Food Chemistry 351 (2021) 129314

Contents lists available at ScienceDirect

Food Chemistry
journal homepage: www.elsevier.com/locate/foodchem

Predicting oil content in ripe Macaw fruits (Acrocomia aculeata) from unripe
ones by near infrared spectroscopy and PLS regression
Ulisses F. Oliveira a, Annanda M. Costa b, Jussara V. Roque a, Wilson Cardoso a,
Sergio Y. Motoike c, Marcio H.P. Barbosa c, Reinaldo F. Teofilo a, *
a
Multivariate Chemical Data Analysis Laboratory, Department of Chemistry, Universidade Federal de Viçosa, 36570-900 Viçosa, MG, Brazil
b
Campus Ponta Porã, Instituto Federal de Mato Grosso do Sul, 79100-510 Campo Grande, MS, Brazil
c
Department of Agronomy, Universidade Federal de Viçosa, 36570-900 Viçosa, MG, Brazil

A R T I C L E I N F O A B S T R A C T

Keywords: A method for early quantification of unripe macaw fruits oil content using near-infrared spectroscopy (NIR) and
Macaw palm partial least squares (PLS) is presented. After harvest, the fruit takes about 30 days to reach its maximum oil
Multivariate regression accumulation. The oil content was quantified thirty days after harvest using Soxhlet extraction. PLS models were
Near-infrared spectroscopy
built using NIR spectra of shell obtained five days after harvest (Shell5). The Shell5 model was compared with
Partial least squares
models built using NIR spectra of the shell (Shell30) and mesocarp thirty days after harvest (Pulp30). Ordered
Early prediction
predictors selection was used to select the most informative variables. The best models presented root mean
square error of prediction and correlation coefficient of prediction of 4.87% and 0.89 for Shell5; 5.83% and 0.85
for Shell30; 4.76% and 0.92 for Pulp30. Thus, the anticipated prediction of oil content could reduce the time and
costs of macaw palm quality control and storage.

1. Introduction
The demand for vegetable oils has increased over the last decades
(Sajjadi, Raman, & Arandiyan, 2016; Senger et al., 2017; Yang, Zhang,
Li, Yu, Mao, Wang, & Zhang, 2018). It may be attributable to food
consumption and biofuel production related to the continual increase in
the world’s population and demand for cleaner and renewable energy
sources (Muscat, de Olde, de Boer, & Ripoll-Bosch, 2019; Prates-Valério,
Celayeta, & Cren, 2019).
The oleaginous palm, Acrocomia aculeata, commonly known as
macaw palm, is found in savanna-like vegetation, semi-deciduous sea­
sonal forests, and deforested areas, among other regions of Central and
South America (Lanes, Motoike, Kuki, Nick, & Freitas, 2015; Machado,
Figueiredo, & Guimarães, 2016). The estimated oil production from this
crop is up to 6200 kg/ha, similar to oil palm (Elaeis guineensis), the major
current source of vegetal oil. However, the macaw palm does not
compete with rainforest and fertile land and presents higher storage
stability (Evaristo et al., 2016; Prates-Valério et al., 2019). The macaw
palm mesocarp oil comprises about 60% oleic acid and 29% palmitic
acid and is rich in carotenoids, which play an essential role as an anti­
oxidant agent (Nunes, Favaro, Galvani, & Miranda, 2015). It can be used
in food, pharmaceutical, cosmetics, and biodiesel production due to its
higher oxidation stability and operation at low temperatures (Montoya,
Motoike, Kuki, & Couto, 2016; Simiqueli, de Resende, Motoike, &
Henriques, 2018). The cake obtained after oil extraction can be used as
animal feed since it contains no toxic compounds, and the shell could be
used for energy generation (Tilahun et al., 2019). Therefore, macaw
palm has the potential to be an economically viable crop.
Macaw palm is still under the process of domestication, and most
studies aim at its breeding to increase its oil yield (Pires, dos Santos

Abbreviations: AutoiOPS, automatic interval ordered predictors selection; FeediOPS, feedback interval ordered predictors selection; hMod, number of latent
variables of the model; hOPS, number of latent variables for ordered predictors selection; LOD, limit of detection; MPGAB, macaw palm germplasm active bank; NIR,
near-infrared spectroscopy; Nlvs, number of latent variables; nVars, number of variables; OC, oil content; OPS, ordered predictors selection; PLS, partial least squares;
Pred, prediction; Pulp30, spectra of macaw mesocarp post-harvest thirty days post-harvest; R, correlation coefficient (Rc) of calibration, (Rcv) of cross-validation, and
(Rp) of prediction; RMSE, root mean square error (RMSEC) of calibration, (RMSECV) of cross-validation and (RMSEP) of prediction; SEL, selectivity; SEN, sensitivity;
Shell5, spectra of macaw shell five days post-harvest of, Shell30, spectra of macaw shell thirty days post-harvest.
* Corresponding author.
E-mail addresses: ulisses.oliveira@ufv.br (U.F. Oliveira), jussara.roque@ufv.br (J.V. Roque), wilson.cardoso@ufv.br (W. Cardoso), motoike@ufv.br
(S.Y. Motoike), barbosa@ufv.br (M.H.P. Barbosa), rteofilo@ufv.br (R.F. Teofilo).

