IoT case study: field stations that identify birds by sound, with set-up cut to under 10 minutes

The challenge: simple field setup and reliable detection outdoors

The client sells outdoor monitoring stations to researchers, conservationists and people who simply enjoy listening to wildlife. The goal was monitoring that's simpler and cheaper than specialist field kit, without a complicated set-up, and that works at scale for scientists and the public alike.

The product needed to meet four requirements:

• Set-up. Most systems for monitoring wildlife need manual calibration and a technical installation, and that holds back how widely they're used. The client asked for stations its own customers could get running in under 10 minutes, with no engineer involved.
• Noise. Wind, rain, insects and other outdoor sounds can pull detection accuracy down, depending on the recording. The system had to analyse audio continuously and still identify what it heard reliably.
• Sync. Audio streams, sensor data and species detections had to stay current across the stations, the cloud, the mobile app and the web dashboards, with no dropped sessions.
• Scale. Every station added means more processing, and performance had to hold as the number of stations grew.

What we built: machine learning for noisy field audio, and a set-up that starts with a QR code

Six pieces do the work:

• Species recognition on adapted BirdNET models. BirdNET models, which can identify thousands of bird species, were adapted to name the birds in a station's live audio as it streams.
• Noise reduction for outdoor audio. Audio filtering and noise reduction went in, so detection holds up better when wind, rain or other sounds overlap the birdsong. With the optimised models, they raised detection accuracy by about 18 to 25% during the test.
• A mobile app for the stations. Built in React Native for iOS and Android, it handles live listening, live monitoring and device management.
• Shopify for buying, a QR code for activating. Customers buy a station on a Shopify site and activate it with a QR code. With a simpler onboarding flow, that brought average set-up from nearly 25 to 30 minutes to under 10.
• A cloud back end that keeps everything in step. Python and FastAPI handle the API, device control, live audio processing and the sync between stations and dashboards. PostgreSQL stores detections, activity, users and monitoring logs. AWS carries the live streams, storage and processing, built to take more stations without affecting monitoring or sync, and infrastructure work cut live-stream interruptions from approximately 8 to 10% to under 3%.
• A dashboard and a listening map. A live monitoring dashboard on the web, and a worldwide map of linked stations, wildlife activity and species detections. The platform also supports adding environmental sensors.

Where the model stops, and who takes over

The model has one job here: naming the bird species in a station's live audio. Keeping the stream up and getting a station running in the first place are engineering problems, and two of the three results on this page come from those.

Detection accuracy, up about 18 to 25% during the test, is the one result that belongs to the models, and they share it with the noise reduction. The quicker set-up came from QR activation and the onboarding flow, and the fewer interruptions from the infrastructure.

Results

Researchers can reach live field data sooner, while wildlife enthusiasts and educators have a simpler way to take part in conservation.

Technologies and tools

Timeline

The build took five months.

If you're weighing whether an existing model could be adapted to your own data, our page for anyone choosing a machine learning development company sets out what to check before you hire one, us included. If the model sits inside something your customers buy and use, our AI product development services start from what each user costs to serve.