Computers and Electronics in Agriculture 122 (2016) 1928

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

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

Original papers

Classification and characterization of blueberry mechanical damage with

time evolution using reflectance, transmittance and interactance
imaging spectroscopy
Meng-Han Hu, Qing-Li Dong , Bao-Lin Liu
School of Medical Instrument and Food Engineering, University of Shanghai for Science and Technology, 516 Jun Gong Rd., Shanghai 200093, PR China

a r t i c l e i n f o a b s t r a c t

Article history: The aim of this work was to evaluate the performance of hyperspectral data coupled with chemometrics
Received 7 April 2015 methods in characterizing and detecting the non-visible mechanical damage of blueberries with time
Received in revised form 1 December 2015 evolution. Reflectance and transmittance as well as interactance hypercubes were automatically seg-
Accepted 9 January 2016
mented by the region growing based algorithms. The maximum-normalized spectra were pretreated
Available online 23 January 2016
by the Standard Normal Variate algorithm, and subsequently the Competitive Adaptive Reweighted
Sampling algorithm was applied to extract the damage-specific wavelengths. Based on confusion matri-
Keywords:
ces and area under Receiver Operating Characteristics curves, transmittance showed relatively superior
Hyperspectral imaging
Mechanical damage
performance to reflectance and interactance. Application of new sample set subjected to impact tests
Physical property with time evolution, results demonstrated that it was especially difficult to distinguish fresh damage
Classification model in blueberry. At 2 days after impacted, several transmittance-based classifiers obtained satisfactory accu-
Chemometrics racies for classifying damaged (sound) blueberries: logistic regression 79.1% (85.7%), multilayer
perceptron-back propagation 74.4% (92.1%) and logistic function tree 72.1% (95.2%). Furthermore, the
physical property preliminarily proved to be more pronounced than the absorbed impact energy for
damage incidence and severity of blueberry via the use of multiple comparison.
2016 Elsevier B.V. All rights reserved.

1. Introduction labor-intensive and fallible manual inspection. Opara and Pathare

Fruits can be injured during a complete farm-to-table chain
such as harvesting, handling, transport and storage (Li and
Thomas, 2014; Van Zeebroeck et al., 2007). The presence of
mechanical damage will accelerate fruit softening (Zhou et al.,
2007) and even increase fruit susceptibility to fungal spoilage
(Mehra et al., 2013), which in turn reduce the income of fruit
industry and bring the possible hazard to the consumer (Opara
and Pathare, 2014). Blueberry belongs to a small soft fruit, and
its production has greatly expanded these years due to the
health-promoting and antimicrobial attributes (Elks et al., 2013;
Retamales, 2012; Shen et al., 2014); however, the highly perishable
nature of blueberry makes it particularly prone to mechanical
damage and afterward leads to amount of economic loss (Li
et al., 2010). To improve blueberry quality, it is necessary to pre-
vent the damaged blueberries for marketing.
Non-destructive measurements in conjunction with chemomet-
rics methods are attempted to estimate fruit damage, instead of
Corresponding authors. Tel./fax: +86 21 5527 1117.
E-mail addresses: qdong@usst.edu.cn (Q.-L. Dong), blliuk@163.com (B.-L. Liu).
(2014) elaborated the progresses in applications of non-
destructive defect detection approaches, containing machine
vision system (Zhang et al., 2015), visible and near infrared (vis/
NIR) spectroscopy (Martinsen et al., 2014; Wu and Wang, 2014),
hyperspectral imaging (Pu et al., 2015), thermal imaging
(Baranowski et al., 2009), magnetic resonance imaging (Mazhar
et al., 2015; Patel et al., 2015). Among these techniques, hyperspec-
tral imaging treats spectral data as a series of images, and therefore
it has been viewed as the combination of machine vision and spec-
troscopy techniques (Manley, 2014; Wu and Sun, 2013). Compared
with thermal imaging and MRI approaches, hyperspectral imaging
has less temperature influence and is more suitable for online
detection (Van Zeebroeck et al., 2007). Thus, the majority of inves-
tigators had applied the hyperspectral imaging for detecting fruit
mechanical damage (ElMasry et al., 2008; Xing et al., 2005).
Siedliska et al. (2014) detected apple bruise based on supervised
classification models which were also utilized by Baranowski
et al. (2013, 2012) to classify the bruised apple using hyperspectral
and thermal imaging techniques. The time evolution of apple
bruise was taken into account by the above investigators
(Baranowski et al., 2013). In regard to damage time evolution,

