sensors
Article
Detecting and Tracking the Positions of Wild Ungulates Using
Sound Recordings
Salem Ibrahim Salem
1,2,
* , Kazuhiko Fujisao
3,4
, Masayasu Maki
5
, Tadanobu Okumura
6
and Kazuo Oki
1,3
Citation: Salem, S.I.; Fujisao, K.;
Maki, M.; Okumura, T.; Oki, K.
Detecting and Tracking the Positions
of Wild Ungulates Using Sound
Recordings. Sensors 2021, 21, 866.
https://doi.org/10.3390/s21030866
Received: 11 December 2020
Accepted: 18 January 2021
Published: 28 January 2021
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2021 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
1
Faculty of Engineering, Kyoto University of Advanced Science, 18 Yamanouchi Gotanda-cho, Ukyo-ku,
Kyoto 615-8577, Japan; [email protected]
2
Faculty of Engineering, Alexandria University, Lotfy El-Sied St. off Gamal Abd El-Naser-Alexandria,
Ash Shatebi, Alexandria Governorate 11432, Egypt
3
Institute of Industrial Science, The University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo 153-8505, Japan;
fujisaok222@affrc.go.jp
4
National Agriculture and Food Research Organization, 3-1-1 Kannondai, Tsukuba, Ibaraki 305-8517, Japan
5
Faculty of Food and Agricultural Sciences, Fukushima University, 1 Kanayagawa,
Fukushima 960-1296, Japan; [email protected]
6
Wildlife Management Office Inc., 922-7 Komiya-cho, Hachiouji, Tokyo 192-0031, Japan; [email protected]
* Correspondence: [email protected] or [email protected]
Abstract:
Monitoring wild ungulates such as deer is a highly challenging issue faced by wildlife
managers. Wild ungulates are increasing in number worldwide, causing damage to ecosystems.
For effective management, the precise estimation of their population size and habitat is essential.
Conventional methods used to estimate the population density of wild ungulates, such as the
light census survey, are time-consuming with low accuracy and difficult to implement in harsh
environments like muddy wetlands. On the other hand, unmanned aerial vehicles are difficult to
use in areas with dense tree cover. Although the passive acoustic monitoring of animal sounds is
commonly used to evaluate their diversity, the potential for detecting animal positions from their
sound has not been sufficiently investigated. This study introduces a new technique for detecting and
tracking deer position in the wild using sound recordings. The technique relies on the time lag among
three recorders to estimate the position. A sound recording system was also developed to overcome
the time drift problem in the internal clock of recorders, by receiving time information from GPS
satellites. Determining deer position enables the elimination of repetitive calls from the same deer,
thus providing a promising tool to track deer movement. The validation results revealed that the
proposed technique can provide reasonable accuracy for the experimental and natural environment.
The identification of deer calls in Oze National Park over a period of two hours emphasizes the great
potential of the proposed technique to detect repetitive deer calls, and track deer movement. Hence,
the technique is the first step toward designing an automated system for estimating the population
of deer or other vocal animals using sound recordings.
Keywords:
wild ungulates; animals count; tracking animals; deer; Oze; sound recognition; repetitive
calls; time drift; sound localization
1. Introduction
Over the past few decades, deer populations have been significantly growing along
with an apparent expansion of their area of distribution in some regions of Europe [
1
], North
America [
2
], and Japan [
3
], causing massive damage to farming and forestry [
4
]. Deer have
also invaded many protected areas such as the Oze marshland [
5
], and Kushiro-shitsugen
National Park [
6
] in Japan, resulting in irreversible effects on the original ecosystem such
as destruction of rare plant species [
7
]. Monitoring the population dynamics of deer is thus
crucial for the effective evaluation and management of ecosystems [8].
One of the methods used commonly to count the number of deer is the spotlight-
census survey [
1
,
9
]; a light-census survey is a traditional technique using artificial light
Sensors 2021, 21, 866. https://doi.org/10.3390/s21030866 https://www.mdpi.com/journal/sensors
Sensors 2021, 21, 866 2 of 15
during nighttime to count the number of deer. Nevertheless, the light-census survey
is strongly influenced by weather conditions, deer movements, and accessibility of the
location, which causes considerable uncertainty in the observed data. Camera traps that are
automatically triggered by movement have also been used to monitor the density of various
wildlife animals such as American marten (Martes americana) [
10
], livestock [
11
], and wild
ungulates [
12
,
13
] to evaluate their impact on the surrounding environment [
14
,
15
]. Thermal
cameras have been also used to automatically detect animals in a grassland habitat [
16
].
The camera’s accuracy relies on the photographing frequency and the animal’s moving
speed. Several other methods rely on walking around the survey region to search for deer
and its sign, such as pellet counts [
17
,
18
] and quadrat methods [
19
,
20
]. However, these
methods are not suitable for harsh environments (e.g., muddy wetlands) and some animals
do not leave indices that can be counted [
21
,
22
]. Aerial surveys were used to survey marsh
and pampas deer in Brazil during the dry season [
23
]. In recent years, the application of
unmanned aerial vehicles (UAVs) for wildlife monitoring has expanded rapidly [
24
–
26
].
However, it is very difficult to use both aerial surveys and UAVs over areas with dense tree
coverage, which obscure the detection of animals in both visible and infrared bands.
Sound recordings have been used by researchers to estimate animal density, such as
counting yellow rails using autonomous recording units [
27
]. Sound-based techniques
are suitable for rugged and difficult-to-reach areas (e.g., muddy or mountainous regions).
Sound recordings have also been used to image the phonotactic behavior of female frogs
in the dark [
28
]. Enari et al. [
2
,
8
] used passive acoustic monitoring to monitor sika deer
in eastern Japan; they built a sound recognizer to automatically detect three types of deer
calls from sound recordings. Their results showed that the fully automated procedure
results in numerous false detections, and manual screening is needed to correct this issue.
Passive acoustic telemetry techniques have been used to track the movement patterns of
fishes and marine organisms [
29
]. The technique uses sound waves emitted from a tagged
transmitter implanted in fishes in order to track them. The need to mark and recapture the
fish or animals limits the application of acoustic telemetry techniques [
2
,
30
], particularly
in open habitats where deer or other vocal animals live. Another limitation related to
the sound-based technique is that repetitive calls from the same vocal animal must be
identified and excluded during the counting process.