https://doi.org/10.1016/j.foodchem.2021.129314
Received 26 May 2020; Received in revised form 28 December 2020; Accepted 5 February 2021
Available online 18 February 2021
0308-8146/© 2021 Elsevier Ltd. This article is made available under the Elsevier license (http://www.elsevier.com/open-access/userlicense/1.0/).
U.F. Oliveira et al.
Fig. 1. Macaw palm (Acrocomia aculeata) (A), NIR set up to acquire the macaw fruit’s shell spectra (B), and NIR set up to acquire the macaw fruit’s mesocarp
spectra (C).
Souza, Kuki, & Motoike, 2013). An important characteristic of this fruit
is that it continues to increase its oil content after harvest (Costa et al.,
2018; de França, Reber, Meireles, Machado, & Brunner, 1999; Mon­
selise, 2018). The current harvest practices based on the collection of
dropped fruits and its oil extraction without adequate handling and
storage have resulted in a low-quality oil. Besides, acceptable practices
can increase the fruit’s oil content by over 20% during 30 days of storage
after harvest (Evaristo et al., 2016; Prates-Valério et al., 2019; Tilahun
et al., 2019).
The oil content is often quantified using solvent extraction (Danlami,
Arsad, Ahmad Zaini, & Sulaiman, 2014; Richter et al., 1996). This
method is laborious, time-consuming, and destructive, and they use a
considerable amount of sample and toxic solvents to perform the
extraction (Richter et al., 1996).
Due to the continual increase in oil content, long maturation time,
and limitations of the conventional procedures for oil quantification, a
method capable of early predicting the oil content of the ripe fruits
would be advantageous for decision-making regarding the selection of
high yield fruits, transportation, and storage. Considering the routine
analyses carried out at breeding programs and quality assessments at
industrial sites, a quick, inexpensive, non-destructive, and reliable
method is required to estimate the oil content of numerous samples
(Evaristo et al., 2016).
Noninvasive and non-destructive methods such as near-infrared
spectroscopy (NIR) coupled with partial least squares (PLS) regression
have been widely used to quantify agronomic properties and assess the
quality of agricultural and food products (Ferreira, Galão, Pallone, &
Poppi, 2014; Tilahun et al., 2018). This technique provides rapid in­
formation on physical and chemical properties (Montoya et al., 2016;
Nunes et al., 2015; Pires et al., 2013) with minimal or no sample
preparation with a high predictive ability for unknown samples (Pas­
quini, 2003).
Several works have presented the use of NIR to quantify the oil
content of plants, such as avocado (Olarewaju, Bertling, & Magwaza,
2016), olive (Altieri, Matera, Genovese, & Di Renzo, 2020; Saha &
Jackson, 2018), palm fruits (Sudarno, 2017), maize (Tallada, Palacios-
Rojas, & Armstrong, 2009), and citrus (Steuer, Schulz, & Läger, 2001).
Previous work reported the application of VIS-NIR region from 30,770
to 9300 cm− 1 to predict the oil content of the macaw palm in different
maturation stages using the mesocarp (Matsimbe, Motoike, Pinto, Leite,
2
Food Chemistry 351 (2021) 129314
& Marcatti, 2015). However, studies involving early prediction from
plant products are still scarce due to their high complexity and possible
matrix changes. Studies about the early prediction of sugarcane features
(de Porto, 2019) and cassava productivity (Vitor, Diniz, Morgante,
Antônio, & de Oliveira, 2019) can be found in the literature. Early
predicting a fruit property is essential to assess the fruit before reaching
its full maturity, improving its quality and saving time and storage costs.
Thus, the goal of this work was to establish for the first time a non-
destructive, fast, inexpensive, and reliable method capable of early
predict the mesocarp oil content of ripe macaw fruits (30 days after
harvest) using NIR spectra of their respective unripe fruits (5 days after
harvest) and PLS regression. Ordered predictors selection (OPS) (Roque,
Cardoso, Peternelli, & Teófilo, 2019; Teófilo, Martins, & Ferreira, 2009)
was used to improve the models’ quality by selecting more informative
and predictive variables. The method proposed was compared with
methods using the NIR spectra of ripe fruits.
2. Experimental
2.1. Macaw palm samples
The macaw fruits used in this study were part of the macaw palm
germplasm active bank (MPGAB) of the Universidade Federal de Viçosa
(UFV) located in the municipality of Araponga, Minas Gerais State (lat
20◦ 40′ 1′′ S, long 42◦ 31′ 15′′ W, alt ~980 m). The MPGAB is registered
under the number 084/2013 – SECEX/CGEN and consists of 253 ac­
cessions, collected in several parts of Brazil. This bank is considered the
largest MPGAB in the world.
The fruits were collected in the MPBAG, consisting of 44 accessions
(half-sib families). Three plants were collected for each access, and five
fruits were chosen from each plant for oil content analysis. In total, 660
fruits were collected; however, some of them were lost in transportation
or damaged; therefore, only 630 fruits were used. The fruits were taken
to the post-harvest laboratory, where they remained stored at room
temperature for ripening. After ripening, the fruits were frozen at
− 20 ◦ C freezer to stop the metabolism for conservation and further oil
quantification.
U.F. Oliveira et al.
2.2. Spectral analysis
Due to the logistics of transporting the macaw fruits, only after five
days, the near-infrared (NIR) spectral analysis was performed on the
shell (Shell5). Spectra were taken from the shell to avoid damage to the
fruit before reaching its full maturity. After harvest, considering good
storage practices, the macaw palm takes thirty days to reach its maturity
and, therefore, its maximum oil accumulation (Tilahun et al., 2019).
The Shell5 model result was compared to models built from, NIR
spectral analysis was performed on the shell (Shell30) and mesocarp
(Pulp30) of the fruits after 30 days of maturation. Fig. 1A shows the
fruits before harvest, and Fig. 1B and C refer to how the NIR spectra were
obtained for the macaw shell and mesocarp, respectively.
The NIR spectra were obtained using a Fourier transform NIR (FT-
NIR) 660 spectrometer (Agilent Technologies, Santa Clara, United
States) with a reflectance-integrating sphere accessory from PIKE
Technologies. The range investigated was 10,000–4000 cm− 1 with an
