Towards automated behaviour monitoring in wildlife: a review of machine learning approaches using accelerometer data

Monitoring animal behaviour provides critical insights into species ecology and offers essential information for guiding management and conservation efforts. Among the various approaches used to study behaviour, bio-logging—the use of animal-borne data recorders—has emerged as a valuable tool for observing animals in their natural habitats while minimising human disturbance. A major advancement in bio-logging has been the integration of accelerometers, which enable high-resolution analysis of movement and activity in free-ranging animals. By identifying movement and postural patterns captured by accelerometers, it becomes possible to associate specific accelerometric sequences with distinct behaviours. This task—automatically classifying repetitive accelerometric patterns linked to behaviours—can be achieved using machine-learning algorithms. The first ecological study applying machine learning to identify animal behaviours from accelerometer data was published in 2009. Since then, numerous studies have expanded this approach across a wide range of species, employing diverse methodological frameworks that make it challenging for new practitioners to identify best practices. The aim of this review is to provide a comprehensive overview of accelerometer-based behavioural identification in ecology using machine learning. We summarise the range of species and applications investigated, highlight key methodological trends, and offer practical guidance for researchers seeking to apply this approach. Based on 125 studies, we show that current practices largely rely on a general framework combined with species-specific adaptations. This has led to a predominance of species-focused methodological studies and limited generalisation across taxa. More importantly, the high diversity of datasets, combined with highly variable validation procedures—sometimes insufficient to assess model generalisation to novel data—prevents robust comparisons, limiting the identification of broadly applicable approaches. We therefore seek to re-establish a clear methodological framework that enables meaningful cross-study comparisons. Although methodological advances have been relatively modest until recently, 2024 marks a turning point, with a growing number of studies applying deep learning approaches that hold promise for improving model generalisation. While deep learning remains less widely adopted in ecology than in human or livestock behaviour recognition, leveraging advances from these fields and fostering interdisciplinary collaboration will be essential to accelerate progress. In particular, developments in real-time monitoring offer strong potential to enhance conservation efforts, an important next step in bio-logging. Nevertheless, despite increasing automation and generalisation, reliable behavioural classification models will continue to depend on robust ground-truth data and strong expertise in the natural history and biology of the study organisms.