Understanding the Northern pike (Esox lucius) through accelerometer, is it possible?

Understanding the Northern

pike (Esox lucius) through

accelerometer, is it possible?

Author: Oskar Andersson

Sammanfattning

Accelerometer biologgers är ett verktyg för att identifiera beteenden hos djur. För att kunna använda accelerometern effektivt är det viktigt att träna en maskin–inlärning algoritm för att kunna identifiera olika beteenden utifrån accelerations data. Denna studie testar om accelerometer går att effektivt använda på gäddan (Esox lucious) för att kunna identifiera de olika beteendena för att senare kunna användas på en stor skalig studie. Fem gäddor användes i studien och varje gädda hade en accelerometer i sin kroppshålighet. Gäddorna var filmade i en pool och data användes för att träna en maskin-inlärnings algoritm för att identifiera åtta beteenden, simmar, simmar tillsammans med andra gäddor, hantering av byte,

inaktivitet, skrämd/flyr, lyckad attack, misslyckad attack och attack mot en annan gädda. Studien visar att det går att se skillnad mellan beteendena till en viss gräns. Inaktivitet, simmar och de olika attackerna kunde skiljas åt. Det gick inte att se skillnad mellan simmar och simmar tillsammans samt se skillnad mellan de olika attackerna. Flykt kan blandas ihop med simmar och attackerna. För att se skillnad mellan lyckad och misslyckad attack kan hantering av byte användas för att identifiera lyckad attack, dock funkar bara då bytet inte sväljs helt. Programmet behöver mer träning för att inte missta flykt och de olika attackerna med hantering av byte. Studien visar att accelerometer kan användas på en större skala, huvudsakligen för att se skillnad mellan inaktivitet, attacker och simmar.

Abstract

Acceleration biologgers are tools to measure activity in animals and to identify behavioural modes. To use this technology efficiently it is important to train a machine learning algorithm to identify behavioural modes from acceleration data. This study test this technology on the Northern pike (Esox

luscious) to determine if it is possible to measure pike behaviour in larger

field’s studies. To do this five Northern pike was caught and implanted with accelerometers. The pikes were filmed in a pool with the accelerometer in their body cavity and the data was used to train a machine learning algorithm to identify eight behaviours, swimming, swimming together with other pikes, attacking other pikes, fleeing/scared, successful attack, unsuccessful attack, prey handling and inactivity (sleeping, waiting etc.). Inactive, swimming and the three attacks can be differentiated. Swimming and swimming together could not as well for the three different attacks. To differentiate successful attack and unsuccessful attack prey handling can help to identify successful

Key words

AcceleRater, accelerometer, animal behaviour, biologging, movement ecology, Esox lucius.

Contents

Acceleration profiles across pike behaviour 10

Discussion 22

References 27

Appendix 29

Acknowledgment

I wish to express my gratitude to my supervisors Markus Zöttl and Oscar Nordahl for all the support and help with the experiment and with their constructive criticism I have learned a great deal about the Northern Pike and biologgers. I also want to thank Carl Tamario for lending me the cameras that were used to film the pikes and Per Larsson for assisting with the capture of the pike.

Introduction

The Baltic Sea is changing, species such as the Northern pike (Esox lucius) (also referred to as pike in this study) have decreased in population size while other species such as three-spine stickleback (Gasterosteus aculeatus) are benefitting (Nilsson, et al., 2019). This is due to overfishing, change in climate and habitat destruction (Sundblad et al., 2014; Österblom et al., 2007). The pike is a keystone piscivore in temperate regions meaning it can influence the dynamics of species composition through top-down control. The species is a large (<130 cm), long lived (>10 years), mainly solitary and iteroparous fish (reproduce more than once during its lifetime) that can be found in a broad range of aquatic habitats such as the Baltic Sea and freshwater systems (Craig, 2009; Forsman, et al., 2015). It is a sit-and-wait predator that selects prey depending on the size ratio between the pike and its prey, and is not affected by shoal size, meaning that they can pick out prey from a shoal (Craig, 2008). Usually they eat relative smaller fish and in the juvenile stage they also forage on invertebrates. Pike populations are characterised by intraspecific competition and can resort to cannibalism (Craig, 2008).

The Northern Pike – a model species

The Northern pike has developed into an important model species for investigating ecological and evolutionary questions (Forsman, et al., 2015). The amount of research on pike has increased substantially and there are several advantages that the pike present when it comes to scientific studies (Forsman, et al., 2015). For instance, their large body size allows implanting individuals with accelerometer biologgers that allow the monitoring of activity and the identification of general behavioural modes (Forsman, et al., 2015; Deurs, et al., 2017). It is also possible to identify growth rings on the otoliths which makes it possible to reconstruct the past growth, which allows quantification of growth trajectories and body size at the level of individuals. The otoliths and cleithra can be used to trace elements to determine place of origin, temporal variation in habitats use. (Forsman, et al., 2015; Larsson et al., 2015).

The Northern pike migrate between breeding/spawning grounds and foraging habitats with a consistent homing behaviour (Tibblin, et al., 2016). In the Baltic Sea the Northern pike consistently return to their natal stream to breed, they return several times over their lifespan (Larsson, et al., 2015; Tibblin, et al., 2016). The timing of arriving back to the natal stream is different for

subsequent and by experience the individual timing can be improved throughout of a pikes life (Tibblin, et al., 2016).

The oceans do usually not create a physical barrier that prevents gene flow between populations. Due to the homing behaviour, many populations of pike are separated from each other with limited or no gene flow between them. This gives potential for evolutionary studies which can be shown in many studies of pike (Nordahl, et al., 2019; Sunde, et al., 2020; Berggren, et al., 2019). One study investigates the gene flow between populations of pike in the Baltic Sea by using two methods, microsatellites and restriction site associated DNA sequencing (RADseq). This study shows that the gene flow between the populations is low and detects signs of selection associated with salinity and temperature (Sunde, et al., 2020). There is a strong genetic difference between populations of pike even if these population are

considered to be in a close proximity (Nordahl, et al., 2019). Another study show that there is differences in egg size between two sympatic

subpopulations of anadromous pike. The populations spawn in separate wetlands but migrate to the same forage habitat in the Baltic Sea. Between these populations females in one lay more and smaller eggs. The females of the other population lays fewer and larger eggs. This study also show that the eggs have different hatching success depending on the environments

(Berggren, et al., 2016). This means that the populations have different parental reproductive strategies.

After the fry hatches they grow in size as they forage near their spawning area and are eventually ready to migrate to foraging habitats such as the Baltic Sea. Here is where the unknown enters. Most studies are related to the spawning/breeding habitats where the pike is more readily captured and sampled and little is known about the behaviour during the marine phase of the pikes life stage. What happens outside the breeding season is mostly unknown except through some glimpses of indirect indication through the otoliths for instance. A more detailed approach to uncover what the pikes are doing in the Baltic Sea would be through an accelerometer.

Acceleration logger

Acceleration logger (ACC for short) are a type of biologger that record acceleration on a tri-axial (X-axis, Y-axis and Z-axis) space and the data can be used to identify general behavioural modes (table 1). The accelerometer sense movement in form of acceleration when the pike is moving and logs the acceleration of the pike on frequency up to 100 Hz. The accelerometers should be able to successfully store data at a frequency at 10 Hz (Deurs, et al., 2017). The accelerometer have an internal clock that can be used to determine time of event with a time lag of one second added per day. Similar studies with biologgers on fish have been made as it is becoming a more common practice for movement ecology. These studies focuses mostly

on behaviours such as moving around, inactive (sleeping, resting etc.) and foraging (Broell, et al., 2013; Brownscombe, et al., 2014; Deurs et al., 2017; Kawabata, et al., 2014). The foraging behaviour is usually categorised as successful forage, unsuccessful forage and what kind of prey that the fish is foraging. In this study eight behaviours were chosen (table 1). The general behaviours of successful attack, unsuccessful attack, prey handling,

swimming and inactive are behaviours usually studied in acceleration studies as mentioned before. It is also the most common behaviours and therefore the most interesting ones. During this study other behaviours such as attacking another pike, scared and swimming together was notices and used to evaluate if these behaviours are separated from other behaviours.