http://dx.doi.org/10.1016/j.compag.2016.01.015
0168-1699/ 2016 Elsevier B.V. All rights reserved.
20 M.-H. Hu et al. / Computers and Electronics in Agriculture 122 (2016) 1928
Fig. 1. The definition of blueberry damage degrees: (a) sound; (b) minor damage; and (c) (major) damage.
Fig. 2. Photographs of hyperspectral reflectance/transmittance/interactance imag-
ing system.
Jimenez-Jimenez et al. (2013) used digital camera to conduct tem-
poral evolution assessment of table olive damage. Huang et al.
(2015) developed an online multispectral imaging system for apple
bruise detection. H. Lee et al. (2014) used support vector machine
(SVM) to classify hyperspectral morphological features for recog-
nizing crack in tomato. In the publication of K.Q. Yu et al. (2014),
the crack in the surface of fresh jujube was discriminated by the
combined use of vis/NIR hyperspectral imaging and least-squares
SVM. The feasibility of using hyperspectral imaging for the early
detection of mechanical damage in mango had been proved by
the previous study of Rivera et al. (2014), and their results demon-
strated that the k-Nearest Neighbors was superior to supervised
classification models. W.-H. Lee et al. (2014) used NIR hyperspec-
tral imaging combining with F-value based band ratio algorithm
to detect pear bruise. A novel hyperspectral imaging system based
on two liquid crystal tuneable filters was developed for the inspec-
tion of citrus fruit decay (Gomez-Sanchis et al., 2014).
Previous literature on determining fruit mechanical damage
usually applied hyperspectral imaging technique in reflectance
mode. The potential use of the other sensing modes, such as trans-
mittance and interactance, is required to be discussed for the
detection of fruit mechanical damage. Transmittance mode had
been effectively used to perform the detections of pickling
cucumber internal defect (Cen et al., 2014) and parasite in cooked
clam (Coelho et al., 2013) as well as insect-damaged soybean
(Huang et al., 2013; Ma et al., 2014). For interactance, despite being
advantageous over the other modes (Wu and Sun, 2013), few stud-
ies had been reported to conduct this mode for food quality evalua-
tion (ElMasry and Wold, 2008; Gou et al., 2013; Sivertsen et al.,
2012) due to the relatively complicated imaging architecture
(Leiva-Valenzuela et al., 2014). The successful application of these
modes had been achieved in the studies of spectroscopy (Nicola
et al., 2007; Wang et al., 2013), owing to no or low spatial resolution.
In the case of non-destructive techniques for blueberry quality
control, Li et al. (2014) and Yang et al. (2014) used traditional com-
puter vision and hyperspectral imaging to identify the blueberry
mature stages, respectively. Leiva-Valenzuela et al. (2014) corre-
lated the hyperspectral reflectance and transmittance data with
the soluble solids content and firmness of blueberry by partial least
squares regression. In order to recognize visible mechanical dam-
aged blueberries, Leiva-Valenzuela and Aguilera (2013) proposed
the classifiers based on a traditional computer vision system, and
the average classification accuracy was beyond 85%. However,
the mechanical damage underneath the blueberry skin was unde-
tectable when using these classifiers. The principle of optical non-
destructive method is based on the fact that the change of fruit
(micro) structure will accompany the change of fruit optical prop-
erties (Allan-Wojtas et al., 2001; Mollazade et al., 2012). Apart
from the optical measurements, Li et al. (2010) used electronic
nose to identify the fungal disease of blueberry. The study of using
electronic nose in classifying the impacted blueberries was carried
out by Demir et al. (2011), and one of their results showed that
blueberry impacted damage was invisible to the naked eye. Unlike
optical methods, the testing principle of electronic nose is on the
basis of the changes in volatile characteristics. From the above lit-
erature evidences, very limited research work had been conducted
on the hyperspectral imaging for detecting the invisible mechani-
cal damage of blueberry.
The objectives of current study are to: (1) establish the classifi-
cation models based on the hyperspectral reflectance and trans-
mittance as well as interactance modes for distinguishing
blueberry mechanical damage; (2) verify the main factors of the
impact damage incidence, e.g. mechanical properties and absorbed
energy, and characterize the spectral behavior during cold storage;
and (3) use the established classification models to identify blue-
berry impact damage with respect to the time evolution.
2. Materials and methods
2.1. Blueberry sample preparation
A total of 737 blueberries (Vaccinium corymbosum) were col-
lected from Frutera San Fernando S.A., Chile during December
2014 to January 2015. After transported to the lab, the blueberries
were stored at environmental temperature of 4 C and relative
 M.-H. Hu et al. / Computers and Electronics in Agriculture 122 (2016) 1928 21

Fig. 3. Hypercube processing and classifier descriptions: (a) reflectance and transmittance hypercubes; (b) procedures of hypercube processing; (c) procedure of spectral
processing; (d) classifier descriptions. Note: 1 = region growing of seed (1, 1) with threshold 400; interactance images are segmented by the same operation of reflectance,
expect that the threshold to eliminate bright pixels is set as 800.

humidity of approximate 85%. Mature samples with dark blue or berries were applied to induce the impact damage, and analyzed
purple skin were selected for further analysis (Yang et al., 2014). by hyperspectral imaging with respect to the time evolution
Among 737 samples, 304 sound and 253 damaged berries were measurements were carried out before impact and immediately
used for establishing classification models. The rest of 180 sound after impact (fresh damage), followed by 12 h, 1 and 2 days.
22 M.-H. Hu et al. / Computers and Electronics in Agriculture 122 (2016) 1928

Fig. 4. Typical curve of three-parameter Lorentz function with constant.

After the experiments, the testing berries were sliced through

equator to guarantee the ground truth information of the samples.
Furthermore, in order to ensure the berries for impact test with no
prior mechanical damage, the classification models established in
current study were applied to eliminate the prior-damaged
samples.
The damage degrees were defined by P. Yu et al. (2014a): the
damaged areas more and less than 25% of cut surface are grouped
as (major) damage and minor damage, respectively (Fig. 1). In cur-
rent work, berries of minor damage were considered to be accept-
able for marketing, and therefore, we categorized them into sound
group when conducting classification. For the reasons that the
blueberry has deep dark pigments in skin and mechanical damage
is located below the skin, the mechanical damage is non-visible to
the naked eye.

2.2. Blueberry impact damage

Blueberries with the equatorial orientation were dropped from

a controlled height of 120 cm which had been applied by P. Yu
et al. (2014b) onto a contact surface made of hard plastic and
therefore impact damage. In this experiment, the mean fruit mass
was 2.095 g, and therefore, the mean absorbed impact energy was
24.635 mJ, which was calculated from the following equation
(Jimnez-Jimnez et al., 2012):
EMgH 1
where E, M and H are the absorbed impact energy (mJ), mean fruit
mass (g) and drop height (m), respectively, and g represents the
earths gravitational acceleration (here is 9.8 m/s2). After the first
impulsive force, most blueberries returned a small height, thus
ignoring the absorbed energy caused by the rebound height.
Fig. 5. Mean sound and damaged spectra of blueberries: (a) reflectance; (b)
2.3. Hyperspectral image acquisition transmittance (dot lines are offset); and (c) interactance (SD = standard deviation).

