Method for Identification of Individual Flamingos

Based on Movement Logs

Riku Okazaki

and Yu Suzuki

Solution Promotion Department, Otas Co., Ltd., KDX Sendai Honcho Building, 2-3-10 Honcho, Aoba-ku, Sendai, Japan

Department of Information and Data Science, Notre Dame Seishin University, 2-16-9 Ifuku-cho, Kita-ku, Okayama, Japan

Keywords: Animal Computer Interaction, Individual Identification.

Abstract: This research aims to enable non-wearing and non-contact identification of individual animals. As an

approach to this goal, we examined the feasibility of an identification method using movement logs, which is

one of the animal behaviors. In this study, flamingos kept in a zoo were targeted. In order to collect the

movement logs of flamingos, we developed a system in YOLO to identify individual flamingos and record

their locations based on videos taken in the zoo's keeping area. In addition, we analyzed the collected

movement logs of multiple individuals using a neural network and found that the movement logs could be

used to detect individuals with higher identification accuracy than random inference for individual

identification. Furthermore, we showed that the location where individuals tend to stay and the posture they

tend to adopt change depending on conditions such as weather (rainy or cloudy) and the time of day (noon

time period). This research indicates the feasibility of identifying individuals by their movements.

1 INTRODUCTION

Researchers examine the age and lifespan of

individual animals in the field of wildlife herpetology

and animal behavior. Animal keepers manage the

health and care of their animals or identify them when

they go missing. Animal identification is important in

the biological study and management of these

animals. Identification methods that involve contact

with the animal or attachment of an instrument are in

widespread use around the world (Silvy, 2012)

(Ahmad, 2022).

The World Organisation for Animal Health

(WOAH) has established five basic principles in

animal welfare, defined as "the physical and mental

state of an animal in relation to the conditions in

which it lives and dies" (WOAH, 2023).

i. freedom from hunger, thirst, and malnutrition

ii. freedom from fear and distress

iii. freedom from physical and thermal discomfort

iv. freedom from pain, injury, and disease

v. freedom to express normal patterns of behavior

Contacting the animal or wearing an instrument for

identification may cause ii through v of these to not

be satisfied. Some previous studies (Silvy, 2012)

(Ahmad, 2022), indicate that contact with animals or

wearing equipment causes “fear”, “distress”,

“physical discomfort”, “pain” or “injury”. In addition,

these factors can cause “disease” and failure to

express “normal patterns of behabiour”. Therefore, it

is necessary to adopt a method that does not interfere

with the animal's body in order to identify individuals

while satisfying the five basic principles.

The goal of this research is to enable non-contact

and non-wearing identification in accordance with the

basic principles of animal welfare. As an approach,

we focus on behavioral features whose effectiveness

has not yet been verified. Among various types of

behavior, behavior that all animals engage in is

migration. Therefore, this study examines the

feasibility of individual identification using animal

movement logs. This paper describes the

development of a system for acquiring movement

logs, and then reports the results of the verification on

the feasibility of individual identification using

movement logs.

438

Okazaki, R. and Suzuki, Y.

Method for Identiﬁcation of Individual Flamingos Based on Movement Logs.

DOI: 10.5220/0013650400003964

In Proceedings of the 20th International Conference on Software Technologies (ICSOFT 2025), pages 438-445

ISBN: 978-989-758-757-3; ISSN: 2184-2833

2 EXISTING RESEARCH ON

ANIMAL IDENTIFICATION

WITH PHYSICAL

INTERFERENCE

Individual identification methods involving physical

interference can be broadly classified into three types:

marking, tagging, and embedding. Marking is a

method of placing visual landmarks on the animal's

body. Examples include freeze branding, tattoos and

notching on ears (Silvy, 2012) (Ahmad, 2022).

Tagging is a method of attaching an artifact to the

animal's body that serves as a symbol. Examples

include ear-tags and bands (Silvy, 2012). Embedding

is a method of implanting a microchip or other device

into the body in which bioinformation to be managed

is written.