Swimming together and attack on pike is social behaviours and if successfully identified can be used in social studies on the pike.

Behaviour Description

Successful attack An attack on prey that end with the

prey caught and eaten.

Unsuccessful attack A failed attempt to catch a prey. Prey handling When the pike is preparing a caught

prey by shaking its head.

Swimming When the pike is swimming around in

the pool without the direct intention of foraging (where no attack is present).

Swimming together When two pikes are swimming

together in form of parallel, synchronised swimming.

Attack on pike When a pike attacks another pike. Scared When a pike is fleeing from a threat,

usually other pikes in this study.

Inactive When the pike is sleeping, resting or

waiting for prey.

In other studies where biologgers where tested on pike, the biologger was glued on top of the head of the pike (Deurs, et al., 2017) due to more accurate reading of the movement of the head in the purpose of prey

Table 1: The biomechanical behaviours that will be investigated and a description of each behaviour.

method, as it has been reasoned in other studies as well (Deurs, et al., 2017; Broell, et al., 2013). The problems that the pike present is that they tend to move around in the vegetation where an external accelerometer can

apprehend movement as well as become lost, therefore this study will use the biologger in the body cavity. A surgical procedure may take a toll on the pike due to physical recovery. On the other hand when the wound has healed the pike should be less affected by the internal tag than by an external tag, therefore an internal procedure should be more beneficial for the pike if the incision does not get infected.

Here we investigate the question whether accelerometer recorded data from implanted biologgers can be used to identify general behavioural modes in captive pikes. We implanted five individuals with miniaturised accelerometer loggers, technosmart, Italy, and filmed their behaviour in the laboratory for 12-14 days while the accelerometer loggers where recording acceleration patterns of different behaviours. Subsequently, I analyse the accelerometer patterns to determine how they differ for several behavioural modes (table 1) and we investigate the question whether machine learning techniques would be able to identify certain behavioural states automatically. This could possibly lead to largescale studies on pike in its natural habitat.

The machine learning program will use different models to label data as behaviours. The program will then evaluate the models performances by using three ratios accuracy, precision and recall (Resheff, et al., 2014). Accuracy gives a ratio of correctly predicted observation to the total number of observation, how often the model will make a correct identification for that behaviour. It reveals how many of the observations was identified as true positive (samples correctly identified as the behaviour) and true negative (samples correctly identified as not that behaviour) out of all observations (figure 13). Accuracy is dependent on the total number of samples put into the program. Precision gives a ratio between the number of correctly

identified observation (true positive) and number of observation mistaken for that behaviour (false positive) (figure 13). In other words how many of the observation identified as a behaviour was actually that behaviour. Recall is the ratio of correctly predicted observation (true positive) of a behaviour to the total number observation of that behaviour (figure 13).

Method

Study animals

We captured the five pikes used in this study at the bay of Nedra Sandby, at the east coast of Öland at 15 January year 2020 between the hours of 09.00-15.00 using gill nets. The caught pikes were put in a tank with around 150 litre water taken from the bay. Subsequently, they were transported to the

laboratory in Linnaeus University at Kalmar where they were put in a pool. The pike were acclimatized for two weeks before use. The water in the pool had a depth of half a meter and the water is taken directly from the Baltic Sea. The pool had a constant inflow and outflow of water. Two barriers divided the pool into three sections two sections for study and one for containing the remaining pike. The sections size was 1.7m × 3m and

contained synthesized seaweed to give some hiding place in the middle of the three rooms. The seaweed was made out of blue plastic rope that was cut into nine one meter long parts and then put in a square three by three with a distance of 30 centimetres between them, this was made for each room. In the pool there was bleak (Alburnus alburnu) present together with roach (Rutilus rutilus), enough to create a school. To trigger attacks live prey is needed as dead prey won’t suffice (Broell, et al., 2013). Camera was put up to record the pike. The two cameras used was IP Cameras made by

AVTECH and contained IR-cameras to record during night. The cameras recorded the entire time the pike were in the pool.

Accelerometer biologgers

The model of the accelerometer was AXY-4 which is small (9 × 15 × 4 mm), light-weight (0.7 gram), ultra-low power, 3-axis accelerometer with

temperature sensor. The accelerometer measures the acceleration in g (g force). The accelerometers were recharged to maximum battery at 2.23V and the settings was put to 50 Hz and a maximum recording of 4g. These setting was used as the memory was predicted to last around 2 weeks (14 days). Before the accelerometers was used, they were put in a flat, oblong silicone casket (4.0 × 1.4 cm) to prevent water damage and movement of the

accelerometer itself when inside the pike. The accelerometer was put in the same way in all the pikes, the incision was made into the abdominal cavity between the pelvic and pectoral fins. The accelerometers location in the cavity was in the ventral side behind the pelvic fin.

The fish was sedated by 20 ml of benzocaine in ethanol (5 mg benzocaine in 100 ml ethanol) diluted in 20 litre of water. The pike was put in a box of 20 litre water with the benzocaine to sedate it. The implantation did not start until the pike stopped responding to touch. It was put on a measuring board with a wet towel to keep it moist. The cut was made vertical to the scale and the accelerometer was pushed in together with a PIT-tag for easy

identification using PIT-scanner. After the cut was stitched together the pike was put in an isolated area to rehabilitate under observation until the effect of the sedation is gone. After the effect of the sedation was gone the pike was put in the pool to begin the study. When the memory of the accelerometer

Data recording

After approximately two weeks the pike was replaced with new ones due to the memory of the accelerometer was full. The first pike nr. 1 (table 3) showed that the memory of the accelerometer with the settings of 50 Hz will last for around 12 days. Therefore, the other set of pikes (pike nr. 2 and pike nr.3; table 3) was in the pool for 12 days instead of 14. After 6 days the barrier between the two pikes was opened to give possibility for the pikes to interact with each other. This gave 6 days of data when the pikes are alone and 6 days of data when the pikes are together. This reoccurred with two other pikes giving a total of five pikes used in the study. The last two pikes had a different setting on the accelerometer, 25 Hz instead of 50. As 25 Hz use less memory the last pair of pikes was in the pool for 14 days.

The recordings of the pike were watched through and all behaviours was written down with the information of what happened in detailed and what time. The notes of the events of the behaviour was then used to take out the behaviours from the raw data into samples.

Statistics

To analyse the raw data from the accelerometer it first needed to be

converted to csv file. All of the observed behaviours was then taken out into their own time-intervals (start of behaviour to the end) to be prepared for machine learning in the system AcceleRater. All pikes behaved differently and therefore the distribution of samples is different of each behaviour. For instance swimming every pike contributed with the same amount of samples. Only three pikes successfully caught a prey (table 17).

AcceleRater is a supervised machine learning algorithm, a program that analyse the intervals and determine the probability through different models to accurately identify that behaviour again (Resheff, et al., 2014). By making a template with examples of the different behaviours AcceleRater helps to determine if it is possible to use the template for accurate readings of other observations. AcceleRater uses several models of your own choosing such as nearest neighbours, linear SVM (Support Vector Machines), RBF SVM (Radial Basis Function Support Vector Machines), decision tree, random forest, naïve bayes, QDA (Quadratic Discriminant Analysis) and ANN (Artificial neural networks) (Resheff, et al., 2014; Nathan, et al., 2012). As the models are complex and works in very different ways the program (AcceleRater) makes and average percent correctness for each model to help evaluate the models performance in general. This makes it easy to decide which model to use as the model with the highest percent correctness performs best. The model with the highest percent correctness will be focused on as that model is most fit for analysing the behaviours of the pike.