The pushbroom vis/NIR hyperspectral reflectance/transmit

tance/interactance imaging system was presented in our previous
study of Hu et al. (2016) (Fig. 2). This system can acquire the spec-
tral range from 328.81 nm to 1113.54 nm with full width at half
maximum of approximately 0.8 nm.
For reflectance and transmittance, the intensities of halogen
light source were set to 44 KLux and 59 KLux, respectively, and
the exposure times were accordingly set as 9 ms and 30 ms to
reduce the saturation. In order to obtain the images with high spa-
tial resolution, the motor speeds of reflectance and transmittance
were set at 6 mm/s and 3 mm/s, respectively. The gain of
electron-magnifying charge-coupled detector (EMCCD) camera
was fixed to 1. The reflectance and transmittance illumination
angles in respect to the vertical direction were about 30 and
20, respectively.
However, the interactance illumination angle was the vertical
direction and hyperspectral data collection line. The light shield
made of black metal plates was used to block the most of reflected
light. Actually, the reflected light cannot be reduce completely, and
 M.-H. Hu et al. / Computers and Electronics in Agriculture 122 (2016) 1928 23

therefore, the obtained data in interactance mode are the mixed

spectra of reflectance and interactance. The interactance imaging
framework in current work was adjusted on the basis of our previ-
ous study (Hu et al., 2016). The light intensities and exposure time
were set to 50 ms and 4.2 Klux. The motor speed was set at
2.5 mm/s. The gain of EMCCD was fixed to 2500 to ensure the pixel
values of the images within 80% EMCCD maximum digital number.
For all imaging modes, each berry was imaged with the orienta-
tion of the stem scar facing vertically toward the camera. Prior to
experiments, the halogen light source should be heated for half
an hour.
The obtained hypercubes were corrected by the reference
images line by line using the following equation:

SR  SD
Sr  10a 2
RW  RD

In equation, Sr is the relative sample image, SR is the sample image,

Fig. 6. Mean weight of 1000 CARS runs. (Note that the rectangle tags indicates the SD is the dark image of operated exposure time, RW is the reference
range of main peaks for reflectance and interactance, but the main two peaks for white image of corrected exposure time, and RD is the reference
transmittance.) dark image of corrected exposure time. Ten reference white and
dark images were acquired each time, and then averaged for correc-
tion. The a in Eq. (2) is used to magnify spectral response values,
Table 1 and it is determined by the priori knowledge which is equivalent
Results of damage-specific wavelengths selected by CARS algorithm. to 4 for reflectance and interactance modes, and 2 for
Imaging mode CARS selected wavelengths Wavelength
transmittance.
number A rectangular white standard (Spectral Imaging Ltd., Finland)
was applied to capture the RW image. The corrected exposure times
Reflectance 795.28; 800.91; 804.13; 815.39; 821.03; 27
832.31; 834.72; 838.75; 842.78; 849.23; of RW and RD images were set to 2.5 ms, 1000 ms and 10 ms for
851.65; 859.71; 867.78; 871.00; 879.07; reflectance, transmittance and interactance, respectively.
887.13; 890.36; 891.97; 895.20; 896.01;
898.43; 899.24; 921.82; 935.53; 939.56;
945.21; 956.49 2.4. Hypercube analysis and classifier establishment
Transmittance 720.73; 737.51; 738.31; 739.12; 796.89; 32
798.50; 799.30; 800.91; 801.71; 802.52; The procedures of hypercube processing are summarized in
803.32; 804.13; 805.74; 806.54; 807.34;
Fig. 3(b). A region growing algorithm was used to create reflec-
861.32; 867.78; 870.20; 871.81; 872.62;
883.1; 884.71; 894.40; 910.53; 953.27;
tance and transmittance mask images at 834.72 nm, allowing the
954.88; 958.91; 1022.45; 1023.26; 1037.69; subsequent acquisition of background-eliminated reflectance
1046.51; 1050.51 images at 834.72 nm and transmittance images at 552.79 nm.
Interactance 804.93; 821.03; 852.45; 854.07; 858.9; 40 The bright pixels in reflectance images are eliminated using simple
860.52; 863.74; 876.65; 878.26; 884.71; threshold method, and the threshold 400 is determined by our pre-
904.88; 919.40; 923.44; 929.89; 933.92; liminary experiments. The saturation region caused by illumina-
934.73; 935.53; 941.98; 950.04; 956.49;
tion angle in transmittance images is removed by applying
958.10; 980.65; 999.16; 1007.19; 1016.83;
1023.26; 1024.06; 1032.08; 1032.88; region growing algorithm again. Interactance images can be seg-
1035.29; 1036.09; 1037.69; 1039.30; mented in the same operations as in reflectance, except that the
1051.31; 1052.11; 1068.11; 1084.88; threshold for the use in bright pixel elimination is set at 800. For
1085.67; 1094.44; 1100.81
each imaging mode, the final mask images mask to the correspond-
ing hypercubes to obtain the resulting segmented images.

Table 2
Confusion matrices of five classifiers for distinguishing sound and damaged blueberries with the use of CARS selected reflectance, transmittance and interactance spectra.

