Postharvest Biology and Technology 161 (2020) 111078

Contents lists available at ScienceDirect

Postharvest Biology and Technology

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

Assessment of avocado fruit dry matter content using portable near infrared T
spectroscopy: Method and instrumentation optimisation
Phul P. Subedi*, Kerry B. Walsh
Institute for Future Farming Systems, Central Queensland University, Rockhampton, 4701, Australia

ARTICLE INFO ABSTRACT

Keywords: Avocado flesh dry matter content (DMC) is an index of eating quality of ripened fruit, and DMC is also related to
Dry matter fruit maturity, with a (cultivar dependant) minimum DMC recommended for harvest. Based on DMC variation
Fruit quality within the fruit, the outer equatorial region of the fruit was chosen for optical and physical sampling. Three
Handheld handheld near infrared spectrophotometers were compared for in-field non-invasive assessment of DMC, with
Harvest decision
the best results for prediction of independent sample sets obtained using an instrument employing an inter-
Non-invasive
SWNIRS
actance optical geometry and the wavelength range 720−975 nm, with mean centred second derivative of
absorbance spectra (e.g., correlation coefficient of determination, R2, for partial least squares regression model
(PLSR) prediction of an independent test set of 0.71, compared to 0.37 and 0.31 for two reflectance geometry
instruments). This performance difference to the reflectance geometry units was less marked for fruit with skin
removed (e.g., prediction set R2 0.88 for the interactance geometry unit and 0.74 and 0.71 for the reflectance
geometry units). PLSR model performance was examined for models based on cumulative combination of fruit
populations across three growing seasons and four growing locations for a single cultivar model and a combined
two cultivar model. Bias corrected root mean square of error of predictions stabilized in the third season at
approximately 1.5 % dw/fw, with bias varying by approximately 1 %. The coefficients of the PLSR model
stabilised as population size increased, making these values a useful guide to model stability. In-field use was
demonstrated, tracking DMC of fruit on tree from between 14 and 27 % over several months to inform a harvest
timing decision. Use on ripening fruit was also demonstrated. Tracking of known (tagged) fruit was re-
commended over assessment of randomly chosen fruit to reduce bias error in estimation of population DMC
change.

1. Introduction (i.e., eating quality). A (cultivar specific) minimum DMC specification

The minimum (‘physiological’) maturity level of avocado fruit is
achieved when the fruit is able to ripen without shrivelling or devel-
oping a watery taste and rubbery texture (Magwaza and Tesfay, 2015).
For some cultivars, fruit maturity can be indexed by changes in external
skin or stem colour, but this does not hold for all cultivars. A change in
seed coat colour to brown is also an indicator of fruit maturity, but this
character requires destructive assessment. More commonly, maturity is
commonly indexed by the related measures of oil, dry matter content
(DMC) or moisture content (Parodi et al., 2007). Oil content and the
related variable of DMC increase with maturation time, and is related to
the flesh creaminess and smoothness, and thus eating quality, of ri-
pened fruit (Hofman et al., 2013).
Arpaia et al. (2003a) and Gamble et al. (2010) established avocado
fruit DMC specifications by cultivar in context of consumer acceptance
has been set in most significant production regions, to ensure a
minimum maturity before harvest and an acceptable eating quality of
ripened fruit (Magwaza and Tesfay, 2015). In some regions a maximum
DMC specification has also been established, to reduce physiological
disorders in stored fruit. Avocado fruit can remain on tree for up to 12
months without ripening, achieving high DMC levels, but the longer
time on tree is associated with shorter shelf life and susceptibility to
disease rancidity and biennial bearing.
Greater attention has been given to harvest maturity in export or-
ientated industries such as South Africa and New Zealand, with harvest
starting only when fruit reach the minimum maturity (DMC) specifi-
cation. Further, harvested batches of fruit should ideally be of pre-
dictable, even ripening, i.e., the fruit should be of a common maturity.
Sorting of populations of avocado fruit by DMC should provide maturity
segregation.

⁎
Corresponding author.
E-mail addresses: p.subedi@cqu.edu.au (P.P. Subedi), k.walsh@cqu.edu.au (K.B. Walsh).

