Behavioural state inference from movement and environmental data using Markovian step selection functions

Abstract

O_LIMovement paths reflect temporal shifts in behavioural states, typically driven by internal and external drivers. However, the inherently multiphasic nature of these trajectories is frequently overlooked in empirical studies, an oversight that can hinder progress in our understanding of movement ecology. While Hidden Markov Models (HMMs) can successfully identify latent states--such as foraging or travelling--they face significant challenges, particularly in determining the appropriate number of states and in interpreting their ecological relevance in the context of both movement patterns and environmental covariates. C_LIO_LIWe present a framework based on Hidden Markov Models with Step Selection Functions (HMM-SSFs) that identifies behavioural states, represented by ecologically meaningful labels linked to explicit hypotheses about animal movement, that best explain observed movement patterns. The framework imposes interpretable conditions and diagnostic criteria on the post-identified behavioural states to ensure ecological coherence. It is grounded in the evaluation of biologically motivated scenarios rather than purely data-driven partitioning. The framework proceeds in two main steps: first, movement-based states are identified using movement-derived covariates only; second, these states are refined by incorporating environmental predictors, such as habitat structure or species interactions (e.g., predator-prey dynamics). This sequential integration enables the detection of ecological responses that are conditional on behavioural context. C_LIO_LISimulations show that the framework effectively recovers behavioural states across most conditions. State decoding accuracy was notably higher when control locations were drawn from an exponential-family distribution, compared to a uniform one. The exponential-family approach improved state separation and reduced mislabelling, especially when few control locations are generated. However, low state persistence--particularly in Encamped behaviours--resulted in an overestimation of the number of states. These findings underscore the influence of transition probabilities on behavioural labelling. Finally, we applied our framework to zebra (Equus quagga) movement data by combining movement predictors with changes in direction toward the nearest preferred habitat. This enabled us to distinguish between habitat-dependent and habitat-independent travelling behaviours, as well as to identify spatially finer-scale such as encamped state. C_LIO_LIThe proposed framework balances complexity and biological interpretability by using basic movement metrics to identify the behavioural states and their sequence that best explain multiphasic movement paths, together with environmental factors directing movement in each state. Unlike traditional methods that predefine the number of states, the framework estimates both state number and labels, offering a flexible and ecologically meaningful approach for behavioural inference. C_LI