increment of 2 cm− 1. The spectra were obtained using the software Pro
Resolutions Version 5.1 and stored as log (1/R), where R is the reflec­
tance collected. For each NIR spectrum (Shell5, Shell30, and Pulp30)
acquired in the instrument, 32 scans were performed to increase the
signal to noise ratio, and the average of these scans stored.
For Shell5 and Shell30, each fruit was directly placed on the in­
strument window without sample preparation and scanned in two po­
sitions. Each position was measured three times, and the averaged
spectrum of the two positions was used to modeling. For Pulp30, the
fruit shell was removed using a knife to expose the mesocarp, and then
the spectrum was obtained, placing the mesocarp directly on the in­
strument window. A spectrum was obtained from three different
mesocarp positions, and the averaged spectrum was used to modeling.
In total, 630 spectra were used for each model, Shell5, Shell30, and
Pulp30. The spectra were acquired in a laboratory with a controlled
temperature at 21 ◦ C.
2.3. Oil content quantification
Oil extraction was performed using n-hexane (Sigma-Aldrich, San
Luis, United States) in a Soxhlet extractor (Marconi, Piracicaba, Brazil)
according to the adapted procedure from Analytical Norms of Adolfo
Lutz Institute (IAL, 2008). The fruit’s mesocarps were dried in a venti­
lated oven at 65 ◦ C for 72 h. After drying, 5 g of sample was placed in a
filter paper cartridge and arranged in the Soxhlet extractor containing
150 mL of n-hexane; the extraction was carried out for eight hours. In
sequence, the n-hexane extract was transferred to a beaker, placed in an
oven at 105 ◦ C for 24 h for n-hexane evaporation. Finally, the remaining
oil was cooled down to room temperature and weighed.
Oil extraction was performed for each fruit, which generated an oil
content (OC) value for each sample (Equation (1)).
( )
M0
(1)
OC = 100
MS
where OC stands for oil content percent, M0 stands for oil mass in grams,
and Ms stands for the total mesocarp mass in grams.
2.4. Multivariate regression models
The y vector (dependent variable) is associated with oil content
values. The X matrix of the NIR spectra (independent variables) was
different for each condition (Shell5, Shell30, and Pulp30), but the y
vector was the same. The Kennard and Stone algorithm (Kennard &
Stone, 1969) was used to split the samples into calibration and predic­
tion sets. For all models, 504 and 126 samples were selected for the
calibration and prediction sets, respectively.
The spectra were imported to the Matlab 2019a environment (The
Mathworks, Natick, USA). An inverse regression model (y = Xb) was
3
Food Chemistry 351 (2021) 129314
built using the PLS regression. Random cross-validation with ten splits
was used. Besides the raw data, two preprocessing methods, mean center
and autoscale, eight transformations, smoothing, first derivative, second
derivative, multiplicative scatter/signal correction (MSC), detrend,
normalize, baseline, standard normal variate scaling (SNV). Their
combinations were tested to find the best regression model.
Two OPS methods (Roque et al., 2019), i.e., AutoiOPS and FeediOPS,
were applied to select more predictive variables to improve the models.
Firstly, the original response variables (X matrix columns) were sub­
divided into intervals, each one containing fifty variables. In each in­
terval, informative vectors that contain information about the location
of the best response variables for prediction were calculated. The orig­
inal variables in each interval were differentiated according to the cor­
responding absolute values of the informative vector elements. The
differentiated variables were sorted in descending order. An initial
subset of two variables (window) for each interval was defined to build
PLS models. The initial subset was extended by adding one variable
(increment) over each window until all variables were taken. The var­
iable subsets in each interval were compared using the cross-validation
parameters calculated during validations, and the best variable subset
was defined for each interval. These variable selection steps were per­
formed using several informative vectors, and for each interval, the best
vector was chosen. This method is called AutoiOPS.
Additionally, after performing the AutoiOPS selection, the selected
variables returned to a new selection, and this procedure is called Fee­
diOPS. This method was carried out until relative differences between
two consecutive RMSECV values were less than 0.02% (Roque et al.,
2019).
Two optimum number of latent variables (nlvs) are employed in this
work: one representing the component number for model building
(hMod) and the other representing the component number employed to
generate the best informative vector in the OPS method (hOPS) (Roque
et al., 2019). The OPS methods were applied using the algorithms
available at www.deq.ufv.br/chemometrics. All the calculations were
performed in Matlab 2019a.
2.5. Figure of merit
The quality of the models was evaluated using the parameters root
mean square error (RMSE) and correlation coefficient (R), which were
calculated by Equations (2) and (3), respectively.
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√ N
√∑
RMSE = √ (yi − y1 )2 /N (2)
i
( )
∑N
̂yi − ̂ y (yi − y)
i=1
R = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
)2̅√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (3)
(
∑N ∑N
̂y i − ̂ y (yi − y)2
i=1 i=1
where ̂ y i and ̂
y are the estimated value and mean of estimated values,
respectively, y andy are the observed values and the mean of the
observed values, respectively, and N represents the number of samples.
When internal cross-validation (CV) is applied, N represents the number
of samples in the cross-validation set, and the error and correlation co­
efficient are called root mean square error of cross-validation (RMSECV)
and correlation coefficient of cross-validation (RCV), respectively. For
the external validation, N represents the number of samples predicted
(P), and, in this case, the error and correlation coefficients are named
correlation root mean square error of prediction (RMSEP) and correla­
tion coefficients of prediction (RP), respectively.
The sensitivity (SEN), selectivity (SEL), and limit of detection (LOD)
were calculated according to Equations (4), 5, and 6, respectively.
SEN = 1/‖b‖ (4)
U.F. Oliveira et al. Food Chemistry 351 (2021) 129314