https://doi.org/10.1016/j.postharvbio.2019.111078
Received 11 September 2018; Received in revised form 8 November 2019; Accepted 19 November 2019
Available online 02 December 2019
0925-5214/ © 2019 Published by Elsevier B.V.
P.P. Subedi and K.B. Walsh
The common method of determination of fruit DMC involves de-
structive sampling for oven drying and is thus time consuming and cost
ineffective, effectively limiting sampling numbers. A number of non-
invasive techniques (e.g., nuclear magnetic resonance, ultrasonic) have
been trialled for avocado DMC assessment, with use of short wave-
length near infrared spectroscopy (SWNIR, approx. 750−1100 nm)
recommended as the most practical method (e.g., Magwaza and Tesfay,
2015).
An early report on the use SWNIR to assess DMC on intact avocado
was made by Schmilovitch et al. (2001) using a reflectance optical
geometry. These authors reported a root mean square of error of pre-
diction (RMSEP) on DMC of 0.9 and 1.3 % (weight basis) for fruit of two
cultivars respectively, with fruit DMC varying between 14 and 23 %.
The lower RMSEP was attributed to the thinner skin of the associated
cultivar. Clark et al. (2003) reported use of an interactance geometry
(using the wavelength range 750−1050 nm) to assess DMC of intact
cultivar Hass fruit, with results for a multiple linear regression (MLR)
model noted to be similar to that for a partial least squares regression
(PLSR) model (correlation coefficient of determination, R2 = 0.88,
RMSEP = 1.8 % compared to R2 > 0.83, RMSEP < 1.8 %; model
based on 12–16 factors). The use of a reflectance geometry was asso-
ciated with a poorer result (R2 < 0.75, RMSEP > 1.9 %, with use of
18–20 factors). Wedding et al. (2010) reported on use of a benchtop
FTNIR spectrometer (wavelength range 830−2500 nm) in assessment
of intact avocado fruit DMC, reporting a R2 = 0.76 and RMSEP = 1.53
%. Li et al. (2018) recently reported use of the a low cost spectrometer
(SCiO from Consumer Physics, Israel) to assess avocado ripeness stage
(with 50–70 % classification accuracy), but DMC was not assessed.
In practical use, a NIRS based model will be used in assessment of
fruit of new (‘independent’) populations, varying in harvest date, lo-
cation and/or cultivar to that of the calibration set. A ‘robust’ model
should perform well despite such sample variation. The majority of
publications either employ fruit from a single population (location,
harvest date, cultivar) or, where data on several populations has been
collected, employ a split of each population to the calibration and va-
lidation sets. For example, Clark et al. (2003) utilised fruit of four
harvests (n = 180), with random splits of each harvest set to create the
calibration (n = 120) and validation (n = 60) sets.
The use of validation sets that are not ‘independent’ of the cali-
bration set will result in over optimistic assessment of the performance
of the technology. For example, Olarewaju et al. (2016) reported on use
of a reflectance geometry and a longer wavelength range
(800−2500 nm) for DMC assessment of intact Hass fruit. When using a
random split of samples to calibration and validation sets, validation
R2 = 0.84; RMSEP = 2.5 % and bias = -0.5 % were noted, but when
the model was based on samples from one year was used in prediction
of samples from the next season, R2 was only 0.53, with RMSEP of 3.8
%. Wedding et al. (2013) considered model robustness across orchards
and seasons, noting a model developed on a single seasons data to
predict samples from subsequent years poorly (R2 between 0.09 and
0.61, RMSEP of 2.6–5.0 %, compared to results for a three season model
in prediction of samples drawn from those years but not used in the
model (R2 = 0.89, RMSEP = 1.4 %). Striving for industry adoption,
Blakely (2016) reported DMC prediction results for a validation set
based on a separate harvest event to that of a calibration set based on
8235 intact Hass fruit assessed using a handheld unit (940−1780 nm).
Validation performance was poor (R2 = 0.40, root mean square of error
of cross validation, RMSECV = 2.9 %).
Several uncertainties are apparent across the body of previous work.
Firstly, the NIRS methods employed are ‘point’ measurements, i.e.,
measurement is made of a point on the fruit. Knowledge of the spatial
variation of the attribute of interest (DMC) within the fruit is therefore
required, in order to assess a volume of the fruit that is represntative of
the whole fruit, or at least that part of the fruit that is of analytical
interest. Secondly, NIRS is a secondary method and must be combined
with a primary reference method. Matching the optically sampled
2
Postharvest Biology and Technology 161 (2020) 111078
volume to the reference method sampled volume is therefore important,
given the variation in DMC within a fruit. This leads to a third con-
sideration, the optical geometry employed in the instrumentation used.
Spectra acquired using a reflectance geometry will carry object surface
information, with diffuse reflectance emanating in majority from within
5 mm depth of apple fruit (e.g., Lammertyn et al., 2000; Peirs et al.,
2002) (and probably from even less depth for avocado fruit given the
denser and thicker skin). Partial transmittance or interactance geome-
tries seem more appropriate for acquisition of useful spectral informa-
tion from greater depths of the fruit (e.g., Schaare and Fraser, 2000).
As a fourth consideration, the literature reports vary in the wave-
length range utilised, probably as a function of instrumentation avail-
able to the researcher. The value of the SWNIR region (750−1100 nm)
is well established in terms of lower absorption by water and thus
gereater effective penetration depth into fruit (e.g., Herold et al., 2009),
however many of the reports are based on instrumentation employing
longer wavelength ranges.
The combined use of a reflectance geometry and a long wavelength
region is likely to result in spectra dominated by absorption features of
the fruit skin, and thus less robust models if skin properties change
between populations. Nonetheless, the ‘proof is in the pudding’, and
any geometry or instrumental wavelength range that results in a good
prediction result of populations independent of the calibration set is
acceptable. In this spirit, Kaur et al. (2017) report a comparison of four
handheld instruments for assessment of apple, kiwifruit and summer-
fruit.
This reasoning leads to a final, fifth, point, that models need to be
tested on independent sets of fruit, not sets drawn from the populations
used in calibration.
In summary, while the potential for assessment of avocado fruit
DMC content is well established, the practicality of the technology re-
mains to be demonstrated in terms of instrumentation type and model
robustness. The current study aims to address the five issues raised
earlier through an assessment of performance of three instruments, in
context of repeatability, wavelength range and optical geometry. Model
robustness is documented across populations varying in cultivar,
growing district and season, and ripening stage. Finally examples of
practical use for in-field estimation of harvest maturity and of DMC of
ripening fruit are given.
2. Materials and methods
2.1. Fruit and reference method
Avocado (Persea americana Mill.) fruit of cultivars Shepard and Hass
were sourced from four farms in north, central and southern
Queensland, Australia, across the 2016, 2017 and 2018 seasons, with a
total of 23 populations, each of approximately 30 fruit (Table 1).
Spectra and associated reference values were generally acquired on
each of two sides of each fruit (Table 1).
Following recording of spectra, a 27 mm diameter, 12 mm deep core
of tissue was extracted from the location on the fruit of spectra acqui-
sition and the skin removed. Dry matter content of the sliced cores was
assessed following drying to constant weight (48 h) in a fan forced oven
operated at 65 °C.
In a separate exercise, ten Hass and five Shepard fruit were sec-
tioned into 24 pieces each and oven dried to allow assessment of DMC
distribution within the fruit. For a further 20 Hass fruit, oven dry matter
content of two outer (0−12 mm) mesocarp cores taken from the
equator of the fruit was compared to that of the rest of the fruit. Spectra
were not acquired of these fruit.
2.2. Acquisition of spectra
Fruit were marked on two sides in an equatorial location, with
spectra and reference cores acquired from these positions. The term
P.P. Subedi and K.B. Walsh Postharvest Biology and Technology 161 (2020) 111078