Walker et al (Walker, 2011) evaluated heart rate,

respiratory rate, and behavior of Steller's sea lions

under general anesthesia when branded with a hot

iron for identification. The results showed that heart

rate and respiratory rate increased and body

movements and tremors increased during branding.

The authors concluded that anesthetized Steller's sea

lions may feel pain during branding.

Lay et al (Lay, 1992) evaluated blood

composition, heart rate, vocal rate, and the difference

between the body temperature at the branding site and

that at the same site opposite the branding site when

calves were subjected to hot-iron branding, freeze

branding, or sham branding at the same temperature

as room temperature for individual identification.

Results showed that plasma epinephrine, a blood

component that increases with psychological stress,

was increased in hot-iron branded calves. Heart rate

was greater in calves that were hot-iron branded or

freeze branded, and calves that were hot-iron branded

or freeze branded vocalized. Differences in regional

body temperatures were reported to be greater

between freeze-branded and hot-iron calves.

Burley et al (Burley, 1982) studied the effects of

the color of plastic tags attached to zebra finches on

reproductive behavior. The authors summarized that

female zebra finches prefer males with red plastic

color bands and tend to avoid light blue and light

green male ones, while male zebra finches prefer

black and pink females and avoid blue and light green

female ones.

Carminato et al. (Carminato, 2011) reported the

presence of a microchip in a tumor surgically

removed from the body of a cat and concluded that

the microchip caused a foreign body reaction,

although very rare, resulting in fibrosarcoma. Thus, it

can be confirmed that implantation can cause disease.

Therefore, it is necessary to use identification

methods that avoid physical interference, since

individual animal identification involving it may not

achieve the basic principles of animal welfare.

3 IDENTIFICATION METHOD

USING BEHAVIORAL

FEATURES

3.1 Comparison of Physical and

Behavioral Features in

Identification

In recent years, AI has begun to be actively used in

non-contact and non-attached identification methods.

There are two main types of AI-based identification

methods: physical feature-based methods and

behavioral feature-based ones.

Methods based on physical features utilize data on

visual features of the animal's body. These include the

identification of sea turtles by the scale patterns on

the sides of their heads by Karun et all (Karun, 2021)

and the identification of chimpanzees using their

faces by Schofield et all. (Schofield, 2019). Basically,

still images or videos are used as training data. In the

case of still images, the input data is used as-is or with

necessary portions cropped. The AI is trained based

on these data to create a classifier for individual

identification. In the case of video, after clipping a

still image for each frame from the video, the same

procedure is used to create a discriminative AI.

Methods based on physical features are limited

because they allow AI to use as training data only a

portion of the moment when individual identification

becomes possible.

In methods based on physical features, individual

identification is based on appearance information,

which is subject to change depending on physical

growth and environmental conditions. When the

appearance of an individual changes, there is a high

possibility that the data that was previously acquired

and used for training will no longer be usable. On the

other hand, in the method based on behavioral

features, individual identification is performed based

on information that reveals the habits, personality,

and customs of the individual. Since these are less

likely to change compared to appearance, the

acquired information can be used for a long period of

time. In addition, there are few studies on individual

identification methods using behavioral features

other than those in the literature (Suzuki, 2021), and

Method for Identiﬁcation of Individual Flamingos Based on Movement Logs

439

there are no studies that summarize their

effectiveness. Therefore, in this study, we use

behavioral features to identify individuals.

3.2 Behavior Used for Individual

Identification

To explore behaviors that could be used for individual

identification, we first observed behaviors of familiar

animals. As a result, we noticed that some breeding

cats prefer to stay in certain places, and that turtles in

ponds come up to the land to sunbathe at certain times

of the day. Since these phenomena are caused by

movement among behaviors, there is a possibility that

individual animals can be identified based on

movement data. Therefore, in this study, we focus on

“movement logs” among behaviors to identify

individual animals.

3.3 Selection of Animals to Be

Identified

With the cooperation of Yagiyama Zoological Park,

we selected animals for identification in this study.

