You are on page 1of 9

Postharvest Biology and Technology 43 (2007) 326–334

Prediction of mango eating quality at harvest using short-wave


near infrared spectrometry
P.P. Subedi a,∗ , K.B. Walsh a , G. Owens b
a Plant Sciences Group, Central Queensland University, Building 9, Rockhampton, Qld 4702, Australia
b Northern Territory Department of Primary Industry, Fisheries and Mines, Berrimah Agricultural Research Centre, Darwin, NT 0881, Australia

Received 19 June 2006; accepted 18 September 2006

Abstract

Short wave near infrared (SWNIR) (400–1100 nm) spectroscopy was trialled in assessment of mango (Mangifera indica L.) fruit maturation
and as a harvest time guide to final eating quality. Fruit maturity was indexed in terms of flesh colour (r2 = 0.79 for Hunter b value against
maturity score), dry matter content (r2 = 0.66 for % DM against maturity score) and a visual ranking of maturity. Fruit eating quality at fully
ripe stage was indexed in terms of total soluble solids content of extracted juice (TSS). Partial least squares (PLS) regression models based
on second derivative of absorbance spectra for DM, TSS, Hunter b and visual maturity ranking were optimised in terms of the wavelength
range of SWNIR. Optimal TSS and DM models used the same wavelength region and also produced similar PLS regression coefficient plots,
suggesting that the models are unable to differentiate between the soluble and insoluble forms of carbohydrate in the fruit. When used in
prediction of new populations, DM and Hunter b models based on several harvest dates were acceptable (e.g. for DM, R2v = 0.74, bias = 1,
bias corrected root mean square error of prediction [SEP] = 1% DM). Calibration models on TSS of ripe fruit, developed using SWNIR
spectra collected (non-destructively) of hard green mango (R2c = 0.90), were useful in prediction of an independent population (R2v = 0.92 with
SEP = 0.67 and bias = 1.25% TSS). We conclude that the SWNIR technique can be used at fruit harvest in assessment of fruit maturity (as
flesh Hunter b or % DM), and also in prediction of the future TSS of fruit after ripening.
© 2006 Elsevier B.V. All rights reserved.

Keywords: Near-infrared; Spectroscopy; Non-invasive; Mango; Internal quality

1. Introduction Fruit maturity is typically based on time since fruit set


and a subjective assessment of skin roughness, fruit firmness,
Mango (Mangifera indica L.) fruit is usually harvested at glossiness, shoulder ‘fullness’, and peel and flesh colour,
the hard green stage, after physiological maturity is reached, although these parameters are known to be potentially mis-
but before the onset of the climacteric respiration rise. Fruit leading, or dependent on a level of skill lacking in a casual
picked before physiological maturity will not ripen properly, workforce. Further, it is recognised that these parameters can-
leading to a poor eating experience. Fruit picked immediately not be used with success on certain cultivars. Prediction of
after physiological maturity will ripen, but a further delay in harvest date can be improved by use of heat sum accumulation
harvest date allows fruit carbohydrate reserves to increase, models.
and such fruit attain a superior eating quality. To achieve high Carbohydrate content (indexed as DM) and flesh colour
early season market prices, however, growers are tempted to are also objective indicators of mango fruit maturity. Assess-
harvest fruit as soon as (or before) physiological maturity is ment of these parameters has required destructive measure-
reached. Strip picking of the crop also results in harvest of ment, but non-destructive measurement using near infrared
fruit of a range of maturity stages. spectroscopy may be possible. However, physiological matu-
rity is indexed by a plateau in the rate of DM accumulation,
∗ Corresponding author. Tel.: +61 7 49232315; fax: +61 7 49306536. rather than a specific DM level per se, and thus attempts
E-mail address: p.subedi@cqu.edu.au (P.P. Subedi). to set a specific % DM maturity standard should be sea-

0925-5214/$ – see front matter © 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.postharvbio.2006.09.012
P.P. Subedi et al. / Postharvest Biology and Technology 43 (2007) 326–334 327

