Journal of Food Measurement and Characterization

https://doi.org/10.1007/s11694-018-9764-x

ORIGINAL PAPER

Hyperspectral imaging as an effective tool for prediction the moisture

content and textural characteristics of roasted pistachio kernels
Toktam Mohammadi‑Moghaddam1,2 · Seyed M. A. Razavi2 · Masoud Taghizadeh2 · Biswajeet Pradhan3 ·
Ameneh Sazgarnia4 · Ahmad Shaker‑Ardekani5

Received: 10 December 2017 / Accepted: 21 February 2018

© Springer Science+Business Media, LLC, part of Springer Nature 2018

Abstract
The objective of this study was to develop calibration models for prediction of moisture content and textural characteristics
(fracture force, hardness, apparent modulus of elasticity and compressive energy) of pistachio kernels roasted in different
conditions (temperatures 90, 120 and 150 °C; times 20, 35 and 50 min and air velocities 0.5, 1.5 and 2.5 m/s) using Vis/NIR
hyperspectral imaging and multivariate analysis. The effects of different pre-processing methods and spectral treatments
such as normalization [multiplicative scatter correction (MSC), standard normal variate transformation (SNV)], smoothing
(median filter, Savitzky–Golay and Wavelet) and differentiation (first derivative, ­D1 and second derivative, ­D2) on the obtained
data were investigated. The prediction models were developed by partial least square regression (PLSR) and artificial neural
network (ANN). The results indicated that ANN models have higher potential to predict moisture content and textural char-
acteristics of roasted pistachio kernels comparing to PLSR models. High correlation was observed between reflectance data
and fracture force ­(R2 = 0.957 and RMSEP = 3.386) using MSC, Savitzky–Golay and ­D1, compressive energy ­(R2 = 0.907
and RMSEP = 15.757) using the combination of MSC, Wavelet and D ­ 1, moisture content (­ R2 = 0.907 and RMSEP = 0.179)
and apparent modulus of elasticity ­(R  = 0.921 and RMSEP = 2.366) employing combination of SNV, Wavelet and ­D1,
2

respectively. Moreover, Vis–NIR data correlated well with hardness (­ R2 = 0.876 and RMSEP = 5.216) using SNV, Wavelet
and ­D2. These results showed the capability of Vis/NIR hyperspectral imaging and the central role of multivariate analysis
in developing accurate models for prediction of moisture content and textural properties of roasted pistachio kernels.

Keywords  Hyperspectral imaging · Non-destructive · Reflectance · Roasting · Texture · Vis–NIR

Introduction

Pistachio nut (Pistacia vera L.) is one of the most delicious

and nutritious nuts in the world and is being used as a salted
* Seyed M. A. Razavi and roasted product or as an ingredient in desserts (baklava,
s.razavi@um.ac.ir halva, paste, butter etc.) [1]. Iran is the main pistachio pro-
1
Department of Food Science and Technology, Neyshabur ducer in the world with about 415,531 Mt in 2014, followed
University of Medical Sciences, Neyshabur, Iran by USA, Turkey and China [2]. Pistachio kernel is a good
2
Department of Food Science and Technology, Ferdowsi source of fat (50–60%) [3] containing considerable amount
University of Mashhad (FUM), P.O. Box: 91775‑1163, of unsaturated fatty acids (linoleic, linolenic and oleic acids).
Mashhad, Iran Roasting is one of the most important food processes, which
3
School of Systems, Management and Leadership, Faculty provides useful attributes to the product. One of the objec-
of Engineering and IT, University of Technology Sydney, tives of nut roasting is to alter and significantly enhance the
Sydney, NSW, Australia flavour, texture, colour and appearance of the product. Roast-
4
Department of Medical Physics, Faculty of Medicine, ing gives nuts a variety of textures and colors, increasing their
Mashhad University of Medical Sciences, Mashhad, Iran crispness and making them crunchier [4–6]. Roasted nuts are
5
Pistachio Research Center, Horticultural Sciences Research widely consumed as an appetizer and used as raw materials
Institute, Agricultural Research, Education and Extension in sweets, confectionery, chocolate and biscuits. Because of
Organization (AREEO), Rafsanjan, Iran

