mesocarp DM
Pulp TSS
Fpen
Peel DM
Peel TSS
X̄ ± SD
L
a
b
a/b
Pulp DM
Pulp TSS
Fpen
Peel DM
Peel TSS
64.9 ± 7.1
1
−6.1 ± 9.85
0.58
1
36.2 ± 8.4
0.67
0.35
1
0.2 ± 0.3
0.59
0.98
0.36
1
24.6 ± 1.5
0.14
0.41
0.04
0.40
1
11.1 ± 7.8
0.66
0.88
0.48
0.88
0.31
1
2.5 ± 1.6
0.64
0.88
0.42
0.90
0.53
0.94
1
10.7 ± 1.2
0.02
0.23
0.02
0.24
0.11
0.25
0.13
1
3.3 ± 1.9
0.01
0.09
0.01
0.09
0.34
0.15
0.05
0.44
1
Fig. 4. Banana: optimisation of wavelength region for partial least squares (PLS) model on total soluble solids of fruit mesocarp (spectra collected of intact fruit) for models
based on fruit across all ripening stages (n = 408). Results shown for (A) optimal number of factors, (B) root mean square error of calibration and (C) root mean square error of
cross-validation for PLS models with wavelength regions starting between 500 and 1050 nm, and ending between 550 and 1100 nm (i.e. results of 50,000 models displayed).
Gray scale bar to right of each panel represents a legend relating to units. Regression coefficients for a 500–1040 nm model are displayed in (D).
Table 5
Banana: prediction statistics for models developed using fruit of ripe and over ripe stages (n = 204), with prediction sets of fruit of green and ripening stages (n = 100 each.
Model prediction of ripe and over ripe populations represents a calibration result. Validation set mean and standard deviation (SD) are given.
Attribute
Maturity stage
Mean
SD
Rp2
RMSEP
Bias
Mesocarp (%DM)
Green
Ripening
Green + ripening
Ripe + over ripe
Green
Ripening
Green +Ripening
Ripe + over ripe
25.65
25.36
25.50
24.53
1.77
3.86
2.85
18.62
0.94
1.19
1.08
1.58
0.33
2.48
2.07
2.06
0.55
0.56
0.52
0.88
0.10
0.00
0.15
0.79
0.69
0.87
0.76
0.54
0.79
1.40
1.15
0.94
0.12
0.13
0.01
0.002
0.34
−0.28
0.03
0.001
Mesocarp (%TSS)
Rp2 = correlation coefficient of determination of prediction.
RMSEP = root mean square error of prediction.
Please cite this article in press as: Subedi, P.P., Walsh, K.B., Assessment of sugar and starch in intact banana and mango fruit by SWNIR
spectroscopy. Postharvest Biol. Technol. (2011), doi:10.1016/j.postharvbio.2011.06.014
G Model
POSTEC-9318; No. of Pages 8
ARTICLE IN PRESS
P.P. Subedi, K.B. Walsh / Postharvest Biology and Technology xxx (2011) xxx–xxx
7
Fig. 5. Banana: total soluble solids prediction results (predicted against actual) of intact banana fruit using a partial least squares regression model (721–974 nm) developed
from data of all ripening stages (green to over ripe stages) in prediction of a validation set of data from across all ripening stages (Rp2 = 0.77, RMSEP = 0.94; Table 5).
Fig. 6. Banana: optimisation of wavelength region for partial least squares regression model on dry matter of fruit mesocarp (spectra collected of intact fruit). Results shown
for (A) optimal number of factors, (B) root mean square error of calibration and (C) root mean square error of cross-validation for PLS models with wavelength regions starting
between 500 and 1050 nm, and ending between 550 and 1100 nm (i.e. results of 50,000 models displayed). Gray scale bars to right of each panel represent a legend relating
to units. Regression coefficients for a 500–1050 nm model are displayed in (D).
Please cite this article in press as: Subedi, P.P., Walsh, K.B., Assessment of sugar and starch in intact banana and mango fruit by SWNIR
spectroscopy. Postharvest Biol. Technol. (2011), doi:10.1016/j.postharvbio.2011.06.014
G Model
POSTEC-9318; No. of Pages 8
ARTICLE IN PRESS
P.P. Subedi, K.B. Walsh / Postharvest Biology and Technology xxx (2011) xxx–xxx
8
However, use of the technology for analysis of TSS across ripening
stages (i.e. in samples with variable starch content) is not recommended. This result is consistent with that of McGlone et al.
(2003a,b) for apple. Application of the technology to the assessment of banana mesocarp DM is compromised by the thickness of
the peel, and the lack of correlation between peel and mescocarp
DM.
Acknowledgements
We appreciate the funding support of Hortical P/L and Horticulture Australia Ltd., and provision of a CQUniversity scholarship to
Phul Subedi.
References
Fig. 7. Banana: dry matter prediction results (predicted vs. actual) of intact banana
fruit using a partial least squares regression model developed from data of all ripening stages (green to over ripe stages) in prediction of a validation set of data from
across all ripening stages (Rp2 = 0.88, RMSEP = 0.54; Table 5).
