AI Animal GPS Track Reconstruction

ISEF Category: Animal Sciences

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Animal GPS collars do not always send perfect lines on a map. A few missing points can turn a clear trail into a broken scribble. You can treat that gap as a puzzle, then ask whether a diffusion model can rebuild the most likely path. That makes this project a mix of wildlife tracking, map reading, and machine learning.

What Is It?

Animal-tracking GPS data records where an animal moves over time. When the signal drops, the trace gets degraded, which means parts of the path go missing or look messy. Think of it like a torn strip of film or a connect-the-dots page with half the dots gone.

A diffusion model is a type of generative AI that learns how to turn noise into a likely pattern. In this case, the pattern is not a face or an image, but a movement path. You train it on public Movebank tracks, then ask it to fill in missing sections of a broken route. The goal is not to guess the exact truth, but to recover a path that fits the animal's movement style and the context around the gap.

Why This Is a Good Topic

This is a strong science fair topic because you can test it with public data, clear metrics, and a real problem. Wildlife researchers deal with missing GPS fixes all the time, so your project connects to animal movement, conservation, and data quality. You also get to learn data cleaning, model training, error analysis, and how to compare AI output with simpler baselines.

Research Questions

• How does the length of missing GPS gaps affect reconstruction error?
• What is the effect of species movement style on diffusion-model accuracy?
• Does adding time-of-day information improve track reconstruction?
• To what extent do habitat labels improve the realism of rebuilt paths?
• Which performs better on degraded traces, diffusion reconstruction or linear interpolation?
• Does training on one Movebank study transfer to a different study with similar sampling rates?

Basic Materials

• Laptop or desktop computer with internet access.
• Public Movebank study data with GPS coordinates and timestamps.
• Python installed with Jupyter Notebook.
• Spreadsheet software for tracking train, validation, and test splits.
• External drive or cloud storage for data backups.
• A basic plotting library such as Matplotlib or Seaborn.

Advanced Materials

• GPU workstation or access to a university compute cluster.
• PyTorch environment with CUDA support.
• QGIS for map-based inspection of animal tracks.
• Higher-resolution telemetry metadata from a research collaboration.
• Statistical software for mixed-effects or spatial analysis.
• Version-controlled storage for model checkpoints and experiment logs.

Software & Tools

• Python: Runs data cleaning, training, and evaluation scripts.
• PyTorch: Trains the diffusion model and baseline neural networks.
• Jupyter Notebook: Lets you inspect tracks, plots, and model outputs interactively.
• GeoPandas: Reads and processes spatial track files.
• QGIS: Displays animal paths on maps and helps spot odd reconstructions.

Experiment Steps

1. Define the reconstruction target by deciding what gap lengths, sampling rates, and species you will include.
2. Split the data so the model trains on some animals or studies and tests on held-out ones.
3. Choose the features that describe movement context, such as time gaps, turning angle, speed, and habitat labels if available.
4. Set a baseline method so you can compare the diffusion model against a simpler interpolator or sequence model.
5. Plan metrics that measure both point error and path realism, then test them across several gap sizes.
6. Check whether the model stays accurate across species, seasons, or movement styles, not just on one easy dataset.

Common Pitfalls

• Training on one animal and testing on nearby points from the same animal, which leaks movement patterns into the test set.
• Using only average position error, which can miss unrealistic jumps, loops, or backtracking.
• Mixing studies with very different sampling rates, which confuses the model and inflates noise.
• Removing GPS points only at random, which does not mimic long collar dropouts or missed fixes in the field.
• Ignoring coordinate projection and cleaning steps, which can bend the map and corrupt distance calculations.

What Makes This Competitive

A stronger entry would not just fill in missing points. It would compare the diffusion model against simple interpolation, then test where each method breaks by gap length, species, and habitat. You can raise the bar by reporting uncertainty, not just best guesses, and by checking whether reconstructed paths still match known movement rules. That kind of analysis shows careful modeling, not just a flashy result.

Project Variations

• Try bird migration tracks instead of mammal movement to see whether long, directional travel changes reconstruction quality.
• Compare clean summer tracks with winter or urban tracks, where signal loss and movement behavior may be harder to model.
• Add habitat or land-cover layers to test whether environmental context helps the model recover more realistic paths.

Learn More

• Movebank: Search public animal movement studies, metadata, and GPS track examples.
• PubMed: Look for review articles on animal telemetry, movement ecology, and missing-data methods.
• USGS Wildlife Ecology Program: Find free publications on wildlife tracking and movement analysis.
• PyTorch Tutorials: Learn how diffusion models are built and trained with open documentation.
• QGIS Documentation: Use the free manual to inspect GPS tracks on maps and check spatial patterns.