13
Vol.:(0123456789)
 T. Mohammadi‑Moghaddam et al.

the deep green color of pistachio kernels, it is favored in the further processing. The average weight ratio of kernel was
ice cream and pastry industries [7]. Temperature and time are 62 g/100 g of split sample. The average moisture content
the main factors controlled during industrial roasting process. of pistachio kernels was measured as 2.9% (w.b.). Kernels
But initial moisture content and air velocity also affect heat were sorted manually to get the uniform sizes for roasting.
transfer rate subsequently drying and physicochemical changes 1 kg pistachio nuts were soaked in 5 L of 20% salt solution
in proteins [4, 8]. During roasting, moisture content of moist for 20 min [23]. After soaking, the salt solution was drained
nuts reduces [9] and texture becomes more crumbly and fragile using a strainer and the excess water was removed using a
(less hard) [10]. cloth filter. For roasting process, 3 temperature levels (90,
Texture characteristics such as firmness of fruits relate to 120 and 150 °C), 3 time intervals (20, 35 and 50 min) and
their mechanical properties. A large number of devices and 3 air velocities (0.5, 1.5 and 2.5 m/s) were applied. Thus,
techniques have been developed to determine the firmness in this research, 27 different treatments of roasted pistachio
of fruits and vegetables; mostly based on puncture, compres- kernels were analyzed. Roasting was performed in an electri-
sion or response to shear stress, creep, or impact. Mechanical cal oven equipped with a controller to adjust the air veloci-
force/deformation techniques applying compression or pen- ties of roasting (CIFS120, Fan Azma Gostar, Tehran, Iran).
etration are rarely appropriate for online sorting of fruits and The air velocities were measured by an electronic anemom-
vegetables since they are of low speed and/or destructive [11]. eter (AM4205, Lutron company, Taiwan) having an accuracy
Mohammadi-Moghaddam et al. [12] studied the effect of time, of 0.1 m/s. After roasting, the whole kernels were allowed
temperature and air velocity of roasting on textural proper- to cool at room temperature (20 ± 2 °C).
ties of pistachio kernels. Increasing the roasting temperature
decreased the fracture force, hardness, apparent modulus of Moisture content determination
elasticity and compressive energy.
In recent years, spectral imaging techniques (i.e. hyper- Moisture content of pistachio kernels was measured in tripli-
spectral and multispectral imaging) have emerged as a power- cate using oven drying method (105 ± 2 °C) until a constant
ful tool for safety and quality inspection of various agricul- weight was reached [24]. Moisture content was measured in
tural commodities. Spectral imaging combines conventional three replications.
imaging and spectroscopy to obtain both spatial and spec-
tral information from the target, which is extremely useful Instrumental texture analysis
for evaluating individual food samples. This technique has
drawn tremendous interest from both academic and industrial Instrumental texture measurements were carried out at room
areas, and has been developed rapidly during the past decade temperature using a texture analyzer (CNS Farnell Com,
[13–15]. U.K.) (Fig. 1). Uniaxial compression test, using a 35 mm
There are some published studies on the potential of hyper- diameter plastic cylinder, was made on the pistachio ker-
spectral imaging to determine physicochemical properties of nels, which were mounted over a platform. The operational
nuts, grains and seeds [16–22]. However, there is no study conditions of texture analyzer were 2  mm deformation
available about the ability of hyperspectral imaging to predict (target value), 20 g trigger point, one cycle and 30 mm/
the moisture content and textural attributes of pistachio ker- min cross head speed. The textural parameters of kernels
nels. The objectives of this study were (i) to apply reflectance were expressed as fracture force (highest peak followed
hyperspectral imaging for non-destructive determination of by a sudden drop [N]), hardness (peak compression force
moisture content and textural attributes (hardness, fracture [N] at target deformation), apparent modulus of elasticity
force, apparent modulus of elasticity and compressive energy) or initial tangent modulus (sample rigidity that is the slope
of pistachio kernels roasted in different conditions (tempera- of linear part of the force–deformation curve [N/mm]) and
tures: 90, 120 and 150 °C; times: 20, 35 and 50 min and air compressive energy (area under the curve for the compres-
velocities: 0.5, 1.5 and 2.5 m/s), (ii) to investigate the effects of sion that is the work [N.mm] required to attain deformation,
different combination of pre-processing methods on predicted indicative of internal strength of bonds within product) [12,
moisture content and textural properties of pistachio kernels. 24]. Texture measurements were performed in at least four
replications.

Materials and methods Vis/NIR hyperspectral imaging measurements

Sample preparation A hyperspectral imaging system (Spectrograph model:

ImSpector V10, Spectral Imaging Ltd., Oulu, Finland) in
Dried O’hadi pistachio nuts were supplied from a local the Vis–NIR range (400–1000 nm) was employed in this
market in Mashhad, Iran in 2012 and stored at 4 °C for study. The main components of this system are: objective

13
Hyperspectral imaging as an effective tool for prediction the moisture content and textural…

Fig. 1  Texture analyzer instrument and typical hyperspectral image and spectrum of roasted pistachio kernels in the range 400–1000 nm

lens, spectrograph (spectroscopic resolution of 5 nm), CCD including noises, uncertainties, variabilities, interactions and
camera (Basler A312f, effective resolution of 580 × 580 pix- unrecognized features. A lot of pre-processing techniques
els by 12 bits), acquisition system, moving table, and illumi- for spectral data have been recently developed [26]. In this
nation via fibre optics. Each scan was taking approximately study, several pre-processing methods including normali-
10–15 s. Figure 1 shows a typical hyperspectral image and zation [multiplicative scatter correction (MSC), standard
spectrum of roasted pistachio kernels. Hyperspectral images normal variate transformation (SNV)], smoothing (median
were obtained in at least five replications. filter, Savitzky–Golay and Wavelet filter level 2) and differ-
entiation (first derivative and second derivative) were used.
Pre‑processing Normalization is designed to remove multiplicative spectral
effects. In normalization, the spectral vector is transformed
The overall mean reflectance spectra of kernels were con- into unit length [27]. Several smoothing techniques have
verted to absorbance values (log1/R). This approach ignores been proposed to remove random noise from spectra, includ-
the fact that light penetration in biological tissue is more ing moving average filters and the Savitzky–Golay algorithm
complicated than structural materials and also involves scat- [28, 29]. It is crucial to select the proper smoothing window
tering [25]. In addition, the spectral pre-processing tech- width. Smoothing is also necessary in order to optimize the
niques are required to remove any irrelevant information signal-to-noise ratio [26, 30]. In general, it is combined with