For the model with highest percent correctness, a confusion table and a table that showcase the accuracy, recall and precision for each behaviour will be used to evaluate the result of the model. The confusion table display the chance of one behaviour being mistaken for another, that is to say that if the model received the acceleration profile of behaviour X how often will the algorithm label this profile as behaviour X or confuse it with other

behaviours.

In table 2 there is a list of the numbers of samples for each behaviour used to supervise the machine learning algorithm. All the observation put in the template needs to be at the same size (same amount of data points), yet the behaviour do not follow a consistent timeframe. Therefore the smaller

behaviours such as the different forms of attack and scared was lengthened to three seconds while the longer behaviours such as swimming, swimming together and inactivity was divided into many three seconds intervals.

Therefore swimming together contains 1199 samples taken from 20 different intervals of two pikes swimming together. This creates an asymmetric dataset where there will be a high amount of true negative compared to true positive for behaviours with fewer samples. As the number of samples is not the same for each behaviour the behaviours with fewer samples and for relatively rare behaviours, precision will be an important indicator capacity to identify the behaviour correctly. As this is the case precision and recall will be the parameters in focus as they do not use true negatives but accuracy will be important to assess the capacity of the algorithm to identify frequent behaviours.

Three of the accelerometer was set to 50 Hz and two others was set to 25 Hz and

therefore the three seconds interval did not contain the same amount of data points. To counter this problem the 50 Hz intervals was converted to 25 Hz by removing every other point to make the data readable for the AcceleRater program.

All forms of movements require energy to be performed. Thereby a proxy for metabolic rate is used in studies using biologgers. This proxy is called overall dynamic body

acceleration (ODBA). ODBA is a metric for predicting energy expenditure for animals. ODBA originates from a correlation between

Behaviour Number of samples Attack on pike 17 Inactive 500 Prey handling 25 Scared 23 Successful attack 8 Swim 500 Swim 1199

Table 2: The amount of samples of each behaviour used in the AcceleRater.

2014), it is a well-established parameter for acceleration studies. Therefore ODBA can be used to determine which behaviour cost more to use without needing to calculate the amount of energy directly as it is not in the interest of this study, what is interesting is the difference and that is what ODBA gives. In this study ODBA will be used as a characteristic feature to determine differences and similarities between the behaviours.

To calculate the ODBA the behaviours was isolated from the raw data. The intervals were from the start of the behaviour until the behaviour stopped, creating a raw data interval. The raw data intervals were then converting to absolute values for each axis. Absolute values give a more clear view on how the acceleration was changing over the event. The absolute values were then averaged on each axis to gain PDBA (Partial Dynamic Body Acceleration) which gave PDBAx, PDBAy and PDBAz for each sample, one PDBA for

each axis. ODBA is the total sum of PDBAx, PDBAy and PDBAz. As the

PDBA and ODBA is an average the length or the frequency of the behaviours are not an effecting factor.

ODBA, PDBAx, PDBAy and PDBAz was used in ANOVA (Tukey) tests to

help highlight direct differences between the behaviours in a pair-wise comparison of mean. ANOVA tests were also used to test differences and similarities of the maximum value of each axis for each samples and the differences in minimum values for each axis for each sample to determine differences and similarities between the behaviours. Maximum and minimum values are used as they can determine unique characteristics of the

behaviours.

The first test made was a two-way ANOVA on ODBA between behaviour and individual to determine if there is a difference in ODBA between behaviour, between individuals and if the ODBA of the behaviour are effected by the individuals. Two-way ANOVA was used as it can determine if there is differences between behaviours, individuals and to determine of individuals have an effect on the ODBA for the behaviours. ODBA is an overall measurement compared to the other characteristics features that are limited to one axis.

Ethic statement

The implantation of the accelerometer was performed by people educated in fish tagging procedures and animal welfare with extensive experience in tagging pike with similar methods. The experiment was conducted with a valid ethical permit (5.2.18-482/14).

Pike ID number

Gender Size[cm] Logger identification number Total time period Time period separated Time period together Couple Frequency Of the logger [Hz] 1 Female 58 1 05/02/2020 – 19/02/2020 NA NA NA 50 2 Male 45 4 19/02/2020 – 02/03/2020 19/02/2020 - 25/02/2020 25/02/2020 – 02/03/2020 1 50 3 Female 59 5 19/02/2020 – 02/03/2020 19/02/2020 - 25/02/2020 25/02/2020 – 02/03/2020 1 50 4 Male 51 3 05/03/2020 – 19/03/2020 05/03/2020 – 13/03/2020 12/03/2020 – 19/03/2020 2 25 5 Female 53 2 05/03/2020 – 19/03/2020 05/03/2020 – 13/03/2020 12/03/2020 – 19/03/2020 2 25 Table 3: Information about the individual pikes, with their sex, size, accelerator biologger, the total time period that they were used during the study, the time period when they were separated into their own room, the time period when they were together and which pikes was together. Pike nr. 1 was alone and therefore not separated or together with another pike.

Results

Acceleration profiles across pike behaviour

Figure 1: Representative acceleration plots for four of the eight behaviours. The behaviours were randomly chosen to give a representative visualisation. The plot show acceleration (g) over time (minutes respectively seconds) during the events of swimming, two pikes swimming together with another pike, inactivity and when scared by the other pike. Blue line represent acceleration surge, orange is sway and brown is heave.

When the pike swims around (top plot) it goes from one place to another and repeats this pattern. Therefore it accelerate, deaccelerate and repeat. This is the same for when they swim together. The third plot from the top the pike is not moving (inactive) and therefore no acceleration. In the bottom plot the pike is fleeing. At the beginning it is inactive for a few seconds as it is the calm before the storm. Then the pike gets scared and usually turns around and swim away quick. After it is out of danger it slows down.

Differences in characteristics of acceleration profiles across behaviours

A two-way ANOVA analysis tested the variance in ODBA between the individuals of pike and the variance in ODBA between the behaviours. There was a significant difference in ODBA between the individuals (Sum Sq = 0.48919, DF = 4, F-value = 11.309, P-value = <0.001) as well as a significant difference between behaviours (Sum Sq = 2.68907, DF = 7, F-value =

35.523, P-value = <0.001). There was a significant interaction between the terms individuals and the behaviours (Sum Sq = 0.63911, DF = 20, F-value =

Figure 2: Representative accelerations plots for four of the eight behaviours. The behaviours were randomly chosen to give a representative visualisation. The plot show the pikes acceleration (g) pattern over time (sec) for the events of a successful attack, an unsuccessful attack, an attack from an pike towards a pike and prey handling. Blue line represent acceleration in the surge axis, orange is sway and brown is heave.

The top plot contains a successful attack where the pike is sitting in waiting until it launches an attack towards a prey (the spike at 38 second mark). If the prey is not swallowed directly the prey is handled in the mouth of the pike and shaken. At the bottom plot prey handling is shown. Every spike represent a head shake. Unsuccessful attack is the same as successful attack but failed to catch the prey. The Third plot from the top visualise the third kind of attack, a pike attacking another pike.

2.955, P-value = <0.001) revealing that the variation between individuals have an effect on the variation in ODBA for the behaviours.

An analysis of variance (ANOVA) of ODBA found significant difference between some behaviours (Sum sq = 2.787, DF = 7, F-value 24.3, P-value = <0.001). A post hoc test revealed (Tukey, table 7 and figure 3) that inactive, prey handling, swimming and swimming together was not significant different from each other. The variance for these four behaviours was overlapping each other for the most part and overlapped parts of scared, successful attack and successful attack (figure 3). Scared, successful attack, unsuccessful attack and attack on pike had a higher ODBA than inactive, prey handling, swimming and swimming alone. Scared also had a significant higher ODBA than unsuccessful attack. Scared overlapped successful attack, unsuccessful attack and attack on pike completely.