Imaging mode SMO Logistic MP-ANN Trees.FT Trees.LMT

Sound Damage Sound Damage Sound Damage Sound Damage Sound Damage
Reflectance
Sound 222/74a 46/16 215/75 19/18 206/72 27/14 214/71 28/15 218/72 28/14
Damage 6/2 144/47 13/1 171/45 22/4 163/49 14/5 162/48 10/4 162/49
Accuracy (%) 97.4/97.4 75.8/74.6 94.3/98.7 90.0/71.4 90.4/94.7 85.8/77.8 93.9/93.4 85.3/76.2 95.6/94.7 85.3/77.8
Transmittance
Sound 219/71 35/0 214/68 22/1 209/72 30/0 217/71 29/0 217/72 26/0
Damage 9/5 155/63 14/8 168/62 19/4 160/63 11/5 161/63 11/4 164/63
Accuracy (%) 96.1/93.4 81.6/100 93.9/89.5 88.4/98.4 91.2/94.7 84.2/100 95.2/93.4 84.7/100 95.2/94.7 86.3/100
Interactance
Sound 215/72 43/0 208/53 23/0 210/65 30/0 205/70 33/0 214/72 37/0
Damage 13/4 147/63 20/23 167/63 18/11 160/63 23/6 157/63 14/4 153/63
Accuracy (%) 94.3/94.7 77.4/100 91.2/69.7 87.9/100 92.1/85.5 84.2/100 89.9/92.1 82.6/100 93.9/94.7 80.5/100
a
Calibration sets/Prediction sets; major diagonal of confusion matrix is actual class.
24 M.-H. Hu et al. / Computers and Electronics in Agriculture 122 (2016) 1928
Fig. 7. ROC curves of five classification models in prediction blueberry set:
(a) reflectance; (b) transmittance; and (c) interactance.
Fig. 3(c) shows the procedures of spectral processing. The
extracted mean spectra are subjected to the maximum normaliza-
tion to remove the variations among fruits and images (Xing and
Guyer, 2008). Subsequently, the maximum-normalized spectra
are pretreated by the Standard Normal Variate algorithm. In order
to select effective wavelength regarding blueberry damage detec-
tion, Competitive Adaptive Reweighted Sampling (CARS) algorithm
is applied to calibration sample set including 228 sound and 190
damaged berries. The CARS algorithm chooses the informative
Fig. 8. Statistical analyses of blueberry predicted hardness and absorbed impact
energy. Note: within the same symbol, the symbols not followed by the same letter
are highly significant difference (p < 0.01).
spectra based on the principle of survival of the fittest of the
Darwins Evolution Theory (Li et al., 2009; Chen et al., 2015). In
terms of every imaging modes, the damage-specific wavelengths
are confirmed by mean CARS weights of full spectrum calculated
from 1000 repeated times. In each CARS run, 10 folds cross
validation is employed. Above spectral operations are suitable for
reflectance and transmittance as well as interactance spectra.
Five classification models viz. Sequential Minimal Optimization
(SMO) (Keerthi et al., 2001), logistic regression (Cessie and
Houwelingen, 1992), multilayer perceptron back-propagation
(MP-ANN), function tree (tree.FT) (Gama, 2004) and logistic model
tree (tree.LMT) (Landwehr et al., 2005) are applied to build classi-
fiers based on damage-specific spectra (Fig. 3(d)). The calibration
set is used for modeling with 10 folds cross validation. Model per-
formances are subsequently evaluated by the prediction set involv-
ing 76 sound and 63 damaged berries, through the application of
confusion matrix of classification accuracies and Receiver Operat-
ing Characteristics (ROC) curve (Siedliska et al., 2014). The num-
bers along major diagonal of confusion matrix indicate the true
decisions made by classifier (Fawcett, 2006). In ROC curve, true
positive rate is served as Y axis and false positive rate as X axis.
These two rates can be derived from the following expression:
( true positives
true positive rate true positivesfalse negatives
false positives
3
false positive rate false positivestrue negatives
where if the sample is positive and the predicted class is positive, it
is counted as true positives; if the predicted class is negative, it is
grouped as false negatives. If the negative sample is classified as
negative, it is counted as true negatives; otherwise counted as false
positives (Fawcett, 2006). The Area Under ROC Curve (AUC) is used
to estimate classifiers (Zhang and Su, 2006).
The hypercube and spectral analyses were implemented using
Matlab R2009b (The Mathworks, Inc., Natick, MA, USA). A software
of Waikato Environment for Knowledge Analysis (Weka) Version 3.
6. 12 (University of Waikato, Hamilton, New Zealand) was used to
execute modeling operations.
2.5. Statistical analysis of spectra and factors for impact incidence
In order to conduct the one-way ANalysis Of VAriance (one-way
ANOVA) and Least Significant Difference (LSD) multiple compar-
isons for spectral data and its time evolution, the raw relative
reflectance and interactance spectra are fitted by modified three-
parameter Lorentz function with constant (Hillmann et al., 1996;
Lu and Peng, 2006):
 M.-H. Hu et al. / Computers and Electronics in Agriculture 122 (2016) 1928 25

  !
2A w
y w2 c 4
p 4x  xc 2

where x denotes waveband vector and its corresponding spectra are

stored in y; A represents the area under Lorentz curve; xc is the cen-
ter of curve; w is the width at half maximal spectral peak and c is a
curve offset.
The descriptions of Lorentz function parameters are shown in
Fig. 4. Comparison of sound and damaged spectra in Fig. 5
(a) and (c) demonstrated that the possibility to use A parameter
to characterize these two blueberry classes; hence, the area param-
eter in Lorentz function is chosen for statistical analyses of reflec-
tance and interactance spectra.
Nonetheless, transmittance spectra cannot be fitted by Lorentz
function. The offset presented in Fig. 5(b) shows a capability of
identifying sound and damaged berries, thus being used as the sta-
tistical parameter for transmittance spectra.
Two main factors affecting blueberry impact incidence, e.g.,
hardness and absorbed impact energy, are considered in our
study. A transmittance-based prediction model of mechanical
properties proposed in our previous study (Hu et al., 2015) is
applied to obtain blueberry hardness. In current work, adding
transmittance wavelengths, this prediction model is improved in
terms of calibration and prediction correlation coefficients of
0.89 and 0.92 as well as the ratio of percent deviation (RPD) of
1.97. The absorbed impact energy of blueberry is calculated from
Eq. (1) in Section 2.2.
The curves were fitted in Matlab R2009b (The Mathworks, Inc.,
Natick, MA, USA), and statistical analyses were carried out in SPSS
Version 22.0 (International Business Machines Corp., New York,
USA).

3. Results and discussion

3.1. Spectral properties of sound and affected blueberries

Mean reflectance and transmittance as well as interactance

Fig. 9. Mean transmittance spectra of blueberries in respect to successive
measurement times: (a) sound; (b) minor damage; and (c) (major) damage.
spectra of 304 sound and 253 damaged berries are shown in
Fig. 5. For each imaging mode, owing to the strong absorption of
the deep dark pigments in blueberry skin, consistently low light
intensities were observed in the spectra below 690.43 nm for both
sound and damaged berries. Beyond 690.43 nm, reflectance and
transmittance of sound berries were higher than those of damaged
berries; however, the reverse spectral phenomenon was found in
transmittance. The possible reason for this might be that the
damaged tissues allowed more transmittance passing through
blueberry, but absorbed more reflected and interacted light.
Results of one-way ANOVA demonstrated that there were highly
significant differences between sound and damaged berries in all
imaging modes (p < 0.01; data not shown), suggesting the
possibility of pattern recognition techniques to identify affected
blueberries. Considering the quantum efficiencies of imaging units
and signal-to-noise ratio of hypercubes, the spectral ranges of
679.31000.76 nm, 690.431050.51 nm and 683.271109.56 nm

Table 3
Results of multiple comparison of sound and affected blueberries using offset in transmittance on successive measurement times.