13
 T. Mohammadi‑Moghaddam et al.

other pre-processing methods to accomplish the de-noising During training, momentum and gradient values were
[26, 31, 32]. Wavelet transforms (WT) is a very popular 0.001–1 × 1010, and 1–1 × 10−10, respectively. The training
kind of operation today due to its widespread application in process was carried on for 1000 epochs. Totally 824 data
chemometrics and signal processing. In this method, the raw were experimentally collected and randomly divided into
signals are decomposed into a series of different frequency three partitions for training (70% of data), validating (15%
signal blocks. The small Wavelet coefficients presenting of data) and testing (15% of data) of the developed network.
noise and background signals are removed [33]. Another In ANN method, pre-processing techniques were done for
pre-processing method known as the first and second deriva- all data. Due to the high volume of hyperspectral data, 1
tive is used to remove background and increase spectral res- neuron in hidden layer was used for prediction of the mois-
olution [26]. First derivatives are applied to remove additive ture content and textural characteristics of pistachio kernels.
baseline effects while second derivatives are used to remove Data from hyperspectral imaging and moisture content from
sloped additive baselines [27]. experimental method were considered as input and output of
neural network respectively. Then R­ 2 and RMSEP were used
Data analysis to compare the performance of different ANN architectures.

In this study, samples’ spectra were isolated from hyperspec-

tral images by ENVI software (version 4.7, Research System Results and discussions
Inc. (RSI), USA). All pre-processing techniques were car-
ried out using ParLeS software (version 3.1, the University Tables 1 and 2 show the moisture content and textural char-
of Sydney, Australia) [34]. To make models at PLSR and acteristics of roasted pistachio kernels and the summary sta-
ANN methods, ParLeS software and Matlab R2009a (ver- tistics for all samples selected in each data set, respectively.
sion 7.8.0.347, MathWorks Inc., USA) were used, respec- In this study, the range of moisture content, fracture force,
tively. Principal component analysis (PCA) was performed hardness, apparent modulus of elasticity and compressive
for a maximum of two principal components. PCA showed energy were 0.5–2.26%, 35.61–81.98 N, 43.52–82.76 N,
that there was one spectral outlier in calibration data set. 21.68–47 N/mm and 117.10–280.73 N.mm, respectively.
This sample was identified and removed before modeling by
PLSR and ANN. PLSR based on ten segments were used as Interpretation of hyperspectral data
the main method for developing of calibration sets. To apply
PLSR, prior to develop determination models, sample data Figure 2a–c show the typical reflectance spectra of differ-
was randomly divided to calibration set (70% of the whole ent parts of roasted pistachio nut at wavelength range of
samples) and validation set (30% of the whole samples). 400–1000 nm, respectively. It can be seen that the spectra
In PLSR method, pre-processing techniques were done on of shell (Fig. 2a), whole kernel (Fig. 2b) and internal part
calibration and validation set separately. Validation set was of kernel (Fig. 2c) show different patterns. As illustrated,
used to evaluate the robustness of the developed models. The the obtained spectra have some reflectance peaks at specific
coefficient of determination ­(R2), the root mean square error wavelengths. Figure 2c shows the internal part of kernel has
of prediction (RMSEP) and the ratio of the standard devia- 2 peaks at 630 and 690 nm, while shell and whole kernel
tion of the response variable to RMSEP [known as relative have 1 peak at 670 and 720 nm, respectively and the peak
performance determinant (RPD)] were calculated. Gener- of whole kernel is sharper than shell (Fig. 2a, b). It can be
ally, a powerful model should have low RMSEP and high ­R2 seen the intensity of reflectance is different for shell, whole
between the predicted and measured values of each property. kernel and internal part of kernel. The highest and the lowest
Moreover, low number of PLSR factors and high RPD are intensity were for internal part of kernel and whole kernel,
desirable. An RPD between 1.5 and 2 means that the model respectively. The spectral slope of internal part is higher than
can discriminate low from high values of the response vari- shell and whole kernel at 500–700 nm (Fig. 2a–c). Kim et al.
able; a value between 2 and 2.5 indicates that course quan- [37] found one peak for apple at 430–930 nm at reflectance
titative predictions are possible, and a value between 2.5 and fluorescence hyperspectral imaging. Weinstoc et al. [17]
and 3 or above corresponds to good and excellent prediction identified 3 peaks at 1000, 1200 and 1450 nm for corn at
accuracy, respectively [29, 35, 36]. reflectance hyperspectral imaging.
To apply artificial neural network (ANN), a two-layer
feed-forward network with sigmoid hidden neurons and Effect of roasting conditions on the spectra
linear output neurons was used. The network was trained of pistachio kernels
with Levenberg–Marquardt back propagation algorithm,
unless there is not enough memory, in which case scaled Figure 3a–c show the effect of temperature, air velocity
conjugate gradient back propagation (trainscg) was used. and time of roasting on the spectra of pistachio kernels. It

13
Hyperspectral imaging as an effective tool for prediction the moisture content and textural…

Table 1  Moisture content and textural properties of pistachio kernels at different temperatures, times and air velocities of roasting
Tempera- Time (min) Air veloc- Moisture content (%) Fracture force (N) Hardness (N) Apparent modulus of Compres-
ture (°C) ity (m/s) elasticity (N/mm) sive energy
(N.mm)