The criteria for selection were that there must be

multiple individuals of the same species, including

related species; that two or more individuals must be

kept or exhibited in the same space; that all target

animals must move frequently within the space; and

that the animals must not be interested in touching or

destroying the set-up equipment. After selecting

animals that met these conditions, we decided to

conduct this study on flamingos.

The flamingos kept at the zoo are European

greater flamingo (Phoenicopterus ruber roseus),

Chilean flamingo (Phoenicopterus chilensis),

Caribbean flamingo (Phoenicopterus ruber), totaling

17. All but the Caribbean flamingo have a color tag

on their right or left leg to uniquely identify them

within the species. Some European flamingos and

Chilean flamingos have overlapping colors on the left

and right sides of their legs and color tags.

We selected four flamingos for this study. Their

species and tag correspondence are shown in Table 1.

Based on the color of the tag and the left and right

Table 1: Correspondence between flamingo species kept at

the zoological garden and color tags.

ecies Left Le

ht Le

European greater flamingo blue none

European greater flamingo none blue

Caribbean flamingo none none

European greater flamingo none black

sides of the legs, we named them “Rblue”, “Lblue”,

“None” and “Rblack”.

4 DEVELOPMENT OF THE

SYSTEM TO OBTAIN

FLAMINGO LOCATION

INFORMATION

4.1 Overview of the System

To achieve the objective of this study, we developed

a system to verify the feasibility of flamingo

identification using movement logs. First, we

captured video of flamingos to obtain movement data.

Next, we created a flamingo detection model using

YOLO. Finally, we recorded location information

from the movement data of each individual flamingo

detected by the model.

4.2 Collection of Video Data

To obtain flamingo movement logs, We captured

video images from inside the breeding space (Figure

1). We selected the shooting location because it was

easy for flamingos to gather and because it was

possible to shoot from a horizontal angle. We used

Samsung Galaxy A7 as our photographic equipment.

The shooting period was from July 4 to October 22,

2023, from 9:00 a.m. to 5:00 p.m. To prevent thermal

runaway of the equipment, we took video in a cycle

of 30 minutes of shooting followed by 10 minutes of

standby.

Figure 1: Angle of view 4 from inside the breeding space.

4.3 Creation of a Custom Learning

Model for Individual Flamingo

Detection Using YOLO

As a preliminary step to identify individuals using

movement logs, it is necessary to record which

individuals are in which locations in the video.

Therefore, we created a flamingo individual detection

ICSOFT 2025 - 20th International Conference on Software Technologies

440

model using YOLO v8 to automatically record the X

and Y coordinates representing the position. To

generate training data to identify each individual, we

manually linked the position of the flamingo in the

video to the individual. We identified and linked

individuals using color tags attached to the left and

right legs. However, some individuals had

overlapping left and right color tags, making manual

identification difficult, so we annotated only those

individuals for which the left and right leg color and

color tag color were completely unique for all species.

Four of the individuals were frequently seen in the

angle of view of the camera, so we created a training

model for them.

To prepare the annotation, we cut one frame per

second from a 30-minute movie (pixel count:

1920px✕1080px, frame rate: 30fps) and saved each

frame as a jpg file. To avoid running out of memory

during training, we downscaled the number of pixels

to 960px✕540px. For annotation, we used labelImg

and labeled each flamingo as a class. The train data

was annotated based on a total of 4282 images from

the videos of October 11-22, 2023, and the validation

data was annotated based on a total of 3167 images

from the data of September 4-13, 2023. Training was

performed for 100 epochs based on the yolo8n.pt

model, and a custom training model was created.

Figure 2 shows the training results. The six items

from the middle to the left are losses, which indicate

the error between the percentage of correct answers

and the predicted value. If they move to the lower

right, we have achieved good learning. The four items

to the right of the middle are “precision”, “recall”,

“mAP50” and “mAP50-95”. If they move to the

upper right, the learning is good. Learning was

generally good except for “precision,” which

indicates the percentage of true positives out of all

positives. What we did to compensate for the low

“precision” is described below in section 4.5.

Figure 2: Result of the custom learning model. ‘loss’ and

‘precision’ moved to the lower right. ‘recall’, ‘mAP50’ and

‘mAP50-95’ moved to the upper right.