Table 1
Assessed fruit populations: fruit and spectra number, DMC statistics, source (farm A to D) and cultivar. Populations are numbered in chronological order of sampling.
Sample # refers to the number of paired spectra – reference analyses, SD refers to standard deviation.
Year Population Sample# Spectra# Mean DM(%) SD (%) Source farm Cultivar

2016 1 60 120 22.89 1.37 A Shepard

2 60 120 22.81 1.51 A Shepard
3 100 200 20.85 2.98 A+B Shepard
4a 50 50 24.15 2.20 A+B Shepard
5 40 80 20.54 3.06 B Shepard
6 60 120 24.97 2.20 B Hass
7b 60 120 23.95 3.33 B Shepard
8b 60 120 25.21 2.23 B Hass
9c 60 240 25.21 2.23 B Hass
2017 10b 100 400 21.24 4.81 A+C Shepard
11 58 116 23.16 4.59 C Shepard
12 60 120 17.83 1.35 C Hass
13a 40 40 21.27 1.94 C Shepard
14 60 120 21.94 2.56 C Hass
15b 64 128 27.77 2.89 C Hass
16b 44 88 28.40 2.43 C Hass
17c 40 80 26.43 1.64 C Hass
2018 18 72 144 19.15 3.23 C Shepard
19 100 200 19.88 2.69 A Shepard+Hass
20 90 180 22.49 1.93 C Shepard
21 92 184 19.04 1.29 C Hass
22 100 200 24.64 3.98 D Shepard+Hass
23b 100 200 25.33 2.01 C Hass
Total 1570 3370 23.11 4.28

a
Single scan per side.
b
Fruit scanned at two temperatures (22 °C and 30−35 °C).
c
Fruit scanned at three temperatures (15–18 °C, 22–25 °C and 30−35 °C).

‘equator’ is used to refer to the center-line of the main body of the fruit. instruments, two spectra were acquired at each fruit location, except as
A F750 handheld spectrometer (#15006, Felix Instruments, Camas, noted Table 1.
WA, USA) was used to acquire spectra of all populations (intact fruit),
while for a subset of populations, spectra were also acquired using
2.3. Chemometric and data analysis
MicroNIR (Viavi, San Jose, USA) and SCiO v1.2 (Consumer Physics,
Sand Hill, USA) instruments. For some populations of fruit, spectra
Chemometric analyses (principal component analysis and partial
were re-acquired of the same fruit after removal of skin, or equilibration
least squares regression, PLSR) were performed using The Unscrambler
at different temperatures (15–18, 20–25 and 30−35 °C; Table 1). Re-
software (version 10.3, CAMO, Oslo, Norway). PLSR models based on
peatability was assessed of all instruments as SD of 20 repeated scans of
F750 and MicroNIR were based on a 9 point Savitsky Golay second
a reference polytetrafluoroethylene (PFTE, also commonly referred to
derivative treatment of absorbance data using the range 729–975 nm
as Teflon™) tile.
(after Guthrie et al., 2005) and 920–1676 nm, respectively. Spectral
In a comparison of instruments, it is useful to compare the instru-
data was mean centred. Cross validation based on 20 random groupings
ments in terms of the optical geometry employed, the referencing
was used for selection of the number of factors used in the PLSR model,
procedure, the wavelength range, the pixel resolution (for array in-
based on the minimum Y variance or a rate of decrease of Y variance of
struments, the average wavelength range covered by a single element of
less than 1 %.
the detector array), the wavelength resolution (typically expressed as
For spectra acquired using the SCiO, the SCiO Developers licence
the full width half maximum, FWHM, of a spectral line as recorded by
was used to create PLSR models on the SCiO cloud site. This licence
the spectrometer) and the repeatability of the instrument (expressed as
does not allow downloading of spectra, precluding analysis with The
the standard deviation of absorbance across the wavelength range
UnScrambler. The method of selection of number of PLS factors is not
monitored of 20 repeated measures of the reference material) (Herold
reported for this software (www.sciolab.consumerphysics.com), and as
et al., 2009).
the software does not report cross validation results, raw calibration Rc2
The F750 instrument operates over the wavelength range of
and root mean square of error of calibration (RMSEC) was reported for
310−1100 nm with a pixel resolution of approximately 3 nm and an
all instruments, to allow comparison. As spectral data was not acces-
optical resolution of 8−13 nm (depending on wavelength). This in-
sible from this package, screen shots of spectra were acquired.
strument employs a partial transmittance geometry and self references
with every sample measurement using an internal shutter. The
MicroNIR operates over the wavelength range of 908−1676 nm with a 3. Results
pixel resolution of approximately 6 nm and an optical resolution of <
12.5 nm. The detector views an illuminated sample surface (reflectance 3.1. Within fruit DMC distribution
geometry). The unit was referenced to an external plate after every 20
samples. The SCiO operates over the wavelength range of Within fruit DMC distribution was similar in the assessed cultivars,
740−1070 nm with a pixel resolution of approximately 1 nm and an and averaged results are presented. Mesocarp DMC content decreased
optical resolution of 30 nm. On this instrument, the detector input is basipetally with the fruit, being higher at the stem than the tip end by
located adjacent (2 mm separation) to the illumination source. The unit between 3.5 and 7.7 %. As a percentage of stem end mesocarp DMC,
was placed onto the surface of the fruit with a gap of a few millimteres mesocarp from the equator of the fruit was 89 + 1.4 % (mean and
between fruit and lamp and detector (i.e., reflectance geometry). For all standard error) and mesocarp from the acropetal end of the fruit was 85
+ 2.5 % for the fruit considered. Outer mesocarp DMC levels varied

