An information theory framework for hierarchical path segmentation and analysis of animal movement

Improved animal tracking technologies provide opportunities for novel segmentation of movement tracks/paths into both behavioral activity modes (BAMs; e.g., foraging, resting, commuting) and finer segments critical to our understanding of the movement ecology of individuals and, hence, the functioning of ecosystems. Current BAM segmentation methods include biological change point analysis (BCPA) and hidden Markov models (HMM). Here we use a bottom-up approach providing a two-tier sub-BAM fixed-length segmentation of animal tracks into μ-step-long “base segments” (ultra-fine tier, “letters” or “symbols”) and m-base-segment-long “words” (fine tier). The base segments are clustered into n statistical movement element (StaME) categories. The word segments, consisting of m base segments, are clustered into k “raw” canonical activity mode (CAM) categories. A rectification process is then implemented so that all word segments coded by the same sequence of m StaMEs are identified with the same “rectified” CAM type. These fixed-length CAM-type segments on being given behavioral interpretations such as “fast, medium or slow directed movement”, “fast or slow random movements”, or “stationary” then provide insight into how different, larger, variable length BAM-type segments, such as “resource gathering” or “bee-line commuting” are made up of a characteristic mix of smaller, fixed-length sequences of CAM types. The percentage of reassignment errors, along with information theory measures associated with our method, is used to compare the efficiencies of coding both simulated and empirical barn owl movement tracks for a selection of parameter values and approaches to clustering at the StaME (ultra-fine “letters”) and CAM (fine) tiers. Once implemented, our methods can be used to provide a refined scale coding scheme for BAMs that themselves have been segmented using BCPA, HMM, and other methods. Our approach thus complements and enriches rather than replaces current animal track segmentation methods to provide a magnifying lens on how different types of BAMs are themselves constituted by various types of CAMs.