Expert Systems with Applications 38 (2011) 1210112106

Contents lists available at ScienceDirect

Expert Systems with Applications

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

High speed detection of potato and clod using an acoustic based intelligent system
Adel Hosainpour a,, Mohammad H. Komarizade b, Asghar Mahmoudi c, Mahrokh G. Shayesteh d
a

Department of Agricultural Machinery, College of Agriculture, University of Ilam, Ilam, Iran

Department of Agricultural Machinery, College of Agriculture, University of Urmia, Urmia, Iran
c
Department of Agricultural Machinery, College of Agriculture, University of Tabriz, Tabriz, Iran
d
Department of Electrical Engineering, Faculty of Engineering, University of Urmia, Urmia, Iran
b

a r t i c l e
Keywords:
Acoustics
Clod
Discriminating
Neural networks
Potato
Separation

i n f o

a b s t r a c t
Discriminating between potato tubers and clods is the rst step in developing an automatic separation
system on potato harvesters. In this study, an acoustic-based intelligent system was developed for high
speed discriminating between potato tubers and soil clods. About 500 kg mixture of potato tubers and
clods were loaded on a belt conveyer and were impacted against a steel plate at four different velocities.
The resulting acoustic signals were recorded, processed and potential features were extracted from the
analysis of sound signals in both time and frequency domains. A multilayer perceptron neural network
with a back propagation algorithm was used for pattern recognition. Altogether, 17 potential discriminating features were selected and fed as input vectors to the articial neural network models. Optimal network was selected based on mean square error, correct detection rate and correlation coefcient. At the
belt velocity of 1 m s 1, detection accuracy of the presented system was about 97.3% and 97.6% for potatoes and clods, respectively. Increasing the belt velocity resulted in the reduction of detection accuracy
and increase in the number of miss classied samples. By using this system, it is expected that a potato
harvester may operate at a capacity of 20 ton hr 1 with the accuracy of about 97%.
2011 Elsevier Ltd. All rights reserved.

1. Introduction
Potato is one of the most prominent crops in the world with an
estimated annual production of approximately 325 million ton
(FAO, 2008). This enormous amount of potato needs fully automated harvesters. For a long time, the agro-industry has attempted
to automate potato harvesting operations. Although the tasks of
digging, picking up potatoes and removing ne clods are performed mechanically, removing clods and probable stones are usually done manually. The clods and stones must be removed before
entry into the potato harvester container; because they occupy a
large space of the container and with their abrasive nature may
mechanically damage potato tubers. Accompanying mixture of potato also reduces the air circulation and prevents proper temperature and humidity control (Main, 1971). Discriminating between
potato tubers and clods is the rst step in developing an automatic
separation system on potato harvesters. Developing an automated
system for separating clods from potato tubers is more challenging,
because potato tubers have a wide diversity in shape, size and color. Moreover, the shape, size, and moisture content of clods are not
Corresponding author. Tel.: +98 4116671294, mobile: +98 9143002974.
E-mail addresses: a_hosainpour12@yahoo.com, a.hosainpour@urmia.ac.ir (A.
Hosainpour), m.komarizade@urmia.ac.ir (M.H. Komarizade), A_mahmoudi@tabrizu.ac.ir (A. Mahmoudi), m.shayesteh@urmia.ac.ir (M.G. Shayesteh).
0957-4174/$ - see front matter 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2011.02.164

