Inferring behavioural states from tracking data with hidden Markov models: a validation study using GPS video-camera collars

Background: Hidden Markov models (HMMs) are increasingly used to infer animal behavioural states from GPS tracking data, yet their interpretation often remains uncertain in the absence of empirical validation. Misinterpretation of statistical states as biologically meaningful behaviours can undermine scientific understanding and conservation decisions. Our objective was to evaluate how well HMM-inferred states correspond to directly observed behaviours and to test how the temporal resolution of GPS sampling influences behavioural inference. Methods: We used GPS collars equipped with video cameras to validate HMM-inferred behavioural states in 81 female migratory caribou (Rangifer tarandus) from two herds. We compared states derived from two- and three-state HMMs to behaviours observed in short collar video clips. To assess the effect of temporal scale, we fit HMMs to GPS data resampled at 20-, 60-, and 120-minute relocation intervals. Results: HMM-inferred behavioural states frequently diverged from video-observed behaviours at the start of observed GPS steps, especially at longer relocation intervals. These mismatches appeared to result from overlapping movement metrics among caribou behaviours (e.g., foraging vs. resting or traveling) and the inability of coarser GPS data to capture behavioural switches occurring at finer temporal scales than the fix rate. Videos of eating were the most misaligned with HMM-inferred states, likely due to high variation in caribou movement while foraging that is often characteristic of mixed-feeding large herbivores. Inferred states for a given location were often inconsistent across temporal scales, indicating that HMM outputs must be interpreted cautiously with respect to the GPS sampling frequency. Conclusions: The predicted HMM state can differ substantially from true behaviour at the start of each step, in particular at coarse temporal scales. Our results serve as a reminder to interpret HMM states over whole steps rather than at observed positions, validate movement-derived states where possible, and align sampling resolution with species-specific behavioural patterns.