4.4 Extraction of Location Information

We used videos taken between September 20 - 27,

2023 as the original data from which we extracted

location information. To avoid memory shortages

during inference, we downscaled the pixel size from

1920px✕1080px to 960px✕540px. We performed

inference on the data using the custom training model

we created, and saved the results in a CSV file and in

a video file with the Bounding Box added. In the

contents of the CSV file shown in Table 2, “x1”, “y1”,

“x2” and “y2” are the coordinates of the Bounding

Box detected by inference (x1 is the X coordinate of

the left side, y1 is the Y coordinate of the upper side,

x2 is the X coordinate of the right side, and y2 is the

Y coordinate of the lower side). “class” is the class

name of the detected object (0 for Rblue, 1 for Lblue,

2 for None, 3 for Rblack). “confidence” is the

probability that the detected object's ”class” is correct.

In this study, this coordinate data is treated as a

location data set.

4.5 Modification of the Location Data

Set

Figure 2 shows that the “precision” value remained at

a low value around 0.1 even after further training. The

current position data set contains a large amount of

erroneous data. Therefore, it may be difficult to

distinguish true positives from false positives. To

verify the feasibility of individual identification more

Table 2: Example of the contents of a saved CSV file.

x1 y1 x2 y2 class confidence

387.57 176.61 426.84 362.69 none 0.51

82.58 130.03 182.07 419.71 Lblue 0.51

218.10 275.05 271.28 388.94 Rblue 0.58

204.60 290.04 270.67 391.59 Rblac

0.65

Method for Identiﬁcation of Individual Flamingos Based on Movement Logs

441

accurately in the subsequent validation, we

modify the data. It reduces the number of erroneous

data in the location data set.

Comparing the location data set and the saved

video obtained in the previous section, we modify the

data by manually deleting the false positives. The data

to be deleted are those with incorrect class and those

whose errors can be easily identified by the color of

the color tag or by the lack of positional coverage. We

did not perform the procedure to re-label the data

correctly. The number of records after data

examination was 427,252.

5 FEASIBILITY OF LOCATION-

BASED FLAMINGO

IDENTIFICATION

5.1 Hypothesis on the Location of

Individual Flamingos

In this study, we hypothesized that each individual

flamingo tends to stay in a certain location. By testing

this hypothesis, we examine the feasibility of

flamingo identification based on location information.

5.2 Method for Testing Hypotheses

Based on the location data set extracted in Chapter 4,

we will verify whether or not a particular individual

tends to stay within the location shown in the video.

We use neural network (NN) as a verification method.

We added “weather” and “time” to the location data

set to account for possible changes in individual

behavior due to weather and time period effects

(Table 3). “weather” is weather information observed

and published by Japan Meteorological Agency.

During the period under verification, “clear”, “sunny”,

“slightly cloudy”, “cloudy” and “rainy” were

observed. Since the weather information was

announced every three hours from 0:00, we entered

the same weather conditions for the three hours after

the observation time. “time” is the time when the

image was taken. We verify by changing these under

which the number of records vary.

From the items in Table 2, we select the objective

variable as “class” and the explanatory variables from

the other items. Among the explanatory variables, “x1”

and “x2” represent the horizontal position. We

removed “x2” because the correlation coefficient was

0.99, which is an extremely strong positive

correlation and multicollinearity was observed. y1”

and ‘y2’ represent the depth position or orientation.

Since their correlation coefficient was 0.26, we

judged that there was no multicollinearity and

decided to use both as explanatory variables. Since

“confidence” is an item unrelated to location

information, we removed it from the explanatory

variables.

The number of codes in the location data set used

for verification is 631,486. The objective variable

“class” is an integer from 0 to 3. The explanatory

variable “x1” is a number between 0 and 960. “y1” is

a number between 0 and 960. “y2” is a number greater

than 0 and less than or equal to 960.

We judge the results by the value of accuracy,

which represents the correctness rate of the

discriminator when using all the data in the location

data set, both training and validation data, and the

value of val_accuracy when using only the validation

data in the location data set. The number of training

epochs is 100.