3
P.P. Subedi and K.B. Walsh
Fig. 1. F750 spectra of hard green avocados (pop# 12 Hass), with (A,C) skin on (B,D) skin off (C,D), presented as absorbance (A,B) and second derivative (C,D).
between two opposed points on the fruit ‘equator’ by between 0.08 and
1.2 % (SD of 0.6 %). There was also a depth DMC gradient, with skin at
approximately 27 % and seed at 31 %, while flesh DMC increased
modestly with depth. Relative to the outer 0.5–1.0 cm layer of equa-
torial mesocarp, DMC content of subsequent layers was 105 and 104 %
with increasing depth (for fruit with an average DMC in the outer
equatorial mesocarp of 17.5 %).
Fig. 2. F750 absorbance spectra of (A) fruit with skin on (top group) and with
skin removed (bottom group) and (B) the difference spectra (intact – skin re-
moved) of individual fruit (pop # 12 Hass).
4
Postharvest Biology and Technology 161 (2020) 111078
Across 20 fruit, the correlation coefficient of determination (R2) of
the average of oven DMC assessed of two 10 mm deep cores taken at the
‘equator’ of the fruit to the DMC of the remainder of the fruit was 0.82,
with slope 1.04.
3.2. Fruit spectra
Fruit skin absorbed visible light strongly, masking differences
Fig. 3. F750 (A) absorbance and (B) second derivative of absorbance (d2A)
spectra of an intact Hass fruit at different stages of ripening.
P.P. Subedi and K.B. Walsh Postharvest Biology and Technology 161 (2020) 111078

Fig. 4. SCiO v1.2 normalised reflectance spectra of (A) intact and (B) skin removed fruit, and second derivative of reflectance spectra of (C) intact and (D) skin off
fruit.

between fruit in this wavelength region (F750 spectra, Fig. 1A, C).
Absorbance spectra of fruit with skin removed demonstrated greater
variation in the visible region, particularly around the 650−700 nm
range. More variation was also apparent around the 960 nm absorbance
peak. Additional features were apparent in the second derivative (d2)
spectra, both in the visible (around 550 nm) and shortwave NIR regions
(around 740, 840 and 910 nm). The difference spectra of intact and skin
removed fruit was dominated by a broad feature between 500 and
700 nm and at 960 nm (Fig. 2).
As fruit ripened, changes in the visible region of spectra were no-
table, particularly around 500−580 nm (Fig. 3). Absorbance increased
at 510 nm and decreased over the 520−580 nm range as fruit ripened.

Table 2
Avocado fruit DMC PLSR calibration statistics for three handheld NIR devices
(F750, MicroNIR and SCiO v1.2), using second derivative spectra of intact fruit
and fruit with skin removed. Populations 6, 8 and 9 (Table 1) were employed
(n = 180; SD = 3.92 % DM. #f refers to number of PLSR factors used in the
PLSR model.
Device Intact fruit Skin removed

R2C RMSEC(%) #f R2C RMSEC(%) #f

F750 0.84 1.03 4 0.89 0.85 4

Micro-NIR 0.82 1.11 12 0.83 1.04 5
SCiO v1.2 0.59 1.72 4 0.82 1.08 5

Table 3
Avocado fruit DMC prediction statistics for three handheld portable NIR de-
vices, using intact fruit and fruit with skin removed. Models were based on data
of populations 6 and 8, n = 120, and used in prediction of population 9, n = 60
(Table 1).
Device Intact fruit Skin removed

Rp2 RMSEP(%) Bias(%) RP 2 RMSEP(%) Bias(%)

F750 0.71 1.37 0.64 0.88 1.96 0.45

Fig. 5. MicroNIR absorbance spectra of (A) intact fruit, (B) fruit with skin re- Micro-NIR 0.37 2.71 1.95 0.74 3.08 −1.74
moved, and (C) difference spectra (intact – skin removed). Scio-v1.2 0.31 2.05 2.34 0.71 2.31 −1.89

