🤖 machine learning

Transformer-Based Wildlife Species Classification from Daily Movement Trajectories

This paper demonstrates that Transformer-based sequence models, when trained on large-scale GPS trajectories with augmented movement features and a unified 1-hour temporal resolution, significantly outperform traditional baselines like LSTMs and CNNs in classifying wildlife species under challenging cross-study and data-limited conditions.

Original authors: Obed Irakoze, Prasenjit Mitra

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Obed Irakoze, Prasenjit Mitra

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guess what kind of animal is walking past you in the dark. You can't see the animal, you don't know where it is, and you can't hear it. All you have is a notebook recording exactly how it moved every hour for a whole day: step, step, pause, turn left, run, stop.

This paper is about teaching a computer to solve that exact puzzle. The researchers asked: Can we identify a specific animal species just by looking at its daily movement patterns, without knowing its name or its location?

Here is a breakdown of how they did it and what they found, using simple analogies.

The Challenge: The "Ghost" in the Data

Usually, if you want to identify an animal, you might look at where it lives (a forest vs. a desert) or what it looks like. But this team wanted to see if the animal's personality (its movement style) was enough to identify it.

They faced a tricky problem: Animals in the same park often move similarly because of fences, roads, or human activity, not just because they are the same species. If you train a computer on animals in one park, it might just learn "this park's style" rather than "this animal's style."

To fix this, they used a "blind test" strategy. They trained the computer on data from several different African regions and then tested it on a completely new region it had never seen before. This ensured the computer was learning the animal's true "gait," not just the geography.

The Tools: The "Super-Reader" vs. The "Old School" Readers

The researchers tested different types of computer brains (models) to see which one could best read these movement stories.

The Old School Readers (LSTM, CNN, TCN): These are like students who read a story one sentence at a time, trying to remember the beginning to understand the end. They are good, but they sometimes get lost if the story is long or complex.
The Super-Reader (The Transformer): This is the star of the show. Imagine a student who can look at the entire story at once, instantly connecting the beginning, middle, and end. It can see the big picture of how an animal moves over a whole day.

The Result: The "Super-Reader" (Transformer) won every time. It was significantly better at guessing the species, often beating the older models by a wide margin (8 to 22% better). For example, when trying to spot an elephant, the Transformer was right 83% of the time, while the others struggled more.

The Secret Sauce: Adding More Details

At first, the computer only looked at the most basic info: How far did the animal move from point A to point B?

The researchers then gave the computer a "magnifying glass" by adding more details to the data:

Speed: Was it a slow stroll or a sprint?
Direction: Was it heading straight or wandering?
Turning: Did it make sharp turns (like hunting) or wide loops (like grazing)?

The Result: This was a game-changer. It's like trying to identify a friend by just their height (hard) versus their height, voice, and walk (easy). Adding these details helped the computer identify tricky or rare animals, like lions and zebras, much better.

The Timing Question: Fast vs. Slow Motion

The researchers also wondered: Does it matter if we check the animal's position every 30 minutes or every hour?

30-Minute Check: This is like watching a movie in high definition. You catch every tiny twitch and sudden turn. This was great for fast, agile animals like lions.
1-Hour Check: This is like watching the movie in standard definition. You miss the tiny twitches, but you get a clearer, smoother picture of the whole day.

The Result: Surprisingly, the 1-hour check worked better overall. Why? Because real-world data is messy. Animals sometimes lose their GPS signal. If you check every 30 minutes, you get a lot of "missing" pages in the story. Checking every hour filled in the gaps better and gave the computer a more consistent story to read, leading to better results across the board.

The Bottom Line

The paper proves that animals have a unique "movement signature." Just like you can recognize a friend by their walk even in a crowd, a computer can now recognize a lion, an elephant, or a baboon just by how they move around their day.

By using a smart "Super-Reader" model (Transformer) and looking at the big picture of movement (speed, turns, and direction) over a consistent 1-hour schedule, the computer can identify wildlife species accurately, even when it has never seen that specific animal in that specific place before. This means we might one day be able to monitor wildlife populations simply by analyzing their movement patterns, without needing to tag every single animal or know exactly where they are.

Technical Summary: Transformer-Based Wildlife Species Classification from Daily Movement Trajectories

Problem Statement

The paper addresses the challenge of inferring wildlife species identity solely from daily movement trajectories, without relying on absolute geographic location, habitat descriptors, or environmental covariates. While animal-borne telemetry data (e.g., from Movebank) has accumulated significantly, existing models often fail to generalize across different studies and regions. This failure is attributed to models learning region-specific artifacts (e.g., fences, roads, human activity) rather than intrinsic species movement behaviors. Furthermore, prior evaluations frequently utilize random train-test splits, leading to "geographic leakage" where models memorize site-specific routines rather than generalizable movement patterns. Additionally, the heterogeneity of temporal sampling resolutions across studies complicates the comparison of movement dynamics, as resampling can suppress ecologically informative short-term patterns or introduce noise.