where ‖b‖ represents the Euclidean norm of the regression

coefficientsΔ(α, β).
Each model was verified for chance correlation. The y vector
(dependent variables) was randomized 10,000 times, and models were
built using these randomizations. The correlation coefficient between
the reference and predicted values was evaluated. If the model’s cor­
relation parameter built using the authentic y was isolated from those of
the models using the randomized y, then the model did not occur by
chance.

3. Results and discussion

3.1. Oil content

The samples’ mesocarp oil content ranged from 29.5 to 74.5%, with
an average of 55.7% and a 10.7 standard deviation. The variability in oil
content is related to the plants’ genetic variability (Costa et al., 2018).
This variability is also due to the different regions where the fruits were
harvested.

3.2. Spectral interpretation

Fig. 2. (A), (C), and (E) NIR spectra for Shell5 (Xshell5), Shell30 (Xshell30), and
Pulp30 (Xpulp30) respectively; (B), (D) and (F) represents the spectra trans­
formed using first derivative and smoothing. Region 1: 9000 to 7700 cm− 1;
Region 2: 7200 to 6500 cm− 1; Region 3: 5800 cm− 1; Region 4: 5670 cm− 1;
Region 5: 5200 cm− 1; Region 6: 4500 to 4000 cm− 1.
where ||b|| is the Euclidean norm of the vector of regression coefficients
of the PLS model (Assis et al., 2017; Roque, Dias, & Teófilo, 2017).
SEL = nasi /‖xi ‖ (5)
where nasi is the absolute scalar value of the net analytical signal for
sample i and || xi || represents the Euclidean norm of the instrumental
response vector for sample i (Assis et al., 2017; Rambo, Amorim, &
Ferreira, 2013; Roque et al., 2017).
For all raw spectra shown in Fig. 2 (A, C, and E), the band from 9000
to 7700 cm− 1 (region 1) is related to the second overtone of C–H
stretching, specifically related to oleic acid. The higher absorbance of
this band is related to an oil rich in unsaturated fatty acids. The band
from 7200 to 6500 cm− 1 (region 2) is related to water and combination
bands. The bands in 5800 (region 3) and 5670 cm− 1 (region 4) are
related to saturated and unsaturated fatty acids, in this case, palmitic
and oleic acids, respectively. The band around 5200 cm− 1 (region 5) is
related to water absorbance and combination bands. The region from
4500 to 4000 cm− 1 (region 6) is related to combinations of bands of
edible oils (Hourant, Baeten, Morales, Meurens, & Aparicio, 2000;
Ozaki, McClure, & Christy, 2006; Sato, Kawano, & Iwamoto, 1991).
The spectra of Shell5 and Shell30 are similar, the latter presenting a
more scattered profile. It may be due to physical changes in the shell
during the maturation. The Pulp30 spectra differ from Shell5 and Sell30
spectra, presenting a more accentuated absorbance in the regions from
8700 to 8000 cm− 1, 7200 to 6500 cm− 1, 6000 to 5500 cm− 1, 5300 to
5000 cm− 1, and 4500 to 4000 cm− 1. This difference may be because the
Shell5 and Shell30 spectra were obtained from the shell, which is not
rich in oil and water, while the Pulp30 spectra were obtained directly on
the pulp, which is rich in oil and water. For that reason, the spectra of
Pulp30 presented a more accentuated profile in the regions related to
fatty acids and water.