90 20 0.5 2.12 ± 0.10 44.90 ± 9.26 76.95 ± 6.49 37.65 ± 2.19 230.55 ± 4.51

90 20 1.5 2.18 ± 0.11 71.29 ± 10.21 71.70 ± 8.38 47 ± 2.99 171.60 ± 13.70
90 20 2.5 2.26 ± 0.10 73.03 ± 5.26 73.59 ± 5.40 38.07 ± 4.33 240.32 ± 20.18
90 35 0.5 1.75 ± 0.08 75.48 ± 1.49 77.50 ± 5.83 39.94 ± 6.65 232.56 ± 21.30
90 35 1.5 1.98 ± 0.07 81.98 ± 5.92 82.76 ± 6.19 43.27 ± 6.44 280.73 ± 28.70
90 35 2.5 1.50 ± 0.07 77.77 ± 7.69 79.21 ± 7.53 38.50 ± 1.24 242.70 ± 25.88
90 50 0.5 1.50 ± 0.08 78.15 ± 5.89 80.42 ± 3.67 41.59 ± 3.60 243.09 ± 24.52
90 50 1.5 2.02 ± 0.12 64.96 ± 9.17 80.42 ± 2.41 41.39 ± 0.03 247.28 ± 11.44
90 50 2.5 1.50 ± 0.09 80.00 ± 10.25 80.76 ± 10.20 30.78 ± 6.15 231.65 ± 2.71
120 20 0.5 2.04 ± 0.10 76.90 ± 5.19 81.28 ± 0.93 40.10 ± 7.44 253.98 ± 5.46
120 20 1.5 1.46 ± 0.08 59.05 ± 9.29 63.93 ± 4.28 41.37 ± 6.94 199.06 ± 14.29
120 20 2.5 2.19 ± 0.04 46.45 ± 5.05 73.27 ± 1.40 36.63 ± 8.61 194.75 ± 13.74
120 35 0.5 1.50 ± 0.06 60.38 ± 7.28 61.91 ± 4.80 31.6 ± 1.66 188.80 ± 1.04
120 35 1.5 1.59 ± 0.07 65.86 ± 10.21 68.06 ± 7.68 24.98 ± 3.87 189.31 ± 10.08
120 35 2.5 0.50 ± 0.02 56.13 ± 1.01 63.13 ± 6.59 21.76 ± 4.45 166.90 ± 5.71
120 50 0.5 1.00 ± 0.02 41.35 ± 8.28 59.86 ± 2.71 28.92 ± 2.83 154.68 ± 6.52
120 50 1.5 1.44 ± 0.01 59.55 ± 2.95 69.86 ± 4.72 33.15 ± 3.09 200.52 ± 5.98
120 50 2.5 0.75 ± 0.01 56.68 ± 9.32 56.97 ± 9.25 21.68 ± 6.49 151.51 ± 8.79
150 20 0.5 1.75 ± 0.10 46.59 ± 4.38 48.04 ± 7.18 28.89 ± 0.94 127.52 ± 14.04
150 20 1.5 1.24 ± 0.11 35.61 ± 2.52 43.52 ± 5.99 22.19 ± 1.23 119.98 ± 2.19
150 20 2.5 1.26 ± 0.08 43.14 ± 7.80 47.17 ± 1.03 24.25 ± 8.08 117.10 ± 6.51
150 35 0.5 1.50 ± 0.10 40.05 ± 6.54 44.97 ± 6.78 23.72 ± 6.87 125.13 ± 9.82
150 35 1.5 1.08 ± 0.10 58.72 ± 4.38 59.69 ± 2.74 23.61 ± 6.84 158.36 ± 8.60
150 35 2.5 1.00 ± 0.09 39.33 ± 6.25 45.26 ± 6.99 22.28 ± 8.65 144.21 ± 13.58
150 50 0.5 0.75 ± 0.05 39.57 ± 2.21 43.62 ± 0.92 23.54 ± 6.66 134.92 ± 12.72
150 50 1.5 0.82 ± 0.06 50.32 ± 8.82 55.10 ± 8.87 27.15 ± 3.35 148.80 ± 9.50
150 50 2.5 0.50 ± 0.04 25.54 ± 5.94 37.59 ± 5.70 21.22 ± 1.75 101.18 ± 13.18

Table 2  Statistical results on both calibration and validation sets for prediction of moisture content and textural properties of roasted pistachio
kernels at PLSR method
Characteristic Calibration set Validation set
Mean SD Median Minimum Maximum Mean SD Median Minimum Maximum

Moisture content (w.b. %) 1.61 0.49 1.50 0.50 2.26 1.26 0.45 1.26 0.75 1.98
Fracture force (N) 58.16 14.15 58.72 35.61 77.77 59.39 17.16 59.09 39.33 81.98
Hardness (N) 65.84 12.5 68.06 43.52 81.28 63.31 15.94 63.93 43.62 82.76
Apparent modulus of elasticity (N/mm) 31.88 8.44 31.6 21.68 47.00 32.45 7.96 30.78 22.28 43.27
Compressive energy (N.mm) 187.38 46.82 188.8 117.10 253.98 190.06 54.44 199.06 127.52 280.73

SD standard deviation

can be seen that roasting increased the reflectance of pis- the roasting, moisture content decreases and amount of
tachio kernels at the range of 400–640 and 670–1000 nm, reflectance increases. Figures 3b and 2c show the effect
whereas the reflectance of pistachio kernels decreased at of roasting air velocity and time on the spectra of pista-
640–670 nm, though these changes were intangible. As chio kernels. It can be observed these parameters cause
shown in Fig. 3a, increasing the roasting temperature led to change the reflectance of pistachio kernels at the
to increase the reflectance of pistachio kernels. During

13
 T. Mohammadi‑Moghaddam et al.

Fig. 2  Original reflectance spectra of a pistachio shell, b whole pista-

chio kernel and c internal part of pistachio kernel

Fig. 3  Effect of roasting temperature (a), air velocity (b) and time (c)
mentioned ranges, although these changes don’t show sen- on the spectra of pistachio kernels
sible difference at various roasting time and air velocity.