This study introduces a new sound-based technique to estimate deer’s position based
on sound recordings of deer calls at three stationary recorders. The aim of the proposed
research was to: (1) introduce a new technique to detect deer positions from their calls,
and to identify repetitive calls of the same deer; (2) validate the technique using measure-
ments from both controlled experiments and wild deer environments; and (3) discuss
the strengths, limitations, and offer further suggestions for the proposed technique to be
implemented for deer surveys.
2. Methods
2.1. Study Area
Two validation experiments (i.e., in the control environment, and in the natural
environment of deer) were performed to provide a reliable validation of the proposed
technique. During the two experiments, a deer whistle was used to emit deer calls at
different locations (i.e., validation points), and that emitted sound was recorded using
sound recorders installed in each experiment’s site (details in Sections 3.1 and 3.2). Deer
whistles have been effectively used to mimic, call, and hunt deer for many years [31,32].
The first validation experiment was conducted at a playground located in the Uni-
versity of Tokyo, Japan (hereafter called UTokyo experiment), with 90 m length and 55 m
width, as shown in Figure 1b. Three recorders were installed and pointed toward the
center of the playground (Figure 1b). Five validation points (A–E, as shown in Figure 1b)
were selected to validate the proposed technique. At each validation point, deer calls were
emitted using deer whistle and recorded by the three recorders. The coordinates of the
three recorders and five validation points are summarized in Table 1.
Sensors 2021, 21, 866 3 of 15
Figure 1.
(
a
) First validation experiment site conducted on 29 November 2017 at University of Tokyo,
Japan, and (
b
) location of the three recorders (i.e., Rec 1–3) and validation points (i.e., A–E) where
deer calls were emitted using a deer whistle. Playground image from Google Earth.
Table 1.
Coordinates of recorders (Rec 1–4) and validation points (A–E) where deer calls were emitted
using a deer whistle at University of Tokyo experiment (Figure 1b) and Oze marshland (Figure 2d).
UTokyo Experiment Site Oze National Park
Lat Long Lat Long
Rec1 35
◦
39
0
38.12
00
N 139
◦
40
0
54.13
00
E Rec1 36
◦
56
0
31.91
00
N 139
◦
13
0
38.43
00
E
Rec2 35
◦
39
0
36.19
00
N 139
◦
40
0
55.57
00
E Rec2 36
◦
56
0
34.92
00
N 139
◦
13
0
22.63
00
E
Rec3 35
◦
39
0
37.63
00
N 139
◦
40
0
57.72
00
E Rec3 36
◦
56
0
23.49
00
N 139
◦
13
0
24.27
00
E
A 35
◦
39
0
37.38
00
N 139
◦
40
0
57.09
00
E Rec4 36
◦
56
0
27.78
00
N 139
◦
13
0
08.86
00
E
B 35
◦
39
0
36.48
00
N 139
◦
40
0
55.63
00
E A 36
◦
56
0
38.62
00
N 139
◦
13
0
18.70
00
E
C 35
◦
39
0
37.61
00
N 139
◦
40
0
54.96
00
E B 36
◦
56
0
30.14
00
N 139
◦
13
0
23.51
00
E
D 35
◦
39
0
37.23
00
N 139
◦
40
0
54.66
00
E C 36
◦
56
0
28.36
00
N 139
◦
13
0
25.95
00
E
E 35
◦
39
0
37.06
00
N 139
◦
40
0
55.76
00
E D 36
◦
56
0
30.29
00
N 139
◦
13
0
29.18
00
E
E 36
◦
56
0
19.48
00
N 139
◦
13
0
25.56
00
E
Figure 2.
Location of recorders installed in Oze marshland. (
a
) Location of Oze National Park
in central Japan; (
b
) location of Oze National Park with respect to four Japanese prefectures (i.e.,
1-Niigata, 2-Fukushima, 3-Gunma and 4-Tochigi); (
c
) special protected area of Oze National Park
(in blue); and (
d
) Google Earth image indicating location of recorders (Rec 1–4), distance between
recorders, locations of two camera traps (Cam 1 and 2) and locations of validation points (A–E) where
deer calls were emitted using a deer whistle.
Sensors 2021, 21, 866 4 of 15
The second validation experiment was conducted in Oze National Park (Figure 2). Oze
National Park is located in central Japan (Figure 2a), with an area of 372 km
2
[
33
]. It spans
four Japanese prefectures (Niigata, Fukushima, Gunma, and Tochigi prefectures), as shown
in
Figure 2b
. Oze is famous for its rare plant species, and a wide variety of wildlife [
34
].
Recently, there has been a significant increase in deer population in Oze, causing irreversible
damage to rare plants. Like the UTokyo experiment, the proposed technique was also
validated in Oze National Park at five points (A–E, as shown in Figure 2d) using a deer
whistle. Four sound recorders were installed in Oze Marshland (
Figure 2d
), which located
in the special protected area (Figure 2c) of Oze National Park. The coordinates of the four
sound recorders and five validation points are listed in Table 1. The real deer calls were
also collected in Oze site during the rutting season (i.e., October 2018). The data of two
camera traps (Ltl Acorn 6310W; LTL Acorn Outdoors, Green Bay, WI, USA) located within
the Oze site were used to validate the retrieved deer positions from deer calls (
Section 3.3
).
The Ltl-6310w camera is equipped with a wide field of view of 100
◦
and a passive infra-red
(PIR) sensor to detect the sudden change of ambient temperature caused by a moving
object. The two cameras were set to capture an image and then record a video for 20 s once
the PIR sensor captures a motion. Shapefiles used to create Figure 2 were downloaded
from the Japanese National Land Numerical Information database [35].
2.2. Description of Proposed Technique
Sounds provide valuable information about the deer producing them, such as sex and
age (buck, doe or fawn), emotion (e.g., snort sound during danger or estrus bleat that is
made to attract bucks) [
31
], and rough information about their position. The proposed
technique aims to determine the accurate position of deer (i.e., latitude and longitude of
deer position) from their sound recordings. By knowing the precise location of deer, we
can identify repetitive calls from the same deer and avoid counting them, as well as track
its movement. Our technique relies on using three synchronous stationary recorders to
record deer calls.
The concept of detecting deer position using the three sound recordings is as follows.
