Classifying Eagle Flight Behaviors

Motivation

In this analysis, we aim to better understand the flight behavior of bald eagles by examining GPS-derived movement metrics and identifying patterns of similarity and difference within the data.

Data

The data used is obtained from GPS biologgers on Bald Eagles in Iowa (bergen, et al.)

• 2,093,022 GPS points from 57 eagles over ~ 4 years

• 13 Variables within the data set, 6 of them being flight variables useful within analysis

• X,Y, and Z variables for location are also present, as well as time recorded for each measurement

Methods

With the massive amounts of data present, a clustering approach is appropriate to find distinct flight patterns in the data.

The goal is:

• Get Data ready (scale, fix skewing issues)

• Determine the number of clusters that does the best job creating distinct groups

• Plot the clusters (Using Principal Component Analysis to plot in 2 dimensions)

• Examine Characteristics for each variable within each cluster

• Determine, for each cluster, the typical flight behaviors or patterns of eagles based on the variables in that cluster.

Methods - Prepping data

Before data for the clustering can be used, some steps need to be taken:

• Data must be scaled to ensure all variables are weighted equally

• Every variable distribution (except for vertical rate) looks like this. To fix the skewing issue, we will square root the data

Methods - Determining cluster amount needed

Next, we need to determine the amount of clusters that are best at creating unique groupings

To do this, we will use 3 approaches:

• Elbow Plot

• Average Silhouette distances plot (bootstrapped to bypass computational limits problem)

• Visual inspection

Methods - Determining cluster amount needed

Elbow (WSS) plots show the total within-cluster variation for different numbers of clusters, with the ‘elbow’ indicating the optimal cluster count.

There seems to be no obvious elbow, so this is not much help

Methods - Determining cluster amount needed

An average silhouette width plot shows how well data points fit within their assigned clusters for different numbers of clusters. Values range from -1 to 1, with 1 being the best

Judging by this plot, we should use 4 clusters to have the most distinct clusters possible

Methods

Visualizing how k amount of clusters would look can also give us a good idea for the appropriate amount of clusters - Which seems to be either 3 or 4

Methods - PCA

Since we have 6 dimensions of data, we will need to reduce our data down in order to properly visualize

To do this, we will use Principal Component Analysis to reduce dimensionality while retaining most of its variability

We will be using K-means to perform this clustering for its ability to handle large amounts of data

Methods - PCA

Results

Results

Results

Examining Horizontal Flight Patterns for each cluster over a Flight Segment

Results

Examining Vertical Flight Patterns for each cluster over time

Conclusion