Effects of different pre‑processing techniques ­R2 and RPD values and high RMSEP value (Tables 3, 4).
Cogdill et al. [16] used transmission hyperspectral imaging
The results of PLSR and ANN models with several combi- (750–1090 nm), PLSR and PCR methods for prediction of
nations of preprocessing techniques for moisture content and moisture content of corn. They reported PLSR has higher
textural characteristics of roasted pistachio kernels are sum- potential in predicting moisture content of corn comparing
marized in Tables 3 and 4. In the absence of pre-processing to PCR.
techniques, low correlation coefficients were observed for
prediction of all studied parameters. However, with the use Moisture content
of pre-processing techniques, in some models, correlation
coefficient and RPD increased and the RMSEP decreased. Table 3 shows the results obtained for PLSR and ANN mod-
Although PLSR models could predict moisture content and els on prediction of moisture content of roasted pistachio
textural properties of pistachio kernels, but it could not kernels. It can be seen that ANN models can predict mois-
make good predictions for these properties, because of low ture content of roasted pistachio kernels better than PLSR

13
Hyperspectral imaging as an effective tool for prediction the moisture content and textural…

Table 3  The prediction Pre-processing method PLSR ANN

results of moisture content
2
with different pre-processing No. of factor RPD RMSEP R RMSEP R2
techniques
Original data 4 0.57 0.787 0.242 0.238 0.816
SNV, median filter, ­D1 5 0.94 0.478 0.473 0.238 0.804
MSC, median filter, D­ 1 5 0.98 0.459 0.478 0.219 0.824
SNV, median filter, ­D2 5 0.46 0.982 0.528 0.376 0.530
MSC, median filter, D­ 2 4 0.45 1.005 0.525 0.292 0.690
1
SNV, Wavelet, ­D 4 0.85 0.524 0.517 0.179 0.907
MSC, Wavelet, ­D1 1 1 0.448 0.251 0.244 0.820
SNV, Wavelet, ­D2 9 0.78 0.572 0.051 0.284 0.742
MSC, Wavelet, ­D2 5 0.39 1.153 0.055 0.314 0.689
SNV, Savitzky–Golay, ­D1 2 0.84 0.535 0.432 0.269 0.767
MSC, Savitzky–Golay, ­D1 2 0.89 0.505 0.333 0.308 0.676
SNV, Savitzky–Golay, ­D2 1 0.74 0.605 0.226 0.199 0.873
MSC, Savitzky–Golay, ­D2 1 1.04 0.432 0.233 0.130 0.767

R2 coefficient of determination, RMSEP the root mean square error of prediction, RPD the ratio of the
standard deviation of the response variable to the RMSEP, SNV standard normal variate transformation,
MSC multiplicative scatter correction, D1 first derivative, D2 second derivative

models. In total, PLSR models showed low RPD and ­R2. For was better against SNV to predict the fracture force for
all samples, RPD was lower than 1.5, indicating that devel- smoothing techniques. Nevertheless, better results were
oped models do not give an accurate prediction for moisture achieved using a combination of MSC, Savitzky–Golay
content. The best results with ANN method were achieved and ­D1 for predicting the fracture force with R­ 2 = 0.957
using a combination of SNV, Wavelet and D ­ 1 for predict- and RMSEP = 3.386 (Table 4). The ANN model preproc-
2
ing moisture content with ­R  = 0.907 and RMSEP = 0.179 essed with the combination of SNV, Savitzky–Golay and
(Table 3). Cogdill et al. [16] developed a technique for creat- ­D1 had the same ­R2, but RMSE value was higher than the
ing calibrations to predict the constituent concentrations of best model (RMSE = 3.405) (Table 4). PLSR models could
single maize kernels from NIR hyperspectral transmission predict the fracture force but they didn’t give an accurate
data in the range of 750–1090 nm. PLSR and PCR were prediction because of high RMSEP and low RPD. The best
used to develop predictive calibration models for moisture PLSR model was obtained with ­R2 = 0.823, RMSEP = 12.37,
and oil contents. Standard normal variate, detrending, multi- RPD = 1.38 and 3 PLSR factors.
plicative scatter correction, wavelength selection by genetic
algorithm, and no preprocessing were compared for their
effect on predictive model performance. The predictive cali- Hardness
bration model for moisture content achieved a best standard
error of cross-validation (SECV) of 1.20%, with RPD of The results indicated that ANN models can predict the hard-
2.74. The best results for prediction of oil content achieved ness better than PLSR models. The best results with PLSR
an SECV of 1.38%, with an RPD of only 1.45. Huang et al. models were achieved using a combination of SNV, Wavelet
[38] used the hyperspectral imaging technology and PLSR and ­D1 with R
­ 2 = 0.643, RMSEP = 10.78, RDP = 1.48 and 2
for prediction the moisture content of maize kernels during PLSR factors (Table 4). However, due to high RMSEP and
the drying. They predicted the moisture content with the uni- low ­R2 and RPD, it can be mentioned that prediction of hard-
formity (indirect) and uniformity directly. Better prediction ness values with ANN model was not sufficiently desirable;
results were achieved using the direct method (RP = 0.848 however it was better than PLSR models. The best results
and RMSEP = 2.73) than the indirect method (RP = 0.521 with ANN models were achieved using a combination of
and RMSEP = 10.96). SNV, Wavelet and ­D2 with ­R2 = 0.876 and RMSEP = 5.216
(Table 4). Williams et al. [19] used the NIR hyperspectral
Fracture force imaging (960–1662 nm and 1000–2498 nm) for distinguish-
ing between hard, intermediate and soft maize samples. They
ANN models could predict the fracture force as well and found that PCA can illustrate a distinct difference between
better than the moisture content and other textural char- glassy and floury endosperm along principal component.
acteristics. In terms of pre-processing methods, MSC Subsequently, they applied partial least squares discriminant