Figure 3: Boxplot over the distribution of ODBA (g) for each behaviour. The behaviours have shorten for easy reading. AOP is attack on pike, IN is inactive, PH is prey handling, SA is

successful attack, SC is scared, SW is swimming, SWT is

swimming together and UA is unsuccessful attack. The significant labels bellow the behaviour describes which of the behaviours have a significant difference between them and which don’t.

An analysis of variance (ANOVA) of PDBAx found significant difference

between some behaviours (sum sq = 1.569, DF = 7, F-value = 32.01, P-value = <0.001). A post hoc test revealed (Tukey, table 8 and figure 4) that

inactive, prey handling, swimming, swimming together was not significant different from each other and was significant lower from the other

behaviours. Attack on pike was significant lower than successful attack yet showed no significant difference between scared and unsuccessful attack. Unsuccessful attack was significant different from all behaviours except for scared and successful attack. The overlapping variance between behaviours was quite large, all behaviours overlapped the other except for inactive and successful attack which did not overlap each other.

Figure 4: Boxplot over the distribution of PDBA (g) of the X-axis for each behaviour. The behaviours have shorten for easy reading. AOP is attack on pike, IN is inactive, PH is prey handling, SA is successful attack, SC is scared, SW is swimming, SWT is

swimming together and UA is unsuccessful attack. The significant labels bellow the behaviour describes which of the behaviours have a significant difference between them and which don’t.

An analysis of variance (ANOVA) of PDBAy found significant difference

Figure 5: Boxplot showing the distribution of PDBA (g) of the Y-axis for each behaviour. The behaviours have shorten for easy reading. AOP is attack on pike, IN is inactive, PH is prey

handling, SA is successful attack, SC is scared, SW is swimming, SWT is swimming together and UA is unsuccessful attack.

Figure 6: Boxplot showing the distribution of PDBA (g) of the Z-axis for each behaviour. The behaviours have shorten for easy reading. AOP is attack on pike, IN is inactive, PH is prey

handling, SA is successful attack, SC is scared, SW is swimming, SWT is swimming together and UA is unsuccessful attack.

between some behaviours (sum sq = 0.550, DF = 7, F-value = 2.621, P-value = 0.0138). A post hoc test revealed (Tukey, table 9 and boxplot figure 5) that scared had a significant higher value than swimming together and was not significant different from the other behaviours. Same goes for swimming together which was not significant from the other behaviours except for scared. All the other behaviours was not significant different from each other. All the behaviours overlapped each other to a great extent (figure 5). Prey handling was the only behaviour to completely overlap other behaviours (scared, successful attack and attack on pike).

An analysis of variance (ANOVA) of PDBAz found no significant difference

between the behaviours (sum sq = 0.0736, DF = 7, F-value = 0.766, P-value = 0.616). A post hoc test (Tukey, table 10, figure 6) was made to further evaluate the differences between the behaviours. All the behaviours overlapped each other (figure 6). Prey handling overlapped all the other behaviours, scared overlapped all the other behaviours except for prey handling, pike attack overlapped all the other behaviours except for prey handling and scared.

An analysis of variance (ANOVA) in maximum values of the X-axis found significant difference between the behaviours (Sum sq = 42.89, DF = 7, F-value = 65.82, P-F-value = <00.1). A post hoc test revealed (Tukey, table 11 and figure 7) that scared, successful attack, attack on pike and unsuccessful attack had a significant higher maximum than inactive, swimming,

swimming alone and prey handling. Within these four behaviours unsuccessful attack was lower than scared and successful attack but no significant difference from attack on pike. Prey handling had a significant higher maximum value than inactive and swimming together. Inactive show no significant difference to swimming together yet was significant lower than swimming. Swimming together and swimming was not significant different. Inactive hardly overlapped swimming and swimming together as the

maximum values was near zero. Swimming and swimming together

overlapped each other for most parts and parts of prey handling, scared and swimming also overlapped unsuccessful attack. Pike attack, prey handling, successful attack and unsuccessful attack partially overlapped each other. Scared overlapped most behaviours (except for inactive) where unsuccessful attack, successful attack and pike attack was overlapped completely.

An analysis of variance (ANOVA) in maximum values of the Y-axis found

Figure 7: Boxplot showing the distribution of maximum value (g) of the X-axis for each behaviour. The behaviours have shorten for easy reading. AOP is attack on pike, IN is inactive, PH is prey handling, SA is successful attack, SC is scared, SW is swimming, SWT is swimming together and UA is unsuccessful attack.

Figure 8: Boxplot showing the distribution of maximum value (g) of the Y-axis for each behaviour. The behaviours have shorten for easy reading. AOP is attack on pike, IN is inactive, PH is prey handling, SA is successful attack, SC is scared, SW is swimming, SWT is swimming together and UA is unsuccessful attack.

significant difference between the behaviours (sum sq= 34.30, DF = 7, F-value = 29.93, P-F-value = <00.1). A post hoc test revealed (Tukey, table 12 and figure 8) scared had the highest maximum value but was not significant different from successful attack and attack on pike. Attack on pike and successful attack was not significant different from each other as well as unsuccessful attack. Unsuccessful attack was not significant different from prey handling as well. Inactive had the lowest values together with

swimming and swimming together. Swimming and prey handling was not significant different from each other. Most of the behaviours overlap each other partly with the whiskers except for scared which completely overlaps successful attack and swimming that completely overlaps swimming together.

An analysis of variance (ANOVA) in maximum values of the Z-axis found significant difference between the behaviours (sum sq =, 14.62, DF = 7, F-value = 26.1, P-F-value = <00.1). A post hoc test revealed (Tukey, table 13 and figure 9) that inactive, swimming and swimming together had the lowest values and was not significant different from each other. Scared, attack on pike and successful attack had the highest values with no significant difference between them. Prey handling and unsuccessful attack was not significant different from each other as well as successful attack. All the behaviours partially overlapped each other except for scared that completely

Figure 9: Boxplot showing the distribution of maximum value (g) of the Z-axis for each behaviour. The behaviours have shorten for

overlapped prey handling, successful attack, pike attack and unsuccessful attack.

An analysis of variance (ANOVA) in minimum values of the X-axis found significant difference between the behaviours (sum sq = 0.02695, DF = 7, F-value = 4.382, P-F-value = 0.000183). A post hoc test revealed (table 14 and figure 10) that almost all the behaviours was not significant different from each other except for inactive that was significant different from the other behaviours except successful attack. For the minimum values for the X-axis there was many outliers (figure 10). Inactive completely overlaps all the other behaviours, successful attack overlaps all behaviours except for unsuccessful and inactive, unsuccessful attack overlap all the other

behaviours except for inactive. All the other behaviours overlaps each other.

An analysis of variance (ANOVA) in minimum values of the Y-axis found significant difference between the behaviours (sq =1.965, DF = 7, F-value= 13.04, P-value= <0.001). A post hoc test revealed (Tukey, table 15 and boxplot figure 11) that inactive was significant higher value than all the other behaviours except for swimming together. Prey handling, scared, swimming and swimming together was not significant different from each other. Attack

Figure 10: Boxplot showing the distribution of minimum value (g) of the X-axis for each behaviour. The behaviours have shorten for easy reading. AOP is attack on pike, IN is inactive, PH is prey handling, SA is successful attack, SC is scared, SW is swimming, SWT is swimming together and UA is unsuccessful attack.

on pike, prey handling, scared, successful attack and unsuccessful attack was not significant different from each other and was lower than the other

behaviours. Inactive completely overlaps all the other behaviours, prey handling overlaps all the other behaviours except for inactive, swimming together overlaps the other behaviours except for inactive and prey handling, swimming overlaps unsuccessful attack, attack on pike, scared and

successful attack completely. Scared overlaps attack on pike and successful attack, unsuccessful attack completely. Attack on pike, successful attack both overlap unsuccessful attack and each other completely.