LOD = (3.3 × nas × ‖b‖ )/max(b) (6)

Table 1
Statistical parameters and figures of merit for the PLS models with all variables (Full) and variables selected using AutoiOPS and FeediOPS.
Shell5 Shell30 Pulp30

Full AutoiOPS FeediOPS Full AutoiOPS FeediOPS Full AutoiOPS FeediOPS

nlvs 6 6 6 7 7 7 6 6 6
nVars* 3113 260 290 3113 200 155 3113 340 220
RMSEC (g/g) 3.81 3.33 3.31 3.65 3.49 3.99 3.14 2.52 2.19
Rc 0.93 0.95 0.96 0.94 0.94 0.93 0.95 0.97 0.98
RMSECV (g/g) 4.97 4.21 4.16 4.97 4.32 4.75 3.89 3.24 2.82
Rcv 0.88 0.91 0.92 0.88 0.91 0.89 0.93 0.95 0.96
RMSEP (g/g) 5.28 5.50 4.87 6.11 5.83 5.95 4.76 5.49 5.47
Rp 0.88 0.88 0.89 0.83 0.85 0.82 0.92 0.88 0.88
SEN 0.54 0.20 0.21 0.50 0.11 0.12 0.83 0.29 0.22
SEL 0.08 0.10 0.10 0.07 0.21 0.20 0.13 0.13 0.13
LOD (g/g) 26.19 11.43 18.32 25.89 13.23 7.01 13.66 8.83 8.66

nlvs: number of latent variables; nVars: number of variables; Rc: correlation coefficient of calibration; Rcv: correlation coefficient of cross-validation; Rp: correlation
coefficient of prediction; RMSEC: root mean square error of calibration; RMSECV: root mean square error of cross-validation; RMSEP: root mean square error of
prediction; SEN: sensitivity; SEL: selectivity; LOD: limit of detection.