13


Table 4  The prediction results of textural characteristics and moisture content of pistachio kernels with different pre-processing techniques
Textural Fracture force (N) Hardness (N) Compressive energy (N.MM) Apparent modulus of elasticity (N/s)
character-

13
istics
Pre- PLSR ANN PLSR ANN PLSR ANN PLSR ANN
processing 2 2 2 2 2 2 2
methods No. RPD RMSEP R RMSEP R No. RPD RMSEP R RMSEP R No. RPD RMSEP R RMSEP R No. RPD RMSEP R RMSEP R2
of of of of
fac- fac- fac- fac-
tor tor tor tor

Original 1 1.16 14.76 0.44 7.48 0.88 4 1.49 10.70 0.51 6.57 0.82 4 1.41 38.49 0.64 23.12 0.79 4 1.03 7.72 0.44 3.58 0.86
data
SNV, 7 1.28 13.37 0.56 14.33 0.45 2 1.33 12.10 0.45 6.74 0.77 1 1.16 46.86 0.55 23.36 0.78 7 1.42 5.61 0.49 3.49 0.82
median
filter, ­D1
MSC, 7 1.25 13.65 0.56 7.11 0.82 2 1.33 12.02 0.46 8.23 0.66 1 1.14 47.85 0.55 35.99 0.67 5 1.39 5.71 0.50 5.47 0.70
median
filter, ­D1
SNV, 6 0.34 50.22 0.31 7.50 0.80 7 0.31 52.09 0.35 8.56 0.63 7 0.34 160.02 0.32 31.85 0.68 6 0.30 26.64 0.59 4.56 0.78
median
filter, ­D2
MSC, 7 0.33 50.99 0.31 7.38 0.78 7 0.30 52.94 0.36 5.26 0.86 6 0.34 162.37 0.33 21.52 0.82 6 0.29 27.07 0.61 3.94 0.77
median
filter, ­D2
SNV, Wave- 3 1.38 12.37 0.82 7.67 0.79 2 1.48 10.78 0.64 7.25 0.74 2 1.33 41.05 0.54 27.32 0.73 2 1.27 6.29 0.30 2.37 0.92
let, ­D1
MSC, 6 1.15 14.89 0.29 8.83 0.70 1 1.15 13.89 0.34 12.77 0.64 1 1.09 49.76 0.45 15.76 0.91 1 1.05 7.56 0.29 3.44 0.85
Wavelet,
­D1
SNV, Wave- 5 0.54 31.39 0.01 9.89 0.61 2 0.47 33.88 0.00 5.22 0.88 5 0.52 102.08 0.01 34.73 0.57 5 0.45 17.80 0.07 4.76 0.70
let, ­D2
MSC, 2 0.32 54.03 0.52 6.29 0.85 4 0.27 58.56 0.54 9.94 0.52 5 0.33 178.66 0.32 35.20 0.55 2 0.27 18.96 0.11 3.04 0.86
Wavelet,
­D2
SNV, 8 1.32 12.95 0.76 3.40 0.96 1 0.94 16.97 0.43 6.19 0.82 1 0.97 55.91 0.54 17.96 0.87 1 0.98 8.14 0.32 4.29 0.75
Savitzky–
Golay, ­D1
MSC, 9 1.49 11.47 0.64 3.39 0.96 2 1.40 11.39 0.45 8.89 0.64 1 1.25 43.69 0.53 24.34 0.77 1 1.17 6.79 0.33 4.83 0.82
Savitzky–
Golay, ­D1
SNV, 1 0.95 17.99 0.49 7.57 0.78 1 0.90 17.71 0.45 6.56 0.79 1 0.94 57.90 0.56 29.98 0.74 1 0.95 8.42 0.34 4.16 0.79
Savitzky–
Golay, ­D2
MSC, 4 1.25 13.62 0.43 6.51 0.84 1 1.13 14.05 0.36 7.70 0.73 1 1.22 44.62 0.48 24.40 0.76 1 0.99 8.04 0.31 3.43 0.84
Savitzky–
Golay, ­D2

R2 coefficient of determination, RMSEP the root mean square error of prediction, RPD the ratio of the standard deviation of the response variable to the RMSEP, SNV standard normal variate
T. Mohammadi‑Moghaddam et al.

transformation, MSC multiplicative scatter correction, D1 first derivative, D2 second derivative)

