Machine learning for inferring animal behavior from location and movement data

Abstract

Movement ecology has rapidly advanced owing to recent developments of animal-attached devices and wide applications of sophisticated statistical and machine learning techniques in analysis of animal movement data. Global Positioning System (GPS) transmitters used for estimating animal locations and tri-axial accelerometers used for measuring the 3-dimensional accelerations of animal's motion aid researchers in collecting location and locomotion data at fine spatial and temporal scales. Machine learning and other advanced statistical methods bridge conceptual models to data, providing insights into ecological and physiological mechanisms underlying animal behavior and movements. This study reviews the general principles and applications of state space models, hidden Markov models, random forests, and support vector machines in the inference of animal behavior from movement data. Unsupervised learning algorithms, including Bayesian state space models implemented for robust correlated random walk models and hidden Markov models, help infer different behavioral modes using GPS location data. State space models can account for measurement error in GPS locations and estimate the true locations of animals; however, without including movement-state switching, state space models do not infer behavioral modes directly. On the contrary, hidden Markov models estimate the probabilities that animals switch between different behavioral modes. Nevertheless, hidden Markov models neither directly estimate animal locations nor account for measurement error explicitly. Supervised learning algorithms integrate data on locations and directional accelerations with synchronized behavioral observations (i.e., labels) to classify behaviors to pre-defined behavioral categories. Unlike unsupervised learning, supervised learning requires behavioral observations to label locations and accelerometer data to train the learning algorithms. However, behavioral observations synchronized with relocations and acceleration records are often missing or unattainable in many species, hindering the applications of supervised learning, making unsupervised learning a suitable tool for behavioral annotation of movement paths in secretive (cryptic) or less studied species. Environmental and behavioral annotations of animal movement paths by machine learning improve understanding the effects of environmental conditions on animal movements and behavioral decisions.