When a deer makes a sound, the sound waves propagate in a series of spherical waves
in all directions and are recorded at the three stationary recorders at times that are based
on the location of the recorders with respect to the position of the deer. From these times,
the time lag (i.e., time-delay) to detect the deer sound among the three recorders can be
estimated. The position of deer can be retrieved from these time lags. As shown in
Figure 3
,
the three recorders (i.e., Rec 1–3) have known positions (i.e.,
Lat
x
,
Lon
x
; where
x
= 1, 2,
and 3), whereas the deer position (i.e.,
Lat
0
,
Lon
0
) is unknown. Rec 3 is the closest to the
deer position, whereas Rec 1 is the farthest. Thus, the deer sound will be first detected by
Rec 3, then Rec 2, and then Rec 1. Of course, the time duration (T
0
) that the deer sound
requires to reach the nearest recorder from the deer position (i.e., Rec 3) is unknown. In
contrast, the time lag of detecting the same sound at Rec 2 (i.e., time delay in detecting
deer sound at Rec 2 after detecting it at Rec 3) can be calculated from sound recordings
(hereafter, time lag
T
2
, as shown in Figure 3a). Similarly, time lag
T
1
represents the time
difference between detecting the same deer sound at Rec 3 and Rec 1 (Figure 3a). The
distance (D
1
) corresponding to the time lag T
1
shown in Figure 3a can be calculated as:
D
1
= T
1
× V (1)
where
V
represents the speed of sound in air, which depends on the air temperature. For
instance, the speed of sound in dry air is 331.20 and 343 m s
−1
at 0 and 20
◦
C, respectively.
Similarly, the distance
D
2
(Figure 3a) corresponding to the time lag T
2
is calculated as
D
2
= T
2
× V
. Thus, the only information that can be estimated from the sound recordings
are time lags T
1
and T
2
, and their corresponding distances D
1
and D
2
.
Sensors 2021, 21, 866 5 of 15
Figure 3.
Schematic diagrams explaining the concept of the proposed technique. (
a
) Available and
unknown information based on the sound recorders’ positions and sound recordings, (
b
) trial-and-
error method to retrieve deer position where the blue point represents the investigated point and the
grid’s intersection points represent the candidate positions of deer, and (c) distances corresponding
to the time lag of deer call (D
1
, D
2
) and investigated point (D
∗
1
, D
∗
2
).
The estimation of the position of the sound source (
Lat
0
,
Lon
0
) relied on a trial-and-
error method. A grid with a tolerance of 0.5 m along both the x and y axes (Figure 3b) was
used to determine the position of the sound source. The grid should provide adequate
coverage for detection of sound at all three recorders. Consequently, the distance between
any recorder and the grid edge (hereafter called edge distance) was set to be at least
750 m (Figure 3b). In the trial-and-error method, each of the grid’s intersection points was
investigated assuming that the deer emitted the sound at that point (e.g., the blue point in
Figure 3b). As the coordinates of the intersection points along with three-recorders were
known, the distances between investigated point and each of the three-recorders can be
calculated (i.e,
D
∗
inv_1
,
D
∗
inv_2
and
D
∗
inv_3
) as shown in Figure 3b. The calculated distances
enabled us to identify the nearest recorder from the investigated point, and calculate the
distances corresponding to the time lag in the other two-recorders (i.e.,
D
∗
1
,
D
∗
2
, shown in
Figure 3c). The root mean square error (RMSE) was then calculated between the distances
corresponding to the time lag of deer calls recorded at Rec 1 and Rec 2 (i.e.,
D
1
and
D
2
),
and time lag distances from the point being investigated (i.e.,
D
∗
1
and
D
∗
2
). By repeating
Sensors 2021, 21, 866 6 of 15
these calculations at all the grid intersection points, the point with the lowest RMSE was
extracted. The point with the lowest RMSE represents the retrieved deer position, which
matches or is very close to the actual deer position.
The following steps explain the details of the calculations required at each intersection
point of the grid, to determine the intersection point nearest to the deer position.
Step (1): For each intersection point of the grid, loop the following steps to be executed.
Let us use the blue intersection point shown in Figure 3b,c as an example, with the following
steps.
Step (2): Calculate the distances between the point under investigation, and each
of the three recorders. The coordinates of the point being investigated (
Lat
∗
0
,
Lon
∗
0
) are
known, as it is a grid intersection point. Applying the Pythagoras’ theorem on a geographic
projection [
36
], the distance between any two points with known coordinates can be
calculated as follows:
d = R
earth
×
q
x
2
+ y
2
(2)
x =
(
Lon
II
− Lon
I
)
×
π
180
× cos
Lat
II
+ Lat
I
2.0
×
π
180
(3)
y =
(
Lat
II
+ Lat
I
)
×
π
180
(4)
where
R
earth
refers to the mean radius of the earth (6,371,000 m), (
Lat
I
,
Lon
I
) are the
coordinates of the first point (
I
), and (
Lat
II
,
Lon
II
) are the coordinates of the second point
(
II
). The distance from the point under investigation to each of the three recorders (i.e.,
D
∗
inv_1
, D
∗
inv_2
, and D
∗
inv_3
, as shown in Figure 3b) are calculated using Equations (2)–(4).
Step (3): Determine the minimum distance (hereafter,
D
∗
min
) among the three distances
(i.e.,
D
∗
inv_1
,
D
∗
inv_2
and D
∗
inv_3
) calculated in step (2). This is to determine which recorder
is the nearest to the point being investigated. For the point investigated in this example,
Rec 3 is the nearest recorder, and D
∗
min
is equal to D
∗
inv_3
.
Step (4): Check if the recorder nearest to the point being investigated is the same
recorder that is nearest to the source of the deer call. If yes, the rest of the steps are
executed. If not, the loop should terminate, and the next intersection point of the grid
should be checked. This condition was set to reduce the computational time. For the point
investigated in this example, the nearest recorder is Rec 3, which is the same recorder
nearest to the sound source (i.e., Rec 3 was the first recorder to detect the deer sound).
Step (5): Calculate the distances corresponding to the time lag at the other two
recorders. For the point under investigation, shown in Figure 3b,c, Rec 1 and Rec 2 are
far from the investigating point compared to Rec 3. Thus, the distance
D
∗
1
(Figure 3c)
is the corresponding time lag distance of Rec 1, which can be calculated by subtracting
the distance
D
∗
min
calculated in Step (3) from the distance
D
∗
inv_1
calculated in Step (2)
(
D
∗
1
= D
∗
inv_1
− D
∗
min
). Similarly, the distance corresponding to the time lag of Rec 2 can be
calculated (D
∗
2
= D
∗
inv_2
− D
∗
min
), as shown in Figure 3c.
