Forest fires spread fast, and the places they start rarely have good internet. That was the starting point for FIRO, our edge AI wildfire detection system.
Why run the model on the device?
Sending every camera frame to the cloud needs bandwidth and power that remote forests don't have. Running the model on the device means the camera only has to send an alert, not a video stream.
What we built
- Model: MobileNetV2 with transfer learning, trained on a custom dataset of 6,000+ fire and non-fire images. It reached 97% classification accuracy.
- Optimisation: the model was converted to TensorFlow Lite with INT8 quantisation, which makes it small and fast enough for a Raspberry Pi 5.
- Monitoring: a dashboard built with Flask, Dash and Plotly shows live status, and alerts are pushed as soon as fire is detected.
Lessons learned
- Start small. A compact model that runs reliably beats a large one that needs a GPU.
- Quantise early. Test the quantised model's accuracy as you go, not only at the end.
- Think about false alarms. Sunsets and red objects look like fire, so a varied dataset matters as much as the architecture.
FIRO started as a final-year research project and is now being prepared for publication. If you need computer vision that runs where the internet doesn't, get in touch.