5
P.P. Subedi and K.B. Walsh Postharvest Biology and Technology 161 (2020) 111078

Table 4 3.4. Model robustness

PLSR calibration statistics of avocado fruit DMC (intact and with skin removed)
across three seasons and two varieties. The Shepard population consisted of Further work was limited to the use of the F750 instrument, with
populations 1–5, 7, 10, 11, 13 and 18–20 and 22 (Table 1, n = 502, SD = 3.56 additional populations of fruit assessed both as intact fruit and with skin
%). The Hass population consisted of populations 6, 8, 9, 12, 14–17 and 21–23
removed (Table 4). Dry matter content models based on fruit with skin
(Table 1, n = 436, SD = 3.01 %). The combined Shepard and Hass population
removed were characterised by a RMSECV < 1.21 % and R2cv > 0.88 for
consisted of populations 1–23 (n = 939, SD = 3.67 %). SD refers to standerd
either cultivar singly or for the two cultivar model (Table 4). For models
deviation. # f refers to number of factors in PLSR model.
based on intact fruit, R2 remained above 0.83 and RMSECV below 1.4 %.
Population Intact fruit Skin removed In another approach, a DMC model was developed in one season
across 6 populations and used in prediction of the remaining three
SD(%) R2cv RMSECV(%) #f R2cv RMSECV(%) #f
populations of that season (Fig. 6, left panels), while another model was
Shepard 3.56 0.83 1.25 10 0.88 1.20 5 developed across three years and 22 populations, and used in prediction
Hass 3.01 0.85 1.12 9 0.91 1.16 4 of a further population (Fig. 6, right panels). Prediction results were
Shepard and Hass 3.67 0.84 1.38 9 0.90 1.21 8
similar (RMSEP around 1.5 %).
The PLSR model regression coefficients of these models were rela-
tively smooth, with the most consistency between the two models and
The SCiO instrument captured spectra over the range
highest weightings in the 890−970 nm range (Fig. 7). Differences in
740−1070 nm, with features apparent at 825 and 950 nm (Fig. 4). The
weightings between the two models were apparent at higher and lower
instrument and associated cloud based software (Developer version)
wavelengths.
does not offer the option of processing Reflectance to Absorbance (i.e., –
As an estimate of the number of populations required to develop a
log of Reflectance), so features are inverted compared to Figs. 1–3.
model robust in prediction of subsequent independent populations, a
The MicroNIR instrument captured spectra over the range
chronologically cumulative model was built across the available po-
908–1675 nm, with absorbance features apparent at 960, 1170 and
pulations, with each cumulative model used in prediction of the next
1440 nm (Fig. 5). Difference spectra were dominated by a
population as an independent test set (i.e., model developed on popu-
1400−1500 nm feature.
lation 1 used in prediction of population 2, 1 + 2 in prediction of 3, 1–3
in prediction of 4, and so on, up to populations 1–22 in prediction of
population 23) (Fig. 8). Bias corrected RMSEP stabilised at around 1.5
3.3. Comparison of handheld instruments
% after aggregation of approximately 16 populations, while R2 varied
in relation with population SD (R2 = 1 – (SEP/SD)2). Biases of around 1
Spectrophotometer repeatability, assessed as the SD of repeated
% were noted even with large calibration sets.
spectra of a PFTE tile at the wavelength of highest signal, was 0.04,
Similar results were obtained for cumulative population DMC
0.22 and 0.40 mAbsorbance units for the F750, MicroNIR and SCiO
models based on a single cultivar (Sheperd) (Fig. 9). SEP plateaued at
instruments respectively.
approximately 1.5 % after aggregation of 8 populations, while biases of
PLSR models on DMC were constructed and tested using the same
around 1 % were again noted even with larger calibration sets.
population sets for the three handheld instruments. Model calibration
The PLSR b coefficients of the Shepard population cumulative
statistics were improved with use of fruit with skin removed. All results
models (Fig. 10) varied initially as populations were added, but a sta-
(intact and skin removed, all instruments) acheived an R2C > 0.8, except
bilisation of the coefficients was reached after combination of ap-
for the SCiO when used with intact fruit (Table 2). When models were
proximately 10 populations (from across three seasons), as demon-
used in prediction of an independent test set, the F750 provided the
strated in a plot of the R2 of the final model (based on populations
best result (lowest bias and SEP, highest R2), particularly with the intact
1–13) on those of preceding models (Fig. 11).
fruit set (Table 3).

Fig. 6. PLSR model cross validation results for NIR

estimated DMC and associated reference values (oven
DM) (panels A and C) and prediction results for an
independent test set (panels B and D). For panel A, the
model was based on data of populations 10–15
(n = 964, SD = 4.2 %), with prediction (panel B) of
populations 16–18 (n = 312; SD = 4.3 %). For panel C,
the model was based on data of populations 1–22
(n = 3118; SD = 3.6 %), with prediction of population
23 (n = 200; SD = 4.0 %).

