A simple method to classify behaviours of aerial predators using tri-axial high-frequency accelerometers
BackgroundRemote monitoring of animals has opened new avenues in behavioural research. The advent of animal-borne accelerometers makes it possible to derive the body posture and activity level, and thereby to identify fine-scale behaviours with no need for direct observation. This is particularly useful for species that forage or hunt beyond human sight in sometimes hard-to-access or remote areas. However, the complexity of numerous methods developed for this purpose may discourage researchers from effectively using acceleration data. ResultsWe present a simple and straightforward approach to identify and classify behaviours based on tri-axial high-frequency (25 Hz) acceleration data. This method relies on constructing a decision tree to assign predefined behaviours to time series of acceleration data using four accelerometer-based parameters that are biologically relevant and easy to interpret. Initially, we manually labelled sequences of acceleration data with the behaviours using paired high-frequency (i.e. 1 Hz) GPS tracking data as ground-truth. Then, threshold values of the acceleration-based parameters were objectively determined based on receiver operating characteristic (ROC) analysis. Finally, we built the decision tree to automatically classify the behaviours of the whole dataset and tested further improvements. As a case study, we used two falcon species (lesser kestrel Falco naumanni and common kestrel F. tinnunculus). Our simple method achieved an overall accuracy of 97.6% in classifying the following behaviours: perching (99.9%), soaring-gliding (99.3%), flapping (93.1%) and hovering (86.4%). We also showed that kestrels prey capture attempts can be identified from the tri-axial high-frequency accelerometers due to diagnostic signatures (the so-called spikes). ConclusionsOur simple method has broad applicability to free-ranging aerial predators. We show that behavioural categories can be accurately classified from a relatively small subset of manually labelled acceleration data (43,075 seconds). Although quantifying hunting success needs further investigation, the detection of prey capture attempts opens new possibilities for studying the hunting behaviour of aerial predators.