Opportunities from remote recording of activity in wild ungulates: classification of behaviour in red deer (Cervus elaphus) using activity sensors in GPS collars

Recent advances in bio-telemetry allow to remotely record and infer behaviour in many wild species. For wild ungulates, collecting spatiotemporal data (i.e. location and time) and, simultaneously, knowing their behaviour (e.g. if the animal is feeding or resting) is critical. Such information could indeed help understand their habitat use, their responses to external cues and other aspects of their biology. The research aimed at evaluating the potentialities offered by GPS collars equipped with activity sensors for inferring behaviour in red deer (Cervus elaphus). Four wild-born animals raised in captivity, 2 hinds and 2 stags, were equipped with GPS collars with built-in tri-axial accelerometers. The animals were equipped with the collars for a period ranging from 5 to 40 days, during the months of May, June, July, August, October and November. The device measured acceleration 4-8 times per second and automatically scaled raw acceleration data into levels ranging from 0 to 255, providing an average activity value every 300 seconds on two axes: the X-axis for forward/backward movements and the Y-axis for sideways and rotary movements. A set of 7 behaviours (bedded, bedded rumination, standing, grooming, feeding, walking, running) was directly observed in synchrony with collar measurements. Behaviours were classified with discriminant analysis (DA) using Gaussian Finite Mixture Models. Activity measurements on the X and Y-axes were predictors. Models were trained on a random subset and validated on the remaining data with cross-validation. We preliminarily tested models classifying both all observed behaviours and several behavioural classes obtained by merging similar behaviours (e.g. inactive behavioural class was obtained by merging bedded and bedded rumination). Overall, walking and running were misclassified in more than 90% of cases and were mainly assigned to feeding behaviour. Feeding activity was correctly classified in 84.1% of cases, while standing only in 35.5%. The model correctly classified bedded behaviour in 79.8% of cases while bedded rumination was mainly mistaken with bedded behaviour, with a rate of correct classification below 12%. Thus, bedded and bedded rumination were merged into an inactive behavioural class and the remaining behaviours into an active behavioural class. DA correctly classified respectively over 88% and 94% of active and inactive intervals. The use of activity sensors in GPS collars could thus represent an effective technique to track movements of wild ungulates and, at the same time, to monitor their activity levels as well as to infer with a reasonable accuracy at least their state of activity (active or inactive). Nevertheless, synchronous observation on sample animals is required to calibrate the predictive model. The scaling and averaging process of acceleration data over a predefined sampling interval probably represents one of the major criticalities that reduces the possibilities to infer a higher set of behaviours.