son/region/cultivar specific. Flesh colour may also an unre- spectra, with a 25 s total measurement time), to assess DM
liable guide to fruit maturity guide (Saranwong et al., 2004), and starch content of hard green fruit (cv. Mahajanaka). For
being sensitive to cultural practice (e.g. nitrogen fertiliser). DM, they reported R2v = 0.92, SEP = 0.41%, S.D. = 1.26%,
Mesocarp starch is converted into soluble sugars during while for starch (% DM basis), R2v = 0.86, SEP = 1.71%,
the ripening process. Final eating quality is linked to soluble S.D. = 4.08% was reported.
sugar content. Thus, total carbohydrate content (or dry matter Walsh et al. (2004) reported model statistics for
(DM) content as a measure of total starch and soluble sugar mango DM prediction of R2c = 0.79, RMSECV = 0.46%
(TSS) content) at fruit harvest is linked to final eating quality. (S.D. = 1.1%), using SWNIR instrumentation that is com-
Short wavelength near infrared spectroscopy (SWNIR) mercially available for fruit pack-lines, grading fruit at ca.
has been used as a non-destructive method of evaluating inter- 7 s−1 .
nal quality of a range of fresh fruit on the basis of TSS or DM Saranwong et al. (2003b) proposed a quality standard for
content (e.g. Kawano et al., 1992; Walsh et al., 2004). In the final eating quality of 19.6% DM. This value was calcu-
following review of the available literature on application to lated as the population mean minus one standard deviation
mango, we have attempted to be clear with respect to the vari- on DM. However, this standard is specific to this popula-
ous statistical measures reported (citing, as available: R2c as a tion. In a subsequent publication, Saranwong et al. (2004)
calibration cross-validation correlation coefficient, and R2v a noted that combining (SWNIRS predicted) estimates of both
validation correlation coefficient; root mean squared error DM and starch at harvest slightly improved prediction of
of cross-validation, RMSECV; root mean square standard ripe stage TSS (R2v = 0.85, SEP = 0.55, bias 0.16% TSS)
error of prediction; RMSEP; bias corrected root mean square over use of harvest stage DM (R2v = 0.74, SEP = 0.73) or
standard error of prediction, SEP; difference between mean starch alone. This observation infers that DM is an index
predicted and actual values, bias; and population standard of a range of attributes other than starch that impact on
deviation, S.D.). TSS.
The first reported application for mango is that of Guthrie From this literature base, we conclude that the SWNIRS
and Walsh (1997), who reported model statistics of R2c = 0.98, technique holds promise for mango fruit quality assessment.
and RMSECV = 1.19 for % DM content of cv. ‘Kensing- However, it is essential that the robustness of the method be
ton Pride’ (KP) (S.D. = 3.65% DM). This work was based established, with the ability of a model to test truly inde-
on 760–2500 nm spectra collected with NIRSystems 6500 pendent sets of fruit tested. Existing literature reports have
using reflectance optics. Reference values were determined involved use of subsets of a given population for calibration
of a 50–60 mm diameter, 20 mm depth core. This work was and validation groups, rather than truly independent valida-
extended by Schmilovitch et al. (2000), who reported model tion sets.
statistics of R2v = 0.86, 0.82 and 0.94 (RMSEP = 1.22%, 17.1 In the current study we aim to test the robustness of
N and 37 d) on TSS, firmness and time to ripeness, respec- SWNIR spectroscopy in prediction of mango fruit DM con-
tively, of cv.‘Tommy Atkins’ (S.D. not reported), based on tent and flesh colour, as a harvest time guide to fruit maturity
1200–2400 nm reflectance spectra. and final eating quality. Robustness was tested across har-
Subsequently, Saranwong et al. (2001) utilised an interac- vest time, region and cultivar. We also sought to correlate
tance probe with a NIRSystems 6500 unit to demonstrate that harvest stage fruit spectra directly with ripe stage TSS,
model results were improved by use of SWNIR (to 1100 nm) rather than using such spectra to predict harvest stage DM
(R2v = 0.94; SEP = 0.55% DM; on S.D. = 1.72% DM), com- and starch, and then relating these attributes to ripe stage
pared to longer wavelengths (to 2500 nm). This result was TSS.
attributed to better penetration of the fruit by the shorter
wavelengths. Saranwong et al. (2003a) reported the trial of a
transportable field spectrometer (‘FT20’ from Fantec), oper- 2. Materials and methods
ating in the SWNIR with a measurement cycle of 3–5 s, for
the assessment of TSS of ripe, intact mango cv. ‘Caraboa’ 2.1. Fruit selection
TSS. Two wavelength windows were used, with a better
result returned from the 850 to 1000 nm region (R2v = 0.96, Representative lots of mango (M. indica L.) fruit of in-
SEP = 0.40, with S.D. = 1.80% TSS) than the 900–1000 nm coming consignments to the W.E. Pack pack-house (Berry
(R2v = 0.88, SEP = 0.55). Saranwong et al. (2003b) reported Spring, Darwin, Northern Territory, Australia), from differ-
on the use of the same device for field measurements of (cv. ent growers in the Darwin and Katherine regions (total of
Nam Dork Mai) fruit DM content. It was necessary to place 1157 mango fruit, being 932, 150, 50 and 25 from cvs. KP,
a light-tight silver bag over fruit to exclude sunlight. Sample Calypso, R2E2 and Celebration, respectively; populations
(fruit) temperature compensation was achieved by including 1–18). These fruit were collected over the 2005 North-
fruit scanned at three temperatures (25, 32 and 39 ◦ C) into the ern Territory mango-harvesting season (20 September to 8
partial least squares regression (PLSR) model. Saranwong et November). Further cv. Calypso fruit were sourced from
al. (2004) again utilised a NIRS6500, operating with interac- Mareeba (population 19) and Bundaberg (population 20)
tance optics in the 700–1100 nm region (average 50 scans per regions in Queensland.
328 P.P. Subedi et al. / Postharvest Biology and Technology 43 (2007) 326–334