4
U.F. Oliveira et al.
Fig. 3. Measured and predicted oil content for (A) Shell5; (D) Shell30; (G) Pulp30 (● represents the regression sample set and the prediction sample set). Relative
error for calibration set. (B) Shell5; (E) Shell30; (H) Pulp30. Relative error for prediction set (C) Shell5; (F) Shell30; (I) Pulp30.
3.3. Modeling
The original NIR spectra and their respective transformations are
shown in Fig. 2. Mathematical transformations in the matrix rows are
essential to increase the signal-to-noise ratio and reduce systematic er­
rors from undesired light scattering variations. The best preprocessing
and transformation combination that produced the smallest RMSECV
and highest RCV was chosen. The transformation that presented the best
results for all data sets was the first derivative, followed by smoothing.
Statistical parameters and figures of merit for the PLS models using
all variables (Full) and selected variables (AutoiOPS and FeediOPS) are
shown in Table 1. For Shell5, the model using the variables selected
using AutoiOPS presented a lower RMSECV value than the Full model.
However, a smaller RMSEP was achieved with the variables selected
using FeediOPS, which means that it is a more predictive model than the
model using all variables. For the Shell30 model, the variables selected
improved both RMSECV and RMSEP values; the AutoiOPS was the best
approach for this model. For the Pulp30 model, the variables selected
improved the RMSECV values but not the RMSEP values, so the Full
model was chosen.
Comparing the models Shell5-FeediOPS, Shell30-AutoiOPS, and
5
Food Chemistry 351 (2021) 129314
Pulp30-Full from the results shown in Table 1, it is possible to observe
that the Shell5-FeediOPS model was quite similar to the models built
with spectra of the ripe fruits, i.e., Shell30-AutoiOPS and Pulp30-Full.
Furthermore, the Shell5-FeediOPS model presented RMSECV and RMSEP
values lower and Rcv and Rp values higher than the Shell30-AutoiOPS
model.
Matsimbe et al. determined the oil content in macaw palm using VIS-
NIR spectra (30,770 to 9300 cm− 1) of the mesocarp (Matsimbe et al.,
2015) and found RMSEP, Rcv, and nlvs values equal to 7.08%, 0.88, 9,
respectively. Considering the Pulp30, the model presented better pa­
rameters, RMSEP of 4.76% and Rcv of 0.92, with six nlvs, a more
parsimonious model. It may be due to the wavelength range used and a
more extensive preprocessing screening, resulting in more corrected
spectra, and therefore a better model.
For all models, low SEN values can be observed. A decrease in their
values is noticed in the models using the selected variables. This may be
due to the use of more interpretative and predictive variables. As ex­
pected, low SEL values were observed for all models as the spectra
contain information about other components, which could be consid­
ered interferents. For both AutoiOPS and FeediOPS models, this param­
eter was higher as more selected regions of the NIR spectra related to the
U.F. Oliveira et al.
Fig. 4. Correlation coefficient of cross-validation (Rcv) versus correlation coefficient of calibration (Rc) (chance correlation) for (A) Shell5; (B) Shell30; (C) Pulp30.
Fig. 5. Variables selected for (A) Shell5 using FeediOPS and (B) Shell30 using AutoiOPS.
property were used. Except for the Pulp30 model, where the selection
did not improve the SEL values. The LOD values also decreased for the
models using the selected variables.
Table 1 shows that FeediOPS models were chosen for Shell5, Auto­
iOPS for Shell30, and the Full model for Pulp30, as variable selection did
not improve the RMSEP value. Fig. 3 presents the measured versus
predicted oil content values (3A, 3D, and 3G) using the NIR spectra for
the best model of each data set. A linear fit can be observed for all
datasets, indicating that the models can accurately predict the oil con­
tent in macaw fruit.
Fig. 3 also shows the relative errors of calibration (Fig. 3B, E, and H)
and prediction (Fig. 3C, F, and I). Most of the relative errors were less
than 10%, showing that the model Shell5 could be used to satisfactory
predict the amount of oil present in the fruit 25 days before its
extraction.
The above models were evaluated in order to verify if there was a
correlation by chance. Fig. 4 shows that the models presented in this
work were distinguished from the models built by randomizing the y
vector. Therefore, it can be concluded that the models were not obtained
by chance.
Fig. 5 presents the variables selected for Shell5 (Fig. 5A) and Shell30
(Fig. 5B). The variables selected for Shell5 and Shell30 were obtained by
FeediOPS and AutoiOPS, respectively. The selected regions, 1000–9000,
8500–7500 cm− 1 6100–5500 cm− 1, were similar for both sets as the
models were built using shell NIR spectra, differing only in the days after
6
Food Chemistry 351 (2021) 129314
harvest. Also, it can be observed that the selected regions are related to
absorbance bands of fatty acids, and the bands related to water absor­
bance were not selected.
Thus, the prediction of the macaw fruit’s oil content was possible
through NIR of the shell or mesocarp and PLS regression. All models
presented reliable predictive capacity. However, the model Shell5 was
the most advantageous, as the ripe fruit’s maximum oil content could be
early predicted 25 days before without damaging the fruit. This method
can improve the harvest and storage practices aimed at improving the
oil’s quality and reducing the time needed for decision-making related
to selecting the best fruits. Besides, this work opens the way to further
studies relating to the early prediction of fruit properties.
4. Conclusions
From this study, NIR spectroscopy and PLS regression proved to be
capable of early predict the maximum oil content of macaw fruits using
spectra from the shell of unripe fruits. Comparing the results of the
Shell5 with the Shell30 and Pulp30 models showed that the models
using the unripe fruits were similar to the ones using the ripe fruits. The