6
P.P. Subedi and K.B. Walsh
Fig. 7. PLS regression b-coefficients for the models based on data from (A) a
single year and from (B) three years, as presented in Fig. 6.
3.5. Ripening fruit
Fruit lost weight during ripening at a linear rate (R2 = 0.99, 0.14 g/
fruit/day) in the storage conditions imposed. As fruit ripened, NIR es-
timated DMC increased to a lesser extent than DMC estimated by ad-
justing the day 1 estimate by fruit weight (Fig. 12). The R2 between NIR
predicted DMC at day 8 of ripening and oven dry DMC was 0.93 and
RMSE 0.89 % (data not shown).
3.6. In-field use
Repeated (n = 20) measures of NIR DMC of a single fruit gave a SD
of 0.13 % on a mean value of 25.5 %. Twenty tagged fruit were mon-
itored weekly in each of 12 and 10 orchards of Shepard and Hass cul-
tivars, respectively. DMC of individual fruit demonstrated a linear in-
crease (average R2 = 0.92 for the set of 30 fruit illustrated in Fig. 13A),
with averaging of fruit per orchard block resulting in smoother time
series data (Fig. 13B, C).
4. Discussion
4.1. Sampling considerations
Design of a sampling strategy requires knowledge of spatial varia-
tion in attribute level. Mesocarp DMC content varied between stem and
tip end of the fruit, with relatively small variation around the ‘equator’
of the fruit. The DMC content of the outer mesocarp of the fruit
‘equator’ was demonstrated to represent the whole fruit. Therefore
sampling on the ‘equator’ of the fruit is recommended for NIR DMC
assessment.
To a lesser extent, DMC also varied with depth of tissue. This var-
iation is problematic in terms of matching the optically and physically
sampled volumes, as the physically sampled core was only an
7
Postharvest Biology and Technology 161 (2020) 111078
Fig. 8. Statistics for population cumulative PLSR models in prediction of DMC
of the next population, across two cultivars, three seasons and two preparation
conditions (intact, skin removed). Results for intact fruit are presented as solid
symbols, while those for fruit with skin removed (populations 17) are presented
as open symbols). (A) Rp2, (B) Bias corrected RMSEP and (C) Bias. Inset in panel
provides variety (H for Hass, S for Shephard) and indication of year for each
population. Population numbering as in Table 1.
approximation of the optically sampled volume for the interactance
geometry and 720−975 nm wavelength region. This depth variation is
likely to be even more problematic in context of reflectance geometries,
which will result in a shallower optically sampled volume, with the
majority of received signal coming from less than 5 mm depth (Guthrie
and Walsh, 1997; Lammertyn et al., 2000). Further, the longer wave-
length range of the MicroNIR is associated with higher levels of ab-
sorbance than for the shorter wavelength NIR (Herschel region), and
thus a lower effective penetration. Thus there is a mismatch between
the optically and physically sampled volumes for the SciO and Micro-
NIR instruments, contributing to poorer calibration and validation
statistics for these instruments.
4.2. The contribution of the skin
The spectra of fruit contained features related to carotenoids
(around 550 nm), chlorophyll (around 680 nm), water and carbohy-
drates (OeH vibrations at 740, 840, 960 and 1440 nm), and to lipids
P.P. Subedi and K.B. Walsh
Fig. 9. Statistics for Shepard population cumulative PLSR models in prediction
of DMC of the next population for intact fruit. (A) Rp2, (B) Bias corrected
RMSEP and (C) Bias.
(CeH vibrations at 910 and 1200 nm). The difference spectrum of in-
tact and skin removed fruit effectively represents the skin. This spec-
trum was characterised by a peak interpreted as due to pigmentation
(carotenoids and chlorophyll), with a smaller contribution at 960 nm
(and at 1440 nm for the MicroNIR spectra) due to water. The con-
tribution of the skin to the spectrum of an intact fruit is large, and thus
changes in skin properties have potential to impact DMC predictions.
Fig. 10. Regression coefficients of each of the 12 cumulative Shepard population PLSR models (as presented in Fig. 9).
8
Postharvest Biology and Technology 161 (2020) 111078
The fruit skin (exocarp) possessed a higher DMC content (at approx.
30 %) than the mesocarp. Obviously, NIR assessment of intact fruit
involves passage of light through the exocarp, into the mesocarp, and
through the exocarp a second time. Any change in skin thickness or
DMC content will thus impact prediction of mesocarp DMC. We hy-
pothesised that changes in skin characteristics associated with growing
conditions contribute to bias in NIR DMC estimation of independent
sets. However, PLSR models based on populations with skin removed
demonstrated a similar level of bias as models based on populations of
fruit with skin removed when used in prediction of new populations
(Fig. 10C). Evidently the bias is associated with changes in mesocarp
rather than in exocarp.
4.3. Comparison of handheld spectrophotometers
The three instruments assessed varied in key characteristics of op-
tical geometry, wavelength range, wavelength resolution and repeat-
ability. Better prediction results were obtained for the unit employing
interactance optics and operating in the visible-SWNIR. The poorer
performance of the units using reflectance rather than interactance
geometry was attributed to a greater contribution of signal from the
skin rather than the mesocarp for the reflectance geometry. Consistent
with this conclusion, the relative performance of the reflectance geo-
metry units was improved in assessment of fruit with skin removed.
Use of a longer wavelength range will accentuate this issue, with the
longer wavelength range characterised by higher levels of absorption in
biological (water based) tissue, thus decreasing the optically sampled
volume. Presumably this issue also contributed to the need to use more
factors in the PLSR model for the MicroNIR unit. The outcome of poor
predictive performance for this longer wavelength range is consistent
with the report of Blakey (2016).
For the SCiO unit, relative performance will have been impacted by
the poorer wavelength resolution and repeatability of this unit, relative
to the other instruments. However, better results may be possible with
this unit with use of a standard chemometrics package rather than the
SCiO on-line PLSR package (as Li et al., 2018, reported for a kiwifruit
TSS model).
4.4. Model robustness
Model performace across multiple populations will be impacted by
P.P. Subedi and K.B. Walsh Postharvest Biology and Technology 161 (2020) 111078

Fig. 11. Correlation coefficient of determination for the linear regression of

Shepherd PLSR model B coefficients for each cumulative model up to 1–12 with
the b coefficients of the model based on populations 1-13.

change in either the instrument (e.g., Acharya et al., 2016) or sample

characteristics. The inclusion of multiple populations is required to
increase model robustness for assessment of intact fruit attributes, with
recommendation for inclusion of three seasons of data offered as a
rough rule of thumb (e.g., Wedding et al., 2013). Factors that could
influence fruit optical characteristics (e.g., cell size, cell packing ar-
rangements, skin thickness, chemical matrix) include cultivar, growing
conditions and stage of ripening. With 23 populations of fruit of dif-
ferent harvest dates, seasons, farms and cultivars, the current study
allows some insight into fruit associated factors that impact model
performance.
Acharya et al. (2014) has previously reported on the population
structure required for a PLSR on fruit DMC to be robust to variation in
fruit temperature, in terms of the ratio of samples scanned at a range of
temperatures. These conclusions were implemented in the current
study. For the 23 population cumulative model predictions, the highest
SEPs were recorded for predictions of the populations (7 and 8) which
introduced temperature variation into the sample set (spectra of po-
pulations 1–6 being acquired at room temperature). Subsequent po-
pulations carrying temperature variation maintained a low SEP.
Fig. 13. Time course of NIR DMC assessed of fruit on tree for (A) individual
A large bias was incurred in prediction of a population of a variety
Shepard fruit, (B) average (n = 20) for each of 12 and 10 orchards of (B)
(Hass, population 6) not included in the cumulative model, and in
Shepard and (C) Hass cultivars, respectively. Fruit were monitored weekly until
prediction of a population with multiple sample temperatures (popu- harvest.
lation 17). The prediction of the first population of the 2018 year also
incurred a large bias.
The cumulative model achieved a stable SEP after addition of ap-
proximately 10 and 15 populations for the one and two cultivar po-
pulation sets, respectively, however DMC bias variation of around 1 %
occurred even in later populations However, other than the broad ad-
vice that a calibration set should include populations from the range of
conditions from which fruit will be predicted in future (growing con-
dition, cultivar), we can offer little insight in terms of rules for popu-
lation structure in the way that Acharya et al. (2014) has provided for
temperature variation. To decrease the time involved in creating a ro-
bust model, Anderson et al. (2017) utilised fruit from one orchard and
one season, but from trees subject a range of growing conditions (im-
posed water stress treatments), and a similar strategy could also be
applied to avocado.
The ‘finalisation’ of development of a cumulative population model
can be guided by the rate of change in model coefficients (Fig. 8).
However, as a secondary method, NIRS models require continued
checking, with potential need for maintenance (updating). The suit-
ability of NIRS models can also be monitored by routines that fit the
Fig. 12. Average fruit weight (dashed line), NIR estimated dry matter content unknown sample spectra to the spectral base of the calibration set (e.g.,
(open circle) and dry matter content estimated from the day 1 NIR DMC and Mahalanobois distance based measures such as leverage or Global H,
daily weight loss (solid line) (n = 12). Results are expressed as a percentage of ‘GH’) (Herold et al., 2009, Garido-Varo et al., 2019).
the day 1 value.

