Development of a method for classifying and transmitting high-resolution feeding behavior of fish using an acceleration pinger

Abstract

Background: Monitoring the feeding behavior of animals in the wild is key to understanding their energetics and the influence of the environment on their survival. Recently, a novel acceleration transmitter that processes acceleration data onboard and outputs identification results has been developed by AquaSound Inc. (Kobe, Japan) to investigate feeding biology in fish. To date, few attempts have been made to identify the feeding behavior of fish using transmitters, and none of these attempts accomplished classification of alternative feeding behaviors according to prey items. The objective of this study was to develop an algorithm that can be incorporated in the acceleration transmitter and can identify alternative feeding behaviors in fish, using red-spotted grouper (Epinephelus akaara) as a model species. Results: Most of the identification algorithms describing feeding behavior in fish developed in previous studies used a combination of acceleration and angular velocity. In this study, we constructed an algorithm based on three-axis accelerometry data alone, since a gyroscope consumes much more electricity and would shorten the battery life of the transmitter. Acceleration data were obtained in tank experiments. Feeding behaviors, induced by feeding three types of live prey (Trachurus japonicus, Metapenaeus ensis and Hemigrapsus sanguineus), as well as other behaviors (routine and escape movements), were simultaneously recorded at 200 Hz by acceleration data loggers, implanted in the abdominal cavities of fish, and by a video camera. A decision tree, including a three-dimensional lookup table, was constructed to classify the behaviors into four behavior classes: shrimp-eating, fish-eating, crab-eating and other behaviors. The classification accuracy was estimated to be 0.77 (F-measure) for shrimp-eating, 0.73 for fish-eating, 0.71 for crab-eating and 0.78 for other movements, using fivefold cross-validation. Conclusions: The algorithm developed in this study could be incorporated into the transmitter, which would record acceleration data at high frequency (200 Hz), process the data onboard and output classification results of behaviors. This method would reveal more aspects of fish biology, such as individual feeding strategies.

Development and application of a machine learning algorithm for classification of elasmobranch behaviour from accelerometry data

“…It may be beneficial for future studies to use mandible accelerometers to address whether there is a correlation between prey size and headshaking duration or intensity and whether prey handling techniques vary with prey type to allow for more quantitative analysis. Accelerometers have been used to identify feeding according to prey type in the white-streaked grouper [( Epinephelus ongus ); Kawabata et al 2014 ] and red-spotted grouper [( Epinephelus akaara ); Horie et al 2017 ] differentiating between shrimp, fish and crab. Although teleosts form most of the juvenile lemon shark diet, they have also been documented with crustaceans, other elasmobranch species [whiptail stingrays (Dasyatidae)], molluscs and annelidas in their stomachs (Newman et al 2010 ).…”

A miniaturized threshold-triggered acceleration data-logger for recording burst movements of aquatic animals

“…Previous studies demonstrated that the acceleration data could be processed into activity level (Payne et al, 2016) or the occurrence of specific behaviors, such as jaw opening, flipper stroking and burrowing (Adachi et al, 2014;De Almeida et al, 2013;Naito et al, 2013). However, few on-board processing methods have been reported to determine the types of burst movements, possibly because of the low sampling frequencies of accelerometers and the limited number of established processing algorithms (but see De Almeida et al, 2013;Horie et al, 2017). In addition, such processing would limit the scope of analysis, as it cannot provide the various kinematic variables simultaneously.…”

“…Recently, Broell et al (2013) demonstrated that an accelerometer with a sampling frequency of at least 30 Hz (ideally, 100 Hz) is required to identify the escape responses and feeding strikes of the sit-and-wait predator Myoxocephalus polyacanthocephalus. Moreover, accelerometers with a sampling frequency of 200 Hz have been demonstrated as useful in distinguishing the feeding behavior of trophic generalist fish on different prey types (Horie et al, 2017;Kawabata et al, 2014). These studies clearly show that high-frequency accelerometers are useful in measuring the burst movements of agile animals; however, such high-frequency sampling rapidly consumes electricity and memory, which inevitably shortens the duration of the recording (e.g.…”

A miniaturized threshold-triggered acceleration data-logger for recording burst movements of aquatic animals

“…Previous studies demonstrated that the acceleration data could be processed into activity level (Payne et al, 2016) or occurrence of specific behaviors, such as opening jaws, stroking flippers, and borrowings (Adachi et al, 2014; De Almeida et al, 2013; Naito et al, 2013). However, few on-board processing methods have been reported to determine the types of burst movements, possibly due to the low sampling frequencies of accelerometers, and due to the limited numbers of established processing algorithms (but see (De Almeida et al, 2013; Horie et al, 2017). In addition, such processing would limit the scope of analysis, since it cannot provide the various kinematic variables simultaneously.…”

“…Recently, Broell et al (2013) demonstrated that an accelerometer with a sampling frequency of at least over 30 Hz (ideally, 100 Hz) is required to identify the escape responses and feeding strikes of the sit-and-wait predator Myoxocephalus polyacanthocephalus . Moreover, accelerometers with a sampling frequency of 200 Hz have been demonstrated as useful in distinguishing the feeding behavior of trophic generalist fish on different prey types (Horie et al, 2017; Kawabata et al, 2014). These studies clearly show that the high-frequency accelerometers are useful in measuring the burst movements of agile animals; however, such high frequency sampling rapidly consumes electricity and memory, which inevitably shortens the duration of the recording, or increases its size (e.g., 10 hours in 100 Hz and 135 mAh, ORI400-D3GT from Little Leonardo Co., Tokyo, Japan).…”