An analysis of variance (ANOVA) in minimum values on the Z-axis found significant difference between the behaviours (sum sq = 6.479, DF = 7, F-value = 20.4, P-F-value = <0.001). A post hoc test revealed (Tukey, table 16 and boxplot figure 12) showed that inactive, swimming and swimming together was significantly higher than the other behaviours (except for prey handling that was not significant different from swimming and swimming together). Attack on pike, scared, successful attack and unsuccessful attack

Figure 11: Boxplot showing the distribution of minimum value (g) of the Y-axis for each behaviour. The behaviours have shorten for easy reading. AOP is attack on pike, IN is inactive, PH is prey handling, SA is successful attack, SC is scared, SW is swimming, SWT is swimming together and UA is unsuccessful attack.

handling, scared, successful attack and unsuccessful attack overlaps each other more than they overlap inactive, swimming and swimming together.

Machine learning approach to pike behaviour

The model with the highest percent correctness is random forest with 80.10 %. Followed by RBF SVM (78.85 %) and Decision Tree (77.85 %) (table 4). As the random forest model had the highest percent correctness that model will performing best to determine if it is possible to use accelerator on pike. To evaluate the models performance a confusion table (table 5) and accuracy, precision and recall table (table 6) was created.

According to the confusion table of the random forest model attack on pike has a 30 % chance for correctly being identified, 20 % chance for being mistaken for prey handing, 20 % for being mistaken for scared and 30 % chance for being mistaken for unsuccessful attack. Inactive has a chance of 99.6 % chance for being correctly identified and 0.4 % chance of being mistaken for swimming together. Prey handling has a 73.3 % chance for being correctly identified, 13.3 % chance for being mistaken for swimming and 6.7 % chance for being mistaken for swimming together and

unsuccessful attack. Scared has a chance of 46.2 % to be correctly identified, 23.1 % chance for being mistaken as attack on pike, 7.7 % on prey handling, 7.7 % on swimming and 15.4 % on unsuccessful attack. Successful attack has a 33.3 % chance of being correctly detected and 66.7 % chance for being

Figure 12: Boxplot showing the distribution of minimum value (g) of the Z-axis for each behaviour. The behaviours have shorten for easy reading. AOP is attack on pike, IN is inactive, PH is prey handling, SA is successful attack, SC is scared, SW is swimming, SWT is swimming together and UA is unsuccessful attack.

mistaken as an unsuccessful attack. Swimming has a 29.4 % chance of being correctly identified, 54.5 % chance for being mistaken as swimming together, 15.3 % chance for being mistaken for inactive and 0.41 % chance for being mistaken for prey handling and unsuccessful attack. Swimming together has a 94.5 % chance of being correctly identified and 4.1 % chance of being identified as inactive and 1.5 % chance to be identified as

swimming. Unsuccessful attack has a 100 % chance of being correctly identified. Attack on pike (99.13 %), inactive (94.46 %), prey handling (99.31 %), scared (99.22 %), successful attack (99.83 %) and unsuccessful attack (99.22 %) had a

relative high accuracy to swimming (83.82 %) and swimming together (85.21 %). Due to the data being asymmetrical (different amount of samples for each behaviour) the other parameters, precision and recall will be more accurate for evaluation of the model performance for rare behaviours. Recall reveals that 30 % of all attack on pike samples was identified correctly. Precision reveals that 50 % of all observation identified as attack on pike was correct. Inactive reveals that 99.57 % of all inactive samples was correctly identified. The precision reveals that 76.33 % of all observation identified as inactive was correct. Prey handling reveals that 73.33 % of all inactive samples was correctly identified. The precision reveals that 73.33% of all observation identified as Prey handling was correct. Scared reveals that 46.15% of all inactive samples was correctly identified. The precision reveals that 75.0 % of all observation identified as scared was correct. Successful attack reveals that 33.33 % of all inactive samples was correctly identified. The precision reveals that 100.0 % of all observation identified as successful was correct. Swimming reveals that 29.44% of all inactive samples was correctly identified. The precision reveals that 85.88% of all observation identified as swimming was correct. Swimming together reveals that 94.45 % of all inactive samples was correctly identified. The precision reveals that 80.57 % of all observation identified as swimming together was correct. Unsuccessful attack reveals that 100.0 % of all inactive samples was

Model name % Correct Nearest Neighbors 72.32 Linear SVM 75.43 RBF SVM 78.03 Decision Tree 77.85 Random Forest 80.10 Naive Bayes 57.44 LDA 71.11 QDA 0.00 ANN 65.66

Table 4: Overall performance of the models used to test the data from acceleRater. Its show the percent correct of each model and the standard

deviation.

Table 4: Overall performance of the models used to test the data from acceleRater. Its show the percent correct of each model and the standard

Discussion

To help understand the random forest model and its reading of the observations several ANOVA test were made to determine possible differences between behaviours to set them apart from each other. The two way ANOVA test reveals that the ODBA is different between individuals and different between behaviours. It does not classify where the differences is. It also reveals that there is a significant interaction between the individuals and behaviour, that the values of the behaviour are dependent on the individuals. The main difference between the pikes was in size (table 3). As the individuals of pike had differences such in sex and size the ODBA

In/Out ATTACK ON PIKE INACTIVE PREY HANDLING SCARED SUCCESSFULL ATTACK SWIM SWIM TOGETHER UNSUCCESS-FULL ATTACK ATTACK ON PIKE 30.0 0.0 20.0 20.0 0.0 0.0 0.0 30.0 INACTIVE 0.0 99.6 0.0 0.0 0.0 0.0 0.4 0.0 PREY HANDLING 0.0 0.0 73.3 0.0 0.0 13.3 6.7 6.7 SCARED 23.1 0.0 7.7 46.2 0.0 7.7 0.0 15.4 SUCCESSFULL ATTACK 0.0 0.0 0.0 0.0 33.3 0.0 0.0 66.7 SWIM 0.0 15.3 0.41 0.0 0.0 29.4 54.4 0.41 SWIM TOGETHER 0.0 4.1 0.0 0.0 0.0 1.5 94.5 0.0 UNSUCCESS-FULL ATTACK 0.0 0.0 0.0 0.0 0.0 0.0 0.0 100.0 Attack on pike Inactive Prey handling Scared Successful Attack Swim Swim Together Unsuccessful Attack Recall 30.00 99.57 73.33 46.15 33.33 29.44 94.45 100.00 Precision 50.00 78.57 73.33 75.00 100.00 85.88 80.87 70.97 Accuracy 99.13 94.46 99.31 99.22 99.83 83.82 85.21 99.22

Table 6: The recall, precision and accuracy of the random forest model for each behaviour.

Table 5: Confusion table of the random forest model. This table show how a behaviour (on the left column) might be mistaken by another (first row) and the chance of that mistake to happen.

should be different between individuals. How this effects the machine learning algorithm should not be a problem as the size and gender are dispersed between the samples. Therefore the samples cover a range of possible patterns for the behaviours. That the program learns how the variance in one behaviour looks like.

For the variance in PDBAx it follows a similar pattern to the variance of

ODBA where swimming, swimming together, inactive and prey handling is significant lower than the other behaviours. This is not the same for PDBAy

and PDBAz. In the X-axis there should be more differences as it is

acceleration surge. For the Y-axis and Z-axis (heave and sway) the acceleration can be very similar. The low difference in heave can be explained by the limiting space of half a meter which gives little room of movement in that axis. The sway was observed to be low for each behaviour due to the pikes physiology, it is hard to swim sideways. As the PDBA is an average there should be little variation in this axis. If compared to the other features such as minimum and maximum this is not the case as it is not an average and therefore gives a more similar evaluation to that of ODBA and PDBAx.

