Efficient statistical inference of turning points in animal movement data

1. Recent years have seen a proliferation of high-frequency animal movement data, often at greater than 1 Hz, allowing us to gain much greater insight into behaviour than with lower frequency data. In particular, it is becoming possible to detect the precise points at which animals are making decisions to turn, thus placing the idea that the animals move in ‘steps and turns’ onto rigorous grounding.2. Despite this, current efforts to ascertain the points at which animals turn tend to rely on the user making pre-determined choices of certain model parameter values. Furthermore, whilst they may give good results, there is often no theory explaining why the inferred turning points are most likely to be correct, for example by maximising a likelihood function.3. Here, we propose a theoretically grounded statistical technique to find turning points in high-frequency movement data that does not require any a priori choices of parameter values. By testing our algorithm on simulated data, we show that our technique is both fast (e.g. 3 s to parse 106 data points) and accurate. For example, when the standard deviation of the noise is less than around �12 radians then our algorithm correctly identifies nearly 100 % of the turning points, providing the noise is not heavily autocorrelated. Additionally, we demonstrate the effectiveness of our technique on magnetometer data from free-ranging Arabian oryx (Oryx leucoryx).4. Overall, our work gives a fast, accurate and statistically grounded algorithm for turning point detection in high-frequency data. The resulting model of straight-line steps and turns provides a biologically meaningful summary of the animal's movement behaviour, which has potential to be used as an input to the wide range of step-and-turn techniques used in movement ecology, such as step selection analysis and hidden Markov models of behavioural states.