Optimal deep learning activity recognition pipeline using animal bio‐logging data

Abstract Animal activity recognition using supervised machine learning on bio‐logging data has revolutionized behavioural studies on wildlife and livestock over the past decade. These studies' analytic pipelines differ substantially though, being based on various sensors, features, algorithms and hyperparameter values, which has caused suboptimal or even erroneous results. Furthermore, this lack of model uniformity hampers straightforward comparisons between studies and generalization of results. However, even though animal species, sensors and activities differ between studies, from a machine learning perspective, the classification task is very similar. This suggests that a priori determined patterns in bio‐logging data can be used to develop an optimal animal activity recognition pipeline. I systematically gathered publicly available bio‐logging data and associated activity labels from 17 studies on a wide range of terrestrial, aquatic and aerial animal species. Then I used a nested 5x2CV training and evaluation protocol with different sensor types, feature sets, window sizes, machine leaning algorithms and hyperparameter values. I compared the resulting performances from these ~800,000 different models using a generalized linear mixed model, to draw dataset‐agnostic conclusions about the interaction between all aforementioned model parameters. Finally, I developed a novel metric ( dMedian ) that quantifies the diminishing returns in added variation over increasing bio‐logging sequence lengths, which can be used to a priori determine a window size that optimizes the trade‐off between model performance and temporal resolution. From this followed that small 1D Convolutional Neural Networks with a fixed architecture and hyperparameter values consistently outperformed or performed on‐par with the best models from a lengthy hyperparameter tuning process, thereby making feature engineering and hyperparameter optimization obsolete. Furthermore, I demonstrate that model performance increased asymptotically with window size for all datasets and that my dMedian metric strongly correlates with it. These findings (and the provided R script) provide researchers with an optimal animal activity recognition pipeline, which will lower the entry barrier to apply these models and can consequently speed up ecological discovery through better model uniformity, compatibility and generalization. Moreover, there is great potential to deploy this small CNN on animal‐borne devices themselves, which will especially benefit wildlife studies.