predictable in harvesting condition. Partially and completely muddy potatoes add another obstacle in developing automated separator systems.
Separating potatoes from its impurities has been investigated in
many research studies. Most researchers have examined mechanical methods to solve the problem (Brantley, Hamann, & Whiteld,
1975; Feller, Morgolin, Zacharin, & Pasternak, 1985; Gan-Mor,
Zacharin, Galili, Feller, & Margolin, 1986; Shyam, Singh, & Singh,
1990). Although some of these methods are being used in some potato harvesters, they cause surface damages and bruises on potato
tubers (McGechan, 1980). Moreover, large space occupancy and
high energy consumption are other main drawbacks of these systems. Story and Raghavan (1973) used infrared reectance to sorting potatoes from stones and soil clods. In recent years, some
machine vision based systems have been developed for discriminating between potato tubers and clods (Al-Mallahi, Kataoka, &
Okamoto, 2008; Marchant, Onyango, & Street, 1988; Morrow, Sommer, Heinemann, & Tao, 1990). Operational problems are most
likely to be faced, by using this system under eld conditions
(i.e. rain, moisture, dust, vibration). High price of the cameras
and other auxiliary devices as well as inefciency in detecting clod
and muddy potatoes are other disadvantages of this method.
Recently, non destructive acoustical experiments have been
increasingly implemented in detection and classication of agricultural products (Diezma-Iglesias, Ruiz-Altisent, & Barreiro,

12102

A. Hosainpour et al. / Expert Systems with Applications 38 (2011) 1210112106

2004; Mahmoudi, 2006; Pearson, 2001; Pearson, Cetin, Tewk, &

Haff, 2007). A good correlation between fruit rmness and resonant frequency has been reported in some investigations (GarsiaRamos, Ortiz-Canavate, Ruiz-Altisent, Diez, & Chavez, 2003; Jivanuwong, 1998; Sugiyama, Katsurai, Hong, Koyama, & Mikuriya, 1998;
Younce & Davis, 1995). Quite recently, Elbatawi (2008) used an
acoustic impact method to detect hollow heart of potato tubers.
Articial neural networks (ANN) offer much faster and more exible approach in classication elds. Also, in noisy obscure pattern,
ANN models are more efcient than statistical pattern classiers
(Kavdir & Guyer, 2008).
The main goal of this study is to investigate the feasibility of
using impact sound signals for rapid detection of clods in potato
harvesters. An articial neural network was used as a decision
making unit for this purpose. Experiment was performed in off-line
and on-line stages. In off-line experiments, potential features were
extracted from analysis of emitted sound signals, whereas in online stage, extracted features were used in real-time detection of
clods and potato tubers. Moreover, the effect of sample velocity
on detection accuracy was investigated.
2. Materials and methods
About 500 kg of a mixture of potato tubers, clods and stones
were collected just from the entrance of a potato harvester container in the 2008 harvesting season. Four varieties of potato tubers namely Marfona, Agria, Kosima and Granola, commonly
grown in Iran were used in this study. All samples of potatoes,
clods and stones were weighed and pieces smaller than 40 g were
neglected. Table 1 shows some physical characteristics of the used
samples. During weighing, mixture was manually sorted into two
discrete groups; potato and non-potato. Since the number of
stones was much lower than the number of clods (Table 1), hereafter in this paper the non-potato category is simply called clods. It
is worth mentioning that hard enough clods were selected as samples and other imsy clods were crumbled and removed before
reaching the potato container.
2.1. Experimental apparatus
A test rig was built to feed samples, drop them onto an impact
plate, acquire sound signal upon impact, process the signal data
and discriminate between potato tubers and clods. The schematic
diagram of the experimental apparatus is shown in Fig. 1. The system consisted of a feeding platform, an impact plate, an acoustic
unit and a PC based data acquisition system. A 5 m long and
0.35 m wide belt conveyor was designed and fabricated as feeding
platform. The belt conveyor was driven by a 1 kW hp electric motor. The velocity of the conveyer belt could be changed by using a
variable frequency drive (Inverter IG5). Some special cups were
mounted on the belt for singulating and better control of the sample movement. The impact plate was a polished block of stainless
steel approximately 22.5  30  1.5 cm. To study the effect of
velocity on detection accuracy, four levels of belt conveyer velocity

Table 1
Numbers and physical characteristic of used samples.
Samples

Variety

Number N

Mean weight (g)

Stda (g)

Potato

Marfona
Agria
Kosima
Granola

378
443
446
625
1483
38

186
158
150
115
123
108

82
73
88
50
72
40

Clod
Stone
a

Standard deviation.

