Transferring decision boundaries onto a geographic space: Agent rules extracted from movement data using classification trees

Abstract

We leverage applied machine learning to determine which environmental features are best associated with the “moving” behaviour(s) of a troop of olive baboons (Papio anubis; collared with GPS trackers at Mpala Research Centre, Kenya). Specifically, we develop a behaviour-selection surface informed by classification trees trained using movement trajectories and remotely sensed environmental features. Atop this surface, we simulate agent movement towards set destinations, constrained by the relative extent to which sets of features are associated with behaviour(s). To achieve our goal, we perform: (a) path segmentation using thresholding to label training data; (b) agent-rule extraction using classification trees to associate the relative Euclidean distance of a point from environmental features with behaviour; and (c) implementation of this information into an agent-based model to provide a data-driven simulation of troop movement. We believe this framework can accommodate intensifications in data velocity, veracity, volume, and variety expected from increasingly sophisticated biologgers and data-fusion techniques.