Energy-aware Scheme for Animal Recognition in Wireless Acoustic

Sensor Networks

Afnan Algobail, Adel Soudani and Saad Alahmadi

Department of Computer Science, College of Computer and Information Science, King Saud University,

Riyadh, Saudi Arabia

Keywords: WASN, Object Recognition, Acoustic Sensing, Feature Extraction, Low-power Recognition.

Abstract: Wireless Acoustic Sensor Networks (WASN) have drawn tremendous attention due to their promising

potential audio-rich applications such as battlefield surveillance, environment monitoring, and ambient

intelligence. In this context, designing an approach for target recognition using sensed audio data represents

a very attractive solution that offers a wide range of deployment opportunities. However, this approach

faces the limited resource’s availability in the wireless sensor. The power consumption is considered to be

the major concern for large data transmission and extensive processing. Thus, the design of successful audio

based solution for target recognition should consider a trade-off between application efficiency and sensor

capabilities. The main contribution of this paper is to design a low-power scheme for target detection and

recognition based on acoustic signal. This scheme, using features extraction, is intended to locally detect a

specific target and to notify a remote server with low energy consumption. This paper details the

specification of the proposed scheme and explores its performances for low-power target recognition. The

results showed the hypothesis' validity, and demonstrate that the proposed approach can produce

classifications as accurate as 96.88% at a very low computational cost.

1 INTRODUCTION

Biodiversity, protection and conservation of

ecosystems are a worldwide concern. Thus,

developing an automated monitoring strategy for

animal recognition and tracking is highly demanded.

In this context, since animals are more often heard

rather than seen, acoustic surveying offer a more

efficient solution for monitoring animals. Compared

to traditional methods, wireless sensors can provide

a cost-effective solution to record acoustic data

across larger spatiotemporal scales. The recorded

acoustic signal provides a rich source of information

for context awareness. However, unlike traditional

scalar sensors, acoustic sensors generate large

volumes of data at relatively high sampling rates,

resulting in a large increase in energy consumption

when the full-recorded signal is transmitted to a

remote server. An extensive transmission of the

entire raw audio signal might question the lifetime

of the communication system.

One of the interesting ideas to reduce the volume

of data is to locally detect and recognize a particular

target using the extracted acoustic signature from the

received audio signal. The remote server will be then

informed using small size notification packets. This

scheme can significantly contribute to decrease the

number of required data transmissions, and

consequently leads to overall improvement in the

network lifetime. However, the success of such

scheme in achieving the desired results depends on

its ability to process the acoustic signal at low-cost

in acoustic motes within limited resources and

capabilities.

In this paper, we focus on the specification of a

low-complexity acoustic sensing scheme for target

recognition using two feature extraction methods:

Zero-Crossing Rate (ZCR) and Root Mean Square

(RMS). A Minimum Mean Distance (MMD)

classifier is used for the recognition. Our main

objective is to demonstrate the trade-off between

power consumption and recognition accuracy.

2 RELATED WORKS

Recognition of animals by their sounds has

motivated many research efforts. Han et al. (Han,

Algobail, A., Soudani, A. and Alahmadi, S.

Energy-aware Scheme for Animal Recognition in Wireless Acoustic Sensor Networks.

DOI: 10.5220/0006604100310038

In Proceedings of the 7th International Conference on Sensor Networks (SENSORNETS 2018), pages 31-38

ISBN: 978-989-758-284-4

Muniandy and Dayou, 2011) extracted spectral

centroid, Shannon entropy, and Re

nyi entropy

features. They used k-Nearest Neighbour (k-NN)

classifier to recognize nine frog species gaining just

a 75.6% accuracy. Wavelet transform was combined

with Mel-Cepstral Fourier Coefficients (MFCCs) in

(Colonna et al., 2012) to classifiy nine species of

anuran using k-NN classifier, whereas Xie (Xie et

al., 2015) used perceptual wavelet packet

decomposition sub-band cepstral coefficients for

identifying ten frog species. All the previous works

used wavelet analysis approach due to its ability to

achieve a high accuracy rate compared to time and

frequency domain features. However, these schemes

are very demanding both in computational power

and buffer size.

