Filtering and interpreting location errors in satellite telemetry of marine animals

Abstract

Argos telemetry offers a powerful means of tracking wild animals in their habitat, yet the delivered locations are subject to complex errors and random coverage. Bayesian filters and statistical models allow for objective trajectory estimates and inference on movement rates. As an alternative to Monte-Carlo methods, we investigate here how classic time series technique, such as the Kalman Filter, can be made robust to uncover patterns in the data. Our approach relies on a composite measurement model to account for outliers, and makes use of all the Location Classes to smooth observations and regularize the track to a regular time grid. Two application examples are presented. Using data from freely-swimming leatherback turtles, we confirm that locations of class A (LCA) are more accurate on average than class 0, and we recommend their use in tracking studies. We further show how measurement errors (and their geometry) interact with the assumed movement model, further modulating the final location error and the discriminating ability of the filter. The choice of the movement model appears important, since a model with no velocity constraint may fit observational errors at the expense of trajectory smoothness, while a speed-based model is better behaved but less forgiving for data fitting and outlier identification. Varying sea surface temperatures also appear to degrade the quality of locations and increase the occurrence of outliers, possibly in relation to thermal stratification and depth behavior. These results have important implications when inferring changes in behavior from long-term movements.