An Integrated Time-Varying Ornstein-Uhlenbeck Process for Jointly Modeling Individual and Population-Level Movement of Golden Eagles

This paper proposes a novel time-varying Ornstein-Uhlenbeck stochastic differential equation model that jointly analyzes individual telemetry and population-level abundance data to efficiently infer spatio-temporal dynamics, enabling improved risk assessment for wind projects and retrospective prediction of golden eagle migration origins.

The Gist

Imagine you are trying to understand the life story of a golden eagle. You have two very different types of clues, but neither tells the whole story on its own.

Clue #1: The GPS Tracker (The "VIP Guest List")
Imagine you put a high-tech GPS watch on 93 golden eagles. You know exactly where those specific birds are every hour. It's like having a detailed diary for 93 VIP guests at a massive party.

• The Problem: You don't know where the other 99,999 eagles are. Maybe the 93 you tagged are all from one specific family, or maybe they just happen to live in one corner of the country. If you only look at them, you might think the whole party is happening in that one corner, when in reality, the crowd is spread out everywhere.

Clue #2: The Bird Watchers (The "Crowd Count")
Now, imagine thousands of regular people using an app (eBird) to report where they see eagles. This gives you a heat map of the entire population. It tells you, "Hey, there are a lot of eagles in Colorado right now, and fewer in Texas."

• The Problem: This data is blurry. It tells you where the crowd is, but it doesn't tell you who is in the crowd. Is that eagle in Colorado a bird that lives there year-round? Or is it a traveler passing through on its way to Mexico? You can't tell the difference, so you can't predict where a specific bird came from or where it's going next.

The Big Idea: Merging the Diaries with the Heat Map

The authors of this paper built a mathematical time machine that combines these two clues into one super-model.

Think of it like this:

• The GPS data is the "microscope." It shows the detailed, individual movements of specific birds.
• The Bird Watcher data is the "wide-angle lens." It shows the big picture of where the whole species is.

The authors created a special equation (a "Time-Varying Ornstein-Uhlenbeck Process"—which is just a fancy name for a magnetic spring) to connect them.

The "Magnetic Spring" Analogy

Imagine every eagle is attached to an invisible spring.

• The Magnet: The spring pulls the bird toward a "home base" (like a winter nest or a summer hunting ground).
• The Jitter: The bird isn't a robot; it wanders around randomly (like a drunk person walking home), but the spring keeps pulling it back toward the magnet.
• The Magic Switch: The model realizes that eagles change their minds. In winter, the "magnet" is in Utah. In summer, the magnet moves to Alaska. In spring and fall, the magnets switch places, and the bird flies between them.

The genius of this paper is that they figured out how to calculate exactly where the entire population is at any given second, even though they only have GPS data on a few birds. They used the GPS birds to learn how the springs work (how fast they fly, how strong the pull is), and then applied those rules to the whole crowd seen by the bird watchers.

Why Does This Matter? (The Wind Turbine Problem)

The researchers used this model to solve a real-world danger: Wind Turbines.

Wind farms are great for clean energy, but spinning blades can kill eagles.

• The Old Way: You look at the crowd heat map. You see eagles near a wind farm in Utah, so you say, "Uh oh, that farm is risky."
• The New Way: You ask, "Which specific eagles are near that farm?"
  • Maybe the eagles near the Utah farm are the ones that winter in Colorado.
  • Maybe the eagles near a farm in Wyoming are the ones that winter in New Mexico.

The model can answer questions like: "If we find a dead eagle at a wind farm in Wyoming, where did it likely come from this past winter?"

This is huge for conservation. If you know that a specific wind farm is killing eagles that winter in a specific state, that state's wildlife managers can step in and say, "We need to shut that down or move it," because those are their birds. Without this model, they wouldn't know which birds were at risk.

The "Time Travel" Test

To prove their model works, they did a cool experiment.

1. They took data from eagles they didn't use to build the model.
2. They showed the model: "Here is where this eagle was in October (near a wind farm)."
3. They asked the model: "Where was this eagle in January?"

The model guessed the January location much better than just looking at the crowd heat map or just looking at the GPS data alone. It was like the model could rewind the clock and say, "Ah, this bird was definitely in the mountains of Utah in January, not in the plains of Texas."

The Takeaway

This paper is about connecting the dots. By mathematically linking the detailed lives of a few individuals with the broad movements of the whole species, the scientists created a tool that can:

1. Predict exactly where migratory birds are at any time of year.
2. Figure out which wind farms are dangerous to which specific groups of birds.
3. Help governments and companies make smarter decisions to protect these majestic birds while still building clean energy.

It turns a blurry, confusing picture of nature into a sharp, high-definition movie that helps us keep our wildlife safe.

Technical Summary

Here is a detailed technical summary of the paper "An Integrated Time-Varying Ornstein-Uhlenbeck Process for Jointly Modeling Individual and Population-Level Movement of Golden Eagles."

1. Problem Statement

Managing migratory species like the Golden Eagle (Aquila chrysaetos) requires understanding spatio-temporal migration patterns across their full annual cycle. Current challenges include:

• Data Limitations: GPS telemetry data provides high-resolution individual movement but is biased (not representative of the whole population) and often sparse. Conversely, citizen science data (e.g., eBird) provides broad population-level distribution but lacks individual-level tracking and behavioral context.
• Modeling Gaps: Existing models often fail to jointly leverage both data types for a full-year cycle. Previous attempts (e.g., Integrated Movement Modeling by Buderman et al., 2025) were limited to short time windows (e.g., 8 weeks) or relied on numerical approximations that are computationally expensive at continental scales.
• Management Needs: Agencies need to assess risks (e.g., wind turbine collisions) for specific subsets of a population (e.g., eagles wintering in a specific state) rather than just the aggregate population. Current methods cannot trace the origin of an individual based on a single late-season observation.