at 1, 1.5, 2 and 2.5 m s 1 were examined throughout the study. The

impact plate position was adjusted by trial and error method for all
velocities (Table 2).
A low cost Panasonic Electret capsule microphone (VM-034CY
model), sensitive to frequencies up to 100 kHz, was used for capturing impact sound signals. The microphone was installed inside
an isolated acoustic chamber to eliminate environmental noise effects. To prevent the chamber from acoustic reections, it was
lled with glass wool. For further environmental noise elimination,
the data acquisition system was triggered by a piezoelectric sensor
mounted on the impact plate. By adding this electronic circuit
(main board and microcontroller) only impact emission sound signals from samples were recorded and the environmental noises did
not interfere with the actual desired signals. Microphone output
was send to a PC based data acquisition system, where it was digitized using a sound card (Intel 82801 BA/BAM AC97Audio controller) at a sampling frequency of 44.1 kHz, with 16 bit
resolution. A personal computer (Pentium IV) was used for acquiring, saving and processing of data as well as controlling the system.
A schematic diagram describing the overall detection system is
shown in Fig. 2. A parallel hardware and software architecture is
used to perform the described duty. In each stage of off-line experiment, about 45 samples of potato or clod were manually loaded in
the cups on the belt. By switching the belt conveyor power on,
samples were made to fall onto the impact plate separately and
one bye one. The emitted sound signals were acquired by the
microphone, digitized by the sound card and saved by using Matlab data acquisition toolbox (MathWorks, 2007). Similar procedures were repeated for all belt velocities. Since the maximum
frequency of the sound card was 44.1 kHz and data acquisition
continued for 11.61 ms after triggering, upon receiving a trigger
signal the computer acquired 512 data points from every sample
in the time-domain. Matlab software was used for data collection
and management.
2.2. Signal processing and feature extraction
For extracting potential features, recorded sound signals were
processed and analyzed in both time and frequency-domains. Typical emitted sound signals from potatoes and clods are shown in
Fig. 3. Although the maximum peak values of clod sound signals
were slightly larger than potato ones, the preliminary attempts
to use only time-domain features were not successful. However,
in order not to lose any useful transient feature, all 512 data point
amplitude (Amp) values were considered as features. A 1024-point
fast Fourier transform (FFT) was computed from each sound signal,
using a Mathwork Windows.
The magnitude, power spectral density (PSD) and peak frequency (the frequency corresponding to the peak value of FFT in
frequency spectrum) of each spectrum was computed and a low
pass lter was applied to remove the jagged spikes in the spectrum. In addition to PSD and peak frequency, the 25 normalized
FFT magnitudes before peak frequency and 25 points afterwards
were saved as another frequency-domain feature. Normalized
spectrum was calculated by dividing the magnitude of FFT at each
frequency by maximum value of spectrum (Fig. 4). After processing
and analysis of emitted sound signals in both time and frequency
domain, following features were considered as potential discriminating features: 512 amplitude (Amp) values in time domain; the
peak frequency (PF); 50 normalized FFT (NFFT) magnitudes centered about the peak frequency and the entire 512 PSD magnitudes
in frequency domain. A total of 1075 features were obtained for
each sample. For real time systems, this dimension of the input
vector is large, but the components of the vectors are highly correlated. Principle Component Analysis (PCA) was used to reduce the
dimension of the input vectors to a maximum of 20 features. Three

A. Hosainpour et al. / Expert Systems with Applications 38 (2011) 1210112106

12103

Fig. 1. (a) Schematics of experimental apparatus. (b) Position of impact plate respect to feeding platform.

number of principle components and percentage of eliminated

component variance.