The model with the highest percent correctness was the model random forest with 80.10 %. This means that random forest is the model that should be the best model to identify behaviours from each other in general. The random forest model works by sets of classification trees that use a form of

stochasticity which randomly chooses which variables to use to determine similarities a differences (Nathan, et al., 2012). Why the random forest model has the highest percent correctness is hard to determine. The models can be better than the other models in different scenarios and in this scenario the random forest model is the better one.

The behaviours of attack on pike, prey handling, scared, successful attack and unsuccessful attack have relative few samples compared to inactive, swimming and swimming together (table 2). This makes the accuracy for the behaviour with fewer samples high (above 90 %) compared to the behaviours with many samples (between 83 % - 84 %). This makes the parameters of the accuracy only a partly useful for evaluating model performance and therefore precision and recall helps evaluate the performance of the model more

effectively.

Successful attack had a very high precision at 100 % and a recall at 33.33 % (table 6). The precision reveals that no other behaviour was mistaken for successful attack. Two thirds of successful attack were mistaken for unsuccessful attack which creates a problem as these two need to be set

can’t be set apart from each other it will be impossible to evaluate the foraging success of the pike with the biologgers. The low precision of unsuccessful attack is mostly due to successful attack being mistaken for unsuccessful attack, 66.7 % of the successful attack to be precise. The only ANOVA test that had a significant difference between these two behaviours was the maximum value of the X-axis and even here the values overlaps (figure 7). As successful attack and unsuccessful attack is the same to the model they can be categorized into the same behaviour “attack on prey”, to determine if an attack on prey was successful prey handling can be the solution.

Prey handling is important to be correctly identify as it can be used as an indirect way to determine if an attack on prey is successful or not. This is also reasoned in another study made on pike written by Deurs, et al (2017), where they came to the conclusion there was no visible difference between successful attack and unsuccessful attack on small prey unless prey handling was spotted.

Prey handling has a precision and recall at 73.33 % (table 6). This means almost a quarter of the behaviour is mistaken for other behaviours. Prey handling was mostly mistaken for swimming (13.3 %) (table 5). According to the ANOVA tests these two behaviours were different in PDBAx, maxx,

miny and minz (figure 4, 7, 9 and 12) where prey handling is significant

higher in maxx and maxz. When shaking its head the maximum acceleration

should be higher than when swimming around. That there is no difference between them in maxy is due to the headshaking is not in the heave axis.

Swimming is moving more regular and in all three axis whereas prey

handling is short burst of movement on place where the pike shakes its head. Therefore more samples of both swimming and prey handling should do the trick to increase the precision of prey handling. One problem with prey handling is that the pike can swallow the prey directly without handling the prey first, a successful attack that cannot be identified by prey handling. However, we had few samples of prey handling (N=25) and an increased sample collection would be likely to improve the precision of the machine learning algorithm.

Swimming has the lowest accuracy (83.82 %) of the behaviours and the lowest recall (29.4 %). The precision for the behaviour is relative high (at around 86.0 %) compared to the other behaviours (table 6). Meaning that there is a low chance for other behaviours to be mistaken for swimming. As the precision is high swimming should be unique in its values and patterns. This is not the case as across the ANOVA tests swimming is usually not different from swimming together and inactive which is the two behaviours swimming is mistaken for (figure 3 to 12). Swimming is mainly mistaken for swimming together at a rate of 54.4 % followed by inactive at 15.3 %. Swimming together has a relative low accuracy (85.21 %) as well relative to

the other behaviours. Its precision and recall is very high at 80.87 %

respectively 94.45 %. The relative high precision together with the pattern of the ANOVA test indicates that these two behaviours are hard to distinguish from each other but easily set apart from the other behaviours. As there is no distinguishable difference between swimming and swimming together swimming together can be considered to be swimming. This explains why swimming together has such a high recall compared to swimming and why swimming was mistaken for it. Due to swimming and swimming together was not different from each other the program responded by identifying most of the swimming and swimming together samples as swimming together. As seen in figure 1 inactive follow a constant pattern of barely no

acceleration. Therefore the behaviour should be easily distinguishable from the other behaviours. Inactive has a high recall (99.6 %), a high precision (78.57 %) and a high accuracy (94.46 %). The accuracy of inactive should be very high as the model can easily identify inactivity. As discussed before inactive has many similarities to swimming and swimming together according to the ANOVA tests. The low precision can be explained by swimming and swimming together may be mistaken for inactive. This is not a major problem as swimming and swimming together can contain small time-periods of no acceleration as the pike tends to move slowly when exploring its surroundings it is possible for it to stand still for a couple of seconds. Therefore the program might actually correctly classify parts of the behaviour. This had to be due to the swimming, inactive and swimming together samples needing to be broken down to three seconds intervals to fit the reading capability of the AcceleRater.

The last two behaviours not discussed yet are scared and attack on pike. Attack on pike has the lowest precision (50 %) out of all the behaviours and the second lowest recall (30 %) second to swimming. Attack on pike can be falsely identified as prey handling (20 %), scared (20 %) and unsuccessful attack (30 %). The ANOVA tests show that for the most parts of these behaviours overlap each other (figure 3 to 12). What prey handling and attack on pike has in common is that both behaviours should have low acceleration in the Y-axis. When handling prey it’s usually on the spot and therefore not moving up. When a pike attacks another pike they are usually on the same level and therefore the attack on another pike should have less acceleration in the heave axis compared to attacks on prey. This is due to prey tends to swim higher up in the pool.

Scared has a high accuracy (99.22 %) a relative high precision (75.0 %) and a low recall (49.15 %). The only behaviour to be mistaken for scared is

3-12). Scared is when the pike is fleeing from a threat (mostly other pikes in this study), after they have fled the pike slows down to regular swimming speed and eventually stops. Scared is defined between moment of scared to moment of still or close to not moving. This means that a part of scared could be identified as swimming, which at a certain point is actually correct.

Therefore the mix-up with swimming is not a major problem. Whereas the problem with scared and attack on pike for that matter is the mix-up with prey handling and unsuccessful attack. When falsely identified as an attack on prey or prey handling it can throw of the ratio between successful attack and unsuccessful attack when tested on a larger scale. Both scared and attack on pike show many similarities to the attacks on prey (figure 3-12) and might be hard to see difference from in general. When the pike was fleeing it was due to the presence of another pike. To develop the reading of a pike fleeing other methods could be tried to frighten it.

To summarize the random forest model, the model is capable at this point to determine when the pike is inactive, when it’s swimming, when it is

attacking and when it’s doing prey handling. As the prey handling could be falsely mistaken for other behaviour in smaller numbers more samples of prey handling and the behaviours that was mistaken for (scared, and the three different attacks) it can help to improve the readings of the model.

The pike has of course more than eight behaviours. Other behaviours that can be looked at is spawning behaviour, when they lay eggs and fertilize them. In other studies (Deurs, et al. 2017; Kawabata, et al. 2014) foraging is broken down to not only the success but also what kind of prey that has been forage. In one study on Ephinephelus ongus it was possible to identify when the fish was foraging on fish and when it was foraging on crabs (Kawabata, et al. 2014). Is it possible to separate successful attack on a small prey from successful attack on large prey when it comes to pike and is it possible to differentiate prey?