Sound Minor damage (Major) Damage

Before impact 1087.44 115.52a 1102.43 146.90a 1194.89 157.67a
Immediately after impact 1111.06 138.27a 1164.93 151.81b 1313.14 187.34b
12 h after impact 1215.55 226.36b 1352.606 215.08c 1525.59 226.24c
1 day after impact 1116.98 146.15a 1222.58 163.67bd 1418.79 229.27d
2 days after impact 1131.72 150.65a 1231.41 174.27d 1464.71 265.70 cd

Note: within the same column, values not followed by the same letter are significantly different (p < 0.05).
26 M.-H. Hu et al. / Computers and Electronics in Agriculture 122 (2016) 1928

Table 4
Confusion matrices of transmittance-based classifiers for detecting sound and damaged blueberries with respect to the times.

Time after impact SMO Logistic MP-ANN Trees.FT Trees.LMT

Sound Damage Sound Damage Sound Damage Sound Damage Sound Damage
Immediately
Sound 125 39 102 18 118 30 124 38 125 33
Damage 1 4 24 25 8 13 2 5 1 10
Accuracy (%) 99.2 9.3 81.0 58.1 93.7 30.2 98.4 11.6 99.2 23.3
12 h
Sound 116 19 20 2 71 5 112 18 101 10
Damage 10 24 106 41 55 38 14 25 25 33
Accuracy (%) 92.1 55.8 18.9 95.3 56.3 88.4 88.9 58.1 80.2 76.7
1 day
Sound 123 29 113 16 120 19 120 28 121 24
Damage 3 14 13 27 6 24 6 15 5 19
Accuracy (%) 97.6 32.6 89.7 62.8 95.2 55.8 95.2 34.9 96.0 44.2
2 days
Sound 123 18 108 9 116 11 119 16 120 12
Damage 3 25 18 34 10 32 7 27 6 31
Accuracy (%) 97.6 58.1 85.7 79.1 92.1 74.4 94.4 62.7 95.2 72.1

Note: major diagonal of confusion matrix is actual class.

