Treed Gaussian processes for animal movement modeling

Abstract

Wildlife telemetry data are widely collected and can be used to answer a diverse range of questions relevant to wildlife ecology and management. While multiple animal movement models exist, current methods face challenges in modeling the nonstationarity of animal movement. Additionally, model implementation often poses barriers to practitioner use. To address these issues, I demonstrated a Bayesian machine learning modeling framework for telemetry data. This framework incorporates Bayesian statistics’ ability to quantify uncertainty and estimate comparable movement descriptors, while machine learning enables near automation of modeling. Specifically, my developed framework utilizes treed Gaussian processes (TGPs), a recently developed machine learning model that is well suited to the intrinsic nonstationarity of telemetry data. To ensure accessibility to practitioners, I utilized an existing R package to implement TGP modeling and outlined in detail the nearly automated use of the package within the movement modeling framework. I used telemetry data from a declining grassland bird, the lesser prairie-chicken (Tympanuchus pallidicinctus), as a case study to demonstrate the ease and applicability of this framework. I obtained model-based estimates of trajectories to compare individual and population estimates for movement descriptors such as distance traveled and residence time and compared these estimates across grazing management treatments. To maintain broad useability, I outlined all steps necessary for practitioners to specify relevant movement descriptors and apply TGP modeling and trajectory comparison to their own telemetry datasets. As well as modeling the nonstationarity present in animal telemetry data, the combined benefits of this framework increase accessibility and applicability of animal movement modeling, allowing practitioners to model trajectories and estimate comparable movement descriptors to answer applied management questions.