Hyperspectral imaging as an effective tool for prediction the moisture content and textural…

analysis (PLS-DA) for model classification. The hardness findings of this study demonstrate the predictive ability of
values were predicted with RMSEP of 0.18 in 960–1662 nm Vis/NIR hyperspectral imaging combined with selected
and RMSEP of 0.29 in 1000–2498 nm wavelength range. chemometric tools. This technology can be successfully
used as a non-destructive tool for measuring moisture
Apparent modulus of elasticity content and textural characteristics of pistachio kernels at
different roasting conditions.
Table 4 shows the results of PLSR and ANN methods for
prediction of the apparent modulus of elasticity. It can be Acknowledgements  The authors would like to thank Dr. Raphael A.
Viscarra Rossel for providing ParLes software.
observed again ANN models predicted the apparent modulus
of elasticity better than PLSR models. In the case of ANN,
the best model was found to be the one with preprocessing
techniques containing combination of SNV, Wavelet and D ­ 1
2
with ­R  = 0.921 and RMSEP = 2.366 (Table 4).
References
1. M. Bayram, Comparison of unsplit in shell and shelled kernel
Compressive energy of the pistachio nuts. J. Food Eng. 107(3–4), 374–378 (2011)
2. FAOSTAT. FAOSTAT database results. http://faost​at.fao.org/
The results of PLSR and ANN methods for prediction of site/339/defau​lt.aspx. Accessed 29 Sept 2017
3. M. Maskan, S. Karatas, Fatty acid oxidation of pistachio nuts
compressive energy are shown in Table 4. The best results stored under various atmospheric conditions & different tem-
with PLSR models were achieved without preprocess- peratures. J. Sci. Food Agric. 77, 334–340 (1998)
ing ­(R2 = 0.643, RMSEP = 38.490 and RPD = 1.41). High 4. S. Saklar, Optimization of hazelnut roasting process by using
RMSEP and low RPD indicated inability of PLSR mod- response surface methodology. PhD thesis, METU, Ankara
(1999)
els in predicting compressive energy (Table 4). In the case 5. M. Özdemir, F.G. Seyhan, A.K. Bakan, G. Ilter, O. Devres, Analy-
of ANN, the best model was found using the combination sis of internal browning of roasted hazelnuts. Food Chem. 73,
of MSC, Wavelet and ­D1 as preprocessing techniques with 191–196 (1999)
­R2 = 0.907 and RMSEP = 15.757 (Table 4). 6. A.D. Demir, K. Cronin, The thermal kinetics of texture change and
the analysis of texture variability for raw and roasted hazelnuts.
IJFST 39, 371–383 (2004)
7. J.G. Woodroof, Tree Nuts: Production Processing Products,
Conclusions 2nd edn. (AVI, Westport, 1979)
8. R.Y.Y. Chiou, T.T. Tsai, Characteristics of peanut proteins during
roasting as affected of initial moisture. J. Agric. Food Chem. 37,
The feasibility of utilizing Vis/NIR hyperspectral imag- 1377–1381 (1989)
ing (400–1000 nm) combined with different preprocessing 9. L.B. Emily, D.B. Terri, A.W. Lester, Effect of cultivar and roasting
methods to nondestructive prediction of moisture content method on composition of roasted soybeans. J. Sci. Food Agric.
and textural characteristics of roasted pistachio kernels 89(5), 821–826 (2009)
10. J.F.V. Vincent, Application of fracture mechanics to the texture
was investigated. Employing pre-processing methods, in of food. Eng. Fail Anal. 11, 695–704 (2004)
some cases, increased correlation coefficient and RPD of 11. J.A. Abbott, Quality measurement of fruits and vegetables. Post-
the developed model while it decreased the RMSEP, sug- harvest Biol. Technol. 5, 207–225 (1999)
gesting that preprocessing techniques would enhance the 12. T. Mohammadi-Moghaddam, S.M.A. Razavi, M. Taghizadeh, A.
Sazgarnia, Sensory and instrumental assessment of roasted pista-
models’performance. Based on the obtained data, SNV chio nut/kernel by partial least square (PLS) regression analysis:
method would generate slightly better results than MSC. effect of roasting conditions. J. Food Sci. Technol. 53(1), 370–380
In all cases, the results revealed that ANN models would (2016)
predict moisture content and textural characteristics of 13. A.A. Gowen, C.P. O’Donnell, P.J. Cullen, G. Downey, J.M. Frias,
Hyperspectral imaging—an emerging process analytical tool for
roasted pistachio kernels better than PLSR models. ANN food quality and safety control. Trends Food Sci. Technol. 18(12),
models could predict fracture force as well and better than 590–598 (2007)
the moisture content and other textural characteristics. The 14. D. Sun (ed.), Hyperspectral Imaging for Food Quality Analysis
best obtained models for moisture content ­( R 2 = 0.907, and Control (Academic Press, Elsevier, San Diego, 2010)
15. D. Lorente, N. Aleixos, J. Gَmez-Sanchis, S. Cubero, O.L. Garcia-
RMSEP = 0.179) and apparent modulus of elasticity Navarrete, J. Blasco, Recent advances and applications of hyper-
­( R 2 = 0.921, RMSEP = 2.366) with the combination of spectral imaging for fruit and vegetable quality assessment. Food
SNV, Wavelet and ­D1 and for fracture force ­(R2 = 0.957, Bioprocess Technol. 5(4), 1121–1142 (2012)
RMSEP = 3.386,) with MSc, Savitzky–Golay and ­D 1, 16. R.P. Cogdill, Jr..C.R. Hurburgh, G.R. Rippke, Single-kernel maize
analysis by near-infrared hyperspectral imaging. TASABE 47(1),
for hardness ­( R 2 = 0.876, RMSEP = 5.216) with SNV, 311–320 (2004)
Wavelet and ­D2 and for compressive energy ­(R2 = 0.907, 17. B.A. Weinstock, J. Janni, L. Hagen, S. Wright, Prediction of oil
RMSEP = 15.757) with MSc, Wavelet and D ­ 1 showed high and oleic acid concentrations in individual corn (Zea mays L.)
prediction ability of Vis/NIR hyperspectral imaging. The