Among all mentioned features, MFCC feature set

has been widely used because it provides more

powerful discrimination capability. The authors in

(Evangelista et al., 2014) computed a total of 64

MFCCs to recognize bird species using Support

Vector Machine (SVM) classifier. In (Dong et al.,

2015), MFCC and a spectral ridge were applied on

24 bird sound samples, achieving an accuracy of

71%. A combination of MFCC and LPC were

adopted in (Yuan and Ramli, 2013) to classify eight

frog species using k-NN with an identification rate

of 98%. While in (Noda, Travieso and Sánchez-

Rodríguez, 2016; Colonna et al., 2016), high

recognition rates have been attained for the

classification of anuran species using MFCC and

other features such as energy. However, most of the

previous works are not well suited for WASNs as

they involve performing some kind of

transformation and extracting a large set of features.

In such studies, it is common to prioritize

recognition accuracy even if it results in higher

energy and storage costs.

Other studies developed an automatic

recognition system based on syllable features. The

idea behind such approach is to segment the stream

of sound into syllable units and then derive syllable

features. Colonna et al. (Colonna et al., 2015)

applied ZCR and an energy entropy on 896 syllables

of seven different anuran species. Alternatively, Xie

et al. (Xie et al., 2016) introduced a method to

recognize twenty-four frog species using a set of

different frequency based features in addition to

syllable, Linear Predictive Coding (LPC), and

MFCC features and using a five different machine

learning algorithms. In (Noda, Travieso and

Sánchez-Rodríguez, 2016), a syllable feature used in

conjunction with MFCC, and LFCC features for the

classification of 199 anuran spices using SVM

classifier. The results showed that the proposed set

outperforms the approaches that used only MFCC,

LFCC, or both of them. Also, in (Xie et al., 2015;

Xie et al., 2016), the authors proved that using

syllable features provide higher classification

accuracy compared to MFCC under different levels

of noise contamination.

Most of the presented schemes in reviewed

studies didn’t address the complexity and the power

consumption to prove algorithms’ efficiency. Such

works have also been conducted in the field of

species classification, but the recognition of different

types of animals is not widely known in the

literature. In addition, many current feature

extraction systems are designed for particular

applications, and hence are only appropriate for

identifying a particular type of species. Thus, it

becomes essential to develop an efficient recognition

scheme to identify different animal sounds.

3 GENERAL APPROACH FOR

ACOUSTIC-BASED

LOW-POWER SENSING

We considered the implementation of an object

recognition scheme in a WASN environment. The

network is composed of a set of smart microphone

nodes that are capable to sense and process audio

streams and scalar data. At the set-up process, the

end-user loads those sensors with the target audio-

feature’s descriptor. We assume that each sensor

node will acquire a new acoustic signal in a periodic

manner (over a time interval t). Then, the energy of

the recorded signal will be computed and compared

with a predefined threshold to generate a detection

decision. Upon detection of the presence of a new

object, the sensor node should extract a set of

representative feature descriptors that can uniquely

describe the object.

In the proposed scheme, instead of flooding the

network with a large amount and perhaps useless

transmitted data, a local classifier at each sensor

node will first make a decision on the type of

detected objects based on their extracted feature

vector and the reference descriptor. When the target

object is detected, a notification will be processed

and sent to the base station according to the end-user

requirements, which can be either a detection

notification or vector of features.

The development of such a scheme that meets

the constraints of WASNs is considered a big

challenge. In fact, the design of this scheme requires

SENSORNETS 2018 - 7th International Conference on Sensor Networks

the use of low complexity processing methods for

audio signal processing enabling target identification

with minimum operating power and storage

capacity. In our scheme, we focused on reducing the

dimensionality of the feature vector to be

represented by a minimum number of bytes, while

maintaining a sufficient number of features to enable

robust object recognition. This approach can reduce

the amount of data to be exchanged, and as per

consequence reduce the transmission energy

consumption.

The success of such scheme in achieving the

desired results depends mainly on two factors:

• The accuracy of the extracted features to

abstract the target.

• The size of the dataset used to represent this

feature.

Efficient features should be capable to

discriminate between animals. However, the huge

overlapping in the feature's vectors of different

animals for the same feature makes almost

impossible to accurately identify a specific animal.

This problem is illustrated in Figure 1, where Root

Mean Square (RMS) and Zero Crossing Rate (ZCR)

values for different animals are showing a high

overlapping intervals. In addition, the small set of

training records is not sufficient enough to develop a

model with high productivity. In specific, the

learning algorithm doesn’t have enough data to

capture the underlying trends in the observed data in

order to define the class boundaries properly.Thus,

we proposed a multi-classification method in which

one animal will be assigned to two class labels. The

assignment will be based on finding the two classes

that have the same value intervals for each feature.