In a larger scale study with the biologgers it should be able to determine what the pike does out in the Baltic Sea. It is possible to find out how much

inactivity and swimming the pike does as well on its attack patterns. For instance if it attacks more often at day or night. The information given should help understand other behaviours the biologger cannot directly read. For instance individuals of pike from the same population have differences in homing behaviour (Tibblin, et al., 2016) as mentioned in the introduction. The large scale study could help answering question like: is there a correlation between foraging success-rate and the timing of homing behaviour? It could also show direct differences in behaviour between populations. In the study written by Berggren, et al. (2016) where two population that spawned in different areas yet forage in the same had differences in egg size due to different spawning habitats. The biologger could give more information about how the population differ in foraging

behaviour and any differences in inactivity and swimming. Social interaction between pikes where not able to differentiate from other behaviours and therefore social studies on pike via accelerometer can become difficult. Conclusion

The Northern pike has developed into an important model species for investigating ecological and evolutionary questions (Forsman, et al., 2015), therefore it is important to understand what the pike does in areas hard to directly monitor as it is a major part of its life-cycle. Most of the behaviours have few samples (only 8 to 40) and with those samples the accuracy is above 80 % and precision above 70 % (except for attack on pike). This indicates that the model can differentiate the behaviours from each other effectively (above 50 %) for most of the behaviours. The method appears promising, that with more practice (more samples) it should be possible to eliminate the smaller false positives and false negatives. By using

accelerometer on a larger scale it should be a helpful tool to understand the Northern Pike.

References

Berggren, H., et al., 2016. Testing for local adaptation to spawning habitat in sympatric subpopulations of northern pike by reciprocal translocation of embryos. Plos One, 11(5), pp.e0154488.

Broell, F., et al., C., 2013. Accelerometer tags: detecting and identifying activities in fish and the effect of sampling frequency. Journal of

Experimental Biology, 216(7), pp.1255–1264.

Brownscombe, J. W., et al. (2014). Foraging behaviour and activity of a marine benthivorous fish estimated using tri-axial accelerometer biologgers. Marine Ecology Progress Series, 505, pp. 241 – 251.

Craig, J., 2008. A short review of pike ecology. Hydrobiologia, 601(1), pp.5– 16.

Deurs, M., et al., 2017. Using accelerometry to quantify prey attack and handling behaviours in piscivorous pike Esox lucius. Journal of Fish Biology, 90(6), pp.2462–2469.

Forsman, A. et al., 2015. Pike Esox lucius as an emerging model organism for studies in ecology and evolutionary biology: a review. Journal of Fish Biology, 87(2), pp.472–479.

Larsson, P et al., 2015. Ecology, evolution, and management strategies of northern pike populations in the Baltic Sea. Ambio, 44(Supplement 3), pp.S451–S461.

Nathan, Ran et al., 2012. Using tri-axial acceleration data to identify behavioral modes of free-ranging animals: general concepts and tools illustrated for griffon vultures. Journal of Experimental Biology, 215(6), pp.986–996.

Nordahl, O et al., 2019. Genetic differentiation between and within ecotypes of pike (Esox lucius) in the Baltic Sea. Aquatic Conservation: Marine and Freshwater Ecosystems, 29(11), pp.1923–1935.

Nilsson, J., Flink, H. & Tibblin, P., 2019. Predator–prey role reversal may impair the recovery of declining pike populations. Journal of Animal Ecology, 88(6), pp.927–939.

Resheff, YS., et al., 2014. AcceleRater: a web application for supervised learning of behavioral modes from acceleration measurements. Movement Ecology, 2(1), p.27.

Sundblad, G. et al., 2014. Nursery habitat availability limits adult stock sizes of predatory coastal fish. ICES Journal of Marine Science, 71(3), pp.672– 680.

Sunde, J. et al., 2020. Comparing the Performance of Microsatellites and RADseq in Population Genetic Studies: Analysis of Data for Pike (Esox lucius) and a Synthesis of Previous Studies. Frontiers in Genetics, 11, p.218. Tibblin, P. et al., 2016. Causes and consequences of repeatability, flexibility and individual fine‐tuning of migratory timing in pike. Journal of Animal Ecology, 85(1), pp.136–145.

Wilson, RP. et al., 2006. Moving towards acceleration for estimates of activity‐specific metabolic rate in free‐living animals: the case of the cormorant. Journal of Animal Ecology, 75(5), pp.1081–1090.

Österblom, H. et al., 2007. Human-induced Trophic Cascades and Ecological Regime Shifts in the Baltic Sea. Ecosystems, 10(6), pp.877–889.

Appendix

Appendix 1: Accuracy, precision and recall equations.

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =

𝑅𝑒𝑐𝑎𝑙𝑙 =

Actual class Positive True positive (TP) False negatives (FN)

Negative False positive (FP) True negative (TN)

Appendix 2: data tables

Compared pair Estimate Std.

error

P-value

ATTACK ON PIKE - INACTIVE 0.26 0.038 < 0.001

PREY HANDLING – INACTIVE 0.071 0.036 0.500

SCARED – INACTIVE 0.36 0.038 < 0.001

SUCCESSFUL ATTACK – INACTIVE

0.31 0.050 < 0.001

SWIM – INACTIVE 0.050 0.036 0.86

SWIM TOGETHER – INACTIVE 0.0041 0.038 1.00000

UNSUCCESSFUL ATTACK – INACTIVE

0.21 0.038 < 0.001

PREY HANDLING - PIKE -0.19 0.038 < 0.001 Table 7: The estimate, std. error and P-value of pairwise comparison of means between behaviours for ODBA (post hoc Tukey test).

Table 7: The estimate, std. error and P-value of pairwise comparison of means between behaviours for ODBA (post hoc Tukey test).

Figure 13: The three equation for the parameters of accuracy, precision and recall together with the possible outcomes of the program: true positive (TP), false negative (FN), false positive (FP) and true negative (TN).

SWIM – ATTACK ON PIKE -0.21 0.038 < 0.001

SWIM TOGETHER - PIKE ATTACK

-0.25 0.040 < 0.001

UNSUCCESSFUL ATTACK – ATTACK ON PIKE

-0.048 0.040 0.93

SCARED - PREY HANDLING 0.29 0.038 < 0.001

SUCCESSFUL ATTACK - PREY HANDLING

0.24 0.050 < 0.001

SWIM - PREY HANDLING -0.021 0.036 0.999

SWIM TOGETHER - PREY HANDLING -0.067 0.038 0.65 UNSUCCESSFUL ATTACK - PREY HANDLING 0.14 0.039 0.011 SUCCESSFUL ATTACK - SCARED -0.045 0.051 0.99 SWIM - SCARED -0.31 0.038 < 0.001

SWIM TOGETHER - SCARED -0.36 0.040 < 0.001

UNSUCCESSFUL ATTACK - SCARED

-0.15 0.040 0.0058

SWIM - SUCCESSFUL ATTACK -0.26 0.050 < 0.001

SWIM TOGETHER - SUCCESSFUL ATTACK -0.31 0.051 < 0.001 UNSUCCESSFUL ATTACK - SUCCESSFUL ATTACK -0.11 0.051 0.43

SWIM TOGETHER - SWIM -0.046 0.038 0.93

UNSUCCESSFUL ATTACK - SWIM 0.16 0.038 0.0014 UNSUCCESSFUL ATTACK - SWIM TOGETHER 0.20 0.040 < 0.001

Compared pair Estimate Std. error p-value

ATTACK ON PIKE - INACTIVE 0.18 0.025 < 0.001 PREY HANDLING – INACTIVE 0.053 0.024 0.34 SCARED – INACTIVE 0.22 0.025102 < 0.001 SUCCESSFUL ATTACK – INACTIVE 0.32 0.033 < 0.001 SWIM – INACTIVE 0.060 0.024 0.19

Table 8: The estimate, std. error and P-value of pair-wise comparison of means between behaviours for PDBA from the X-axis (post hoc Tukey test).