Maturity levels of the incoming cv. KP fruit were deter- Systems, Bacchus Marsh, Australia) or a prototype handheld
mined at the pack-house by a trained person into five extension of this technology (‘iQ’, Integrated Spectronics,
categories based on visual and sensory criteria as follows: Sydney, Australia). Both technologies are based on a silicon
diode detector and a tungsten halogen light source (100 W for
• Stage 1 (immature)—narrow cheeks, rough, hard skin.
in-line, 20 W for hand-held) in the interactance optical con-
• Stage 2 (pre-mature)—cheeks narrow, but slightly fuller
figuration reported by Greensill and Walsh (2000) (0◦ angle
compared to stage 1, skin slightly smoother but hard.
between illumination and detected light rays, with detection
• Stage 3 (early maturity stage)—full shoulder, down in top,
probe viewing a shadow cast by the probe onto the fruit). The
smooth skin, slightly beaky.
integration time was varied (typically 40–60 ms), with signal
• Stage 4 (optimum maturity stage)—fully formed, down in
level maintained below saturation but above 66% of detec-
top, full nose, smooth skin slight give in skin.
tor saturation value. Referencing was undertaken daily for
• Stage 5 (ripening stage, too ripe for packing)—fully
the Insight unit, and with every sample acquisition for the iQ
formed, breaking colour, started softening.
unit. Spectra were collected from both sides of each fruit, at
In another exercise, fruit of cv. KP were transported to the equator of the fruit, equidistant from proximal and distal
the laboratory within 48 h of harvest (as hard green mature ends.
fruit), sprayed with 0.2% ethephon 480 and allowed to
ripen at 25 ◦ C. At intervals of approximately 1 day, a ran- 2.4. Chemometrics
domly selected set of fruit were sampled for DM, TSS and
firmness, with the remainder allowed to continue ripening The chemometrics software package, The Unscrambler
at 25 ◦ C. V. 9.1 (Camo), was used for partial least squares regres-
sion (PLSR) calibration model development (prediction of
2.2. Reference analyses a given attribute from spectral data). Calibration model per-
formance was assessed in terms of Rc , RMSECV and SDR
Dry matter content of fruit was determined by taking a (SD/RMSECV), while performance in validation was con-
27 mm diameter, 10 mm deep core of fruit flesh from the sidered in terms of Rv , RMSEP and SDR, slope and bias of
equatorial part of each cheek of each fruit, after removing the validation sets. Cross-validation was performed using 20
skin (1–2 mm thickness) using a potato peeler. Dry matter was segments with random selection of samples. PLSR model
determined by weight loss following 48 h drying in a forced results are influenced by a number of parameters, includ-
air oven (populations 5–20), or 20 min low power microwave ing wavelength range, data pre-processing, and number of
followed by 5 h drying in a Napco 5831 vacuum oven (pop- PLS factors used. In this study, all data was pre-processed
ulations 1–4), at 70 ◦ C. Smaller pieces of tissue were used in terms of transformation to a second derivative, using a
to assess the DM distribution pattern within fruit. For ripe linear calculation method across nine pixels of the diode
fruit, juice was extracted from the 27 mm diameter, 10 mm array (i.e. approximately 27 nm). The number of PLS factors
deep core using a garlic press, and juice TSS was determined used in model development was determined by the com-
using a temperature compensated refractometer (Bellingham parison of root mean squared error of calibration (RMSEC)
and Stanley RMF 320). and RMSECV (using a 20 group validation procedure), with
Internal flesh colour was measured with a Chroma meter the Unscrambler chemometric software suggested number of
CR-400/410 (Minolta, Osaka, Japan) as Hunter L, a and b factors adopted. The optimal wavelength range for the PLSR
values. Hunter L is a measure of luminosity, while a and b model was established using a MatLab based script by creat-
values measure the red–green and yellow–blue space, respec- ing models varying in wavelength start and stop wavelength
tively. Each side of each fruit was cut longitudinally midway at 3 nm increments, and creating a ‘map’ of model perfor-
between the seed and skin, and then recut to take a 1 cm thick mance (RMSEC) by start and end wavelength (Guthrie et al.,
slice. A single colourimeter reading was taken of each slice 2005).
from the region where spectra were acquired. Following the chemometric adage ‘no prediction without
A destructive measurement of fruit firmness was made interpretation and no interpretation without prediction’, we
using an Effegi penetrometer employed with a HortPlus consider model robustness in prediction of independent data
(New Zealand) mechanical drive stand. The force (New- sets, and we consider the spectroscopic basis of these models.
tons) required to penetrate a 6 mm diameter probe 8 mm
into the peeled fruit surface at a speed of 20 mm/min was
recorded. 3. Results and discussion