Table 2
Position of impact plate at different velocity of belt conveyer.
Belt velocity
Vms 1

Horizontal distance
X cm

Vertical distance
Y cm

Angle a
()

1
1.5
2
2.5

25
27
30
34

9
8
7
5

30
30
35
35

sets of features were utilized for PCA purposes: Amp, PSD and
NFFT. After normalization of data, PCA analysis was performed
on data using Matlab software. Table 3 shows the relation between

2.3. Back propagation neural network

The multilayer perceptron (MLP) is one of the most widely
implemented neural network topologies that are being used in
the classication applications. MLPs are normally trained with
the back propagation algorithm (BPA). The back propagation rule
propagates the errors through the network and allows adaptation
of the hidden processing elements (PEs). The MLP is trained with
error correction learning, which means that the desired response
for the system must be known a priori. After adequate training,

12104

A. Hosainpour et al. / Expert Systems with Applications 38 (2011) 1210112106

Data
acquisition

Signal
processing

Feature
extraction

Detection
system

Interpretation
system

Output
Fig. 2. The block diagram of the detection system.

0.2

Table 3
Relation among eliminated components variance and number of selected principle
components.

Potato

Amplitude (V)

0.1

Percentage of eliminated components variance (%)

0
-0.1

2
0.5
0.1

-0.2
-0.3
a

-0.4

10

15

20

Time (ms)

Number of selected
features
Ampa

PSDb

NFFTc

9
28
98

1
4
34

1
3
11

Amplitude.
Power spectral density.
Normalized FFT magnitudes centered about the peak frequency.

0.3
Clod

Amplitude (V)

0.2
0.1
0
-0.1
-0.2
-0.3
0

10

15

20

Time (ms)
Fig. 3. Typical digitized impact sound signals from potato and clod.

In order to minimize ANN training time, only one hidden layer

was considered in the network. The number of neuron in hidden
layer was determined using an exhaustive search from 1 to 40
nodes. The neural network with 29 nodes in hidden layer had
the least standard deviation error as well as high stability. In developing ANN models, the linear function at the input layer and the
non-linear hyperbolic tangent function at both hidden and output
layer were used as transfer functions. Learning rate was 0.7
throughout the momentum learning rule. As an additional guard
against over-tting, the data sets were divided into three randomly
selected data sets; 70% of data were used for training, 15% for testing and the remaining 15% were used for cross validation. NeuroSolutions 5.0 software was used for designing and testing of ANN
models (NeuroSolutions, 2005).

1.2

Potato
Clod

Normalized FFT magnitude

0.8

0.6

0.4

0.2

0
0

840

1680
2520
Frequency (Hz)

3360

4200

Fig. 4. Normalized frequency spectra obtained by FFT for potato and clod.

the network weights are adapted and employed for cross validation in order to determine overall performance of the ANN model.
Gradient descent with momentum (GDM) learning rule is an
improvement to the straight GD rule in the sense that a momentum term is used to speed up learning and stabilizing convergence
(Haykin, 1999). Therefore, the GDM method of learning was used
throughout this study.

3. Results and discussion

To nd the best combinations of potential features and optimal
ANN conguration, 31 different combinations of principle component features were selected and tested by neural network. These
features were fed to the ANN models and their performances were
determined by evaluation of the mean square error (MSE), correct
detection rate (CDR) and correlation coefcient (r). Table 4 shows
the summery of off-line experiment using the belt velocity of
1 m s 1. In summary, the best combination was nine amplitudes,
4 PSD, 3 NFFT and one peak frequency feature. Table 5 shows the
performance of optimal ANN conguration. The nal structure of
network was 17 input nodes, 29 hidden nodes and 2 output nodes.
This relatively low number of input vector was used to guard
against the possibility of over-tting the neural network. The peak
frequency feature had more potential to detect clods and potato tubers (Table 6). However, there are overlaps in some cases which
are mainly due to existence of muddy potatoes. High standard
deviation of clods peak frequency values (Table 6) is due to diversity of moisture content and texture of clods.
After extracting potential features in off-line stage, on-line
experiment was performed to evaluate real-time detection accuracy. In this stage, about 1000 samples of potatoes and clods were
randomly selected and off-line procedures were repeated for them.