were selected for reflectance, transmittance and interactance,
respectively.
CARS algorithm was used to select the wavelengths informative
with mechanical damage to reduce the redundant data in spectra,
thus enhancing accuracy and robustness of classifier. Mean CARS
weights are illustrated in Fig. 6. The CARS weight peaks or valleys
for reflectance concentrated on the spectral range from 795.28 nm
to 899.24 nm, indicating that this very Near Infrared in reflectance
could interpret most of damage information. However, the interac-
tance range of 1016.831085.67 nm might be sensitive to blue-
berry mechanical damage. Particularly, in the case of
transmittance, the sharp valley and peak occurred at 802.52 nm
and 803.32 nm, respectively. According to the mean weight of
every wavelength, a total of 27, 32 and 40 wavelengths were cho-
sen as damaged-specific spectra for reflectance, transmittance and
interactance, respectively (Table 1).
3.2. Classification models based on selected spectra
Confusion matrices of five classification models are summarized
in Table 2. For all classifiers, the discrimination accuracies of sound
blueberries in prediction (calibration) sample sets were beyond
93.4% (90.4%) and 89.5% (91.2%) for reflectance and transmittance,
respectively. In addition to 69.7% accuracy for logistic model, inter-
actance could also predict sound blueberries with accuracy beyond
85.5%. In terms of reflectance, the best classification accuracy of
77.8% (85.3%) were observed for MP-ANN and trees.LMT models
in damaged sample sets of prediction (calibration), followed by
trees.FT 76.2% (85.3%), SMO 74.6% (75.8%) and logistic
71.4% (90.0%). Compared with reflectance, four transmittance
models viz. SMO, MP-ANN, trees.FT and trees.LMT achieved 100%
accuracy for recognizing damaged samples in prediction set. Mean-
while, accuracies of these models in calibration set were accept-
able (>81.6%). For interactance, no misclassification was found for
five established classifiers in damaged prediction set; however,
these model performances (>77.4%) in calibration set were not as
good as transmittance models. In conclusion, as shown in Table 2,
transmittance-based classifiers showed superior performances
than reflectance and interactance.
AUC is an important statistical parameter for evaluating classi-
fier performance: the closer AUC is to 1, the better overall diagnos-
tic performance of established classifier (Fawcett, 2006; Siedliska
et al., 2014). Fig. 7 presents ROC curves of classifiers and the corre-
sponding AUC values. Overall, compared to transmittance and
interactance models, reflectance models had relatively low AUC
values (between 0.85 and 0.92). AUC values of transmittance and
interactance models were very high and ranged from 0.96 to
1.00. Logistic model in interactance and tree.LMT model in trans-
mittance as well as interactance obtained the AUC value of 1.00.
This observation referring to AUC could explain the fact that trans-
mitted and interacted spectra contain more internal information
than reflected spectra (Fu et al., 2007; Wu and Sun, 2013). Based
on AUC values in combination with confusion matrices, classifiers
in transmittance were considered to outperform these in
reflectance and interactance.
3.3. Factors for impact incidence of blueberries
The mechanical damage incidence and severity of fruit are
affected by several factors including physical properties, drop
height and compression degree (Knee, 2002; Opara and Pathare,
2014). In present work, physical property and absorbed impact
energy was analyzed, and the other factors such as variety and
maturity were to great extent kept consistent from berry to berry.
Before impact experiments, when the testing blueberries were
simultaneously identified as damage by four transmittance-based
classifiers built in Section 3.2, they were regarded as prior-
damaged samples. A total of 11 prior-damaged samples were elim-
inated from the following analysis. After impact treatments, due to
the individual variability among blueberries, the dropped berries
were categorized into three groups viz. damage (43 cases), minor
damage (67 cases) and sound (59 cases) by manual evaluation.
A double Y axis plot of these two factors is displayed in Fig. 8 for
three affected groups. No significant difference was observed
between absorbed impact energy and severity of impact damage
(p > 0.05). Highly significant effect of physical property was found
for impact damage occurrence and severity of blueberry (p < 0.01).
As could be seen from Fig. 8, the increase in physical property, as
expressed by spectral predicted hardness, would lead to high dam-
age susceptibility of blueberry. Previous studies of Abedi and
Ahmadi (2014) and Ahmadi et al. (2010) also reported that the
physical properties affected apple and peach bruise susceptibilities
significantly. From this point of view, above results demonstrated
the potential of using physical property parameters to estimate
the damage susceptibility of blueberry in further study. Conse-
quently, we preliminarily concluded that the physical properties
i.e. hardness is the main factor affecting blueberry damage suscep-
tibility, particularly for impact damage.
 M.-H. Hu et al. / Computers and Electronics in Agriculture 122 (2016) 1928
3.4. Spectral behavior of sound and impacted blueberries
As results reported in Section 3.2, only transmittance behavior
of affected blueberry was evaluated during cold storage. Mean
transmittance spectra measured on successive times were shown
in Fig. 9 for three affected groups. To quantitatively analyze spec-
tral changes, the multiple comparison was conducted and the
results were listed in Table 3. In the case of sound group, the signif-
icant differences were found between 12 h measurement time and
the others (p < 0.05). This was also observed in minor damage
group; however, there were no significant difference between
1 day and 2 days measurements, either between 1 day and imme-
diate measurements (p > 0.05). For damage group, no notable dif-
ference existed between 2 days and 1 day measurements, either
between 2 days and 12 h measurements (p < 0.05). Overall, the
results presented in Table 3 were consistent with change tenden-
cies in Fig. 9 by visual inspection. Further experiments are required
for interpreting why the considerable spectral amplitudes occurred
in 12 h measurement time after impacted.
3.5. Classified accuracies of damaged blueberries with time evolution
Table 4 shows confusion matrices of transmittance-based mod-
els for identifying sound and damaged blueberries over a period of
time after mechanically impacted. Generally, sound blueberries
were better classified than damaged except at 12 h measurement
where 106 and 55 sound berries were identified as damaged
respectively for logistic and MP-ANN models. The possible reason
for this might be that more berries of aged damages were used dur-
ing the modeling procedures. On immediate measurement time, in
addition to 58.1% accuracy for logistic, the other models did not
perform well for damaged blueberries with prediction accuracies
below 30.2%. Though prediction accuracies higher than 88% were
obtained for logistic and MP-ANN models on 12 h measurement,