2.3. Spectroscopy 3.1. Sampling strategy in relation to distribution of DM


in fruit
Spectra were collected over the wavelength range
300–1150 nm using ‘Insight 023137’, an interactance optics Most previous studies have removed fruit skin before
unit manufactured for on-line applications (Colour Vision assessment of DM, although Saranwong et al. (2003b)
P.P. Subedi et al. / Postharvest Biology and Technology 43 (2007) 326–334 329

Table 1
Population and PLSR calibration model statistics based on second derivative of absorbance (d2 A) spectra across 20 populations of hard green mango for the
attributes of DM, visual maturity score and Hunter b value
Cultivar Population Date of harvest n Mean S.D. #Factors #Outliers R2c RMSECV
% DM
KP 1 28 September 2005 17 19.1 0.6 6 0 0.98 0.13
KP 2 29 September 2005 44 19.1 0.8 4 1 0.90 0.56
KP 3 30 September 2005 55 19.2 1.7 6 0 0.90 0.49
KP 4 1 October 2005 94 18.7 2.0 6 5 0.90 0.57
KP 5 2 October 2005 32 18.9 2.2 6 0 0.88 0.45
KP 6 4 October 2005 48 20.9 1.3 7 5 0.74 0.74
KP 7 6 October 2005 40 21.4 1.5 5 0 0.96 0.40
KP 8 7 October 2005 38 19.3 2.4 9 0 0.92 0.54
KP 9 10 October 2005 188 19.5 2.6 8 3 0.92 0.67
KP 10 11 October 2005 30 16.9 2.1 8 0 0.98 0.34
KP 11 13 October 2005 120 19.0 1.8 8 4 0.90 0.56
KP 12 19 October 2005 168 17.3 2.7 7 6 0.90 0.61
KP 13 22 October 2005 48 18.9 2.2 6 6 0.88 0.57
KP 1–13 30 October 2005 971 16.7 2.2 10 26 0.88 0.71
R2E2 + celebration 14 17 October 2005 68 16.2 2.0 6 2 0.92 0.36
Calypso 15 25 October 2005 32 15.0 1.2 7 1 0.88 0.29
Calypso 16 26 October 2005 68 15.6 1.3 8 3 0.90 0.43
KP + R2E2 + calypso 1–14 + 17 5 November 2005 1832 16.1 2.2 8 124 0.88 0.71
Calypso 17 19 December 2005 204 14.7 1.7 4 1 0.92 0.48
Calypso 18 25 January 2006 68 16.7 3.2 5 0 0.96 0.66
Scores
KP 6 4 October 2005 47 3.10 1.76 5 2 0.88 0.26
KP 6+7 6 October 2005 60 3.10 2.12 6 2 0.94 0.26
KP 12 19 October 2005 168 2.73 1.11 5 2 0.76 0.55
All KP 6–13 30 October 2005 728 2.84 1.26 8 17 0.72 0.65
Hunter b
KP 6 4 October 2005 48 49.9 21.9 8 2 0.94 5.52
KP 11 13 October 2005 120 43.8 22.5 7 6 0.88 7.84
KP 12 19 October 2005 168 39.8 15.6 5 11 0.90 4.74
KP 13 22 October 2005 288 40.8 19.0 6 10 0.90 5.86
Calypso 16 + 17 19 December 2005 272 43.3 14.0 5 6 0.90 4.50
KP + calypso 12 + 17 21 December 2005 372 44.1 14.3 7 30 0.88 4.33