Methodology

Data Source and Preprocessing

The study utilizes GPS telemetry data from Movebank, covering seven African wildlife species: baboon, buffalo, caracal, zebra, elephant, lion, and wildebeest. The dataset spans 16 distinct studies conducted between 1998 and 2023 across diverse regions.

Temporal Standardization: Trajectories were resampled to two fixed grids: 1-hour and 30-minute intervals. Linear interpolation was applied only when a single intermediate point was missing; larger gaps were excluded to avoid artificial movement inference.
Daily Segmentation: Trajectories were segmented into daily sequences (UTC) to prevent temporal leakage. Days with fewer than 12 (1-hour) or 25 (30-minute) observations were discarded.
Encoding: To avoid reliance on absolute coordinates, latitude and longitude were projected onto a unit sphere to compute displacement vectors. Time was encoded cyclically (sine/cosine) to capture circadian rhythms.

Feature Engineering

Two feature sets were evaluated:

Baseline (5D): Spatial displacement vectors ( $\Delta x, \Delta y, \Delta z$ ) and cyclic time-of-day encodings.
Augmented (10D): The baseline plus kinematic descriptors: step length/speed, movement direction (bearing), and turning angles. These were encoded cyclically to handle discontinuities.

Gap Awareness: A masking strategy was employed where movement-derived features were marked as undefined (masked) if the time gap between observations exceeded the nominal sampling interval, preventing the model from learning spurious patterns from interpolation.

Model Architecture

The primary model is a Transformer-based sequence encoder adapted for daily trajectory classification.

Input: Daily sequences of up to 24 (1-hour) or 48 (30-minute) timesteps.
Mechanism: Input timesteps are projected to an embedding space, augmented with sinusoidal positional encodings, and processed through a learnable [CLS] token and multi-head self-attention layers.
Baselines: The Transformer was compared against LSTM, 1D CNN (with parallel branches), and Temporal Convolutional Networks (TCN).
Evaluation Protocol: A rigorous study-level holdout strategy was used. For each species, one entire telemetry study was held out for testing, while the remaining data was split into training and validation sets at the individual-animal level. This ensures the model is tested on unseen regions and sampling protocols.

Key Contributions and Results

1. Superiority of Transformer Models

The Transformer architecture consistently outperformed LSTM, CNN, and TCN baselines across most species and experimental settings.

Performance Gains: The Transformer achieved balanced accuracy improvements of approximately 8 to 22 percentage points over baselines.
Elephant Classification: In a binary classification task for elephants (1-hour resolution), the Transformer achieved a balanced accuracy of 0.83 and an AUC of 0.92, significantly outperforming baselines (which ranged from 0.68–0.77 balanced accuracy).
Generalization: The model demonstrated robust generalization across distinct movement syndromes and geographic regions, suggesting it successfully captured intrinsic species-level kinematic signatures.

2. Impact of Feature Augmentation

The study demonstrated that minimal displacement data is often insufficient for distinguishing complex behavioral patterns.

Augmentation Benefits: Adding speed, bearing, and turning angle features resulted in substantial performance gains. For instance, the balanced accuracy for the Lion classifier improved from 0.59 to 0.85, and for Wildebeest from 0.70 to 0.84.
Underrepresented Species: Feature augmentation was particularly critical for underrepresented or sparsely represented species (e.g., large carnivores), where data scarcity made learning from raw displacement alone difficult.

3. Temporal Resolution Analysis

The paper compared 1-hour and 30-minute resolutions.

Trade-offs: While 30-minute resolution captured short-term dynamics beneficial for highly mobile species (e.g., Lions, Caracals), the 1-hour resolution yielded more consistent and promising performance across the multi-study dataset.
Reasoning: The 1-hour resolution reduced missing data issues inherent in heterogeneous sampling protocols and provided more stable daily patterns for herd-based species (Elephant, Buffalo, Wildebeest).

Significance and Claims

The paper claims that daily movement trajectories encode discriminative, species-specific patterns that remain informative across different regions and telemetry studies, independent of location-specific routines.

Validation Rigor: By enforcing study-level holdouts, the authors argue that their results reflect true species-level movement signatures rather than artifacts of study-specific sampling or geography.
Practical Utility: The findings suggest that attention-based models are well-suited for wildlife monitoring in areas where labeled data is scarce or where new monitoring locations are added, as the models rely on intrinsic behavior rather than site-specific idiosyncrasies.
Methodological Shift: The work highlights the necessity of moving beyond random cross-validation in movement ecology to prevent geographic leakage and overestimation of model performance.

The authors conclude that while finer sampling (30-min) has niche benefits, a unified 1-hour resolution combined with kinematic feature augmentation and Transformer-based modeling offers the most robust approach for cross-study species classification.