12105

A. Hosainpour et al. / Expert Systems with Applications 38 (2011) 1210112106

Table 4
Performance of off-line experiment for subsets of features.

a
b
c
d

Features

Number of features

Ampa
Amp
PSDb
NFFTc
PFd
Amp + PSD
Amp + NFFT
PSD + NFFT
Amp + PSD + NFFT
Amp + PSD + PF
Amp + NFFT + PF
Amp + PSD + NFFT + PF

28
9
4
3
1
9+4
9+3
4+3
9+4+3
9+4+1
4+3+1
9+4+3+1

Correct detection rate (%)

Potato

Clod

90.2
87.4
89.3
91.8
88.6
92.2
92.7
94.3
95.4
93.4
95.2
97.3

84.2
73.6
91.4
80.1
92.5
91.2
86.8
92.2
93.3
94.5
95.4
97.6

0.068
0.163
0.032
0.070
0.018
0.019
0.067
0.019
0.016
0.013
0.013
0.011

Amplitude.
Power spectral density.
Normalized FFT magnitudes centered about the peak frequency.
Peak frequency.

Table 5
Performance of optimal ANN model.

0.25
Training MSE

Performance

Potato

Clod

MSE
CDRa (%)
r

0.0113
97.3%
0.989

0.0105
97.6%
0.992

Correct detection rate.

Cross validation MSE

0.2

0.15
MSE

Cross validation MSE

0.1

Table 6
Distribution of peak frequency values.
Samples

Potato
Clod
a

0.05

Peak frequency (Hz)

Maximum

Minimum

Mean

Stda

1455
3055

1180
1429

1260
1911

75
205

150

300

450
600
Number of epochs

750

900

Fig. 5. Learning curve with GDM algorithm for 1000 epochs.

Standard deviation.

Table 7
Detection accuracy of on-line system at different belt velocity.
Belt velocity (m s

Correct detection
Potato

1
1.5
2
2.5

Cross validation
MSE

Clod

Ratio

Ratio

176/
181
174/
182
162/
178
157/
184

97.2

79/
81
81/
84
77/
82
71/
81

97.5

0.011

96.4

0.018

93.9

0.045

87.6

0.165

95.6
91
85.2

Fig. 5 shows the neural network training procedure. It illustrates that in the rst 100 epochs of the training and cross validation, MSE reaches to an acceptable magnitude. Therefore, for
decreasing the training time we can consider only 100 epochs.
The number of used features in relation to the size and diversity
of the training sets indicated that, over-tting of the neural network was not possible.
The signal acquisition and computational procedures took
about 25 ms for a sample at throughput rate of 40 samples per second. However there were some mechanical disturbances that limited the detection capacity under 15 samples per second. A potato
harvester by using this system can perform at the eld capacity of
7.2 ton hr 1.
4. Conclusion

In this stage, data acquisition, signal processing and samples detection were performed at the same time. Table 7 shows the performance of on-line system at different belt conveyer velocities.
Increasing the belt velocity resulted in the reduction of detection
accuracy and increase in the number of miss classied samples.
Low detection accuracy in high velocity was caused by high sound
signal intensity. Although the distance between feeder and impact
plate had been adjusted before, saturated and weak sound signals
were likely at the high velocity of belt conveyor. Using only one
microphone does not seem to be sufcient in high speed
separation.

In the present study, an acoustic-based intelligent system was

used for discriminating between potato tubers and clods. Main
advantages of this system are high accuracy, large capacity and
practicality as well as requiring small space and low cost of equipment and computation. In order to maintain detection accuracy,
belt velocities greater than 1.5 m s 1 are not recommended for
the conveyor. By devising four potato stream lines on the belt at
the velocity of 1 m s 1, detection capacity of 20 ton hr 1 with the
accuracy of 97% is possible. A rejection system is needed to study
the real separation accuracy. So, currently the design and development of a high speed rejecter is being conducted by the authors.