9
P.P. Subedi and K.B. Walsh
4.5. Ripening fruit
The ripening of mango fruit is associated with changes in light pe-
netration into the fruit, and thus the optically sampled volume. As a
likely consequence, NIR DMC models are sensitive to the stage of ri-
pening (Subedi and Walsh, 2011). Clark et al. (2003) hinted at a similar
issue for avocado fruit when they attributed poorer R2 and RMSEP of
avocado DMC models based on late season fruit to changes in fruit
texture. However, avocado fruit mesocarp has less visual change in
optical properties during ripening than mango fruit, and a NIR DMC
model developed using unripened fruit predicted DMC of ripened fruit
well (R2 of 0.93). An estimate of fruit DMC with time of ripening was
based on an NIR DMC value assessed at day 1 and adjusted for weight
(water) loss. This calculation, however, does not account for DMC de-
crease due to respiratory C loss. Avocado fruit respiratory rate (CO2
evolution) is reported at approximately 50 mL/g DW/hour for Hass
fruit (Eaks, 1978), equivalent to a weight loss of 1.2 % over a 12 d
period. A further study could measure actual respiration rate to resolve
discrepancy between the measured NIR DMC and the DMC estimated
from weight change and estimated respiration rate (Fig. 12).
4.6. Conclusions - practical use
The DMC of the outer equatorial region of an avocado fruit was
demonstrated to represent DMC of the whole fruit, and thus this portion
is recommended for spectrometer out performed reflectance geometry
based units and was used in field contitions.
The NIR predicted DMC RMSEP of 1.5 % and bias of 1 % was higher
than that seen in thinner skinned fruit such as mango (e.g., RMSEP ap-
proximately 1.0 %, Subedi et al., 2007), but avocado fruit demonstrate a
large range in DMC during maturation (from 14 to 27 % observed in this
study, Fig. 13), and thus the technology can be used in determining order
of harvest of orchard blocks. Knowledge of the rate of DMC increase
allows for forecast of when a crop will reach a target specification. In the
current study, an average rate of DMC increase of 0.44 % DMC per week
was observed, with an average R2 of 0.98 on the linear fit.
To estimate rate of DMC increment, the use of repeated measure-
ment of of tagged fruit over time, rather than random slection of fruit
each week. This protocol is recommended to avoid differences in pre-
diction bias between fruit, allowing a lower sampling number to re-
present the population. The number of samples required to estimate the
mean of a population with a known SD can be estimated as:
t . SD 2 determining avocado fruit maturity. Food Bioproc. Technol. https://doi.org/10.1007/
n=
e (1)
where t is the t statistic (1.96 for n > 30 and a confidence probability of
0.95) and e is the acceptable error (eqn. 5.17, Steel and Torrie, 1980).
Given a typical population lot SD on DMC of 3 % (Table 1) and as-
suming an acceptable error of 1.5 % (equivalent to NIR DMC SEP),
n = 16. Therefore the practice of measuring 20 fruit per lot (orchard
block) is reasonable. This is a manageable number of assessments, e.g.,
for a farm making weekly assessments of 50 blocks of 1000 trees each.
Such data can also feed models of variation during maturation such as
that proposed by such as Jordan and Loeffen (2013).
Declaration of Competing Interest
Co-author Walsh acknowledges involvement in the development of
the F750 instrument with Felix Instruments. We believe this potential
bias is balanced by the input of the other author.
Acknowledgements
This work was supported by project ST15005, Multiscale mon-
itoring of tropical fruit trees, funded by Hort Innovation (Australia) and
10
Postharvest Biology and Technology 161 (2020) 111078
the Federal Department of Agriculture and Water Resources under the R
&D for Profit program. We thank producers Groves Grown, Simpsons,
MMM Mangoes and Tim Heath for supply of fruit. We acknowledge
involvement in the development of the F750 instrument (e.g., Greensill
and Walsh, 2000) and linkage to Felix Instruments, manufacturer of the
F750 instrument.
References
Acharya, U.K., Walsh, K.B., Hayes, C., Subedi, P.P., 2016. Spectrophotometer aging and pre-
diction of fruit attributes. J. Near Infrared Spectrosc. 24, 337–344.
Acharya, U.K., Walsh, K.B., Subedi, P.P., 2014. Robustness of partial least squares models to
change in sample temperature: II. Application to fruit attributes. J. Near Infrared Spectrosc.
22, 287–295.
Anderson, N., Subedi, P.P., Walsh, K.B., 2017. Manipulation of mango fruit dry matter content
to improve eating quality. Sci. Hortic. 226, 316–321.
Arpaia, M.L., Boreham, D., Collin, S., Field, K., Sievert, J., 2003a. Hass maturity project.
Proceedings of the California Avocado Research Symposium 120–124.
