Inferring Behavioral States of American Redstarts Using Automated Telemetry and Machine Learning

Akansha Goswami

Mentor: Dr. Bryant C. Dossman PhD, Director and Principal Scientist, Audubon Everglades Research Station.

Date/Time: August 25th, 2026 at 1:45 PM.

Abstract: The Motus system is an invaluable tool to track animals at a volume and length that would not be possible by direct observations. The tags are small, uniquely coded radio transmitters that make it possible to collect large quantities of data from many individuals. However, the resulting dataset is often millions of detections that can be time and computationally intensive. Currently, most literature focuses on quantifying proportions of active and inactive time periods from calculated signal thresholds. However, this kind of coarse classification can lose valuable temporal information and behavioral transitions. The aim of our project was to create an automated Python pipeline that cleans and prepares raw Motus tag data to be analyzed using hmmTMB in R, a hidden Markov model package specifically for animal movement. The pipeline removes duplicate and invalid detections, and calculates standard deviation of signal strength as a proxy for movement. These data are used in a hidden Markov model to identify resting and active states and estimate transitions between them. Unlike threshold-based approaches, the HMM framework allows behavioral states to be inferred from the temporal structure of the signal while incorporating individual-level and environmental covariates that can change over time. We applied this workflow to American Redstarts (Setophaga ruticilla) to evaluate patterns of activity and inactivity during their nonbreeding season. This approach provides a reproducible framework for transforming large Motus datasets into biologically meaningful behavioral analysis.