12106

A. Hosainpour et al. / Expert Systems with Applications 38 (2011) 1210112106

Moreover, the effect of partially muddy potatoes on detection

accuracy is being investigated too. Installing the mentioned system
on a potato harvester is needed to expand the application of this
method.
References
Al-Mallahi, A., Kataoka, T., & Okamoto, H. (2008). Discrimination between potato
tubers and clods by detecting the signicant wavebands. Biosystems Engineering,
100, 329337.
Brantley, S. A., Hamann, D. D., & Whiteld, J. K. (1975). A multiple belt adjustable
Vee size grader for sweet potatoes and cucumbers. Transactions of the ASAE, 18,
350362.
Diezma-Iglesias, B., Ruiz-Altisent, M., & Barreiro, P. (2004). Detection of internal
quality in seedless watermelon by acoustic impulse response. Biosystems
Engineering, 88, 221230.
Elbatawi, I. E. (2008). An acoustic impact method to detect hollow heart of potato
tubers. Biosystems Engineering, 100, 206213.
Food and agriculture Organization of the United Nations. (2008). <http://
www.fao.org/corp/statistics/en>.
Feller, R., Morgolin, E., Zacharin, A., & Pasternak, H. (1985). Development of a clod
separator for potato packing houses. Transactions of the ASAE, 28, 10191023.
Gan-Mor, S., Zacharin, A., Galili, N., Feller, R., & Margolin, E. (1986). Absorbing stone
impact to enable separation from potatoes. Transactions of the ASAE, 29,
15261529.
Garsia-Ramos, F. J., Ortiz-Canavate, J., Ruiz-Altisent, M., Diez, J., & Chavez, M. (2003).
Development and implementation of an on-line impact sensor for rmness
sensing of fruits. Journal of Food Engineering, 58, 5357.
Haykin, S. (1999). Neural networks: A comprehensive foundation. Prentice-Hall.
Jivanuwong, S. (1998). Nondestructive detection of hollow heart in potatoes using
ultrasonics. Master of Science thesis. Submitted to the faculty of Virginia
Polytechnic Institute and Stat University, Blacksburg, Virginia, pp. 76.

Kavdir, I., & Guyer, D. E. (2008). Evaluation of different pattern recognition

techniques for apple sorting. Biosystems Engineering, 99, 211219.
Mahmoudi, A. (2006). Development of a suitable algorithm using Articial Neural
Networks for sorting of pistachio nuts with closed shells using impact acoustics, PhD
thesis, Department of Agricultural Machinery, University of Tehran, Karaj: Iran,
p. 144.
Main, J. C. (1971). Potatoes in the UK, Developments in production and marketing.
Span, 14, 8790.
Marchant, J.A., Onyango, C.M., & Street, M.J. (1988). High speed sorting of potatoes
using computer vision. ASAE Paper No. 883540.
MathWorks, MATLAB Users Guide, the Math Works, Inc., (2007).
McGechan, M. B. (1980). An investigation into the damage sustained by different
varieties of potatoes during riddling to remove soil. Journal of Agricultural
Engineering Research, 25, 345353.
Morrow, C.T., Sommer, J.H., Heinemann, P.H., & Tao, Y. (1990). Automated machine
vision inspection of potatoes. ASAE Paper No. 903531.
NeuroSolutions for Excel, NeuroDimension, Inc., (2005).
Pearson, T. C. (2001). Detection of pistachio nuts with closed shells using impact
acoustics. Applied Engineering in Agriculture, 17, 249253.
Pearson, T. C., Cetin, A. E., Tewk, A. H., & Haff, R. P. (2007). Feasibility of impactacoustic emissions for detection of damaged wheat kernels. Digital Signal
Processing, 17, 617633.
Shyam, M., Singh, V., & Singh, R. (1990). Design and development of potato grader.
AMA, 21, 4049.
Story, A. G., & Raghavan, G. S. V. (1973). Sorting potatoes from stones and soil clods
by infrared reectance. Transactions of the ASAE, 16, 304309.
Sugiyama, J., Katsurai, T., Hong, J., Koyama, H., & Mikuriya, K. (1998). Melon ripeness
monitoring by a portable rmness tester. Transactions of the ASAE, 41, 121127.
Younce, F. L., & Davis, F. D. C. (1995). A dynamic sensor for cherry rmness.
Transactions of the ASAE, 38, 14671476.