TY - JOUR

T1 - Detecting narwhal foraging behaviour from accelerometer and depth data using mixed-effects logistic regression

AU - Jensen, Frederik H.

AU - Tervo, Outi M.

AU - Heide-ørgensen, Mads Peter

AU - Ditlevsen, Susanne

N1 - Correction: https://doi.org/10.1186/s40317-023-00330-5

PY - 2023

Y1 - 2023

N2 - Background: Due to their Arctic habitat and elusive nature, little is known about the narwhal (Monodon monoceros) and its foraging behaviour. Understanding its ability to catch prey is essential for understanding its ecological role, but also to assess its ability to withstand climate changes and anthropogenic activities. Narwhals produce echolocation clicks and buzzing sounds as part of their foraging behaviour and these can be used as indicators of prey capture attempts. However, acoustic data are expensive to store on the tagging devices and require complicated post-processing. The main goal of this paper is to predict prey capture attempts directly from acceleration and depth data. The aim is to apply broadly used statistical models with interpretable parameters. The ultimate goal is to be able to estimate prey consumption without the more demanding acoustic data. Results: We predict narwhal buzzing activity using mixed-effects logistic regression models with 83 features extracted from acceleration and depth data as explanatory variables. The features encompass both instantaneous values as well as delayed values to capture behavioural patterns lasting several seconds. The data correlations were not strong enough to predict the exact timing of the buzzes, but were reliably able to detect buzzes within a few seconds. Most of the of the buzz predictions were within 2 s of an observed buzz (68%), increasing to 94% within 30 s. Conversely, 46% of the observed buzzes were within 2 s of a predicted buzz, increasing to 82% within 30 s. Additionally, the model performed well, although with a tendency towards underestimation of the number of buzzes per dive. In total, we predicted 17, 557 buzzes versus 25, 543 observed across data from 10 narwhals. Classifying foraging and non-foraging dives yielded a precision of 86% and a recall of 91%. Conclusion: We conclude that narwhal foraging estimation through acceleration and depth data is a valid alternative or supplement to buzz recordings, even when using somewhat simple statistical methods, such as logistic regression. The methods in this paper can be extended to foraging detection in similar marine species and can aid instrument development.

AB - Background: Due to their Arctic habitat and elusive nature, little is known about the narwhal (Monodon monoceros) and its foraging behaviour. Understanding its ability to catch prey is essential for understanding its ecological role, but also to assess its ability to withstand climate changes and anthropogenic activities. Narwhals produce echolocation clicks and buzzing sounds as part of their foraging behaviour and these can be used as indicators of prey capture attempts. However, acoustic data are expensive to store on the tagging devices and require complicated post-processing. The main goal of this paper is to predict prey capture attempts directly from acceleration and depth data. The aim is to apply broadly used statistical models with interpretable parameters. The ultimate goal is to be able to estimate prey consumption without the more demanding acoustic data. Results: We predict narwhal buzzing activity using mixed-effects logistic regression models with 83 features extracted from acceleration and depth data as explanatory variables. The features encompass both instantaneous values as well as delayed values to capture behavioural patterns lasting several seconds. The data correlations were not strong enough to predict the exact timing of the buzzes, but were reliably able to detect buzzes within a few seconds. Most of the of the buzz predictions were within 2 s of an observed buzz (68%), increasing to 94% within 30 s. Conversely, 46% of the observed buzzes were within 2 s of a predicted buzz, increasing to 82% within 30 s. Additionally, the model performed well, although with a tendency towards underestimation of the number of buzzes per dive. In total, we predicted 17, 557 buzzes versus 25, 543 observed across data from 10 narwhals. Classifying foraging and non-foraging dives yielded a precision of 86% and a recall of 91%. Conclusion: We conclude that narwhal foraging estimation through acceleration and depth data is a valid alternative or supplement to buzz recordings, even when using somewhat simple statistical methods, such as logistic regression. The methods in this paper can be extended to foraging detection in similar marine species and can aid instrument development.

KW - Accelerometer data

KW - Buzz

KW - Foraging

KW - Imbalanced data

KW - Jerks

KW - Logistic regression

KW - Mixed-effects model

KW - Narwhal

U2 - 10.1186/s40317-023-00325-2

DO - 10.1186/s40317-023-00325-2

M3 - Journal article

AN - SCOPUS:85151091203

VL - 11

JO - Animal Biotelemetry

JF - Animal Biotelemetry

SN - 2050-3385

IS - 1

M1 - 14

ER -