13
 T. Mohammadi‑Moghaddam et al.

kernels using near infrared reflectance hyperspectral imaging and 29. B.M. Nicolaï, K. Beullens, E. Bobelyn, A. Peirs, W. Saeys, K.I.
multivariate analysis. Appl. Spectrosc. 60(1), 9–16 (2006) Theron, J. Lammertyn, Nondestructive measurement of fruit and
18. B. Zhu, L. Jiang, F. Jin, L. Qin, A. Vogel, Y. Tao, Walnut shell vegetable quality by means of NIR spectroscopy: a review. Post-
and meat differentiation using fluorescence hyperspectral imagery harvest Biol. Technol. 46, 99–118 (2007)
with ICA-kNN optimal wavelength optimal wavelength selection. 30. J.H. Perkin, B. Tenge, D.E. Honigs, Resolution enhancement
Sens. Instrum. Food Qual. Saf. 1, 123–131 (2007) using an approximate-inverse Savitzky-Golay smooth. Spectro-
19. P. Williams, P. Geladi, G. Fox, M. Manley, Maize kernel hardness chim. Acta B 43, 575–603 (1998)
classification by near infrared (NIR) hyperspectral imaging and 31. M.J.C. Pontes, S.R.B. Santos, M.C.U. Araujo, L.F. Almeida,
multivariate data analysis. Anal. Chim. Acta 653, 121–130 (2009) R.A.C. Lima, E.N. Gaiao, U.T.C.P. Souto, Classification of dis-
20. S.R. Delwiche, I.C. Yang, M.S. Kim, Hyperspectral imaging for tilled alcoholic beverages and verification of adulteration by near
detection of black tip damage in wheat kernels, in Proceedings infrared spectrometry. Food Res. Int. 39, 182–189 (2006)
of SPIE–The International Society for Optical Engineering, vol 32. L. Wang, F.S.C. Lee, X.R. Wang, Y. He, Feasibility study of quan-
7315, (2009) tifying and discriminating soybean oil adulteration in camellia
21. J. Xing, S. Symons, M. Shahin, D. Hatcher, Detection of sprout oils by attenuated total reflectance MIR and fiber optic diffuse
damage in Canada Western Red Spring wheat with multiple wave- reflectance NIR. Food Chem. 95, 529–536 (2006)
bands using visible/near-infrared hyperspectral imaging. Biosyst. 33. J. Yoon, B. Lee, C. Han, Calibration transfer of near-infrared spec-
Eng. 106, 188–194 (2010) tra based on compression of wavelet coefficients. Chemom. Intell.
22. X. Zhang, F. Liu, Y. He, X. Li, Sensor application of hyperspectral Lab. Syst. 64, 1–14 (2002)
imaging and chemometric calibrations for variety discrimination 34. R.A.V Rossel, ParLeS: software for chemometric analysis of spec-
of maize seeds. Sensors 12, 17234–17246 (2012) troscopic data. ‎Chemom. Intell. Lab. Syst., 90, 72–83 (2008)
23. F. Goktas Seyhan, Effect of soaking on salting and moisture 35. P.C. William, D.G. Sobering, How do we do it: a brief summary
uptake of pistachio nuts (Pistachia vera L.) from Turkiye. GIDA of the method we use in developing near infrared calibrations, in
28(4), 395–400 (2003) Near Infrared Spectroscopy: The Future Waves, ed. by A.M.C.
24. S.M.A. Razavi, M.R. Edalatian, Effect of moisture content and Davies, P. Williams (NIR Publications, Chichester, 1996), pp.
compression axis on physical and mechanical properties of pista- 185–188
chio kernel. Int. J. Food Prop. 15, 507–517 (2012) 36. W. Saeys, A.M. Mouazen, H. Ramon, Potential for onsite and
25. A.M. Davies, A. Grant, Review: near-infra-red analysis of food. online analysis of pig manure using visible and near infrared
IJFST 22, 191–207 (1987) reflectance spectroscopy. Biosyst. Eng. 91, 393–402 (2005)
26. H. Cen, Y. He, Theory and application of near infrared reflectance 37. M.S. Kim, Y.R. Chen, P.M. Mehl, Hyperspectral reflectance and
spectroscopy in determination of food quality. Trends Food Sci. fluorescence imaging system for food quality and safety. ASABE
Technol. 18, 72–83 (2007) 44(3), 721–729 (2001)
27. H. Swierenga, A.P. De Weijer, R.J. Van Wijk, L.M.C. Buydens, 38. M. Huang, W. Zhao, Q. Wang, M. Zhang, Q. Zhu, Prediction of
Strategy for constructing robust multivariate calibration models. moisture content uniformity using hyperspectral imaging tech-
Chemom. Intell. Lab. Syst. 49, 1–17 (1999) nology during the drying of maize kernel. Int. Agrophys. 29(1),
28. T. Næs, T. Isaksson, T. Fearn, T. Davies, A User-Friendly Guide 39–46 (2015)
to Multivariate Calibration and Classification (NIR Publications,
Charlton, 2004)

13