assessed DM of fruit cores with skin. The DM content of 3.2. Inter-relation of attributes at fruit maturity
a 27 mm diameter, 10 mm deep core was 23.8 ± 2.1% with
skin left attached, and 19.4 ± 1.7% with skin removed. Given Of the various measures of mango quality (L, a, b, DM),
the ‘bias’ effect of inclusion of skin, and the lack of relevance b value and DM demonstrated the highest correlation with
of skin DM content to the consumer, all further work in this maturity score (Tables 2 and 3). Fruit DM at hard green stage
study was based on cores with skin removed. was also best indexed by Hunter b value.
The DM content of the flesh was higher near to the skin
and gradually decreased towards the seed (after skin removal,
0–10 mm depth: 18.8 ± 2.5; 10–20 mm depth: 15.8 ± 1.8). Table 2
However, the DM content of the 0–10 mm core was correlated Linear correlation coefficient of determination (r2 ) of selected attributes to
with that of the 10–20 mm core (r2 = 0.93). A 0–10 mm deep visual maturity score and DM [data of populations 11 and 12 combined,
core was used in the reference assessment of TSS and DM refer Table 1 (n = 288)]
estimates as the optical geometry used is expected to optically Attribute r2
sample to this depth. To score To DM
In mature KP mango fruit at harvest (combined popula-
L 0.63 0.53
tions 11 and 12 of Table 1), DM content was 18.8 ± 1.7% a 0.49 0.41
(mean ± S.D.) (0–10 mm depth core). Flesh DM (0–10 mm) b 0.79 0.74
was higher at the stem end (19.03 ± 1.27% DM) and lower
at the acropetal end (17.26 ± 0.89). The equatorial region of DM 0.66 –
the fruit was chosen for sampling (spectra acquisition and Spectra 0.79 0.94
reference analyses) as representative of the whole fruit for Results for attribute ‘spectra’ refers to the calibration outcome of a PLS
DM content. model using second derivative of absorbance spectra.
330 P.P. Subedi et al. / Postharvest Biology and Technology 43 (2007) 326–334

Table 3 possible start and stop wavelength combinations within the


Mean value and standard deviation of attributes Hunter L, a, b, and % DM set wavelength region used (see Guthrie et al., 2005), sim-
for fruit of different maturity levels (n = 17 per maturity level; being part of
population 12 of Table 1)
ilar to the ‘searching window’ method described by Du
et al. (2003). Model RMSEC was used to select the best
Maturity L a b DM
start and end wavelengths (Fig. 1). Thus, for example, the
1 87.3 ± 1.7 −5.71 ± 1.2 26.3 ± 9.1 15.0 ± 1.1 optimal wavelength range for modelling DM (i.e. lowest
2 87.3 ± 1.4 −5.32 ± 1.1 23.2 ± 4.8 15.2 ± 1.4
3 84.8 ± 2.4 −6.02 ± 1.2 43.1 ± 10.2 17.8 ± 1.1
RMSEC) involved a start wavelength of 700–880 nm and an
4 80.1 ± 5.8 −2.00 ± 5.0 56.7 ± 11.0 18.7 ± 2.1 end wavelength of 920–1100 nm. The optimal wavelength
5 67.3 ± 6.4 7.54 ± 5.6 65.6 ± 6.7 19.7 ± 1.1 range for modelling TSS involved a start wavelength of
740–850 nm and an end wavelength of 950–1100 nm. The
optimal wavelength range for modelling Hunter b involved
Flesh colour, as indexed by b value was a better index of a start wavelength of 760–870 nm and an end wavelength of
fruit maturation (r2 = 0.79) than DM content (r2 = 0.66). This 1000–1100 nm. The optimal wavelength range for modelling
result is ascribed to variation in fruit carbohydrate load with maturity score involved a start wavelength of 700–850 nm
position in the tree canopy, and between tree positions in an and an end wavelength of 1000–1100 nm. Thus the optimal
orchard (S.D. on DM was 2.7% in this population, Table 1). wavelength region was similar for the TSS, DM and Hunter
Thus, while a plateau in DM accumulation may index mat- b models, and different to that for the maturity score model.
uration of a starchy fruit like mango, there is a range in the Carbohydrate (starch, sugar) models are expected to
plateau level for DM within a given orchard. weight features related to water and sugar OH stretching
Therefore, we focused our SWNIR work onto prediction (second overtone of OH stretch at 960 nm, third combination
of Hunter b value and DM of flesh, as indices of fruit maturity. overtone at 840 nm) and CH stretching (third overtone stretch
at 910 nm) (e.g. Golic et al., 2003). The observed optimal
3.3. Instrumentation wavelength region for the DM and TSS models are consis-
tent with involvement of these features. The PLSR model B
The in-line hardware differed from the hand-held hard- coefficients are difficult to interpret in terms of these absorp-
ware in having temperature regulation (Peltier cooling of the tion features (Fig. 2). However, the consistency between the B
spectrometer and fan cooling of the lamp). The hand-held coefficients of the DM and SSC models indicates that models
hardware compensated with referencing after every sample for both attributes are based on similar spectral features.
(the equipment is intended for use in field conditions). In Of course, DM is a measure of both starch and soluble
practice, under open pack-house conditions, the two units per- sugar, while the TSS reference method measures soluble
formed comparably (e.g. DM model results on population 11 sugar but not starch. Calibration models for DM were accept-
and 12 (n = 288) of R2c = 0.92, RMSECV = 0.61 for handheld able at any stage of ripening (R2c > 0.81, data not shown) while
and R2c = 0.92, RMSECV = 0.62 for in-line units, respec- those for TSS were acceptable at fully ripe stage (R2c = 0.85)
tively). Therefore, results for only one format of the equip- but not at harvest stage (R2c = 0.62, data not shown). These
ment (handheld) are presented in the remainder of this report. observations indicate that DM models may be robust in ripen-
Two drying methods were used, due to the availability ing fruit (with conversion of starch to sugar), while TSS
of equipment. Care was taken to use the same drying tem- models are not.
perature and to ensure sample weight had stabilised before The optimal region for the model on Hunter b did not
removal from the ovens, however the two methods could include the visible wavelengths upon which the b value is
result a bias shift between sets of readings. Data has been based. This result is attributed to the low penetration of wave-
analysed with this variable in mind, with robustness of a PLS lengths in the visible region through fruit tissue (and thus high
DM developed using one drying reference method tested on noise in the absorbance spectra). The Hunter b models based
a set based on the other reference method (see later discus- on longer (SWNIR) wavelengths must be based on secondary
sion). There was no indication of a higher bias error in this correlations with other characters, although it is not clear
situation, relative to use of a models developed using data what physico-chemical features this correlation is based on.
from both drying methods. Presumably the PLSR model is not simply assessing fruit DM
in a secondary correlation to Hunter b, as the PLSR model
3.4. SWNIR PLSR modelling of intact fruit parameters results on maturity score were superior to regression of DM
at fruit maturity on maturity score (Table 2). The PLSR model B coefficients
for Hunter b were also quite different to those for DM (Fig. 2).
Inclusion of non-informative wavelengths in a PLS model Similarly, it is not clear what physico-chemical features
can decrease predictive performance. Previous studies have are responsible for the optimal wavelength range noted for the
simply used a set range, or at best compared two or three maturity score model. Notably, this region was quite different
wavelength ranges. The optimal wavelength region for each to that for the Hunter b model, suggesting that the PLSR
attribute was determined through use of a MatLab-PLS tool- model for maturity was not based on an indirect assessment
box script that developed (47,000) PLSR models for all of Hunter b. The PLSR model B coefficients for maturity
P.P. Subedi et al. / Postharvest Biology and Technology 43 (2007) 326–334 331

