DSCI 415: Classifying and Clustering Bald Eagle Flight Behavior

Background & Motivation

• GPS telemetry enables detailed tracking of eagle movement.
  
• Understanding flight behavior helps ecologists study migration, foraging, and energy use.
  
• Our goal: classify perching vs. flight, then explore variation in flight patterns.

Data Overview

• ~600,000 GPS points from bald eagles
  
• Variables include speed, altitude, vertical rate, turning angle
  
• Data grouped by animal and flight segment
  
• Preprocessing filters out noise and short segments

Classification Objective

• Use heuristics to label points as “perch” or “in flight”
  
• Train a random forest classifier to predict behavior
  
• Evaluate performance using accuracy, ROC AUC, and F1 score

Perch vs. In Flight Birds

• Heuristic labeling: points are classified as “perch” if speed and vertical movement are near zero.

• Flight label creation: each GPS record is assigned either “perch” or “in_flight”.

• Purpose: establishes a ground truth for training and evaluating the classifier.

in_flight     perch 
  1876842     83113 

Confusion matrix

• Purpose: compares the model’s predicted labels against the true labels

• Interpretation: shows how well the classifier distinguishes “perch” vs. “in flight”

           Predicted
Truth       in_flight  perch
  perch             0  29022
  in_flight    566505      0

What the Classifier Reveals

• The model distinguishes perching from flight with 92% accuracy
• ROC AUC of 0.97 shows strong separation between classes
• F1 score confirms balanced precision and recall for flight detection
• Most important features:
  • sn_ms (horizontal speed)
    
  • kph_ms (converted speed)
    
  • abs_vr (vertical rate)

Variable Importance

Clustering Objective

• Focus on in-flight points only
  
• Use k-means to explore behavioral variation
  
• Choose optimal K using silhouette scores

# A tibble: 6 × 2
      k silhouette
  <int>      <dbl>
1     2      0.299
2     4      0.252
3     3      0.235
4     7      0.228
5     6      0.220
6     5      0.219

Silhouette Scores

# A tibble: 2 × 7
  cluster kph_ms sn_ms   agl abs_angle vertical_rate abs_vr
  <fct>    <dbl> <dbl> <dbl>     <dbl>         <dbl>  <dbl>
1 1        15.2  14.9  252.       0.16         -1.02   1.29
2 2         8.59  7.52  74.6      1.2           0.33   0.59

Cluster Feature Distributions

What the Clusters Reveal

• K-means clustering (K=2) separates flight into two distinct modes:
  • Cluster 1: fast, high-altitude, consistent movement
    
  • Cluster 2: slower, lower, more variable — possibly foraging or circling
    
• Silhouette scores confirm good separation
  
• Feature distributions show clear behavioral differences

PCA Biplot

What PCA Shows

• PCA biplot shows partial overlap between clusters
  
• This is expected — PCA only shows two dimensions
  
• In full feature space, silhouette confirms separation
  
• Variable arrows indicate which features drive cluster separation

Exemplar: Cluster 1 Time Series

Exemplar: Cluster 1

Exemplar: Cluster 1 Summary

• Fast, high-altitude flight
  
• Long, smooth trajectory
  
• Consistent vertical movement

Exemplar: Cluster 2 Time Series

Exemplar: Cluster 2

Exemplar: Cluster 2 Summary

• Slower, lower altitude
  
• More variable vertical rate
  
• Possibly foraging or circling behavior

Final Cluster Summary

# A tibble: 3 × 2
  metric   value
  <chr>    <dbl>
1 Accuracy     0
2 ROC-AUC      1
3 F1           1

# A tibble: 12 × 3
   cluster feature       median
   <fct>   <chr>          <dbl>
 1 1       abs_angle       0.16
 2 1       abs_vr          1.29
 3 1       agl           252.  
 4 1       kph_ms         15.2 
 5 1       sn_ms          14.9 
 6 1       vertical_rate  -1.02
 7 2       abs_angle       1.2 
 8 2       abs_vr          0.59
 9 2       agl            74.6 
10 2       kph_ms          8.59
11 2       sn_ms           7.52
12 2       vertical_rate   0.33

Summary Tables

Summary Tables con.

Summary of Findings

• Classification model performs well and identifies key predictors of flight
  
• Clustering reveals meaningful variation within flight behavior
  
• PCA and silhouette scores support the validity of clusters
  
• Exemplars provide intuitive, visual confirmation of behavioral differences

Key Takeaways

• Classifier distinguishes perch vs. flight with high accuracy
• K-means reveals two distinct flight behaviors
• PCA shows overlap, but silhouette confirms separation
• Exemplars illustrate behavioral differences clearly

Future Directions

• Incorporate time-of-day and weather data
• Explore seasonal or migratory patterns
• Apply clustering to other species or regions

Thank You

Questions?