Smartphone Mosquito Wingbeat Detection

ISEF Category: Computational Biology and Bioinformatics

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

The Hook

A mosquito can sound like a tiny drone before you ever see it. That means your phone may catch clues your eyes miss. If you can turn that sound into a species label, you can help map dengue risk faster. This project mixes audio, machine learning, and public health.

What Is It?

Mosquitoes beat their wings at species-specific speeds, so their buzz carries a kind of acoustic fingerprint. Your phone records sound. A model then looks for patterns in that sound, just like it might spot a melody or a bird call. In this project, you would test whether a foundation model trained on bird sounds can still learn mosquito wingbeats after fine-tuning on mosquito recordings.

Think of the model as a student who already knows how to listen for structure in audio. BirdNET embeddings turn each sound clip into a map of features, which are hidden measurements the model uses to tell sounds apart. Contrastive fine-tuning teaches the model which mosquito clips should be close together and which should be far apart. If it works, you can compare species, test recording conditions, and see whether the same system could support citizen-science dengue monitoring.

Why This Is a Good Topic

This is a strong science fair topic because you can test a real machine-learning idea with a clear yes-or-no outcome. You can measure accuracy, confusion between species, and how well the model handles noisy phone recordings. The topic connects to mosquito control and dengue risk mapping, which gives it real-world value. You can also learn data labeling, audio preprocessing, model evaluation, and basic deployment thinking.

Research Questions

• How does contrastive fine-tuning change mosquito species classification accuracy compared with using BirdNET embeddings alone?
• What is the effect of smartphone microphone quality on mosquito wingbeat detection accuracy?
• Does adding background-noise augmentation improve model performance on real-world field recordings?
• To what extent can a model trained on one mosquito species group generalize to a different group from the same genus?
• Which feature set, raw audio, BirdNET embeddings, or handcrafted acoustic features, separates mosquito species best?
• How does recording distance from the mosquito change the model's confidence and error rate?
• What is the effect of training set size on the model's ability to detect rare mosquito species?

Basic Materials

• Smartphone with voice recording app.
• Laptop or desktop computer.
• Free audio annotation tool such as Audacity.
• Spreadsheet software for tracking samples and labels.
• External microphone or phone with a known microphone model.
• Quiet indoor recording space.
• Mosquito audio dataset from a public research source or your own supervised recordings.
• Headphones for checking clip quality.

Advanced Materials

• High-quality USB microphone.
• Microphone stand or fixed recording enclosure.
• Computer with GPU access through a university lab or cloud credits.
• Python environment with audio and machine learning libraries.
• Annotated mosquito wingbeat dataset with species labels.
• Sound level meter for recording environment checks.
• Controlled mosquito rearing or access to preserved reference recordings through a lab.
• GPS-enabled field data collection app for deployment testing.

Software & Tools

• Python: Runs audio preprocessing, feature extraction, model training, and evaluation.
• Audacity: Lets you inspect clips, trim noise, and verify label quality.
• ImageJ: Can help inspect spectrogram exports when you need quick visual checks.
• scikit-learn: Supports baseline classifiers, cross-validation, and confusion matrices.
• librosa: Extracts audio features such as spectra, pitch-related measures, and embeddings.

Experiment Steps

1. Define the mosquito species, recording conditions, and classification task you want to test.
2. Choose one benchmark pipeline, then decide whether BirdNET embeddings, raw audio, or handcrafted features will be your main comparison.
3. Plan a labeling rule for clips so your training data stays consistent across speakers, devices, and noise conditions.
4. Build a validation strategy that separates training, tuning, and test data by location, device, or collection day.
5. Decide which metrics will matter most, such as accuracy, recall for rare species, and confusion between similar wingbeat patterns.
6. Plan a field-test or citizen-science simulation that checks whether the model still works on messy real-world recordings.

Common Pitfalls

• Using recordings from the same phone session for both training and testing, which inflates accuracy.
• Labeling clips from background noise instead of clear wingbeats, which teaches the model the wrong signal.
• Mixing species with very different recording distances, which makes the model learn volume instead of wingbeat structure.
• Ignoring class imbalance, which lets common species dominate the results and hide poor rare-species performance.
• Skipping a site-separated test set, which makes the field deployment look stronger than it really is.

What Makes This Competitive

A competitive version of this project would test more than one model path and compare them with honest held-out data. You could ask whether transfer learning from bird audio really helps, or whether it breaks under noisy phone recordings. Strong projects also include a harder validation setup, like testing on a different device, location, or collection day. If you connect the classifier to a realistic risk-mapping workflow, your project feels closer to a usable tool than a class demo.

Project Variations

• Compare BirdNET embeddings against a small custom CNN trained directly on mosquito spectrograms.
• Test whether the model works better on single-species recordings or mixed-field recordings with overlapping insect sounds.
• Build a city-block level dengue-risk prototype that maps predicted mosquito presence from citizen-submitted phone audio.

Learn More

• PubMed: Search for review articles on mosquito wingbeat acoustics, species classification, and bioacoustic monitoring.
• NIH NCBI Bookshelf: Find background chapters on vector-borne disease and data analysis methods.
• NOAA National Centers for Environmental Information: Explore public guidance on environmental data handling and mapping workflows.
• MIT OpenCourseWare: Look for free machine learning and signal processing course notes to review model evaluation and audio features.
• BirdNET papers in peer-reviewed journals: Search Google Scholar or PubMed for the original BirdNET foundation-model articles and related transfer learning studies.
• USGS Nonindigenous Aquatic Species database: Use it as a model for how biodiversity monitoring databases structure species records and locations.