SWIM TOGETHER – INACTIVE 0.033 0.025 0.89 UNSUCCESSFUL ATTACK – INACTIVE 0.22 0.025 < 0.001 PREY HANDLING - PIKE ATTACK -0.13 0.025 < 0.001 SCARED – ATTACK ON PIKE 0.038 0.026 0.84 SUCCESSFUL ATTACK – ATTACK ON PIKE 0.14 0.034 0.0020 SWIM – ATTACK ON PIKE -0.13 0.025 < 0.001 SWIM TOGETHER - PIKE ATTACK -0.15 0.026 < 0.001 UNSUCCESSFUL ATTACK – ATTACK ON PIKE 0.034 0.026 0.90 SCARED - PREY HANDLING 0.17 0.025 < 0.001 SUCCESSFUL ATTACK - PREY HANDLING 0.27 0.033 < 0.001 SWIM - PREY HANDLING 0.0072 0.024 1.0 SWIM TOGETHER - PREY HANDLING -0.020 0.025 1.0 UNSUCCESSFUL ATTACK - PREY HANDLING 0.17 0.025 < 0.001 SUCCESSFUL ATTACK - SCARED 0.098 0.034 0.072 SWIM - SCARED -0.16 0.025 < 0.001 SWIM TOGETHER - SCARED -0.19 0.026 < 0.001 UNSUCCESSFUL ATTACK - SCARED -0.0035 0.026 1.00

SWIM TOGETHER - SUCCESSFUL ATTACK -0.29 0.034 < 0.001 UNSUCCESSFUL ATTACK - SUCCESSFUL ATTACK -0.10 0.034 0.055 SWIM TOGETHER - SWIM -0.027 0.025 0.96 UNSUCCESSFUL ATTACK - SWIM 0.16 0.025 < 0.001 UNSUCCESSFUL ATTACK - SWIM TOGETHER 0.19 0.026 < 0.001

Compare pair Estimate Std. error p-value

ATTACK ON PIKE - INACTIVE 0.12 0.052 0.31 PREY HANDLING – INACTIVE 0.0011 0.049 1.0 SCARED – INACTIVE 0.14 0.052 0.14 SUCCESSFUL ATTACK – INACTIVE 0.044 0.067 1.0 SWIM – INACTIVE 0.0060 0.049 1.0 SWIM TOGETHER – INACTIVE -0.038 0.052 1.0 UNSUCCESSFUL ATTACK – INACTIVE 0.0083 0.052 1.00 PREY HANDLING - PIKE ATTACK -0.12 0.052 0.32

Table 9: The estimate, std. error and P-value of pair-wise comparison of means between behaviours for PDBA from the Y-axis (post hoc Tukey test).

SCARED – ATTACK ON PIKE 0.020 0.055 1.0 SUCCESSFUL ATTACK – ATTACK ON PIKE -0.074 0.069 0.96 SWIM – ATTACK ON PIKE -0.11 0.052 0.38 SWIM TOGETHER - PIKE ATTACK -0.16 0.055 0.089 UNSUCCESSFUL ATTACK – ATTACK ON PIKE -0.11 0.055 0.48 SCARED - PREY HANDLING 0.14 0.052 0.15 SUCCESSFUL ATTACK - PREY HANDLING 0.043 0.067 1.0 SWIM - PREY HANDLING 0.0049 0.049 1.0 SWIM TOGETHER - PREY HANDLING -0.039 0.052 1.0 UNSUCCESSFUL ATTACK - PREY HANDLING 0.0073 0.052 1.0 SUCCESSFUL ATTACK - SCARED -0.095 0.069 0.87 SWIM - SCARED -0.13 0.052 0.18 SWIM TOGETHER - SCARED -0.18 0.055 0.033 UNSUCCESSFUL ATTACK - SCARED -0.13 0.055 0.26 SWIM - SUCCESSFUL ATTACK -0.038 0.067 1.0 SWIM TOGETHER - SUCCESSFUL ATTACK -0.082 0.069 0.94 UNSUCCESSFUL ATTACK - SUCCESSFUL ATTACK -0.035 0.069 1.0 SWIM TOGETHER - -0.044 0.052 0.99

UNSUCCESSFUL ATTACK - SWIM TOGETHER

0.046 0.055 0.99

Compared pair Estimate Std. error p-value

ATTACK ON PIKE - INACTIVE -0.046 0.035 0.89 PREY HANDLING – INACTIVE 0.017 0.033 1.0 SCARED – INACTIVE -0.00088 0.035 1.0 SUCCESSFUL ATTACK – INACTIVE -0.050 0.046 0.96 SWIM – INACTIVE -0.016 0.033 1.0 SWIM TOGETHER – INACTIVE 0.0089 0.035 1.0 UNSUCCESSFUL ATTACK – INACTIVE -0.019 0.035 1.0 PREY HANDLING - PIKE ATTACK 0.064 0.035 0.60 SCARED – ATTACK ON PIKE 0.046 0.037 0.92 SUCCESSFUL ATTACK – ATTACK ON PIKE -0.0036 0.047 1.0 SWIM – ATTACK ON PIKE 0.031 0.035 0.99 SWIM TOGETHER - PIKE ATTACK 0.055 0.037 0.81 UNSUCCESSFUL ATTACK – 0.027 0.037 1.0

Table 10: The estimate, std. error and P-value of pair-wise comparison of means between behaviours for PDBA from the Z-axis (post hoc Tukey test).

ATTACK ON PIKE SCARED - PREY HANDLING -0.018 0.035 1.0 SUCCESSFUL ATTACK - PREY HANDLING -0.068 0.046 0.81 SWIM - PREY HANDLING -0.033 0.033 0.97 SWIM TOGETHER - PREY HANDLING -0.0085 0.035 1.0 UNSUCCESSFUL ATTACK - PREY HANDLING -0.037 0.035 1.0 SUCCESSFUL ATTACK - SCARED -0.049 0.047 1.0 SWIM - SCARED -0.015 0.035 1.0 SWIM TOGETHER - SCARED 0.0098 0.037 1.0 UNSUCCESSFUL ATTACK - SCARED -0.018 0.037 1.0 SWIM - SUCCESSFUL ATTACK 0.034 0.046 1.0 SWIM TOGETHER - SUCCESSFUL ATTACK 0.059 0.047 0.9 UNSUCCESSFUL ATTACK - SUCCESSFUL ATTACK 0.031 0.047 1.0 SWIM TOGETHER - SWIM 0.025 0.035 1.0 UNSUCCESSFUL ATTACK - SWIM -0.0036 0.035 1.0 UNSUCCESSFUL ATTACK - SWIM TOGETHER -0.028 0.037 1.0

Compared pair Estimate Std. error p-value ATTACK ON PIKE - INACTIVE 1.1 0.092 <0.001 PREY HANDLING – INACTIVE 0.53 0.086 <0.001 SCARED – INACTIVE 1.4 0.092 <0.001 SUCCESSFUL ATTACK – INACTIVE 1.4 0.12 <0.001 SWIM – INACTIVE 0.29 0.086 0.023 SWIM TOGETHER – INACTIVE 0.14 0.092 0.78 UNSUCCESSFUL ATTACK – INACTIVE 1.0 0.092 <0.001 PREY HANDLING - PIKE ATTACK -0.57 0.092 <0.001 SCARED – ATTACK ON PIKE 0.32 0.097 0.022 SUCCESSFUL ATTACK – ATTACK ON PIKE 0.32 0.12 0.15 SWIM – ATTACK ON PIKE -0.81 0.092 <0.001 SWIM TOGETHER - PIKE ATTACK -0.95 0.096 <0.001 UNSUCCESSFUL ATTACK – ATTACK ON PIKE -0.088 0.096 0.98 SCARED - PREY HANDLING 0.89 0.092 <0.001 SUCCESSFUL ATTACK - PREY HANDLING 0.89 0.12 <0.001 SWIM - PREY HANDLING -0.24 0.086 0.10

Table 11: The estimate, std. error and P-value of pair-wise comparison of means between behaviours for maximum value from the X-axis (post hoc Tukey test).