Figure 1: Two-dimensional feature space for animal

dataset.

4 SPECIFICATION OF THE

SENSING SCHEME FOR

TARGET RECOGNITION

The main contribution of this paper is to design a

low-complexity scheme for object recognition in

order to be used in the context of WASN. However,

achieving this goal is directly dependent on the

appropriateness of the feature extraction and

classification algorithms. Two main factors should

be taken into consideration in this task, which are:

• Network resource constraints: wireless sensor

networks are generally characterized by their

limited resources, such as small memory size,

limited power supplies, and low communication

bandwidth. Therefore, it is important to keep the

feature vector that represents the object

descriptor as short as possible. In addition, the

classification algorithm should rely on a model

that is capable of making recognition decisions

with a minimum number of data set.

• Computational complexity: The power

consumption level in the sensor node is directly

affected by the complexity of the feature

extraction and classification method. This

comes from the fact that a large number of

mathematical operations need a large number of

clock cycles to be executed, and hence increase

the demands in energy consumption. Therefore,

it is essential to adopt a low-complexity method

with as less number of arithmetic operations as

possible.

The proposed approach is composed of two

sequentially phases Figure 2: Reference vectors

extraction and object recognition. In the first phase,

different feature vectors will be extracted from each

training records and then the mean for each feature

and each object will be computed. Based on these, a

unique descriptor (one vector per animal) will be

constructed and stored. These descriptors are loaded

into the sensor memory during the set-up process.

Afterward, in the second phase, these descriptors

will be compared with the detected object feature

vector for the classification purpose. In general, each

stage is composed of multiple steps, as detailed in

the following sections.

4.1 Object Detection

The detection of a new object is based on a certain

received signal power threshold level. The purpose

of detection is to differentiate the acoustic events

from the background noise, without determining

Energy-aware Scheme for Animal Recognition in Wireless Acoustic Sensor Networks

Figure 2: Flowchart of the overall object recognition scheme.

whether the target object is present or not. The

sensor will acquire a new acoustic signal in a

periodic manner (over a time interval T). Then, the

received signal energy (E) during that time interval

will be measured using the root mean square feature.

The energy should be checked, if it is greater than a

predetermined threshold value (Tthre), then a new

object is supposed to be detected. The detection

function D can be defined as follows (1)

D = 

  

  

(1)

4.2 Feature’s Extraction

The feature extraction can play an important role in

saving sensor resources in terms of energy, memory

requirements and processing time. In addition, it is

an essential step in any audio based-recognition

process or classification task. In fact, the capability

of the sensor to discriminate the target object from

different auditory objects is a challenging task,

especially when it involves objects with relatively

similar sound waves. Therefore, these characteristic

features offer precise information that can maximize

discrimination between different objects, and hence

increase the recognition accuracy. The proposed

feature extraction process will be based on the

following sequential steps:

I. Divide the captured audio stream into (T)

frames; each frame has 1024 sequential samples

in which 50% of them are overlapped between

two successive frames.

II. Calculate the characteristic’s features for each

frame. In our approach, we have focused on two

temporal feature extraction methods: RMS and

ZCR (Lerch, 2012). RMS is an important

measurement that provides information about

the amplitude of a signal, while ZCR

demonstrates the characteristics of dominant

frequency.

III. Obtain the global feature vector by computing

the mean for each features’ dimension d =

1,2,..,D of N-dimensional features vector F that

is obtained from all the frames.

4.2.1 Root Mean Square Feature

It is a time domain feature denotes the average

signal strength, which can be expressed as

 



















 (2)

where x

is the ith sample value of the frame, and N

denoted as the frame length (Lerch, 2012).

4.2.2 Zero Crossing Rate Feature

This is a time domain feature that measures the

number of times the signal changes its sign within a

time window, from positive to negative and vice

versa. It is one of the most commonly used methods

for computing the frequency content of the signal

(Lerch, 2012), which can be calculated according to

the following equations (3)

 

















  



  



















(3)

where x(m) is the value of the m

sampled signal

and sgn[] is the sign function, which defined as:

SENSORNETS 2018 - 7th International Conference on Sensor Networks

sgn[x(n)] = 





















(4)

4.3 Classification

There are several classification techniques that can

be used for acoustic target recognition. These

techniques can be grouped into two general

categories: Supervised (e.g. Maximum Likelihood,

MMD, k-NN) and machine learning (e.g. artificial

neural network, classification tree, SVM)

algorithms. The machine learning classification

methods have been widely and successfully used in

many studies in the literature to recognize a single

target. However, most these classification methods

cannot be directly applied in the area of WASNs

because they significantly need large resources and

computation power than available on a low-power

sensor node. To overcome these limitations, the

developed algorithm must be computationally light

and its classification model should consume limited

storage space.

