top of page

Exploring the use of machine learning models for precise classification of pitviper movement and foraging behaviors


Thompson, Morgan



Department of Biological and Environmental Sciences

Northern Arizona University

Flagstaff, Arizona USA


Forsythe, Jeremy

The Center for Ecosystem Science and Society

Northern Arizona University

Flagstaff, Arizona USA


Ryan Hanscom

Department of Biology

San Diego State University

San Diego, California USA


DeSantis, Dominic L.

Department of Biological and Environmental Sciences

Georgia College & State University

Milledgeville, Georgia USA


Nowak, Erika M.

School of Earth and Sustainability

Northern Arizona University

Flagstaff, Arizona USA


Miniaturized animal-borne data loggers can be used to remotely and continuously record information about an animal's physiology, behavior, and environment. One such type of data logger is an accelerometer (ACT). Accelerometers are small (<2g) piezo-electric sensors that record static and dynamic acceleration along the X,Y,Z axes (X [surge], Y [heave], Z [sway]). When paired with machine learning, these signals can then be related to distinct behavioral classes, enabling real time classification of animal movement and foraging behaviors. Previously, this technology has been limited to larger-bodied organisms most pliant to the technique but recent miniaturization of devices has expanded the breadth of organisms amenable to the external attachment of such devices. We have begun to explore the use of externally attached ACTs and machine learning models to classify locomotion (l), stillness (s), and striking (t) in pit vipers. A series of captive validation trials were performed to capture instances of stillness and striking at a frequency of 25Hz. These behaviors were then compiled into a larger validation dataset that also included instances of visually verified field locomotion from wild snakes equipped with externally attached ACTs. This validation dataset has been used in the model building and testing in R to create a robust (accuracy >95%) model for the labeling of field datasets with wild-caught Timber Rattlesnakes (Crotalus horridus). Model exploration has shown promising results with multiple model types including Random Forest, Neural Networks, Decision Trees, and Support Vector Machines.

1/2
bottom of page