Accelerometer sampling requirements for animal behaviour classification and estimation of energy expenditure

Abstract

Background Biologgers have contributed greatly to studies of animal movement, behaviours and physiology. Accelerometers, among the various on-board sensors of biologgers, have mainly been used for animal behaviour classification and energy expenditure estimation. However, a general principle for the combined sampling duration and frequency for different taxa is lacking. In this study, we evaluated whether Nyquist–Shannon sampling theorem applies to accelerometer-based classification of animal behaviour and energy expenditure approximation. To evaluate the influence of accelerometer sampling frequency on behaviour classification, we annotated accelerometer data from seven European pied flycatchers (Ficedula hypoleuca) freely moving in aviaries. We also used simulated data to systematically evaluate the combined effect of sampling duration and sampling frequency on the performance of estimating signal frequency and amplitude. Results We found that a sampling frequency higher than Nyquist frequency at 100 Hz was needed to classify fast, short-burst behavioural movements of pied flycatcher, such as swallowing food with a mean frequency of 28 Hz. In contrast, high frequency movements with longer durations such as flight could be characterized adequately using much lower sampling frequency of 12.5 Hz. To identify rapid transient prey catching manoeuvres within these flight bouts, again a high frequency sampling at 100 Hz was needed. For both the experimental data of the flycatchers and the simulated data, the combination of sampling frequency and sampling duration affected the accuracy of signal frequency and amplitude estimation. For long sampling durations, the sampling frequency equal to the Nyquist frequency was adequate for accurate signal frequency and amplitude estimation. Accuracy declined with decreasing sampling duration, especially for signal amplitude estimation with up to 40% standard deviation of normalized amplitude difference. To accurately estimate signal amplitude at low sampling duration, a sampling frequency of four times the signal frequency was necessary (two times the Nyquist frequency). Conclusions The appropriate sampling frequency of accelerometers depends on the objective of the specific study and the characteristics of the behaviour. For studies with no constraints on device battery and storage, a sampling frequency of at least two times the Nyquist frequency will achieve relative optimal representative of signal information (i.e., frequency and amplitude). For classification and energy expenditure estimation of short-burst behaviours, 1.4 times the Nyquist frequency of behaviour is required.

Identifying animal behaviours from accelerometers: Improving predictive accuracy of machine learning by refining the variables selected, data frequency, and sample duration

“…It may be indicative of the variety of grooming motion frequencies and postures adopted by cats to groom their whole body and, while these variations can be visually identified by the researcher using a decision tree, the RF models struggled to deal with the inconsistency in this behaviour. The period over which the mean is taken should also be considered, especially for larger animals that might have a slower stride frequency; for example, Alvarenga et al (2016) found for sheep, that a mean calculated over 5 or 10 s led to a higher accuracy than over 3 s. Supporting this hypothesis, European pied flycatchers (Ficedula hypoleuca) catching prey at high speeds required a frequency of over 100 Hz for accurate identification whereas slower flight required 12.5 Hz (using the 'rabc' behaviour classification R package; Yu et al, 2023). Despite these behavioural considerations, study logistics including battery life will also influence decisions on the frequency of data collection.…”

Evaluating the effects of moonlight on the vertical flight profiles of three western palaearctic swifts

“…wing beats. The values reported are relative activity (normalized by tag between 0 and 100, unitless) and are a trade-off of technical constraints, measurement resolution and species specific characteristics [33,34]. To ensure temporal consistency across loggers, we resampled the 5 min GDL activity and pressure measurements to half-hourly intervals.…”

“…For all MDL measurements, we only considered stationary periods. We did not include activity explicitly, as activity monitoring is logger profile and species dependent [34], making the reuse of activity data from various loggers and species in the same model challenging. Activity was also not measured in all tags, which would reduce our dataset significantly (electronic supplementary material, appendix table S1).…”