Step (6): The RMSE between the time lag distances corresponding to the point under
investigation (i.e.,
D
∗
1
and
D
∗
2
), and the distances calculated based on the deer sound (i.e.,
D
1
and D
2
) is calculated as [37]:
RMSE =
s
D
∗
1
− D
1
+
D
∗
1
− D
2
2
2
(5)
Step (7): The above-mentioned procedure should be performed for all the intersection
points of the grid (Figure 3b), as explained for the blue intersection point. The point that
provides the minimum RMSE represents the position of the deer.
In short, the proposed technique relies on the time lag among three recorders and a
trial-and-error method to estimate the emitted sound position. The proposed technique
uses three recorders in a triangle layout; however, more recorders with different architecture
Sensors 2021, 21, 866 7 of 15
could be used. The influence of wind speed and relative humidity on sound propagation
was not considered in this study.
2.3. Development of Sound Recording System
A sound recording system was designed to meet the requirements of our proposed
technique. The sound data were captured using an external microphone (i.e., AT8538 Power
Module, Audio-Technica, Tokyo, Japan) attached to a Zoom H4n Pro digital recorder (Zoom
North America Inc., Hauppauge, NY, USA), with a sampling rate of 44.1 kHz. The acoustic
data were saved into an SD memory card in MP3 format, with a 32-bit amplitude resolution.
The external microphone was installed approximately 1.5 m above the ground level and
pointed towards the region being investigated, as clearly shown in Figure 4b. Solar panels
were used to supply power to the whole system including recorders (
Figure 4a–c
). The
latitude and longitude information were obtained from a Garmin eTrex 20 GPS receiver
(Garmin International Incorporated Company, Olathe, KS, USA).
Figure 4.
Three sound recorders and solar panels in Oze marshland. (
a
) Recorder 1 (Rec 1),
(b) Recorder 2 (Rec 2), (c) Recorder 3 (Rec 3), and (d) components of the sound recording system.
There were two main challenges during the collection of sound data. The first chal-
lenge was the time drift problem with the sound recorders (i.e., the clock of each recorder
does not run at the same rate as the reference clock). For instance, even if the three recorders
are manually synchronized at a certain time, they will show different times after a while
(e.g., one day). As the proposed technique mainly relies on estimating the time lag among
the three recorders, the three recorders should start collecting data at exactly the same time.
Because of the time drift problem of the recorders, it was difficult to use the internal clock
of the recorder to start sound recording at the same time (i.e., synchronization process).
Consequently, a unique sound recording system was developed which consists of a GPS
synchronizer, a controlling unit, and a data logger, as shown in Figure 4d. The GPS syn-
chronizer was responsible for receiving accurate time information from GPS satellites every
hour. This precise time information was transferred to the controlling unit. The controlling
unit started/stopped collecting sound data every hour. The GPS synchronizer also passed
a signal to the data logger showing the synchronization status (i.e., whether receiving a
signal from GPS satellites or not).
The other main challenge was the power supply in the forest. The amount of sunlight
that reaches the solar panel was insufficient in some areas due to the dense tree cover.
Sensors 2021, 21, 866 8 of 15
Consequently, two solar panels (Figure 4b) were used in areas with dense tree cover. On
the other hand, the ideal time for recording deer calls is nighttime because that is when the
deer are active [
38
]. Accordingly, a timer device (Figure 4d) was added to stop the system
during the daytime (i.e., between 4:00 and 16:00) to save power.
2.4. Distance between Sound Recording Units
The distance between the sound recording units is a critical factor for our proposed
technique because the deer sound should be recorded by three recorders. In addition, it is
important to know the maximum distance at which the deer sound is detectable in Oze,
to reduce the total number of devices. The detectable range for deer sounds may vary
based on many conditions such as tree density, and recorder sensitivity. For instance, Enari
et al. [
2
] observed that the detection range for passive acoustic monitoring was 140 m in
defoliated forests, for deer calls during the rutting season. Masato Minami [
31
] reported
that there are two types of male deer voices in the rutting season; the maximum distances
at which these two voices were detectable were 400 and 700 m.
We also conducted an experiment to check the maximum distance at which sound
could be detected in Oze. During the experiment, sounds from deer whistles were created at
different distances from Rec 1, up to 761 m. Table 2 shows that the sound could be detected
within 400 m. The deer whistle was hardly audible at 545 m, and the sound was not detected
at all at 761 m. Consequently, the distance between sound recording units was adjusted
to be within 450 m, as shown in Figure 2d. The Garmin eTrex GPS receiver (Garmin
Ltd., Olathe, KS, USA) was used to determine the position information (i.e., latitude
and longitude) of the emitted sound from deer whistle. Praat software (University of
Amsterdam, Amsterdam, The Netherlands) was used to analyze the recorded sounds [
39
].
Table 2.
Sound from deer whistle was created at different distances from Rec 1. Symbols O,
∆
and X
represent the audibility status namely adequately detected, barely detected, and cannot be detected,
respectively.
Distance from Rec 1 181 m 297 m 423 m 545 m 761 m
Audibility of sound O O O ∆ X
3. Results and Discussion
3.1. Validation of the UTokyo Experiment
The UTokyo experiment was conducted on 29 November 2017. In the UTokyo ex-
periment, the microphones of the three recorders were installed facing the center of the
playground, as shown in Figure 1b. A deer whistle was used to create a sound at five
different points in the playground (i.e., A–E, shown in Figure 1b). The measured positions
(Table 1) of the five validation points were compared with the positions derived from the
proposed technique. The grid size of the trial-and-error method was
1.6 km × 1.6 km
, to
cover the playground and recorders. The sound editor windows of Praat (Figure 5a–c) illus-
trate the characteristics of the sound recorded by the three recorders when the deer whistle
was used at point E (Figure 5d). The sound recordings were monaural (i.e., one channel) as
one external microphone was connected to each recorder (Figure 5a–c). The blue curves
shown inside the spectrograms represent the pitch (i.e., the fundamental frequency of
sound waves). The pitch curves are one of the factors that can be used to determine the
time when the deer sound was first detected in each recorder. The pitch range (i.e., the
range in which the fundamental frequency analysis is carried out) is an important setting
for pitch analysis and changes based on the target sound. For example, the pitch range for
a male voice ranges from 75 to 300 Hz, whereas it is between 100 and 500 Hz for a female
voice. By testing different pitch ranges for both the UTokyo experiment (i.e., deer whistle
sound) and Oze (i.e., real deer sound), the lower and upper limits of the pitch range were
set at approximately 500 and 5000 Hz, respectively.