5.3 Results of Hypothesis Testing

5.3.1 Results for All Data

The accuracy at the 100th epoch was approximately

0.53, and the val_accuracy was approximately 0.56.

To confirm that there is a significant difference

between the small value of 0.53 and the random

inference value of 0.25, we tested the null hypothesis

that “there is no significant difference between the

percentage of correct answers (0.53) and the random

inference value (0.25)” and tested the difference in

mother proportions. As a result, we were able to show

the correctness of the hypothesis because the p-value

was 0.000049220, a significant difference at a

significance level of less than 1%.

Table 3: Example of location data set plus information on shooting time and weather.

x1 y1 x2 y2 class confidence weathe

time

387.57 176.61 426.84 362.69 2 0.51 clea

1350

82.58 130.03 182.07 419.71 1 0.51 sunny 950

218.10 275.05 271.28 388.94 0 0.58 cloud

1630

204.60 290.04 270.67 391.59 3 0.65 rain

1150

ICSOFT 2025 - 20th International Conference on Software Technologies

442

5.3.2 Results with Data by Weather

We verified six patterns of weather. The results are

shown in Table 4. The results show that the highest

accuracy is approximately 0.60 for validation number

6, which indicates rainy weather, and the highest

val_accuracy is approximately 0.64 for validation

number 3, which indicates cloudy weather. In the

same way as for the NN test using all the data, we

confirmed that there was a significant difference

between the accuracy (0.54) for validation number 2,

which had the lowest value, and the random inference

value (0.25), by testing the difference in the mother

proportions. As a result of testing the null hypothesis

that “there is no significant difference between the

percentage of correct answers (0.54) and the random

inference value (0.25),” the p-value was 0.00002731,

showing that there was a significant difference at a

significance level of less than 1%.

5.3.3 Results with Data by Time Period

We verified three patterns of time period. The results

are shown in Table 5. The results show that the

highest accuracy is approximately 0.59 for validation

number 1, which indicates the time period from

morning to noon, and the highest val_accuracy is

approximately 0.63 for validation number 2, which

indicates the time period during the daytime. In the

same way as for the validation using all the data, a test

of difference in proportions was conducted to confirm

that there was a significant difference between the

accuracy (0.58) for validation numbers 2 and 3, which

were the lowest values among the results obtained,

and the random inference value (0.25). As a result of

testing the null hypothesis that “there is no significant

difference between the percentage of correct answers

(0.58) and the random inference value (0.25),” the p-

value was 0.000002182, showing that there was a

significant difference at a significance level of less

than 1%.

5.4 Analysis of Verification Results

Based on the results obtained in the previous section,

we inferred that under certain conditions, the location

at which individuals tend to remain and the position

at which they tend to take up positions may change.

Therefore, in this section, we analyze the tendency of

the position where individuals tend to stay, the

posture they tend to take, and the individuals that can

be identified under each condition. In the weather

verification, the percentage of correct answers was

higher in rainy or cloudy weather. In the time of day

verification, the percentage of correct answers was

higher during noon time period. Therefore, we

analyze the results for these conditions.

We describe the analysis method. First, we

divide the X-coordinate into 10 segments of 96px

each. Next, we measure the number of x1 in the

location data set that fall within the segmented area.

We analyze by finding the first most frequent region

(mode) and the second most frequent region (second

mode) for each individual under each condition, and

comparing them with the mode and second mode of

all the data. After dividing the Y-coordinate by 54px,

we find the mode and the second mode by the same

procedure as for the x-coordinate. Table 6 lists the

number of x1 for each condition of Rblue as an

example of data to support the analysis.

Table 4: Verification results considering weather conditions.

Verification numbe

Record count Weathe

accurac

val_accurac

1 43,638 clear, sunn

rox. 0.58 A

rox. 0.62

2 103,128 clear, sunn

, sli

htl

cloud

rox. 0.54 A

rox. 0.62

3 213,952 sli

htl

cloud

, cloud

rox. 0.59 A

rox. 0.64

4 383,615 slightly cloudy, cloudy, rain

Approx. 0.55 Approx. 0.57

5 324,125 cloudy, rain

Approx. 0.58 Approx. 0.60

6 169,663 rain

Approx. 0.60 Approx. 0.62

Table 5: Verification results considering time period.