2 = 0.97, RMSECV = 1.77. Note that for the TSS model created
Rcv
across all ripening stages (i.e. green to fully ripe fruit); R2 was
high, courtesy of a high SD; however RMSECV was also high, at
1.77% (Table 3). Such a model could be used in prediction of the
stage of ripening (Fig. 5), but would have little value in screening
within a stage (e.g. SD of TSS content of fully ripe fruit was 2.06%)
(Table 5).
A PLS model for banana mesocarp DM content based on the
full available spectrum (500–1050 nm) also gave weight to features in the visible spectrum, presumably skin pigments (Fig. 6D).
However, the best calibration models (lowest RMSECV, low RMSEC
and number of factors) were based on a SWNIR wavelength region
(R2 = 0.88; 730–930 nm, Fig. 6), with the PLS model regression coefficients heavily weighted in the 870–930 nm region. With pulp
DM poorly correlated to peel DM (R2 = 0.11, Table 4), the model
may be achieving a direct assessment of mesocarp DM content.
It is also possible that the optical method is assessing a character
of the skin that is related to mesocarp DM, although admittedly
mesocarp DM was weakly related to skin Hunter a/b (R2 = 0.40)
or other measured peel attributes. Certainly, prediction across
ripening stages was poor (Table 5), although prediction of a subset population drawn across all ripening stages was reasonable
(Fig. 7). This latter result highlights the danger of using a less than
totally independent population to test model prediction performance.
4. Conclusions
The use of SWNIRS for analysis of DM (total carbohydrate content) of intact mango across all ripening stages is recommended.
Greensill, C.V., Walsh, K.B., 2000. Optimisation of instrumentation precision and
wavelength resolution for the performance of NIR calibrations of sucrose in a
water–cellulose matrix. J. Appl. Spectrosc. 54, 426–430.
Guthrie, J.A., Walsh, K.B., 1997. Non-invasive assessment of pineapple and mango
fruit quality using near infra-red spectroscopy. Austr. J. Exp. Agric. 37,
253–263.
Guthrie, J.A., Walsh, K.B., Reid, D.J., Liebenberg, C.J., 2005. Assessment of internal
quality attributes of mandarin fruit. 1. NIR calibration model development.
Austr. J. Exp. Agric. 56, 405–416.
Jha, S.N., Narsaiah, K., Sharma, A.D., Singh, M., Bansal, S., Kumar, R., 2010. Quality parameters of mango and potential of non-destructive techniques for their
measurement – a review. J. Food Sci. Technol. 47, 1–14.
Lammertyn, J., Peirs, A., De Baerdemaeker, J., Nicolai, B., 2000. Light penetration
properties of NIR radiation in fruit with respect to non-destructive quality
assessment. Postharvest Biol. Technol. 18, 121–132.
Marriott, J., 1980. Bananas – physiology and biochemistry of storage and ripening
for optimum quality. Crit. Rev. Food Sci. Nutr. 13, 41–88.
McGlone, V.A., Clark, C.J., Jordan, R.B., 2007. Comparing density and VNIR models
for predicting quality parameters of yellow-flesh kiwifruit (Actinidia chinensis).
Postharvest Biol. Technol. 46, 1–9.
McGlone, V.A., Jordan, R.B., Seelye, R.J., Clark, C.J., 2003a. Dry-matter – a better predictor of the post-storage soluble solids in apples? Postharvest Biol. Technol. 28,
431–435.
McGlone, V.A., Fraser, D.G., Jordan, R.B., Kunnemeyer, R., 2003b. Internal quality
assessment of mandarin fruit by vis/NIR spectroscopy. J. Near Infrared Spectrosc.
11, 323–332.
Ramaswamy, H.S., Tung, M.A., 1989. Technical note: textural changes as related to
colour of ripening banana. Int. J. Food Sci. Tech. 24, 217–221.
Saranwong, S., Sornsrivichai, J., Kawano, S., 2004. Prediction of ripe-stage
eating quality of mango fruit from its harvest quality measured nondestructively by near infrared spectroscopy. Postharvest Biol. Technol. 31,
137–145.
Shenk, J.S., Westerhaus, M.O., 1993. Analysis of agriculture and food products by near
infrared reflectance spectroscopy. Infrasoft International, LLC, Infrasoft International, Port Matilda, PA, USA.
Subedi, P.P., Walsh, K.B., Owens, G., 2006. Prediction of mango eating quality at
harvest using short-wave near infrared spectroscopy. Postharvest Biol. Technol.
43, 326–334.
Walsh, K.B., Golic, M., Greensill, C.V., 2004. Sorting of fruit using near infrared spectroscopy: application to a range of fruit and vegetables for soluble solids and dry
matter content. J. Near Infrared Spectrosc. 12, 141–148.
Walsh, K.B., Guthrie, J.A., Burney, J.W., 2000. Application of commercially available,
low-cost, miniaturised NIR spectrometers to the assessment of the sugar content
of intact fruit. Aust. J. Plant Physiol. 27, 1175–1186.
Zude, M., 2003. Non-destructive prediction of banana fruit quality using VIS/NIR
spectroscopy. Fruits 58, 135–142.
Please cite this article in press as: Subedi, P.P., Walsh, K.B., Assessment of sugar and starch in intact banana and mango fruit by SWNIR
spectroscopy. Postharvest Biol. Technol. (2011), doi:10.1016/j.postharvbio.2011.06.014