Sensors 2021, 21, 866 9 of 15
This paragraph explains the determination of time lag among the three recorders using
the measurements carried out at point E of the UTokyo experiment. The whistle sound
was emitted at point E, and the sound was first detected at Rec 2 (i.e., the nearest recorder
from point E, as shown in Figure 5d) at time 106.316 (i.e., the relative time measured with
respect to the starting time of the recordings). The same sound was later detected at Rec 1
and Rec 3 at 106.416 and 106.417, respectively. The time lag for both Rec 1 and Rec 3 from
Rec 2 (i.e., the first recorder that detected the sound) was calculated. The time lag at Rec 1
was 0.100 s (106.416–106.316, as shown in Figure 5a), and the time lag at Rec 3 was 0.101 s
(106.417–106.316, as shown in Figure 5c). The mean air temperature on 29 November 2017
was 18
◦
C, and the speed of sound was assumed to be 340 m s
−1
. The Noise Reduction tool
of Audacity software [
40
] was used to reduce the background noise (e.g., rustling leaves
and wind noises) in sound recordings, and thus improving the recorded sounds’ quality
and facilitating deer calls’ extraction. Noise reduction is a two-step procedure. In the first
step, the user provides a sample noise (at least 0.05 s at a sample rate of 44,100 Hz) to train
Audacity. Then, Noise Reduction is applied to the entire sound recording [41].
Figure 5.
Sound editor windows of Praat software show the sound waves and spectrogram of a
recorded sound at (
a
) Rec 1, (
b
) Rec 2, and (
c
) Rec 3 in UTokyo experiment. The recorded sound at
the three recorders was emitted from a whistle located at point E whose position is shown in (d).
Figure 6a illustrates the positions of the five reference points (A–E) as well as the
retrieved points (A_s, B_s, C_s, D_s, and E_s) from the proposed technique, from the
UTokyo experiment. The proposed technique introduced a mean error and RMSE of
4.60 and 4.80 m, respectively. In general, these retrieval accuracies are acceptable for
determining deer positions to estimate their population density in Oze.
Sensors 2021, 21, 866 10 of 15
Figure 6.
Google Earth images illustrate the positions of the validation points at (
a
) UTokyo ex-
periment and (
b
) Oze marshland. The letters (i.e., A–E) indicate the reference positions of deer
whistle and letters with “_s” represent the retrieved positions using the proposed technique. The
distance between the reference and retrieve positions represents the retrieval error using the proposed
technique.
The main factor that can increase error in the retrieved position is the difficulty in
manually determining the starting time of deer calls. For instance, the maximum error
between the reference and retrieved position was 6.17 m for point E, as shown in Figure 6a.
This distance error means that the time error was 0.018 s (i.e., 6.17 m/340 m s
−1
). The time
error of 0.018 s is mainly due to lack of human precision to manually determine the time at
which the sound was first detected at each recorder. Figure 7 illustrates the sound waves
and spectrogram of Rec 2 at time 106.316. The highlighted part in Figure 7 represents the
time frame of 0.018 s. Figure 7 emphasizes the challenge of determining the initial detection
time for sound within 0.018 s using sound waves and pitch curve, as the time at which the
magnitude of sound waves started to increase (i.e., T
2
in Figure 7) was different from the
time at which the pitch curve was detected from sound (i.e., T
1
in Figure 7).
Figure 7.
The sound waves and spectrogram at Rec 2 of the UTokyo experiment for detecting the
position of Point E (Figure 1b). The highlighted part represents time frame of 0.018 s.
3.2. Validation of the Oze Marshland Experiment
The validation experiment was conducted in Oze National Park on 14 September 2018.
During the validation experiment, the three recorders (Rec 1–3) were used to recorded
deer calls emitted from sound whistle at five reference validation points (A–E, shown
in
Figure 6b
). Three out of the five validation points (B–D, hereafter called inner points)
were located within the triangle created by the three recorders, whereas the other two
points (A and E, hereafter called outer points) were located outside the triangle, as shown
in Figure 6b. The latitude and longitude of the five validation points were determined
using the Garmin eTrex GPS receiver, and are listed in Table 1. The Noise Reduction tool
(Audacity software) was not used in the case of the Oze marshland experiment because of
the presence of many sources of irregular background noise. Instead, the sound recordings
Sensors 2021, 21, 866 11 of 15
were amplified to increase the sound volume using the Amplify tool of Audacity software.
The grid size for the trial-and-error method in the Oze marshland experiment, and for
tracking deer calls in Oze (Section 3.3) was 3 km × 3 km.
Figure 6b illustrates the positions of deer whistle at five reference validation locations
(A–E, Table 1) as well as the retrieved locations (A_s, B_s, C_s, D_s, and E_s) from the
proposed technique. The retrieval accuracy of the inner points (i.e., points B–D, Figure 6b)
introduced lower error compared with the outer points (i.e., points A and E, Figure 6b),
with a mean error of 15.97 and 44.85 m for the inner and outer points, respectively. The
reason for the higher error at the outer points may be related to their distant location from
the three microphones. For example, Figure 8 shows a comparison between the inner point
(point C) and outer point (point E). At point C, the distance to all recorders was less than
400 m and the sound intensity for all recorders was high. In contrast, the distance between
point E and Rec 1 (i.e., the farthest recorder from point E among the three recorders) was
495 m, and the sound intensity at Rec 1 was very low, which caused an error in detecting the
starting time of the whistle sound at Rec 1. Additionally, increasing the distance between
the sound source and the sound recordings along with the presence of other environmental
features (e.g., high density of trees or altitude difference) attenuate sound recordings and
thereby diminishing the retrieval accuracy, which is consistent with the findings of Castro
et al. [
42
] and Priyadarshani et al. [
43
]. This conclusion matches our inference in
Section 2.4
,
that the probability of sound detection decreases from approximately 500 m.
Figure 8.
Sound waves and spectrogram of sound recordings from whistle C and whistle E at
(
a
) Rec 1, (
b
) Rec 2, and (
c
) Rec 3, during the validation of the Oze Marshland Experiment. The
locations of both whistles are shown in Figure 6b. Images were generated using Praat software.