their performances were inferior in classifying sound blueberries.
Interestingly, trees.LMT model achieved acceptable accuracies of
80.2% and 76.7%, respectively, for sound and damaged blueberries
on 12 h after impacted. Nonetheless, models performances became
inferior at 1 day measurement time with accuracies not exceeding
62.8% in damaged sample classifications. At 2 days after impacted,
the satisfactory results were obtained for all models in classifying
damaged (sound) blueberries and listed in descending order based
on damaged prediction accuracies: logistic 79.1% (85.7%), MP-
ANN 74.4% (92.1%), trees.LMT 72.1% (95.2%), trees.FT 62.7%
(94.4%) and SMO 58.1% (97.6%). Among them, 3 sound and 9
damaged samples were misclassified by SMO and logistic models,
respectively. The MP-ANN and trees.LMT classifiers were suitable
for both sample groups owing to the relatively high sound and
damaged prediction accuracies.
4. Conclusion
In conclusion, hyperspectral transmittance spectra were identi-
fied to be more sensitive to blueberry mechanical damage than
reflectance and interactance. Classifiers based on CARS selected
transmittance showed superior performances in classifying sound
and damaged blueberries according to the relatively high classifi-
cation accuracies and AUC values. In impact damage experiments,
the physical property preliminarily proved to be more pronounced
than the absorbed impact energy for damage incidence and sever-
ity of blueberry. In addition, established transmittance-based clas-
sifiers were tested with new sample sets with time evolution, and
results demonstrated that it was especially difficult to distinguish
damaged blueberries on the immediate measurement time. At
2 days after impacted, all transmittance classifiers obtained
27
satisfactory accuracies for predicting both sound and damaged
blueberries. This research will greatly benefit the blueberry indus-
try such as the elimination of the damaged berries allowing the
berries up to one year in the freezer.
Acknowledgements
This study was supported by the National Natural Science Foun-
dation of China (NSFC31271896), the Innovation Fund Project for
Graduate Student of Shanghai (JWCXSL1401) and the Shanghai
Municipal Natural Science Foundation (12ZR1420500). The authors
gratefully acknowledge Dr. Renfu Lu for his help in experimental
design.
References
Abedi, G., Ahmadi, E., 2014. Bruise susceptibilities of Golden Delicious apples as
affected by mechanical impact and fruit properties. J. Agric. Sci. 152, 439447.
Ahmadi, E., Ghassemzadeh, H.R., Sadeghi, M., Moghaddam, M., Neshat, S.Z., 2010.
The effect of impact and fruit properties on the bruising of peach. J. Food Eng.
97, 110117.
Allan-Wojtas, P.M., Forney, C.F., Carbyn, S.E., Nicholas, K., 2001. Microstructural
indicators of quality-related characteristics of blueberries an integrated
approach. LWT Food Sci. Technol. 34, 2332.
Baranowski, P., Mazurek, W., Pastuszka-Wozniak, J., 2013. Supervised classification
of bruised apples with respect to the time after bruising on the basis of
hyperspectral imaging data. Postharvest Biol. Technol. 86, 249258.
Baranowski, P., Mazurek, W., Witkowska-Walczak, B., Sawinski, C., 2009. Detection
of early apple bruises using pulsed-phase thermography. Postharvest Biol.
Technol. 53, 91100.
Baranowski, P., Mazurek, W., Wozniak, J., Majewska, U., 2012. Detection of early
bruises in apples using hyperspectral data and thermal imaging. J. Food Eng.
110, 345355.
Cen, H., Lu, R., Ariana, D.P., Mendoza, F., 2014. Hyperspectral imaging-based
classification and wavebands selection for internal defect detection of pickling
cucumbers. Food Bioprocess Technol. 7, 16891700.
Cessie, L.S., Houwelingen, J.C.V., 1992. Ridge estimators in logistic regression. Appl.
Stat. 41, 191201.
Coelho, P.A., Soto, M.E., Torres, S.N., Sbarbaro, D.G., Pezoa, J.E., 2013. Hyperspectral
transmittance imaging of the shell-free cooked clam Mulinia edulis for parasite
detection. J. Food Eng. 117, 408416.
Chen, X., Liu, K., Cai, J., Zhu, D., Chen, H., 2015. Identification of heavy metal-
contaminated Tegillarca granosa using infrared spectroscopy. Anal. Methods 7,
21722181.
Demir, N., Ferraz, A.C.O., Sargent, S.A., Balaban, M.O., 2011. Classification of
impacted blueberries during storage using an electronic nose. J. Sci. Food
Agric. 91, 17221727.
Elks, C.M., Francis, J., Stull, A.J., Cefalu, W.T., Shukitt-Hale, B., Ingram, D.K., 2013.
Overview of the Health Properties of Blueberries, Bioactives in Fruit. John Wiley
& Sons, Ltd., pp. 251271.
ElMasry, G., Wang, N., Vigneault, C., Qiao, J., ElSayed, A., 2008. Early detection of
apple bruises on different background colors using hyperspectral imaging. LWT
Food Sci. Technol. 41, 337345.
ElMasry, G., Wold, J.P., 2008. High-speed assessment of fat and water content
distribution in fish fillets using online imaging spectroscopy. J. Agric. Food
Chem. 56, 76727677.
Fawcett, T., 2006. An introduction to ROC analysis. Pattern Recogn. Lett. 27, 861
874.
Fu, X., Ying, Y., Lu, H., Xu, H., 2007. Comparison of diffuse reflectance and
transmission mode of visible-near infrared spectroscopy for detecting brown
heart of pear. J. Food Eng. 83, 317323.
Gama, J., 2004. Functional trees. Mach. Learn. 55, 219250.
Gomez-Sanchis, J., Lorente, D., Soria-Olivas, E., Aleixos, N., Cubero, S., Blasco, J., 2014.
Development of a hyperspectral computer vision system based on two liquid
crystal tuneable filters for fruit inspection. Application to detect citrus fruits
decay. Food Bioprocess Technol. 7, 10471056.
Gou, P., Santos-Garcs, E., Hy, M., Wold, J., Liland, K., Fulladosa, E., 2013. Feasibility
of NIR interactance hyperspectral imaging for on-line measurement of crude
composition in vacuum packed dry-cured ham slices. Meat Sci. 95, 250255.
Hillmann, U., Schimmack, W., Jacob, P., Bunzl, K., 1996. In situ c-spectrometry
several years after deposition of radiocesium. Radiat. Environ. Biophys. 35, 297
303.
Huang, M., Wan, X., Zhang, M., Zhu, Q., 2013. Detection of insect-damaged vegetable
soybeans using hyperspectral transmittance image. J. Food Eng. 116, 4549.
Hu, M.H., Dong, Q.L., Liu, B.L., Opara, U.L., Chen, L., 2015. Estimating blueberry
mechanical properties based on random frog selected hyperspectral data.
Postharvest Biol. Technol. 106, 110.
Hu, M.H., Dong, Q.L., Liu, B.L., Opara, U.L., 2016. Prediction of mechanical properties
of blueberry using hyperspectral interactance imaging. Postharvest Biol.
Technol. 115, 122131.
28 M.-H. Hu et al. / Computers and Electronics in Agriculture 122 (2016) 1928
Huang, W.Q., Li, J.B., Wang, Q.Y., Chen, L.P., 2015. Development of a multispectral
imaging system for online detection of bruises on apples. J. Food Eng. 146, 62
71.
Jimnez-Jimnez, F., Castro-Garca, S., Blanco-Roldn, G.L., Agera-Vega, J., Gil-
Ribes, J.A., 2012. Non-destructive determination of impact bruising on table
olives using VisNIR spectroscopy. Biosyst. Eng. 113, 371378.
Jimenez-Jimenez, F., Castro-Garcia, S., Blanco-Roldan, G.L., Gonzalez-Sanchez, E.J.,
Gil-Ribes, J.A., 2013. Isolation of table olive damage causes and bruise time
evolution during fruit detachment with trunk shaker. Span. J. Agric. Res. 11, 65
71.
Keerthi, S.S., Shevade, S.K., Bhattacharyya, C., Murthy, K.R.K., 2001. Improvements to