Fig. 1. PLSR calibration model performance, in terms of root mean square standard error of calibration (RMSEC) for different NIR wavelength windows
(varying start and end point) for DM (A); TSS at fruit ripeness (B); visual maturity score (C) and Hunter b value (D). Models based on second derivative
absorbance spectra (n = 168 for (A), (C) and (D); n = 218 for (B)). Spectra were taken of fruit at harvest (hard green stage) in all cases. Bar on right of each
panel displays colour scale for RMSEC levels.

score were quite dissimilar to those for DM and TSS, and the fourth population (R2v = 0.74, bias = < 1, % DM corrected
also differed to those for Hunter b (Fig. 2). SEP = < 1% DM). Continued updating of the model had little
effect in terms of R2v , SEP and bias relative to use of a model
3.5. PLSR model robustness—maturity indices based on data of the first four populations (Fig. 3). Model
updating improved results over use of a model based on four
While calibration results for PLSR models on SWNIR populations of cv. KP when used in prediction of different
spectra on the attributes of DM and Hunter b were acceptable varieties (e.g. SEP in prediction of cv. Calypso, population
(R2c = 0.81), the true test of such models lies in their predictive 18, was 1.26 using the KP model, and 0.18 using the model
ability with populations unrelated to the calibration set. based on populations 1–17; Fig. 3).
Two approaches were used to test DM model robustness: Hunter b models were also reasonably robust within a
(a) a model updating approach was adopted in which the cultivar. For example, a model developed on KP fruit (popu-
model was iteratively updated with new data and then used lation 6, Table 1) yielded R2v = 0.79 and 0.87, SEP = 7.2 and
in prediction of only the next population before updating with 8.5, and bias = 0.2 and 1.7 when used in prediction of two
that data set and (b) a model created using the first four data subsequent populations of the same cultivar (populations 11
sets was used to predict all subsequent data sets (Fig. 3). and 12 (typical population mean = 40 and S.D. = 18) (data
Note that a different drying method was used for populations not shown). However, when a model developed on Calypso
1–4, relative to subsequent populations. However, there was fruit (populations 16 and 17) was used in prediction of KP
no step change in bias of the predicted DM between these fruit (populations 11and 12), results were poor (R2v = 0.16,
two sets of populations, consistent with a level of conformity SEP = 16.6, bias = −15.3 on population mean = 39.8, and
between the two reference methods. When used in prediction S.D. = 15.7) (data not shown).
of population 5–18 (combined), the population 1–4 model We conclude that the PLSR models for DM and Hunter b
yielded a R2v = 0.79, RMSECV = 0.97 and bias = 0.04 (Fig. 4). are acceptably robust when used with fruit of a single cultivar,
The model updating process produced unstable results but attention must be given to model updating when used in
at first, but a level of stability was reached by addition of prediction of different cultivars.
332 P.P. Subedi et al. / Postharvest Biology and Technology 43 (2007) 326–334

