Resolving platypus behaviour from accelerometry: frequency-domain features improve detection of rhythmic behaviours in hydrodynamically challenging aquatic environments
Developing robust methods to quantify how animals allocate time across behaviours is essential for understanding energy use, habitat requirements, and responses to environmental change. For cryptic, semi-aquatic mammals such as the platypus, direct observation is difficult, creating a reliance on remote biologging approaches that can reliably infer behaviour in the wild. However, aquatic environments can both smooth acceleration signals through hydrodynamic damping and introduce noise from water movement, turbulence, and drag, potentially obscuring behavioural differences of similar magnitudes. We tested whether progressively incorporating biomechanical and frequency-domain (FFT-derived) predictors improved behavioural classification in hydrodynamically challenging aquatic environments. Tri-axial accelerometers were deployed on four ex situ platypuses, with synchronised video observations used to validate behaviour. From the acceleration data, we derived three predictor classes of increasing complexity: summary statistics describing activity level, engineered biomechanical variables capturing posture and body orientation, and FFT-derived features describing movement rhythm. These predictors were progressively incorporated into Random Forest models to classify five behaviours: burrow resting, surface resting, grooming, travelling/foraging, and diving. Model performance improved with increasing predictor complexity, although gains were behaviour specific. FFT-derived features substantially improved classification of rhythmic behaviours such as diving and foraging, while engineered biomechanical predictors improved grooming detection. In contrast, resting behaviours, particularly surface resting, showed little improvement. Overall accuracy increased from [~]75% to [~]88% when frequency-domain features were included. Misclassification was greatest among behaviours with overlapping or low-amplitude signals, and cross-individual validation revealed reduced model generalisability, indicating that individual variation in movement patterns constrained transferability. Incorporating frequency-domain features substantially improved behavioural classification in platypuses, particularly for rhythmic behaviours such as diving and foraging. This study provides the first validated accelerometry-based behavioural classification framework for the species and highlights the importance of matching predictor selection to behavioural mechanics. More broadly, the approach offers a transferable framework for aquatic and semi-aquatic taxa.