In our approach, the classification process will be

based on MMD classifier (Rudrapatna and Sowmya,

2006). The main idea of this approach is to check the

objects’ similarity by finding the class mean, which

has the minimum distance from the detected object

feature vector. In spite of its simplicity, some animal

acoustic classification approaches have adopted

MMD classifier because it can provide an acceptable

degree of accuracy in most situations (Huang et al.,

2009; Luque et al., 2016). However, since the

overlapping between animals is large in our case as

previously explained (figure 1), we used two-classes

for the classification approaches in the proposed

scheme, in which an object will yield an estimated

probability of belonging to two classes. The

classification process is composed of two phases:

I. Classification (learning) model construction:

The mean of the feature vectors of training

records will be calculated to construct the

descriptor vector for each specific object. Each

descriptor vector Refi = {μ

RMS

, μ

ZCR

}, contains

two values, which are the mean of (RMS and

ZCR) features. This vector will be stored in the

database and will be loaded into the sensor

memory to be used later in the similarity

matching stage.

II. Similarity matching: The similarity between the

newly extracted unknown object feature vector

and the reference vectors Ref

will be measured

and stored in a distance vector D

= {d

,…d

}

using the Euclidean distance metric (Xie et al.,

2016); The Euclidean distance d between two

vectors can be calculated as follows:

 





















i = 1,..,D (5)

where f

is the detected object vector, and 





is the

mean vector of classes. Then, we will find the two

minimum values in the distance vector D

, which

represent the classes that the test object belongs to.

4.4 Notification to the End User

When the target object is detected in a certain

cluster, a notification will be processed by each node

and will be sent to the cluster head. The notification

will be processed at the node level according to the

end user requirements, which can be either: few bits’

notification or vector of features. The cluster head

will find the packet with the highest RMS value and

pass it to the end user for further classification.

Then, at the end user, the feature vector will be used

to assign the target object to a single animal class.

5 IMPLEMENTATION AND

PERFORMANCES ANALYSIS

5.1 Performance Analysis at the

Application Level

The capabilities of the proposed scheme to classify a

specific target object to two classes are tested and

evaluated using MATLAB tool.

5.1.1 Pre-processing

Collected audio records have to be pre-processed

before being used. The pre-processing involves data

cleaning, audio segmentation, and audio re-

sampling. Data cleaning process aims to remove the

records that have a bad quality sound or contains

high noises. Then, the records that contain mixed

animal sounds should be segmented to produce a

record that has only single animal call. Each record

is made up of either one or multiple continuous calls

of one animal, with the duration ranging from one to

two seconds. Finally, the audio records need to be

re-sampled into a unified sampling frequency in

order to not affect the recognition accuracy. After

the pre-processing, the dataset will be divided into

two sets: training set and test set. Almost 70% of the

total collected samples are used as training data to

build the classification model, while the rest are used

Energy-aware Scheme for Animal Recognition in Wireless Acoustic Sensor Networks

for testing.

5.1.2 Dataset

The dataset for object recognition was built from

audio collected from Animals & Birds Sound Effects

CD (Sound-ideas.com, 2017). The rest of the

records were collected from different animal sound

library websites. The audio recordings are stored in

WAV format with a sample rate of 44.1 kHz and 16

bits per sample. The dataset contains seven animals

with 114 samples: Cat, Cow, Dog, Donkey, Horse,

Parrot, and Sheep. A typical example of two

different dog bark sounds is depicted in Figure 3.

Figure 3: Audio sample (wave form) for dog bark.

5.1.3 Performance Evaluation

The main metric used to examine the performance of

the proposed scheme is the successful recognition

rate. The ratio of success represents how many

objects that the algorithm was able to classify

correctly, which can be calculated as follows











 (6)

where N

is the number of records which were

recognized correctly and N

is the total number of

test samples.