3.3. Tracking Deer Calls in Oze
Sound recordings of real deer calls in Oze were collected between 8 October and 26
October 2018. October is the rutting season (breeding season), in which the males increase
their call range and frequency to attract females. The determination of deer position in
an actual deer environment (i.e., Oze) was conducted using four recorders, as shown in
Figure 2d (this includes the three recorders used during the Oze marshland validation
experiment), in which the distance between recorders remained within 450 m. In order to
demonstrate the importance of detecting deer positions, sound recordings on 13 October
2018, between 18:00 and 20:00 (i.e., two-hours) were processed. There were 72 deer calls
during the two hours, with an average and standard deviation of 6 calls and 3.59 calls for
Sensors 2021, 21, 866 12 of 15
the 10 min interval, respectively. The animation for the accumulation of deer calls over the
two hours considering 30 s intervals is available as Supplementary Material.
Figure 9 shows the accumulation of deer calls from the start time (i.e., 18:00) to the
time posted in each image, over two-hours, and illustrates the changes within a 10 min
frame. For instance, there were six deer calls from 18:00 to 18:10 (Figure 9a) increased to
14 deer calls by 18:20 (Figure 9b). The orange circle in Figure 9d highlights the first call of
a deer. After 50 min, the same deer made another call, as shown in Figure 9i. Similarly,
the yellow circles in Figure 9h,i highlight the first and second calls from the same deer,
respectively. The red arrow in Figure 9a highlights the first call from a deer, whereas the
red arrow in Figure 9l indicates multiple calls from mostly the same deer. The main reason
for assuming that these multiple calls are from the same deer is that the male deer is a
territorial animal (i.e., defends a certain area from other males), particularly during the
rutting season (from late September to early March) [
8
]. Consequently, there should be
a distance between two males, and if the sound comes from adjacent positions, they are
likely to be from the same male. The green arrow in Figure 9a illustrates the first call from
a deer. By checking the variation of this deer call along the 12 images, the movement of
this deer toward Rec 2 can be clearly seen as highlighted in Figure 9l. Likewise, the blue
arrows (Figure 9d,l) and red arrows (Figure 9a,l) clearly show the start and subsequent
calls, respectively, from most likely the same deer.
Sensors 2021, 21, x FOR PEER REVIEW 15 of 18
for assuming that these multiple calls are from the same deer is that the male deer is a
territorial animal (i.e., defends a certain area from other males), particularly during the
rutting season (from late September to early March) [8]. Consequently, there should be a
distance between two males, and if the sound comes from adjacent positions, they are
likely to be from the same male. The green arrow in Figure 9a illustrates the first call from
a deer. By checking the variation of this deer call along the 12 images, the movement of
this deer toward Rec 2 can be clearly seen as highlighted in Figure 9l. Likewise, the blue
arrows (Figure 9d,l) and red arrows (Figure 9a,l) clearly show the start and subsequent
calls, respectively, from most likely the same deer.
Figure 9. Images show the accumulation of deer call positions from starting time (8:00:00 on 13 October 2018) at 10 min
interval for 2 h, as shown in (a–l). Deer symbols represent the position of deer calls retrieved by the proposed technique.
Yellow and orange circles highlight the occurrence of deer calls from the same deer at two different times. The green, red,
and blue arrows highlight the repetition and potential path of the same deer with multiple calls. The No. in each figure
represents the total accumulated deer calls from the starting time (18:00 on 13 October 2018).
Additionally, there are two camera traps located within our validation site, as shown
in Figure 10a. Two deer call positions were retrieved using the proposed technique at
19:11:00 and 19:58:46 (the yellow circle in Figure 10a), located within 25 m from camera
trap 1. Camera trap 1 captured a male deer at 19:30:46 (Figure 10b). The positions of four
deer calls retrieved using the proposed technique were within 50 m from camera trap 2
between 18:11:26 and 18:53:45 (the orange circle in Figure 10a). Camera trap 2 captured
male deer at 18:15:02 (Figure 10c). As both images of the camera traps are for male deer
Figure 9.
Images show the accumulation of deer call positions from starting time (8:00:00 on 13
October 2018) at 10 min interval for 2 h, as shown in (
a
–
l
). Deer symbols represent the position of
deer calls retrieved by the proposed technique. Yellow and orange circles highlight the occurrence of
deer calls from the same deer at two different times. The green, red, and blue arrows highlight the
repetition and potential path of the same deer with multiple calls. The No. in each figure represents
the total accumulated deer calls from the starting time (18:00 on 13 October 2018).
Sensors 2021, 21, 866 13 of 15
Additionally, there are two camera traps located within our validation site, as shown
in Figure 10a. Two deer call positions were retrieved using the proposed technique at
19:11:00 and 19:58:46 (the yellow circle in Figure 10a), located within 25 m from camera
trap 1. Camera trap 1 captured a male deer at 19:30:46 (Figure 10b). The positions of four
deer calls retrieved using the proposed technique were within 50 m from camera trap 2
between 18:11:26 and 18:53:45 (the orange circle in Figure 10a). Camera trap 2 captured
male deer at 18:15:02 (Figure 10c). As both images of the camera traps are for male deer
who are territorial animals, we deduce that the deer positions retrieved using our technique
correspond to the deer captured by the two camera traps. Further research using marked
individuals with tracking devices combined with sound recordings and camera traps are
needed to confirm this speculation.
These outcomes underline the significance of the proposed technique for detecting
repetitive calls from the same deer and excluding them, when using sound recordings for
deer counting, in addition to observing deer movement by determining the position of
deer sounds. Accordingly, the proposed method can be a basis for sound-based techniques
for counting deer and other vocal animals in Oze and other regions.
Figure 10.
(
a
) Shows the deer call positions retrieved between 18:00 to 20:00 on 13 October 2018 using
the proposed technique, along with locations of two camera traps. The deer captured by camera
traps 1 and 2 are shown in (b,c), respectively.
4. Conclusions and Recommendations
In this study, a new technique to detect deer position (i.e., latitude and longitude) from
sound recordings at three recorders was introduced. In the proposed technique, the sound
is recorded by three stationary recorders that receive the same sound at three different
times based on the distance of the sound source from the three recorders. Then, the time
lag in detecting the same sound among the three recorders is estimated. A sound recording
system was also developed to receive accurate time information from GPS satellites to
overcome the time drift problem of the sound recorders (i.e., the clock of each recorder does
not run at exactly the same rate as the reference clock used); therefore, the time lag among
the three recorders was accurately estimated. The distance between sound recording units
along with the presence of obstacles (e.g., high density of trees) influence sound recordings
and hence the retrieval accuracy of the proposed technique. The retrieval accuracy’s mean
error of the experimental and natural environment was 4.60 and 15.97 m, respectively.
