Table 1.
Ethogram of behavior used in simple and complex behavioral classification of a trained golden eagle outfitted with an accelerometer.
Fig 1.
Out of Bag (OOB) errors versus number of predictors, by node, from random forest classification of accelerometer data collected from a trained golden eagle.
Number of nodes (mtry) ranged from 0–28 and number of trees (ntree) from 500 to 5000. We classified data to (a) three behavioral classes: flapping, sitting and soaring and (b) five behavioral classes: flapping banking, flapping straight, sitting, soaring banking and soaring straight. Boxes identify combinations of mtry and ntree values resulting in the lowest OOB error.
Fig 2.
Variable importance plots for predictor variables (described in main text) from random forest classification of accelerometry data collected from a trained golden eagle.
We classified data to (a) three behavioral classes: flapping, sitting and soaring and (b) five behavior classes: flapping banking, flapping straight, sitting, soaring banking and soaring straight. Higher values of mean decrease in accuracy indicate that the variables are more important to the classification process.
Table 2.
Model accuracy parameters for supervised classification of accelerometer data collected from a trained golden eagle.
Parameters reported are sensitivity, specificity, positive predicted value, negative predicted value, prevalence and balanced accuracy as estimated for random forest (RF) model and K-nearest neighbor (KNN) models.
Table 3.
Ten-fold cross validation error rates for classification of accelerometer data collected from a trained golden eagle.
Models tested were random forest (RF) and K-nearest neighbor (KNN).
Fig 3.
Overall classification accuracy using a K-nearest neighbor model to classify acceleration data from a trained golden eagle.
We used both (a) a simple ethogram (three behavioral classes: flapping, sitting and soaring) and (b) a complex ethogram (five behavior classes: flapping banking, flapping straight, sitting, soaring banking and soaring straight). We incrementally increased values of K by 5 until classification accuracy declined and then incrementally adjusted values of K by 1 to identify peak accuracy, indicated with a box.
Fig 4.
Accuracy of behavioral classification accuracy when sampling acceleration data from a trained golden eagle from 5 to 40Hz.
Data were classified to three behavioral classes (flapping, sitting and soaring) and modeled with (a) a random forest classification model and (b) a K-nearest neighbor model.
Fig 5.
Flight behavior of 5 free-ranging golden eagles interpreted from acceleration data.
Plots show percentage of time spent in flapping or soaring flight at different times of a day (a,b) and flight behavior as a function of flight altitude (c,d). Behavior was classified with (a, c) a random forest model and (b, d) a K-nearest neighbor model.