Date Published: April 12, 2017
Publisher: Public Library of Science
Author(s): Maitreyi Sur, Tony Suffredini, Stephen M. Wessells, Peter H. Bloom, Michael Lanzone, Sheldon Blackshire, Srisarguru Sridhar, Todd Katzner, Sergio A Lambertucci.
Soaring birds can balance the energetic costs of movement by switching between flapping, soaring and gliding flight. Accelerometers can allow quantification of flight behavior and thus a context to interpret these energetic costs. However, models to interpret accelerometry data are still being developed, rarely trained with supervised datasets, and difficult to apply. We collected accelerometry data at 140Hz from a trained golden eagle (Aquila chrysaetos) whose flight we recorded with video that we used to characterize behavior. We applied two forms of supervised classifications, random forest (RF) models and K-nearest neighbor (KNN) models. The KNN model was substantially easier to implement than the RF approach but both were highly accurate in classifying basic behaviors such as flapping (85.5% and 83.6% accurate, respectively), soaring (92.8% and 87.6%) and sitting (84.1% and 88.9%) with overall accuracies of 86.6% and 92.3% respectively. More detailed classification schemes, with specific behaviors such as banking and straight flights were well classified only by the KNN model (91.24% accurate; RF = 61.64% accurate). The RF model maintained its accuracy of classifying basic behavior classification accuracy of basic behaviors at sampling frequencies as low as 10Hz, the KNN at sampling frequencies as low as 20Hz. Classification of accelerometer data collected from free ranging birds demonstrated a strong dependence of predicted behavior on the type of classification model used. Our analyses demonstrate the consequence of different approaches to classification of accelerometry data, the potential to optimize classification algorithms with validated flight behaviors to improve classification accuracy, ideal sampling frequencies for different classification algorithms, and a number of ways to improve commonly used analytical techniques and best practices for classification of accelerometry data.
Behavioral constraints can play a substantial role in determining ecological interactions , ecosystem processes , organismal distributions  and animal movements . For flying birds, whose energetic requirements may be substantially greater than those of non-flying animals, costs of locomotion may be so great that they can determine when and how individuals move . Given the great costs birds incur when moving, many species have evolved to specialize on either flapping flight (e.g., hummingbirds, geese; [6,7]) or on soaring and gliding (e.g., oceanic seabirds, vultures; [8,9]). Most species though switch between soaring and flapping flight modes on a fairly regular basis  and time spent in these behaviors is a key determinant of the total energetic costs of transport.
The acceleration based behavior classification approach we used required us to relate the statistical properties of the acceleration data to observable behavioral categories of the animal for model training and validation [9, 13, 14]. Our approach had four phases: (1) collection of accelerometry and behavioral observations (validation data), (2) data processing, including matching accelerometry to behavioral observations, (3) statistical modelling and optimizing classification algorithms, and (4) model application. Subsequently, we subsampled our high frequency accelerometer data to identify optimal sampling rates of such data and we applied our final models to classify raw accelerometry data from five wild golden eagles. Each of these steps is detailed below.
We collected a total of 2 hours, 53 minutes and 7 seconds of accelerometer data at ~140 Hz from the trained golden eagle. We also collected a total of 30 min and 21 sec of video in which the bird was visible in the video frame. Of this time, the bird was sitting on the trainer’s glove for 15 min and 11 sec; we annotated the behavior in the remaining 15 min 10 sec of video (S1 Table). The change point model framework identified 1557 distinct behavioral segments that we classified. The average number of data points for all segments, classified and unclassified, was ~ 49. For the simple ethogram the total number of segments in flapping, sitting and soaring behavior was 438, 219 and 900 respectively. The total number of segments in flapping straight, flapping banking, sitting, soaring straight, and soaring banking were 177, 261, 218, 598 and 303 respectively.
Our analyses demonstrate the utility of accelerometry data to classify flight behavior, the consequence of different approaches to classification of accelerometry data, the potential to optimize classification algorithms with validated flight behaviors, and ideal sampling frequencies for different classification algorithms. Furthermore, we illustrate a number of ways to advance commonly used analytical techniques and that may form the basis for best practices for classification of accelerometry data.
Accelometry provides ecologists a tool to monitor detailed changes in behavior of free-ranging animals. However, collection and use of accelerometry data are challenging because existing protocols are relatively new and not highly refined, and the entire process presents a series of logistical and computational challenges. The protocols we provide here allow species- and model-specific optimization via ground-truthed data that would improve confidence in behavioral, and ultimately energetic, classification from accelerometer data. This will allow further understanding of species-specific behavioral mechanisms and energetics of free ranging wild animals and, furthermore, standardize analytics for inter-specific comparisons that can inform as to the evolution and function of specific behaviors.