Improved supervised classification of accelerometry data to distinguish behaviors of soaring birds

Abstract

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.

Challenges of machine learning model validation using correlated behaviour data: Evaluation of cross-validation strategies and accuracy measures

“…Different types of data and different goals in mind might mean that for different problems different separation techniques and different metrics might be optimal. The data used for animal behaviour classification usually comes from very few animals (often less then 10, sometimes even one, and the largest sample of subjects consisted of 40 individuals [27,33,34]), while each animal contributes hundreds or thousands of data points. This means, that when randomly selecting data for testing, these points will very likely not be independent from a large part of the training set.…”

Development and application of a machine learning algorithm for classification of elasmobranch behaviour from accelerometry data

“…Once the input variables can be appropriately mapped to the outcome, the algorithm can be used to make predictions from new input data (Hastie et al 2009 ). Examples of these techniques include decision trees, random forest (RF), K-nearest neighbour and linear discriminant analysis (Kiani et al 1998 ; Staudenmayer et al 2009 ; Nathan et al 2012 ; Soltis et al 2012 ; Campbell et al 2013 ; Bidder et al 2014 ; Resheff et al 2014 ; Williams et al 2015 ; Sur et al 2017 ). Supervised ML has been applied to classify acceleration data in many studies and has the advantage of clearly defined behaviours and simple interpretation (Leos-Barajas et al 2017 ).…”

Activity budgets for the sedentary Argentine sea bass Acanthistius patachonicus inferred from accelerometer data loggers

“…; Sur et al . ), this work did not find one classification method (KNN versus TREE) consistently better than the other, and some mismatch were observed, in particular for the ‘Hovering’ and ‘Static on bottom’ behaviours (Appendix ). Both behaviours produce a low amplitude signal for all the variables, which may explain why the classification methods confused them.…”