Bird Call AI Mapping Backyard Biodiversity

ISEF Category: Robotics and Intelligent Machines

Subcategory: Machine Learning  ·  Difficulty: Advanced  ·  Setup: School Lab  ·  Time: Full Year

The Hook

Birds can sound like random background noise until you map them. Then the forest starts to look like a moving network of signals. You can turn those calls into species IDs, direction estimates, and a seasonal biodiversity map. That gives you a real AI project with ecology behind it.

What Is It?

This project combines two ideas. First, you train or use a bird-call classifier, which is a model that guesses the species from a sound clip. Second, you use a microphone array to estimate where the call came from. That lets you not only hear the bird, but also place it on a rough map.

Think of it like giving your project ears and a compass. The classifier answers, “What bird is that?” The microphone array answers, “Where was it?” TDOA means time difference of arrival. That is the tiny timing gap between when sound hits each microphone. If you measure those gaps well, you can estimate bearing, which is the direction of the call.

Once you collect detections over a season, you can compare your backyard map to public bird observations from eBird hotspot data. That comparison can show which species appear often, which ones cluster in certain parts of your yard, and how your local pattern matches the larger area.

Why This Is a Good Topic

This makes a strong science fair topic because you can test both machine learning and spatial signal processing with real-world data. You have clear variables, like classifier accuracy, direction estimate error, and species diversity over time. You also connect to a real ecological problem, which is monitoring local biodiversity with passive sensing. A student can learn data labeling, model evaluation, noise handling, and basic statistics without needing a biology lab.

Research Questions

• How does microphone spacing affect bearing error in backyard bird-call localization? ?
• What is the effect of different background noise levels on species classification accuracy? ?
• Does an on-device Coral TPU model match a laptop model for bird-call species identification? ?
• To what extent does adding direction estimates improve the accuracy of backyard biodiversity maps? ?
• Which bird species are easiest to classify from short outdoor audio clips? ?
• How does your seasonal species list compare with eBird hotspot observations in the same area? ?

Basic Materials

• Coral TPU dev board or compatible on-device AI board.
• 4 INMP441 microphones.
• Microcontroller or single-board computer with audio input support.
• Microphone array mounting frame.
• SD card or external storage for audio logs.
• Tripod or fixed outdoor mounting pole.
• Laptop for data labeling and analysis.
• Weather-safe enclosure for outdoor recording hardware.
• Measuring tape for microphone spacing and site layout.
• Notebook or digital log for timestamps, weather, and species notes.

Advanced Materials

• Coral TPU accelerator and host computer.
• 4-channel synchronized audio acquisition system.
• Calibrated sound source for localization checks.
• Outdoor microphone wind protection.
• GPS or site mapping tool for ground-truth location logging.
• ImageJ or similar tool for visualizing map overlays if needed.
• Python environment for signal processing and model evaluation.
• Access to historical eBird checklist or hotspot data.
• Basic calibration equipment for timing and amplitude checks.
• Optional reference microphone for validation.

Software & Tools

• Python: Processes audio, runs evaluation scripts, and calculates localization and classification metrics.
• Audacity: Lets you inspect recordings, trim clips, and spot noise problems.
• Google Earth: Helps you mark observation points and compare detections across your yard.
• R: Supports statistical tests and seasonal comparison plots.
• eBird: Provides public hotspot and checklist data for comparison with your local observations.

Experiment Steps

1. Define the exact species list, recording site, and seasonal window you will study.
2. Choose one audio clip length and one detection rule so your results stay comparable.
3. Plan a calibration method for both species labels and bearing estimates before field collection.
4. Build a clean comparison baseline against a single-microphone setup or a classifier-only setup.
5. Decide how you will score performance with confusion matrices, localization error, and biodiversity indices.
6. Set up a seasonal logging plan that keeps weather, time, and habitat conditions attached to each recording.

Common Pitfalls

• Using unsynchronized microphones, which breaks TDOA timing and makes direction estimates look random.
• Training on too few bird species, which causes the model to overfit a tiny label set and fail outdoors.
• Recording near wind, traffic, or HVAC noise, which masks short bird calls and lowers classifier confidence.
• Mixing manual bird sightings with audio detections without a clear rule, which makes the biodiversity map hard to interpret.
• Comparing your backyard data to eBird without matching season, time of day, or habitat, which creates a misleading benchmark.

What Makes This Competitive

A strong version of this project does more than count detections. You can compare classifier-only, localization-only, and combined pipelines, then show which setup actually improves ecological mapping. Strong entries also include calibration data, error bars, and a clear test for how noise, distance, or weather changes performance. That gives your project a real engineering angle, not just a demo.

Project Variations

• Compare dawn versus midday bird-call detection accuracy in the same yard.
• Test whether adding a fifth microphone improves bearing estimates enough to justify the extra hardware.
• Replace the backyard site with a school courtyard and compare biodiversity maps between two habitats.

Learn More

• eBird from Cornell Lab of Ornithology: Search the site for hotspot data, checklists, and species range tools.
• Macaulay Library: Search for reference bird sounds and labeled recordings from Cornell Lab of Ornithology.
• NOAA National Weather Service: Use local weather records to compare noise, wind, and detection quality.
• PubMed: Search for review articles on passive acoustic monitoring, bird acoustics, and sound localization.
• USGS Bird Banding Laboratory: Read about bird monitoring methods and field identification context.
• MIT OpenCourseWare Signals and Systems: Find lecture material on sampling, time delay, and filtering.