Fig. 2. Mean second derivative of absorbance spectra [d2 ABS] (A), B coef-
ficients for a PLSR model for TSS (open circles) and DM (closed circles)
based on d2 ABS (B), and B coefficients for a PLSR model for Hunter b (open
circles) and visual maturity score (closed circles) based on d2 ABS (C). Data Fig. 3. Flesh DM predictive ability [bias (A), SEP (B) and R (C)] of cali-
from populations 11 and 12, Table 1. bration models across 18 populations. Open circles: populations 1–4 used
for calibration model to predict population 5 and onward; closed circles:
calibration model continually updated (i.e. model on population 1 used to
predict population 2; model on population 1 + 2 used to predict population
3.6. PLS model robustness—fruit ripening 3, and so on).

During ripening, DM content of fruit remained relatively


constant, while TSS increased, consistent with the conversion 3.7. PLSR prediction of eating stage TSS from harvest
of starch to sugar (Fig. 5). Penetrometer readings decreased time spectra
during ripening as flesh firmness declined.
As expected from the similarity of B coefficients for Saranwong et al. (2004) indicated that final (ripe stage)
DM and TSS models, PLSR models on DM worked TSS was more accurately calculated from SWNIRS predicted
well in validation at any stage of fruit ripening (e.g. values of DM and starch than from either variable alone.
R2v = 0.94, RMSEP = 0.90), but models on TSS developed Given that a linear model was used to combine these two
at a given stage of ripeness performed poorly in pre- estimates, it is logical that a PLSR model could be developed
diction of other stages (e.g. R2v = 0.32, RMSEP = 1.14; directly on ripe stage TSS. This approach has the obvious
data not shown). Thus the SWNIR technique is not use- advantage of obviating the need to develop a starch model
ful for assessment of fruit ripening (as opposed to fruit (given the relatively difficulty of the reference technique for
maturity). this character).
P.P. Subedi et al. / Postharvest Biology and Technology 43 (2007) 326–334 333

Table 4
PLSR model calibration statistics for estimation of final TSS (in ripe fruit) from spectra collected of hard green fruit. PLSR models were based on d2 A spectra
Cultivar Population #Sample Mean S.D. r2 DM to TSS #Outliers #Factors R2c RMSECV
KP 19 240 15.9 3.0 0.85 9 5 0.90 0.98
Calypso 20 78 11.9 1.9 0.86 6 6 0.90 0.60
Calypso and KP 19 + 20 318 14.6 3.3 0.55 6 5 0.92 0.84
Calypso, Mareeba 17* 318 14.1 1.8 0.80 7 4 0.90 0.57
Calypso, Bundaberg 18* 63 14.7 2.3 0.95 5 4 0.86 0.81
The correlation coefficient of determination of the regression between NIR predicted DM at hard green stage and final TSS of ripe fruit is also presented.
Populations 19 and 20 were harvested in the NT on 29 October 2005. Populations 17* and 18* were harvested concurrently to 17 and 18 (Table 1).

developed on population 17 data used in prediction of popu-


lation 18 (Table 4) yielded a R2v = 0.92 with SEP = 0.67% and
bias = 1.25%. A cv. KP model developed on population 19
data used in prediction of cv. Calypso population 20 (Table 4)
yielded a R2v = 0.85 with RMSEP = 0.89% and bias = 0.47%.

4. Conclusions and practical implementation

The number of fruit required to estimate the mean of a


consignment can be estimated as the square root of the t
statistic multiplied by the S.D. of the population and divided
by the desired precision of measurement (e.g. Valero and
Ruiz-Altisent, 2000). Given t = 1.96 (desired level of sig-
nificance = 0.025, degrees of freedom = α), and a typical
Fig. 4. Prediction results (predicted vs. actual) for a PLS model based on consignment standard deviation of 2.0% DM (Table 1), in
population 1–4 used in prediction of population 5–18. excess of 16 fruit must be assessed to gain an estimate of
the mean with a precision of only 1% DM. Thus rapid,
non-invasive measurement strategies are essential to the
PLSR calibration models on TSS of ripe fruit, developed
implementation of a quality control program.
using SWNIR spectra collected (non-destructively) of hard
The SWNIRS method can be used to assess ‘maturity’
green mango, were excellent (e.g. R2c = 0.90; Table 4). These
(as indexed by a maturity score) directly, or used to predict
results were generally slightly superior (i.e. higher R2 ) to
other indices of maturity such as DM content or flesh Hunter
the correlation achieved by relating predicted DM in hard
b value. This method thus allows for a practical separation
green fruit to ripe stage TSS. The PLSR models also per-
of fruit on maturity level at harvest, and thus on ripening
formed well in prediction. For example, a cv. Calypso model
time. Further, the SWNIR technique can be used at harvest
to predict the TSS of ripe fruit, and thus the eating quality
of the fruit. The technique can be employed either on-line, to
asses every item of fruit, or using a handheld format, for batch
assessment in the field (prior to harvest) or in pack-house,
in quality control. Further work to assess model robustness
across seasons and growing conditions is anticipated.

