Abstract
Animal movement analysis aims to understand how organisms navigate and exploit complex, varying environments. We introduce Tortuosity Entropy (TorEn), a metric based on symbolic dynamics that quantifies behavioral changes in movement trajectories. By transforming pairwise differences between track points into symbolic sequences and calculating their Shannon entropy, TorEn provides a fine-scaled measure of movement complexity. Validated on both synthetic trajectories and real-world albatross GPS data, TorEn accurately identifies transitions between behavioral states, such as the shift from direct migration to localized foraging. This approach is computationally efficient and applicable to various real-world data types, making it a promising tool for revealing behavioral patterns in movement ecology.
Introduction: The Challenge of Noisy Movement Data
Understanding animal movement is central to ecology, but movement data is notoriously multi-modal and often "noisy." An animal might fly in a straight line for hundreds of miles (low tortuosity) and then suddenly engage in "area-restricted search" (high tortuosity) to find food.
Traditional metrics like the straightness index or fractal dimension often provide a single, static value for an entire trip, failing to capture these critical behavioral transitions. Other methods might focus solely on displacement while ignoring changes in speed or direction. To address this, researchers have developed the TorEn algorithm, which borrows concepts from information theory and symbolic dynamics to reveal the underlying behavioral states of an organism in a fine-scaled, dynamic way.
Methodology: The TorEn Algorithm
The core of the TorEn approach is a three-step process that converts continuous movement parameters into discrete symbolic patterns to measure predictability and complexity.
- Quantification and Symbolization: Raw trajectory data (e.g., a time series of headings, bearings, or speeds) is converted into a sequence of symbols. For example, a simple binary system (0 or 1) or a three-symbol system (0 for stable, 1 for increasing, 2 for decreasing) is used based on a defined threshold ($\lambda$).
- Pattern Vector Mapping: These symbols are grouped into "pattern vectors" using an overlapping embedding dimension ($m$). If $m=3$, a sequence like
01011is mapped into vectors(0,1,0),(1,0,1), and(0,1,1). - Entropy Calculation: The algorithm calculates the Shannon entropy of these pattern vectors. A highly predictable, repetitive movement (like a straight line) results in low entropy, while erratic, unpredictable movement (like intensive foraging) results in high entropy.
A sliding window approach ($M$) allows the entropy value to change over time, creating a "dynamical TorEn" that can track behavioral shifts as they happen throughout a journey.
Validation and Results
The researchers validated TorEn using both synthetic and real-world datasets to ensure its robustness across different movement types.
Synthetic Trajectories
Using Biased Random Walks (BRWs), the authors created paths with varying "concentrations" ($\kappa$), representing different levels of directional persistence. As shown in the simulation results, trajectories with high persistence ($\kappa = 20$) maintain low entropy, while those representing random search ($\kappa = 0$) exhibit high, sustained entropy levels. TorEn successfully distinguished between these states and accurately identified the exact points where a simulated animal switched from "migrating" to "searching."
Real-World Albatross Mapping
When applied to GPS data from the Galapagos albatross (Phoebastria irrorata) sourced from Movebank, TorEn clearly segmented the birds' journeys into three distinct phases: the initial direct flight to the foraging area, the intensive search phase near the Peruvian coast (characterized by high entropy), and the return flight.
A key discovery was the identification of "feeding at situ." While ground speed might drop to near zero (suggesting rest), high TorEn in the bird's heading suggested it was actually actively reorienting and feeding in place rather than just floating idly.
Discussion and Insights
TorEn offers several advantages over traditional movement metrics:
- Symbolic Efficiency: By focusing on the pattern of change rather than raw magnitudes, TorEn is computationally light and robust against minor sensor noise.
- Parameter Flexibility: The metric can be applied to any movement parameter. Applying TorEn to bearing (direction to a destination) can reveal different behavioral nuances than applying it to speed or heading.
- Scale Sensitivity: The choice of window size acts as a lens; small windows identify abrupt changes, while larger windows smooth data to show broader trends. For angular parameters, a threshold between $20^\circ$ and $30^\circ$ was found to be the "sweet spot" for identifying meaningful changes.
Conclusion and Future Work
While TorEn is a powerful tool, its effectiveness depends on proper parameter tuning ($m$, $M$, and $\lambda$) based on the specific species and data resolution. Future iterations of the model will need to address irregular sampling rates, which are common in low-power GPS tags, to ensure the algorithm remains accurate across diverse tracking hardware. Overall, TorEn represents a significant step forward in our ability to decode the complex behavioral language of animal movement.
Summary updated