Verification numbe

Record count Time perio

accurac

val_accurac

1 211,339 9:10 - 12:20 Approx. 0.59 Approx. 0.61

2 141,068 11:10 - 13:00 Approx. 0.58 Approx. 0.63

3 215,912 12:20 - 17:00 Approx. 0.58 Approx. 0.61

Method for Identiﬁcation of Individual Flamingos Based on Movement Logs

443

Table 6: Number of x1 in each condition for Rblue.

X-coordinate all cloudy rainy noon time period

0≦x1≦96

1,415 812 551 0

96＜x1≦192

4,540 703 2,895 1,233

192＜x1≦288

11,957 6,330 5,187 3,851

288＜x1≦384

11,701 4,955 6,717 5,262

384＜x1≦480

23,513 12,069 8,682 8,982

480＜x1≦576

6,648 939 3,084 3,447

576＜x1≦672

12,741 6,760 5,841 2,068

672＜x1≦768

12,010 2,958 7,465 3,859

5.4.1 Analisis of All Data

Figure 3 shows the distribution of the positions and

postures at which each flamingo tends to stay in all

the data.

Comparing the images of the distribution for all

the data with each condition, we describe the changes

in the position and posture that tend to stay under each

condition. If x1 which is the left edge of the body

changes, it indicates a change in the horizontal

position. If y1 which is the upper edge of the head or

body changes, it indicates a change in the posture that

is easy to take. If y2 which is the toe of the foot

changes, it indicates a change in the vertical position.

Figure 3: The location and posture of each flamingo in all

the data.

5.4.2 Analysis of Cloudy Weather

The left panel of Figure 4 is the distribution of the

most likely positions and postures of individuals for

all data. The right panel of Figure 4 shows them in

cloudy weather. 4 individuals are almost completely

uncovered, so all individuals can be identified when

it is cloudy.

“Rblue” changes easier posture to take. “Lblue”

changes easier posture to take. “none” changes

horizontal and vertical position. “Rblack” changes

easier posture to take.

Figure 4: Comparison of distribution between all data and

cloudy weather. Blue circles are Rblue, red triangles are

Lblue, green squares are None, yellow hexagons are Rblack,

1 decimal place is the mode and 2 is the second mode.

5.4.3 Analysis of Rainy Weather

The left panel of Figure 5 is the distribution of the

most likely positions and postures of individuals for

all data. The right panel of Figure 5 shows them in

rainy weather. Because the distribution of individuals

covers some areas, only some individuals (Rblue and

Rblack) can be identified when it is rainy.

“Rblue” changes horizontal and vertical positions

and easier posture to take. “Lblue” changes horizontal

and vertical positions and easier posture to take.

“none” changes the horizontal and vertical positions.

“Rblack” changes horizontal position and easier

posture to take.

Figure 5: Comparison of distribution between all data and

rainy weather. Blue circles are Rblue, red triangles are

Lblue, green squares are None, yellow hexagons are Rblack,

1 decimal place is the mode and 2 is the second mode.

5.4.4 Analysis in the Noon Time Period

The left panel of Figure 6 is the distribution of the

most likely positions and postures of individuals for

all data. The right panel of Figure 6 shows them

during noon time period. Because the distribution of

individuals covers some areas, only some individuals

(Rblue, Lblue and none) can be identified when using

the data from rainy conditions.

ICSOFT 2025 - 20th International Conference on Software Technologies

444

“Rblue” changes horizontal position. “Lblue”

changes horizontal position and easier posture to take.

“none” changes horizontal position. “Rblack”

changes horizontal and vertical positions and easier

posture to take.

Figure 6: Comparison of distribution between all data and

noon time period. Blue circles are Rblue, red triangles are

Lblue, green squares are None, yellow hexagons are Rblack,

1 decimal place is the mode and 2 is the second mode.