5.1.4 Recognition Ratio

We studied the capability of two features (RMS and

ZCR) to discriminate between different animals

using MMD and different machine learning

classifiers, namely, Decision Tree (DT), Gaussian

Mixtures Model (GMM), and K-NN (Table 1). The

machine learning classifiers were tested to decide on

which classifier best suits the single animal

classification at the end user. The results show that

the proposed scheme was capable to correctly

classify 31 out of 32 targets (accuracy of 96.88%).

We also note from Table 1 that the GNN classifier

was able to identify the target animal with better

accuracy than the DT and K-NN classifiers, gaining

71.88% recognition accuracy. Nevertheless, the

three classifiers performed poorly for Horse and

Cow. Compared to other animals, the number of

training set of these two animals is not enough for

classifiers to enable complete learning, which caused

under-fitting problem.

Table 1: The comparison of different classification

algorithm results.

Animal

MMD

GMM

KNN

Cat

100%

75%

Cow

100%

25%

Dog

100%

40%

60%

40%

Donkey

100%

80%

100%

Horse

50%

Parrot

100%

Sheep

100%

66.67%

83.33%

66.67%

Average

96.88%

62.50%

71.88%

65.63%

5.2 Energy Efficiency of the Proposed

Scheme for Sensor Node

We investigated the energy consumption of our

scheme using AVRORA (Titzer, K. Lee and

Palsberg, 2005). AVRORA is an instruction-level

sensor network simulator that emulates the Mica

family platforms (MICA2dot, MICA2, MICAz).

These motes are based on a low-power

ATmega128L micro-controller with 4KB of RAM.

The main purpose of such framework is to execute

TinyOS applications prior their deployment in actual

WSN nodes. The AVRORA helps to estimate the

number of clock cycles, power consumption, and

processing time. We focused mainly in measuring

the performance of a sensor mote during recognition

and notification task. Table 2 summarizes the cost of

the recognition task per sample, assuming a

sampling frequency of 44.1 kHz. As shown in Table

2, feature extraction is the most energy requiring and

time consuming steps of object recognition scheme,

but nevertheless it is very essential to reduce the

communication overhead and prolong the network

lifetime. Such task involves performing several

expensive computations on large sets of data

compared to other steps.

SENSORNETS 2018 - 7th International Conference on Sensor Networks

Table 2: Evaluation of the recognition cost on MICA2.

Sound size

Per 1 second (44100 samples)

Measured

Attribute

Clock

cycles

Time

(ms)

Energy

(mj)

New object

detection

352

0.044

0.0009

Feature

extraction

1527

0.19

0.004

Classification

12025

1.5

0.03

Whole

scheme

13904

1.734

0.0349

In Table 3, we considered measuring the notification

cost for three different scenarios, in which a sensor

node can either send a packet that includes a

detection notification, object feature vector, or

unprocessed acoustic sample. Here we assumed

using a 16-bit ADC acoustic sensor to collect audio

signals at a rate of 44100 samples per second. The

sensor will send data to the server side at a rate of 16

bits/sample * 44100 samples/second = 705600

bits/second. Thus, the power consumed during the

radio transmission for a total of 705,600 bits per

second / 8 = 88,200 bytes per second is 29106 mj.

These latter results prove that the proposed scheme

can dramatically reduce the total energy

consumption in the network.

Table 3: Evaluation of the notification cost on MICA2.

Measured Attribute

Time (s)

Energy

(mj)

Transmit detection

notification (1 byte)

0.01

0.33

Transmit 2D feature

vector (2 bytes per

feature)

0.04

1.32

Transmit raw signal (2

bytes per sample)

0.02

0.66

6 CONCLUSIONS

We presented a low-complexity scheme for target

recognition in an energy-constrained acoustic sensor

network and applied it to audio-based animal's

classification scenario. The proposed scheme uses

low-complexity feature extraction techniques, in

which the goal is to minimize the heavy processing

burden and transmission overhead on the network.

This was achieved by the detection of event of

interest locally using RMS and ZCR features, and

then reporting the event to the server with small-size

packets. Results indicate that the adopted feature

extraction methods were able to generate a unique

signature, which was successfully used to

discriminate between different auditory objects in

the recognition process. Moreover, the performed

simulations have demonstrated that the proposed

scheme can save up to 70% of power consumption,

while guaranteeing high recognition accuracy. As a

future work, we are interested to study a possible

real implementation of the proposed algorithm on

MICA2 mote.