OPS improved the models Shell5 and Shell30 by selecting more pre­
dictive and informative variables. The method presented is faster, non-
destructive, solvent-free, and low-cost when compared with reference
methods.
U.F. Oliveira et al.
CRediT authorship contribution statement
Ulisses F. Oliveira: Data curation, Formal analysis, Investigation,
Methodology, Validation, Visualization, Writing - original draft, Writing
- review & editing. Annanda M. Costa: Conceptualization, Data cura­
tion, Formal analysis, Funding acquisition, Investigation, Methodology.
Jussara V. Roque: Supervision, Validation, Visualization, Writing -
review & editing, Writing - original draft. Wilson Cardoso: Supervision,
Validation, Visualization, Writing - review & editing, Writing - original
draft. Sergio Y. Motoike: Supervision, Funding acquisition, Project
administration, Resources. Marcio H.P. Barbosa: Supervision, Funding
acquisition, Project administration, Resources. Reinaldo F. Teofilo:
Conceptualization, Funding acquisition, Project administration, Re­
sources, Supervision, Validation, Visualization, Writing - review &
editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
Acknowledgments
This study was financed in part by the Coordenação de Aperfeiçoa­
mento de Pessoal de Nível Superior - Brasil (CAPES) - Finance Code 001
and Fundação de Amparo à Pesquisa do Estado de Minas Gerais
(FAPEMIG) (Project: CEX - APQ-02254-15).
References
Altieri, G., Matera, A., Genovese, F., & Di Renzo, G. C. (2020). Models for the rapid
assessment of water and oil content in olive pomace by near-infrared spectrometry.
Journal of the Science of Food and Agriculture, 100(7), 3236–3245. https://doi.org/
10.1002/jsfa.10361.
Assis, C., Ramos, R. S., Silva, L. A., Kist, V., Barbosa, M. H. P., & Teófilo, R. F. (2017).
Prediction of Lignin Content in Different Parts of Sugarcane Using Near-Infrared
Spectroscopy (NIR), Ordered Predictors Selection (OPS), and Partial Least Squares
(PLS). Applied Spectroscopy, 71(8), 2001–2012. https://doi.org/10.1177/
0003702817704147.
Costa, A. M., Motoike, S. Y., Corrêa, T. R., Silva, T. C., Coser, S. M., de Resende, M. D. V.,
& Teófilo, R. F. (2018). Genetic parameters and selection of macaw palm (Acrocomia
aculeata) accessions: An alternative crop for biofuels. Crop Breeding and Applied
Biotechnology, 18(3), 259–266. https://doi.org/10.1590/1984-70332018v18n3a39.
Danlami, J. M., Arsad, A., Ahmad Zaini, M. A., & Sulaiman, H. (2014). A comparative
study of various oil extraction techniques from plants. Reviews in Chemical
Engineering, 30(6), 605. https://doi.org/10.1515/revce-2013-0038.
Evaristo, A. B., Grossi, J. A. S., Pimentel, L. D., de Melo Goulart, S., Martins, A. D., dos
Santos, V. L., & Motoike, S. (2016). Harvest and post-harvest conditions influencing
macauba (Acrocomia aculeata) oil quality attributes. Industrial Crops and Products,
85, 63–73. https://doi.org/10.1016/j.indcrop.2016.02.052.
Ferreira, D. S., Galão, O. F., Pallone, J. A. L., & Poppi, R. J. (2014). Comparison and
application of near-infrared (NIR) and mid-infrared (MIR) spectroscopy for
determination of quality parameters in soybean samples. Food Control, 35(1),
227–232. https://doi.org/10.1016/j.foodcont.2013.07.010.
de França, L. F., Reber, G., Meireles, M. A. A., Machado, N. T., & Brunner, G. (1999).
Supercritical extraction of carotenoids and lipids from buriti (Mauritia flexuosa), a
fruit from the Amazon region. The Journal of Supercritical Fluids, 14(3), 247–256.
https://doi.org/10.1016/S0896-8446(98)00122-3.
Hourant, P., Baeten, V., Morales, M. T., Meurens, M., & Aparicio, R. (2000). Oil and Fat
Classification by Selected Bands of Near-Infrared Spectroscopy. Applied Spectroscopy,
54(8), 1168–1174. https://doi.org/10.1366/0003702001950733.
IAL. (2008). Métodos químicos e físicos para análise de alimentos. In O. Zenebon, N. S.
Pascuet, & P. Tiglea (Eds.), Normas Analíticas do Instituto Adolf Lutz (4th ed., pp.
116–118). Instituto Adolfo Lutz.
Kennard, R. W., & Stone, L. A. (1969). Computer Aided Design of Experiments.
Technometrics, 11(1), 137–148. https://doi.org/10.1080/00401706.1969.10490666.
Lanes, E. C. M., Motoike, S. Y., Kuki, K. N., Nick, C., & Freitas, R. D. (2015). Molecular
Characterization and Population Structure of the Macaw Palm, Acrocomia aculeata
(Arecaceae), Ex Situ Germplasm Collection Using Microsatellites Markers. Journal of
Heredity, 106(1), 102–112. https://doi.org/10.1093/jhered/esu073.
Machado, W., Figueiredo, A., & Guimarães, M. F. (2016). Initial development of
seedlings of macauba palm (Acrocomia aculeata). Industrial Crops and Products, 87,
14–19. https://doi.org/10.1016/j.indcrop.2016.04.022.
Matsimbe, S. F. S., Motoike, S. Y., Pinto, F. A. de C., Leite, H. G., & Marcatti, G. E. (2015).
Prediction of oil content in the mesocarp of fruit from the macauba palm using
7
Food Chemistry 351 (2021) 129314
spectrometry1. Revista Ciência Agronômica, 46(1), 21–28. https://doi.org/10.1590/
S1806-66902015000100003.
Monselise, S. (2018). Handbook of Fruit Set and Development (Boca Raton). CRC Press.
https://doi.org/10.1201/9781351073042.
Montoya, S. G., Motoike, S. Y., Kuki, K. N., & Couto, A. D. (2016). Fruit development,
growth, and stored reserves in macauba palm (Acrocomia aculeata), an alternative
bioenergy crop. Planta, 244(4), 927–938. https://doi.org/10.1007/s00425-016-
2558-7.
Muscat, A., de Olde, E. M., de Boer, I. J. M., & Ripoll-Bosch, R. (2019). The battle for
biomass: A systematic review of food-feed-fuel competition. Global Food Security,
April, 100330. https://doi.org/10.1016/j.gfs.2019.100330.
Nunes, A. A., Favaro, S. P., Galvani, F., & Miranda, C. H. B. (2015). Good practices of
harvest and processing provide high quality Macauba pulp oil. European Journal of
Lipid Science and Technology, 117(12), 2036–2043. https://doi.org/10.1002/
ejlt.201400577.
Olarewaju, O. O., Bertling, I., & Magwaza, L. S. (2016). Non-destructive evaluation of
avocado fruit maturity using near infrared spectroscopy and PLS regression models.