Blakey, R.J., 2016. Evaluation of avocado fruit maturity with a portable near-infrared spec-
trometer. Postharvest Biol. Technol. 121, 101–105.
Clark, C.J., McGlone, V.A., Requejo, C., White, A., Woolf, A.B., 2003. Dry matter determination
in ‘Hass’ avocado by NIR spectroscopy. Postharvest Biol. Technol. 29, 301–308.
Eaks, I.L., 1978. Ripening, respiration and ethylene production of ‘Hass’ avocado fruits at 20° to
40°C. J. Am. Soc. Hortic. Sci. 103, 576–578.
Gamble, J., Harker, F.R., Jaeger, S.R., White, A., Bava, C., Beresford, M., et al., 2010. The
impact of dry matter, ripeness and internal defects on consumer perceptions of avocado
quality and intentions to purchase. Postharvest Biol. Technol. 57, 35–43.
Garido-Varo, A., Garcia-Olmo, J., Fearn, T., 2019. A note on Mahalanobis and related distance
measures in WinISI and The Unscrambler. J. Near Infrared Spectrosc. 27, 253–258.
Greensill, C.V., Walsh, K.B., 2000. Optimization of instrumentation precision and wavelength
resolution for the performance of NIR calibrations of sucrose in a water—cellulose matrix.
Appl. Spectrosc. 54, 426–430.
Guthrie, J.A., Walsh, K.B., Reid, D.J., Liebenberg, C.G., 2005. Assessment of internal quality
attributes of mandarin fruit. 1. NIR calibration model development. Aust. J. Agric. Res. 56
(405-), 416.
Guthrie, J.A., Walsh, K.B., 1997. Non-invasive assessment of pineapple and mango fruit quality
using near infra-red spectroscopy. Aust. J. Exp. Agric. 37 (253), -263.
Herold, B., Kawano, S., Sumpf, B., Tillmann, P., Walsh, K.B., 2009. Chapter 3. VIS/NIR spec-
troscopy. In: Zude, M. (Ed.), Optical Monitoring of Fresh and Processed Agricultural Crops.
CRC Press, Boca Raton USA, pp. 141–249 ISBN 978-1-4200-5402-6.
Hofman, P.J., Bower, J., Woolf, A., 2013. Harvesting, packing, postharvest technology, trans-
port and processing. In: Schaffer, S., Wolstenholme, B.N., Whiley, A.W. (Eds.), The
Avocado: Botany, Production and Uses. CAB International, Oxfordshire, Oxford, pp.
489–540.
Jordan, R.B., Loeffen, M.P.F., 2013. A new method for modelling biological variation using
quantile functions. Postharvest Biol. Technol. 86, 387–401.
Kaur, H., Kunnemeyer, R., McGlone, A., 2017. Comparison of hand-held near infrared spec-
trophotometers for fruit dry matter assessment. J. Near Infrared Spectrosc. 25, 267–277.
Lammertyn, J., Peirs, A., Baerdemaeker, J.D., Nicolai, B., 2000. Light penetration properties of
NIR radiation in fruit with respect to non-destructive quality assessment. Postharvest Biol.
Technol. 18, 121–132.
Li, M., Qian, Q., Shi, B., Medlicott, J., East, A., 2018. Evaluating the performance of a consumer
scale SCiO molecular sensor to predict quality of consumer products. Postharvest Biol.
Technol. 145, 183–192.
Magwaza, L.S., Tesfay, S.Z., 2015. A review of destructive and non-destructive methods for
s11947-015-1568-y.
Olarewaju, O.O., Bertlinga, I., Magwazab, L.S., 2016. Non-destructive evaluation of avocado
fruit maturity using near infrared spectroscopy and PLS regression models. Sci. Hortic. 199,
229–236.
Parodi, G., Sanchez, M., Daga1, W., 2007. Correlation of oil content, dry matter and pulp
moisture as harvest indicators in Hass avocado fruits (Persea americana Mill) growing
under two conditions of orchards in Chincha-Peru. Proceedings VI World Avocado
Congress (Actas VI Congreso Mundial del Aguacate) 2007 Viña Del Mar, Chile. 12-16 Nov.
2007. ISBN No 978-956-17-0413-8.
Peirs, A., Tirry, J., Verlinden, B., Darius, P., Nicolai, B.M., 2002. Effect of biological variability
on the robustness of NIR models for soluble solids content of apples. Postharvest Biol.
Technol. 28, 269–280.
Schaare, P.N., Fraser, D.G., 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, 175–184.
Schmilovitch, Z., Hoffman, A., Egozi, H., El-Batzi, R., Degani, C., 2001. Determination of
avocado maturity by near infrared spectrometry. Acta Hortic. 562, 175–179.
Steel, R.G.D., Torrie, J.H., 1980. Principles and Procedures of Statistics : A Biometrical
Approach, 2nd edition. McGraw-Hill Kogakusha Ltd, Tokyo.
Subedi, P.P., Walsh, K.B., Owens, G., 2007. Prediction of mango eating quality at harvest using
short-wave near infrared spectroscopy. Postharvest Biol. Technol. 43, 326–334.
Subedi, P.P., Walsh, K.B., 2011. Assessment of sugar and starch in intact banana and mango
fruit by SWNIR spectroscopy. Postharvest Biol. Technol. 62, 238–245.
Wedding, B.B., Wright, C., Grauf, S., White, R.D., Tilse, B., Gadek, P., 2013. Effects of seasonal
variability on FT-NIR prediction of dry matter content for whole ‘Hass’ avocado fruit.
Postharvest Biol. Technol. 75, 9–16.
Wedding, B.B., White, R.D., Grauf, S., Wright, C., Tilse, B., Hofman, P., et al., 2010. Non-de-
structive prediction of ‘Hass’ avocado dry matter via FT-NIR spec-troscopy. J. Sci. Food
Agric. 91, 233–238.