Platts SMO algorithm for SVM classifier design. Neural Comput. 13, 637649.
Knee, M., 2002. Fruit Quality and its Biological Basis. CRC Press.
Landwehr, N., Hall, M., Frank, E., 2005. Logistic model trees. Mach. Learn. 59, 161
205.
Li, H., Liang, Y., Xu, Q., Cao, D., 2009. Key wavelengths screening using competitive
adaptive reweighted sampling method for multivariate calibration. Anal. Chim.
Acta 648, 7784.
Lee, H., Kim, M.S., Jeong, D., Delwiche, S.R., Chao, K., Cho, B.K., 2014. Detection of
cracks on tomatoes using a hyperspectral near-infrared reflectance imaging
system. Sensors 14, 1883718850.
Lee, W.-H., Kim, M.S., Lee, H., Delwiche, S.R., Bae, H., Kim, D.-Y., Cho, B.-K., 2014.
Hyperspectral near-infrared imaging for the detection of physical damages of
pear. J. Food Eng. 130, 17.
Leiva-Valenzuela, G.A., Aguilera, J.M., 2013. Automatic detection of orientation and
diseases in blueberries using image analysis to improve their postharvest
storage quality. Food Control 33, 166173.
Leiva-Valenzuela, G.A., Lu, R., Aguilera, J.M., 2014. Assessment of internal quality of
blueberries using hyperspectral transmittance and reflectance images with
whole spectra or selected wavelengths. Innov. Food Sci. Emerg. Technol. 24, 2
13.
Li, C., Krewer, G.W., Ji, P., Scherm, H., Kays, S.J., 2010. Gas sensor array for blueberry
fruit disease detection and classification. Postharvest Biol. Technol. 55, 144
149.
Li, H., Lee, W.S., Wang, K., 2014. Identifying blueberry fruit of different growth
stages using natural outdoor color images. Comput. Electron. Agric. 106, 91
101.
Li, Z., Thomas, C., 2014. Quantitative evaluation of mechanical damage to fresh
fruits. Trends Food Sci. Technol. 35, 138150.
Lu, R., Peng, Y., 2006. Hyperspectral scattering for assessing peach fruit firmness.
Biosyst. Eng. 93, 161171.
Ma, Y., Huang, M., Yang, B., Zhu, Q., 2014. Automatic threshold method and optimal
wavelength selection for insect-damaged vegetable soybean detection using
hyperspectral images. Comput. Electron. Agric. 106, 102110.
Manley, M., 2014. Near-infrared spectroscopy and hyperspectral imaging: non-
destructive analysis of biological materials. Chem. Soc. Rev. 43, 82008214.
Martinsen, P., Oliver, R., Seelye, R., McGlone, V.A., Holmes, T., Davy, M., Johnston, J.,
Hallett, I., Moynihan, K., 2014. Quantifying the diffuse reflectance change
caused by fresh bruises on apples. Trans. ASABE 57, 565572.
Mazhar, M., Joyce, D., Cowin, G., Brereton, I., Hofman, P., Collins, R., Gupta, M., 2015.
Non-destructive 1H-MRI assessment of flesh bruising in avocado (Persea
americana M.) cv. Hass. Postharvest Biol. Technol. 100, 3340.
Mehra, L., MacLean, D., Savelle, A., Scherm, H., 2013. Postharvest disease
development on southern highbush blueberry fruit in relation to berry flesh
type and harvest method. Plant Dis. 97, 213221.
Mollazade, K., Omid, M., Tab, F.A., Mohtasebi, S.S., 2012. Principles and applications
of light backscattering imaging in quality evaluation of agro-food products: a
review. Food Bioprocess Technol. 5, 14651485.
Nicola, B.M., Beullens, K., Bobelyn, E., Peirs, A., Saeys, W., Theron, K.I., Lammertyn, J.,
2007. Nondestructive measurement of fruit and vegetable quality by means of
NIR spectroscopy: a review. Postharvest Biol. Technol. 46, 99118.
Opara, U.L., Pathare, P.B., 2014. Bruise damage measurement and analysis of fresh
horticultural produce a review. Postharvest Biol. Technol. 91, 924.
Patel, K.K., Khan, M.A., Kar, A., 2015. Recent developments in applications of MRI
techniques for foods and agricultural producean overview. J. Food Sci.
Technol. 52, 126.
Pu, Y.-Y., Feng, Y.-Z., Sun, D.-W., 2015. Recent progress of hyperspectral imaging on
quality and safety inspection of fruits and vegetables: a review. Compr. Rev.
Food Sci. Food Safety 00, 113.
Retamales, J., 2012. Blueberries. Crop Production Science in Horticulture 21. CABI.
Rivera, N.V., Gomez-Sanchis, J., Chanona-Perez, J., Carrasco, J.J., Millan-Giralolo, M.,
Lorente, D., Cubero, S., Blasco, J., 2014. Early detection of mechanical damage in
mango using NIR hyperspectral images and machine learning. Biosyst. Eng. 122,
9198.
Shen, X., Sun, X., Xie, Q., Liu, H., Zhao, Y., Pan, Y., Hwang, C.-A., Wu, V.C.H., 2014.
Antimicrobial effect of blueberry (Vaccinium corymbosum L.) extracts against
the growth of Listeria monocytogenes and Salmonella Enteritidis. Food Control
35, 159165.
Siedliska, A., Baranowski, P., Mazurek, W., 2014. Classification models of bruise and
cultivar detection on the basis of hyperspectral imaging data. Comput. Electron.
Agric. 106, 6674.
Sivertsen, A.H., Heia, K., Hindberg, K., Godtliebsen, F., 2012. Automatic nematode
detection in cod fillets (Gadus morhua L.) by hyperspectral imaging. J. Food Eng.
111, 675681.
Van Zeebroeck, M., Van linden, V., Ramon, H., De Baerdemaeker, J., Nicolai, B.M.,
Tijskens, E., 2007. Impact damage of apples during transport and handling.
Postharvest Biol. Technol. 45, 157167.
Wang, H., Li, C., Wang, M., 2013. Quantitative determination of onion internal
quality using reflectance, interactance, and transmittance modes of
hyperspectral imaging. Trans. ASABE 56, 16231635.
Wu, D., Sun, D.-W., 2013. Advanced applications of hyperspectral imaging
technology for food quality and safety analysis and assessment: a review
Part I: fundamentals. Innov. Food Sci. Emerg. Technol. 19, 114.
Wu, G.F., Wang, C.G., 2014. Investigating the effects of simulated transport vibration
on tomato tissue damage based on vis/NIR spectroscopy. Postharvest Biol.
Technol. 98, 4147.
Xing, J., Bravo, C., Jancsk, P.T., Ramon, H., De Baerdemaeker, J., 2005. Detecting
bruises on Golden Delicious apples using hyperspectral imaging with multiple
wavebands. Biosyst. Eng. 90, 2736.
Xing, J., Guyer, D., 2008. Detecting internal insect infestation in tart cherry using
transmittance spectroscopy. Postharvest Biol. Technol. 49, 411416.
Yang, C., Lee, W.S., Gader, P., 2014. Hyperspectral band selection for detecting
different blueberry fruit maturity stages. Comput. Electron. Agric. 109, 2331.
Yu, K.Q., Zhao, Y.R., Li, X.L., Shao, Y.N., Zhu, F.L., He, Y., 2014. Identification of crack
features in fresh jujube using Vis/NIR hyperspectral imaging combined with
image processing. Comput. Electron. Agric. 103, 110.
Yu, P., Li, C., Takeda, F., Krewer, G., 2014a. Visual bruise assessment and analysis of
mechanical impact measurement in southern highbush blueberries. Appl. Eng.
Agric. 30, 2937.
Yu, P., Li, C., Takeda, F., Krewer, G., Rains, G., Hamrita, T., 2014b. Measurement of
mechanical impacts created by rotary, slapper, and sway blueberry mechanical
harvesters. Comput. Electron. Agric. 101, 8492.
Zhang, B.H., Huang, W.Q., Gong, L., Li, J.B., Zhao, C.J., Liu, C.L., Huang, D.F., 2015.
Computer vision detection of defective apples using automatic lightness
correction and weighted RVM classifier. J. Food Eng. 146, 143151.
Zhang, H., Su, J., 2006. Learning probabilistic decision trees for AUC. Pattern Recogn.
Lett. 27, 892899.
Zhou, R., Su, S., Yan, L., Li, Y., 2007. Effect of transport vibration levels on mechanical
damage and physiological responses of Huanghua pears (Pyrus pyrifolia Nakai,
cv. Huanghua). Postharvest Biol. Technol. 46, 2028.