Scientia Horticulturae, 199, 229–236. https://doi.org/10.1016/j.
scienta.2015.12.047.
Ozaki, Y., McClure, W. F., & Christy, A. A. (2006). Near-infrared spectroscopy in food
science and technology. John Wiley & Sons.
Pasquini, C. (2003). Near Infrared Spectroscopy: Fundamentals, practical aspects and
analytical applications. Journal of the Brazilian Chemical Society, 14(2), 198–219.
https://doi.org/10.1590/S0103-50532003000200006.
Pires, T. P., dos Santos Souza, E., Kuki, K. N., & Motoike, S. Y. (2013). Ecophysiological
traits of the macaw palm: A contribution towards the domestication of a novel oil
crop. Industrial Crops and Products, 44, 200–210. https://doi.org/10.1016/j.
indcrop.2012.09.029.
Porto, N. de A., Roque, J. V., Wartha, C. A., Cardoso, W., Peternelli, L. A., Barbosa, M. H.
P., & Teófilo, R. F. (2019). Early prediction of sugarcane genotypes susceptible and
resistant to Diatraea saccharalis using spectroscopies and classification techniques.
Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 218, 69–75.
https://doi.org/10.1016/j.saa.2019.03.114.
Prates-Valério, P., Celayeta, J. M. F., & Cren, E. C. (2019). Quality Parameters of
Mechanically Extracted Edible Macauba Oils (Acrocomia aculeata) for Potential
Food and Alternative Industrial Feedstock Application. European Journal of Lipid
Science and Technology, 121(5), 1800329. https://doi.org/10.1002/ejlt.201800329.
Rambo, M. K. D., Amorim, E. P., & Ferreira, M. M. C. (2013). Potential of visible-near
infrared spectroscopy combined with chemometrics for analysis of some constituents
of coffee and banana residues. Analytica Chimica Acta, 775, 41–49. https://doi.org/
10.1016/j.aca.2013.03.015.
Richter, B. E., Jones, B. A., Ezzell, J. L., Porter, N. L., Avdalovic, N., & Pohl, C. (1996).
Accelerated Solvent Extraction: A Technique for Sample Preparation. Analytical
Chemistry, 68(6), 1033–1039. https://doi.org/10.1021/ac9508199.
Roque, J. V., Cardoso, W., Peternelli, L. A., & Teófilo, R. F. (2019). Comprehensive new
approaches for variable selection using ordered predictors selection. Analytica
Chimica Acta, 1075, 57–70. https://doi.org/10.1016/j.aca.2019.05.039.
Roque, J. V., Dias, L. A. S., & Teófilo, R. F. (2017). Multivariate Calibration to Determine
Phorbol Esters in Seeds of Jatropha curcas L. Using Near Infrared and Ultraviolet
Spectroscopies. Journal of the Brazilian Chemical Society, 28(8), 1506–1516.
Saha, U., & Jackson, D. (2018). Analysis of moisture, oil, and fatty acid composition of
olives by near-infrared spectroscopy: Development and validation calibration
models. Journal of the Science of Food and Agriculture, 98(5), 1821–1831. https://doi.
org/10.1002/jsfa.8658.
Sajjadi, B., Raman, A. A. A., & Arandiyan, H. (2016). A comprehensive review on
properties of edible and non-edible vegetable oil-based biodiesel: Composition,
specifications and prediction models. Renewable and Sustainable Energy Reviews, 63,
62–92. https://doi.org/10.1016/j.rser.2016.05.035.
Sato, T., Kawano, S., & Iwamoto, M. (1991). Near infrared spectral patterns of fatty acid
analysis from fats and oils. Journal of the American Oil Chemists’ Society, 68(11),
827–833. https://doi.org/10.1007/BF02660596.
Senger, E., Bohlinger, B., Esgaib, S., Hernández-Cubero, L. C., Montes, J. M., & Becker, K.
(2017). Chuta (edible Jatropha curcas L.), the newcomer among underutilized crops:
A rich source of vegetable oil and protein for human consumption. European Food
Research and Technology, 243(6), 987–997. https://doi.org/10.1007/s00217-016-
2814-x.
Simiqueli, G. F., de Resende, M. D. V., Motoike, S. Y., & Henriques, E. (2018). Inbreeding
depression as a cause of fruit abortion in structured populations of macaw palm
(Acrocomia aculeata): Implications for breeding programs. Industrial Crops and
Products, 112, 652–659. https://doi.org/10.1016/j.indcrop.2017.12.068.
Steuer, B., Schulz, H., & Läger, E. (2001). Classification and analysis of citrus oils by NIR
spectroscopy. Food Chemistry, 72(1), 113–117. https://doi.org/10.1016/S0308-8146
(00)00209-0.
Sudarno, Silalahi, D. D., Risman, T., Widyastuti, B. L., Davrieux, F., Yuan, Y. Y., &
Caliman, J. P. (2017). Rapid determination of oil content in dried-ground oil palm
mesocarp and kernel using near infrared spectroscopy. Journal of Near Infrared
Spectroscopy, 25(5), 338–347. https://doi.org/10.1177/0967033517732679.
Tallada, J. G., Palacios-Rojas, N., & Armstrong, P. R. (2009). Prediction of maize seed
attributes using a rapid single kernel near infrared instrument. Journal of Cereal
Science, 50(3), 381–387. https://doi.org/10.1016/j.jcs.2009.08.003.
Teófilo, R. F., Martins, J. P. A., & Ferreira, M. M. C. (2009). Sorting variables by using
informative vectors as a strategy for feature selection in multivariate regression.
Journal of Chemometrics, 23(1), 32–48. https://doi.org/10.1002/cem.1192.
Tilahun, S., Park, D. S., Seo, M. H., Hwang, I. G., Kim, S. H., Choi, H. R., & Jeong, C. S.
(2018). Prediction of lycopene and β-carotene in tomatoes by portable chroma-meter
U.F. Oliveira et al.
and VIS/NIR spectra. Postharvest Biology and Technology, 136, 50–56. https://doi.
org/10.1016/j.postharvbio.2017.10.007.
Tilahun, W. W., Grossi, J. A. S., Favaro, S. P., Sediyama, C. S., Goulart, S. D. M.,
Pimentel, L. D., & Motoike, S. Y. (2019). Increase in oil content and changes in
quality of macauba mesocarp oil along storage. OCL, 26, 20. https://doi.org/
10.1051/ocl/2019014.
Food Chemistry 351 (2021) 129314
Vitor, A. B., Diniz, R. P., Morgante, C. V., Antônio, R. P., & de Oliveira, E. J. (2019). Early
prediction models for cassava root yield in different water regimes. Field Crops
Research, 239, 149–158. https://doi.org/10.1016/j.fcr.2019.05.017.
Yang, R., Zhang, L., Li, P., Yu, L., Mao, J., Wang, X., & Zhang, Q. (2018). A review of
chemical composition and nutritional properties of minor vegetable oils in China.
Trends in Food Science and Technology, 74(May 2017), 26–32. https://doi.org/
10.1016/j.tifs.2018.01.013.