REFERENCES

Colonna, J. G., Ribas, A. D., Santos, E. M. d., and

Nakamura, E. F. (2012). Feature sub-set selection for

automatically classifying anuran calls using sensor

networks. In: International Joint Conference on

Neural Networks (IJCNN). pp.1-8. IEEE.

Colonna, J., Peet, T., Ferreira, C., Jorge, A., Gomes, E.

and Gama, J. (2016). Automatic Classification of

Anuran Sounds Using Convolutional Neural

Networks. In: C3S2E '16 Proceedings of the Ninth

International Conference on Computer Science &

Software Engineering. ACM.

Colonna, J., Cristo, M., Salvatierra, M. and Nakamura, E.

(2015). An incremental technique for real-time

bioacoustic signal segmentation. Expert Systems with

Applications, 42(21), pp.7367-7374.

Dong, X., Xie, J., Towsey, M., Zhang, J. and Roe, P.

(2015). Generalised Features for Bird Vocalisation

Retrieval in Acoustic Recordings. In: 17th

International Workshop on Multimedia Signal

Processing (MMSP). IEEE.

Evangelista, T., Priolli, T., Silla Jr., C., Angelico, B. and

Kaestner, C. (2014). Automatic Segmentation of

Audio Signals for Bird Species Identification. In:

International Symposium on Multimedia. IEEE.

Han, N., Muniandy, S. and Dayou, J. (2011). Acoustic

classification of Australian anurans based on hybrid

spectral-entropy approach. Applied Acoustics, 72(9),

pp.639-645.

Huang, C., Yang, Y., Yang, D. and Chen, Y. (2009). Frog

classification using machine learning techniques.

Expert Systems with Applications, 36(2), pp.3737-

3743.

Lerch, A. (2012). An Introduction to Audio Content

Analysis: Applications in Signal. 1st ed. Somerset:

John Wiley & Sons.

Luque, J., Larios, D., Personal, E., Barbancho, J. and

León, C. (2016). Evaluation of MPEG-7-Based Audio

Descriptors for Animal Voice Recognition over

Wireless Acoustic Sensor Networks. Sensors, 16(5).

Noda, J., Travieso, C. and Sánchez-Rodríguez, D. (2016).

Methodology for automatic bioacoustic classification

of anurans based on feature fusion. Expert Systems

with Applications, 50, pp.100-106.

Rudrapatna M., Sowmya A. (2006) Feature Weighted

Minimum Distance Classifier with Multi-class

Energy-aware Scheme for Animal Recognition in Wireless Acoustic Sensor Networks

Confidence Estimation. In: Advances in Artificial

Intelligence. Springer.

Sound-ideas.com. (2017). HD – Animals & Birds Sound

Effects. [online] Available at: https://www.sound-

ideas.com/Product/380/HD-–-Animals-Birds-Sound-

Effects [Accessed 15 Jul. 2017].

Titzer, B., K. Lee, D. and Palsberg, J. (2005). Avrora:

scalable sensor network simulation with precise

timing. In: Fourth International Symposium on

Information Processing in Sensor Networks (IPSN).

IEEE.

Xie, J., Towsey, M., Eichinski, P., Zhang, J. and Roe, P.

(2015). Acoustic Feature Extraction using Perceptual

Wavelet Packet Decomposition for Frog Call

Classification. In: 11th International Conference on

eScience. IEEE.

Xie, J., Towsey, M., Truskinger, A., Eichinski, P., Zhang,

J. and Roe, P. (2015). Acoustic classification of

Australian anurans using syllable features. In: Tenth

International Conference on Intelligent Sensors,

Sensor Networks and Information Processing

(ISSNIP). IEEE, pp.7-9.

Xie, J., Towsey, M., Zhang, J. and Roe, P. (2016).

Acoustic classification of Australian frogs based on

enhanced features and machine learning algorithms.

Applied Acoustics, 113, pp.193-201.

Xie J., Towsey M., Zhang L., Zhang J., Roe P. (2016).

Feature Extraction Based on Bandpass Filtering for

Frog Call Classification. In: Image and Signal

Processing (ICISP). Springer.

Yuan, C. L. T., Ramli D. A. (2013) Frog Sound

Identification System for Frog Species Recognition.

In: Context-Aware Systems and Applications

(ICCASA). Springer.

SENSORNETS 2018 - 7th International Conference on Sensor Networks