Thus, the developed technique is most suitable for open habitat and difficult-to-access
terrain (e.g., muddy wetlands) with sparse tree cover and relatively flat terrain. The
proposed technique tracks the movement of deer and can be applied to other vocal animals,
Sensors 2021, 21, 866 14 of 15
and therefore monitor ecological changes. Hence, the technique is the first step toward
designing an automated system for estimating deer or other vocal animals’ population
using sound recordings, as the position information can be used to exclude repetitive calls
from the same animal. The proposed technique was validated using a deer whistle and
camera traps. It is highly recommended to validate the proposed technique using marked
individuals with tracking devices (e.g., GPS collars) in future applications, particularly
when applying the technique with other vocal species.
Supplementary Materials:
The following are available online at https://www.mdpi.com/1424-822
0/21/3/866/s1, Video S1: The animation for the accumulation of deer calls over two hours.
Author Contributions:
Conceptualization, S.I.S.; data curation, S.I.S. and K.F.; formal analysis, S.I.S.;
funding acquisition, M.M., T.O. and K.O.; investigation, S.I.S. and K.F.; methodology, S.I.S. and
K.O.; project administration, K.O.; resources, K.O.; software, S.I.S.; supervision, M.M., T.O. and
K.O.; validation, S.I.S. and K.F.; visualization, S.I.S.; writing—original draft, S.I.S.; writing—review &
editing, S.I.S. All authors have read and agreed to the published version of the manuscript.
Funding:
This research was supported by the Environment Research and Technology Development
Fund (grant number JPMEERF20174006) of the Environmental Restoration and Conservation Agency
of Japan.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Acknowledgments:
The authors would like to thank the three anonymous reviewers for their
constructive comments to improve the manuscript.
Conflicts of Interest: The authors declare no conflict of interest.
References
1.
Acevedo, P.; Ruiz-Fons, F.; Vicente, J.; Reyes-García, A.R.; Alzaga, V.; Gortázar, C. Estimating red deer abundance in a wide range
of management situations in Mediterranean habitats. J. Zool. 2008, 276, 37–47. [CrossRef]
2.
Enari, H.; Enari, H.; Okuda, K.; Yoshita, M.; Kuno, T.; Okuda, K. Feasibility assessment of active and passive acoustic monitoring
of sika deer populations. Ecol. Indic. 2017, 79, 155–162. [CrossRef]
3.
Ando, M.; Shibata, E. Bark-Stripping Preference of Sika Deer and Its Seasonality on Mt. Ohdaigahara, Central Japan. In Sika Deer;
Springer: Berlin/Heidelberg, Germany, 2009; pp. 207–216.
4.
Morellet, N.; Gaillard, J.; Hewison, A.J.M.; Ballon, P.; Boscardin, Y.; Duncan, P.; Klein, F.; Maillard, D. Indicators of ecological
change: New tools for managing populations of large herbivores. J. Appl. Ecol. 2007, 44, 634–643. [CrossRef]
5.
Naito, T.; Kimura, Y. Sika deer in the Oze area. In A Comprehensive Study of the Oze Area; Oze Research Group: Maebashi, Japan,
1998; pp. 725–739.
6.
Takafumi, H.; Kamii, T.; Murai, T.; Yoshida, R.; Sato, A.; Tachiki, Y.; Akamatsu, R.; Yoshida, T. Seasonal and year-round use of the
Kushiro Wetland, Hokkaido, Japan by sika deer (Cervus nippon yesoensis). PeerJ 2017, 5, e3869. [CrossRef] [PubMed]
7. Takatsuki, S. Effects of sika deer on vegetation in Japan: A review. Biol. Conserv. 2009, 142, 1922–1929. [CrossRef]
8.
Enari, H.; Enari, H.S.; Okuda, K.; Maruyama, T.; Okuda, K.N. An evaluation of the efficiency of passive acoustic monitoring in
detecting deer and primates in comparison with camera traps. Ecol. Indic. 2019, 98, 753–762. [CrossRef]
9.
Takayama, K.; Sonoda, A.; Hayashida, Y.; Ishii, D.; Yanagita, D.; Matsumoto, S.; Katahira, K.; Oshima, I.; Nakanishi, Y.; Inaome, T.;
et al. Wild Sika Deer Invasions of the Grasslands within Livestock Farms. J. Warm Reg. Soc. Anim. Sci. 2017, 60, 21–26.
10.
Clare, J.; McKinney, S.T.; DePue, J.E.; Loftin, C.S. Pairing field methods to improve inference in wildlife surveys while accommo-
dating detection covariance. Ecol. Appl. 2017, 27, 2031–2047. [CrossRef]
11.
Nakano, T.; Bavuudorj, G.; Iijima, Y.; Ito, T.Y. Quantitative evaluation of grazing effect on nomadically grazed grassland
ecosystems by using time-lapse cameras. Agric. Ecosyst. Environ. 2020, 287, 106685. [CrossRef]
12.
Nakashima, Y.; Hongo, S.; Akomo-Okoue, E.F. Landscape-scale estimation of forest ungulate density and biomass using camera
traps: Applying the REST model. Biol. Conserv. 2020, 241, 108381. [CrossRef]
13.
Lahkar, D.; Ahmed, M.F.; Begum, R.H.; Das, S.K.; Harihar, A. Responses of a wild ungulate assemblage to anthropogenic
influences in Manas National Park, India. Biol. Conserv. 2020, 243, 108425. [CrossRef]
14.
Igarashi, T.; Takatsuki, S. Effects of defoliation and digging caused by sika deer on the Oze mires of central Japan. Biosph. Conserv.
Nat. Wildl. Hum. 2008, 9, 9–16.
15.
Abdalla, M.; Hastings, A.; Chadwick, D.R.; Jones, D.L.; Evans, C.D.; Jones, M.B.; Rees, R.M.; Smith, P. Critical review of the
impacts of grazing intensity on soil organic carbon storage and other soil quality indicators in extensively managed grasslands.
Agric. Ecosyst. Environ. 2018, 253, 62–81. [CrossRef] [PubMed]
Sensors 2021, 21, 866 15 of 15
16.
Steen, K.A.; Villa-Henriksen, A.; Therkildsen, O.R.; Green, O. Automatic detection of animals in mowing operations using
thermal cameras. Sensors 2012, 12, 7587–7597. [CrossRef] [PubMed]
17.
