Federated Bird Migration Analysis

ISEF Category: Animal Sciences

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: School Lab  ·  Time: Full Year

The Hook

Bird migration is like a living calendar, and climate change can shift that calendar by days or weeks. If you track those shifts early, you can spot ecological change before it shows up in bigger damage. Your challenge is to do that without pulling every photo into one giant database. Federated learning gives you a way to train across many phones while keeping the raw images local.

What Is It?

This phenomenon asks you to use federated learning, a method where devices train a shared model without sending raw data to one central server. Each phone learns from its own bird photos, then shares only model updates. Think of it like many students studying the same chapter in different rooms, then combining notes instead of handing over their notebooks.

The science target is bird migration timing. That means the first spring sighting, the last fall sighting, or the week when a species appears in a new place. If your model can pick up those timing shifts from citizen-scientist images, you can estimate whether birds are arriving earlier or later across seasons or regions.

Why This Is a Good Topic

This is a strong science fair topic because you can test a clear signal, compare model designs, and measure how well privacy-preserving training performs against central training. It connects to climate change, wildlife monitoring, and citizen science, so the real-world value is easy to explain. You can learn practical machine-learning skills, data cleaning, and experimental design without needing a wet lab.

Research Questions

• How does federated learning compare with centralized training for detecting seasonal changes in bird sighting dates?
• What is the effect of the number of participating phones on model accuracy for bird-migration timing prediction?
• Does adding location metadata improve early migration-shift detection compared with image data alone?
• To what extent does class imbalance across bird species affect the model's ability to detect timing shifts?
• Which image features matter most when predicting whether a bird photo belongs to an early or late migration window?
• How does client-to-client variation in lighting, device quality, and background affect federated model performance?

Basic Materials

• Laptop or desktop computer with at least 16 GB RAM.
• Smartphone or tablet for collecting sample images or testing a mobile workflow.
• Public bird image datasets from citizen-science platforms or open repositories.
• Spreadsheet software for tracking labels, dates, and locations.
• Python installed with a free notebook environment.
• External hard drive or cloud storage for backups.
• Stable internet access for downloading datasets and documentation.

Advanced Materials

• GPU-enabled workstation or university compute cluster.
• Multiple test phones or emulators for simulating client devices.
• Python machine-learning stack with TensorFlow Federated or PySyft.
• Image annotation tool for species and date labels.
• Version-controlled dataset splits with train, validation, and holdout sets.
• Secure server or local federated-learning coordinator.
• Statistical software for mixed-effects modeling and significance tests.

Software & Tools

• Python: Builds data pipelines, trains models, and runs analysis scripts.
• Jupyter Notebook: Lets you explore images, labels, and results in one place.
• TensorFlow Federated: Trains models across devices without moving raw images.
• ImageJ: Measures image quality, brightness, and other visual properties.
• R: Runs statistical tests and season-by-season comparisons.

Experiment Steps

1. Define the bird species, region, and seasonal window you want to study.
2. Decide what each phone will learn locally, and what summary update it will send.
3. Build a clean label scheme for species, date, location, and migration window.
4. Choose a baseline central model so you can compare federated performance against it.
5. Plan controls that separate true timing shifts from changes in photo quality or sampling bias.
6. Set evaluation metrics that reward both prediction quality and privacy-preserving design.

Common Pitfalls

• Mixing bird species with different migration patterns, which hides the timing signal you want to measure.
• Using photos from the same location in every split, which makes the model look better than it really is.
• Letting one phone contribute far more images than the others, which skews the federated updates.
• Ignoring lighting and background differences, which can make the model learn camera conditions instead of bird timing.
• Treating first sighting dates as ground truth without checking observer effort, which can create fake migration shifts.

What Makes This Competitive

A stronger version of this project would do more than train a model. You would test whether federated learning stays accurate under uneven data, noisy labels, and different phone types. You could also compare multiple aggregation rules, then use a harder metric like calibration or early-warning lead time, not just accuracy. That kind of design shows real control over both the ecology and the machine-learning side.

Project Variations

• Track spring arrival timing for one migratory species across two regions with different climate patterns.
• Compare image-only models with image plus location models to see whether metadata improves timing detection.
• Test whether federated learning still works when client phones have very uneven numbers of bird photos.

Learn More

• Cornell Lab of Ornithology eBird: Search for migration timing, range maps, and citizen-science bird observations.
• USGS Bird Banding Laboratory: Find migration and movement resources from a federal bird research source.
• NOAA Climate Data Online: Look up temperature and seasonal data to compare against migration shifts.
• NASA Earthdata: Search for remote-sensing data that may help explain habitat or seasonal change.
• PubMed: Search review articles on bird migration phenology, citizen science, and machine learning.
• NIH Open Access PubMed Central: Find free full-text papers on federated learning and ecological monitoring.