Acknowledgements

Our thanks are extended to Tim Ellis and Garry R.


Waldron of W.E. pack-house, David Hamilton, Northern Ter-
ritory Department of Primary Industry, Fisheries and Mines,
NT DPIFM, and Mark Nolan, Northern Territory Depart-
ment of Business, Economic and Regional Development, NT
Fig. 5. Change in fruit attributes with ripening: DM (open circles), TSS
DBERD for their enthusiasm and support of this project. We
(closed circles) and penetrometer (open circles). Population of intact fruit cv.
KP mango, with destructive sampling of 10 fruit at each sampling time (data also acknowledge funding support from Horticulture Aus-
points represent mean and associated standard error). Fruit were exposed to tralia Ltd and Northern Territory Department of Primary
ethylene and ripened at 25 ◦ C. Industry, Fisheries and Mines, NT DPIFM.
334 P.P. Subedi et al. / Postharvest Biology and Technology 43 (2007) 326–334

References Saranwong, S., Sornsrivichai, J., Kawano, S., 2001. Improvement of PLS
calibration for Bix value and dry matter of mango using information
Du, Y.P., Liang, Y.Z., Jiang, J.H., Berry, R.J., Ozaki, Y., 2003. Spectral from MLR calibration. J. Near Infrared Spectrosc. 9, 287–295.
regions selection to improve prediction ability of PLS models by change- Saranwong, S., Sornsrivichai, J., Kawano, S., 2003a. On-tree evaluation of
able size moving window partial least squares and searching combination harvesting quality of mango fruit using a hand-held NIR instrument. J.
moving window partial least squares. Anal. Chem. Acta 501, 183– Near Infrared Spectrosc. 11, 283–293.
191. Saranwong, S., Sornsrivichai, J., Kawano, S., 2003b. Performance of a
Golic, M., Walsh, K.B., Lawson, P., 2003. Short-wavelength near-infrared portable near infrared instrument for Brix value determination of intact
spectra of sucrose, glucose, and fructose with respect to sugar concen- mango fruit. J. Near Infrared Spectrosc. 11, 175–181.
tration and temperature. Appl. Spectrosc. 57, 139–145. Saranwong, S., Sornsrivichai, J., Kawano, S., 2004. Prediction of ripe-stage
Greensill, C.V., Walsh, K.B., 2000. A remote acceptance probe and illumina- eating quality of mango fruit from its harvest quality measured nonde-
tion configuration for spectral assessment of internal attributes of intact structively by near infrared spectroscopy. Postharvest Biol. Technol. 31,
fruit. Meas. Sci. Technol. 11, 1674–1684. 137–145.
Guthrie, J.A., Walsh, K.B., 1997. Non-invasive assessment of pineapple Schmilovitch, Z., Mizrach, A., Hoffman, A., Egozi, H., Fuchs, Y., 2000.
and mango fruit quality using near infra-red spectoscopy. Austr. J. Exp. Determination of mango physiological indices by near-infrared spec-
Agric. 37, 253–263. trometry. Postharvest Biol. Technol. 19, 245–252.
Guthrie, J.A., Walsh, K.B., Reid, D.J., Liebenberg, C.J., 2005. Assessment Valero, C., Ruiz-Altisent, M., 2000. Design guidelines for a quality assess-
of internal quality attributes of mandarin fruit. 1. NIR calibration model ment system of fresh fruit in fruit centers and hypermarkets. Agric. Eng.
development. Austr. J. Agric. Res. 56, 405–416. Int.: CIGR J. Sci. Res. Dev. II, 1–20.
Kawano, S., Watanabe, H., Iwamoto, M., 1992. Determination of Walsh, K.B., Golic, M., Greensill, C.V., 2004. Sorting of fruit using near
sugar content in intact peaches by near infrared spectroscopy with infrared spectroscopy: application to a range of fruit and vegetables for
fibre optics in interactance mode. J. Jpn. Soc. Hort. Sci. 61, 445– soluble solids and dry matter content. J. Near Infrared Spectrosc. 12,
451. 141–148.