2. Methodology

The authors propose a Joint Integrated Movement Model (IMM) that combines individual telemetry data with population-level species distribution data using a Bayesian framework.

A. Data Sources

• Telemetry: 215 "bird-years" of GPS data from 93 Golden Eagles in western North America (2011–2018). Data was aggregated to daily locations.
• Species Distribution: Weekly relative abundance data from eBird's adaSTEM (Adaptive Spatio-Temporal Exploratory Model) at a 2.8 km resolution.
• Risk Data: Locations of 18,456 wind turbines aggregated into 396 wind projects.

B. The Core Model: Time-Varying Ornstein-Uhlenbeck (OU) Process

The authors develop a stochastic differential equation (SDE) model for individual movement that allows for a full-year analytic solution.

• Individual Dynamics: The movement of an individual $i$ in subpopulation $p$ is modeled as a time-varying OU process:
  $$dx_{i,t} = -\theta_p(x_{i,t} - m_{i,t})dt + \sigma_p dW_t$$
  Where the "attractive point" $m_{i,t}$ shifts based on the time of year:
  • Winter: Attraction to wintering grounds ($m_{i,1}$ or $m_{i,3}$).
  • Summer: Attraction to breeding grounds ($m_{i,2}$).
  • Migration: Transition times ($b_{i,1}$ for spring, $b_{i,2}$ for fall) are modeled as random variables following a scaled Beta distribution.
• Analytic Solution: Unlike previous numerical approaches, this model derives an exact conditional distribution for the bird's location at time $t+\Delta$ given time $t$. This is a weighted average of the previous location and the seasonal attractive points, with weights determined by the time elapsed and migration timing.

C. Modeling Heterogeneity

• Mixture Model: The population is divided into 4 subpopulations (clusters) based on migratory behavior (long-distance, moderate-distance, short-distance, and resident) using k-means clustering on displacement data.
• Hierarchical Priors:
  • Attractive points ($m$) follow an OU process over years to capture site fidelity.
  • Migration timings ($b$) follow a subpopulation-specific Beta distribution.
• Population-Level Scaling: By integrating the individual SDE over the mixture model priors and the initial spatial distribution, the authors derive an analytic Gaussian mixture for the population density $\pi_t(x)$. This avoids numerical integration of the SDE, enabling fast computation.

D. Likelihood and Inference

• Species Distribution Likelihood: Instead of a Poisson likelihood (which suffers from identifiability issues with rounded relative abundance data), the authors propose a Gaussian likelihood where the observed eBird relative abundance is modeled as a noisy observation of the normalized population density.
• Bayesian Inference: Parameters are estimated using Markov Chain Monte Carlo (MCMC) with Metropolis-Hastings updates.
• Risk Assessment: Risk is quantified as the Occupancy Time ($Q_r$), calculated by integrating the probability of an eagle being within a wind project's buffer zone over the year.
• Back-Casting: The model allows for conditional inference: predicting an eagle's location at time $t_1$ given its location at a later time $t_2$ (useful for determining the winter range of a bird killed by a turbine).

3. Key Contributions

1. Full-Annual-Cycle Analytic SDE: The first SDE model for migratory animals that covers the entire year and possesses a closed-form analytic solution, making continental-scale inference computationally efficient.
2. Joint Data Integration: A formal framework that simultaneously leverages sparse individual telemetry and dense population survey data (eBird) to correct for sampling bias and improve predictive power.
3. Subpopulation Risk Assessment: The ability to estimate risk for specific demographic subsets (e.g., "eagles wintering in Utah") rather than just the aggregate population, which is critical for state-level wildlife management.
4. Back-Casting Capability: Demonstrated ability to predict past locations based on future observations, a capability not possible with static distribution models.

4. Results

• Model Fit: The posterior mean of the modeled relative abundance closely matched the eBird adaSTEM data across the year.
• Migration Patterns: The model successfully identified four distinct migratory behaviors (long-distance, moderate, short, resident) and estimated their specific migration timing and attractive points.
• Wind Project Risk:
  • Risk profiles varied significantly by subpopulation. For example, eagles wintering in Santa Fe, NM, faced the highest risk from wind projects in eastern Wyoming, not necessarily those closest to their wintering grounds.
  • The model revealed temporal dynamics in risk, showing spikes during spring and fall migrations.
• Predictive Performance (Validation):
  • The authors validated the model by predicting the location of an eagle at the start of the year ($t_1$) given a location near a wind project later in the year ($t_2$).
  • Conditional Distribution Method: Using the joint model to back-cast yielded a Median Absolute Predictive Error (MAPE) of 748 km.
  • Comparison: This was significantly better than using the posterior mean of the population distribution alone (MAPE 1129 km) or using raw eBird data (MAPE 897 km). The joint model reduced error by ~33% compared to the population mean approach.

5. Significance

• Conservation Management: The model provides a principled tool for Eagle Conservation Plans (ECPs). It allows wind energy developers and regulators to place turbines in locations that minimize collision risk for specific local populations, rather than just the general species.
• Scalability: The analytic solution allows this approach to be applied to other migratory species at continental scales without prohibitive computational costs.
• Policy Impact: By quantifying risk for specific wintering populations, the model facilitates cooperation between different government agencies (e.g., breeding grounds vs. wintering grounds) to manage migratory species effectively.
• Methodological Advancement: It bridges the gap between individual movement ecology and population-level distribution modeling, offering a template for integrating heterogeneous data sources in ecological statistics.