Alves, J.; da Silva, A.A.; Soares, A.M.V.M.; Fonseca, C. Pellet group count methods to estimate red deer densities: Precision,
potential accuracy and efficiency. Mamm. Biol. 2013, 78, 134–141. [CrossRef]
18.
Kimuyu, D.M.; Veblen, K.E.; Riginos, C.; Chira, R.M.; Githaiga, J.M.; Young, T.P. Influence of cattle on browsing and grazing
wildlife varies with rainfall and presence of megaherbivores. Ecol. Appl. 2017, 27, 786–798. [CrossRef]
19.
Kufeld, R.C.; Olterman, J.H.; Bowden, D.C. A helicopter quadrat census for mule deer on Uncompahgre Plateau, Colorado. J.
Wildl. Manag. 1980, 632–639. [CrossRef]
20. Krebs, C.J. Ecological Methodology; Harper & Row: New York, NY, USA, 1989.
21.
Poulsen, J.R.; Clark, C.J.; Bolker, B.M. Decoupling the effects of logging and hunting on an Afrotropical animal community. Ecol.
Appl. 2011, 21, 1819–1836. [CrossRef]
22.
Walsh, P.D.; White, L.J.T. Evaluating the steady state assumption: Simulations of gorilla nest decay. Ecol. Appl.
2005
, 15, 1342–1350.
[CrossRef]
23.
Mourão, G.; Coutinho, M.; Mauro, R.; Campos, Z.; Tomás, W.; Magnusson, W. Aerial surveys of caiman, marsh deer and pampas
deer in the Pantanal Wetland of Brazil. Biol. Conserv. 2000, 92, 175–183. [CrossRef]
24.
Hodgson, J.C.; Baylis, S.M.; Mott, R.; Herrod, A.; Clarke, R.H. Precision wildlife monitoring using unmanned aerial vehicles. Sci.
Rep. 2016, 6, 22574. [CrossRef] [PubMed]
25.
Wang, D.; Shao, Q.; Yue, H. Surveying Wild Animals from Satellites, Manned Aircraft and Unmanned Aerial Systems (UASs): A
Review. Remote Sens. 2019, 11, 1308. [CrossRef]
26.
Barbedo, J.G.A.; Koenigkan, L.V.; Santos, T.T.; Santos, P.M. A Study on the Detection of Cattle in UAV Images Using Deep
Learning. Sensors 2019, 19, 5436. [CrossRef] [PubMed]
27.
Drake, K.L.; Frey, M.; Hogan, D.; Hedley, R. Using digital recordings and sonogram analysis to obtain counts of yellow rails.
Wildl. Soc. Bull. 2016, 40, 346–354. [CrossRef]
28.
Aihara, I.; Bishop, P.J.; Ohmer, M.E.B.; Awano, H.; Mizumoto, T.; Okuno, H.G.; Narins, P.M.; Hero, J.-M. Visualizing phonotactic
behavior of female frogs in darkness. Sci. Rep. 2017, 7, 10539. [CrossRef]
29.
Kessel, S.T.; Cooke, S.J.; Heupel, M.R.; Hussey, N.E.; Simpfendorfer, C.A.; Vagle, S.; Fisk, A.T. A review of detection range testing
in aquatic passive acoustic telemetry studies. Rev. Fish Biol. Fish. 2014, 24, 199–218. [CrossRef]
30.
Marques, T.A.; Thomas, L.; Martin, S.W.; Mellinger, D.K.; Ward, J.A.; Moretti, D.J.; Harris, D.; Tyack, P.L. Estimating animal
population density using passive acoustics. Biol. Rev. 2013, 88, 287–309. [CrossRef]
31.
Minami, M. Vocal Repertoire and the Functions of Vocalization in the Rutting Season in Sika Deer, Cervus Nippon; Osaka City University:
Osaka, Japan, 1997.
32.
Tayanin, D.; Lindell, K. Hunting and Fishing in a Kammu Village: Revisiting a Classic Study in Southeast Asian Ethnography; NIAS
Press: Copenhagen, Denmark, 2012.
33.
Ninan, K.N.; Inoue, M. Valuing forest ecosystem services: Case study of a forest reserve in Japan. Ecosyst. Serv.
2013
, 5, 78–87.
[CrossRef]
34.
Goto, T. Wildlife Management Office (WMO) Field Note No.141. Available online: http://wmo.co.jp/field_note/no-141 (accessed
on 1 February 2020).
35.
NLNI Lakes Data, National Land Numerical Information, Japan. Available online: http://nlftp.mlit.go.jp/ksj/ (accessed on 14
March 2017).
36.
Veness, C. Calculate Distance and Bearing between Two Latitude/Longitude Points. Available online: https://www.movable-
type.co.uk/scripts/latlong.html (accessed on 10 April 2018).
37.
Chai, T.; Draxler, R.R. Root mean square error (RMSE) or mean absolute error (MAE)?—Arguments against Avoiding RMSE in
the Literature. Geosci. Model Dev. 2014, 7, 1247–1250. [CrossRef]
38.
Kamei, T.; Takeda, K.; Izumiyama, S.; Ohshima, K. The effect of hunting on the behavior and habitat utilization of sika deer
(Cervus nippon). Mammal Study 2010, 35, 235–241. [CrossRef]
39. Boersma, P. Praat, a system for doing phonetics by computer. Glot Int. 2001, 2, 341–347.
40.
Audacity. Audacity (R): Free Audio Editor and Recorder, Version 2.0.6. Available online: http://windinthew.users.sourceforge.
net/?lang=en (accessed on 26 January 2021).
41.
Lavi; Ahuja, A.K. Speech Signal De-noising using Wavelet Transform and Different Standard Softwares: Performance Evaluation
and Comparisons Study. In Proceedings of the 2018 5th IEEE Uttar Pradesh Section International Conference on Electrical,
Electronics and Computer Engineering (UPCON), Gorakhpur, India, 2–4 November 2018; pp. 1–6.
42.
Castro, I.; De Rosa, A.; Priyadarshani, N.; Bradbury, L.; Marsland, S. Experimental test of birdcall detection by autonomous
recorder units and by human observers using broadcast. Ecol. Evol. 2019, 9, 2376–2397. [CrossRef] [PubMed]
43.
Priyadarshani, N.; Castro, I.; Marsland, S. The impact of environmental factors in birdsong acquisition using automated recorders.
Ecol. Evol. 2018, 8, 5016–5033. [CrossRef] [PubMed]