6 SUMMARY AND FUTURE

ISSUES

To show the feasibility of individual identification

based on behavioral features of animals, we verified

that individual identification is possible with

movement logs of flamingos. By creating a system to

collect location data set and testing hypotheses using

a neural network, we demonstrated the feasibility of

individual identification based on flamingo location

data. In addition, we were able to show that the

location where each individual tends to stay and the

posture it tends to take change according to each

condition. On the other hand, there is room for

improvement in this study because some data were

detected incorrectly when the location data set was

created by individual detection in YOLO, which may

have affected the accuracy of the data and affected

subsequent validation.

In this study, we identified individual flamingos

using two-dimensional images extracted from videos.

Although it is difficult to obtain three-dimensional

data in a complex natural environment, the method in

this study may be able to obtain data and identify

individuals in such a situation.

There is room for improvement in this study

because the existence of false positive data when

creating the position data set by individual detection

in YOLO may have affected the accuracy of the data,

which may have affected subsequent validation. In

addition, the number of individuals and flamingo

species studied were 4 and 2, respectively, limiting

the generalizability of the findings.

As future tasks, it is necessary to verify the

accuracy of the location data set, to show that other

individuals can be identified using the method in this

study, to show what postures they take under different

conditions, to show that it is possible to identify

individuals by movement for other animals, and to

show that it is possible to identify individual animals

by other behaviors. We will also increase the number

of individuals and flamingo species studied for a

more extensive and longitudinal validation.

ACKNOWLEDGEMENTS

This work was supported by JSPS KAKENHI Grant

Number 22K12112. We would like to express our

deepest gratitude to Yagiyama Zoological Park and

all those involved for their cooperation in obtaining

the necessary data and for their appropriate advice in

conducting this study.

REFERENCES

Ahmad, M., Ghazal, M. T., & Aziz, N. (2022). A survey on

Animal Identification Techniques Past and Present.

International Journal of Computational and Innovative

Sciences, 1(2), 27-32.

Burley, N., Krantzberg, G., & Radman, P. (1982). Influence

of colour banding on the conspecific preferences of

zebra finches. Animal Behaviour, 30, 444‒455.

Carminato A., Vascellari M., Marchioro W., Melchiotti E., &

Mutinelli F. (2011). Microchip-associated fibro-sarcoma

in a cat. Veterinary Dermatology, 22(6), 565-569

Karun K. R., Lars C. G., & Juan, P. M. P. (2021). Daniela

Alarcón-Ruales, Ricardo B. R. Azevedo. Sea Turtle

Facial Recognition Using Map Graphs of Scales. Cold

Spring Harbor Laboratory.

Lay, D. C. Jr., Randel, T. H., Bowers, C. L., Grissom, K.

K., & C., Jenkins, O. K. (1992). Behavioral and

physiological effects of freeze or hot-iron branding on

crossbred cattle. Journal of Animal Science, 70, 330-6.

Schofield, D., Nagrani, A., Zisserman, A., Hayashi, M.,

Matsuzawa, T., Biro D., & Carvalho, S. (2019).

Chimpanzee face recognition from videos in the wild

using deep learning. Science Advances, 5(9), eaaw0736.

Silvy, N. J., Lopez, R. R., & Peterson, J. M. (2012).

Techniques for marking wildlife. The wildlife techniques

manual, 230–257

Suzuki, Y., Osawa, A. (2021). Identifying Individual Cats by

Their Chewing Sounds Using Deep Learning. In

Proceedings of HCI International 2021, 556-560.

Walker, K.A., Mellish, J.E. & Weary, D.M. (2011). Effects

of hot-iron branding on heart rate, breathing rate and

behaviour of anaesthetised Steller sea lions. The

Veterinary Record, Vol.169(14), 363.

WOAH - World Organization for Animal Health Founded as

OIE. (2023, July 18). Terrestrial Code Online Access.

https://www.woah.org/en/what-we-do/ standards/codes-

and-manuals/

Method for Identiﬁcation of Individual Flamingos Based on Movement Logs

445