Real-Time Classification of Forest Fire Sound Using CNN on SONY Spresense Microcontroller

This paper introduces an autonomous system for fire detection using real-time audio processing and artificial intelligence on the Sony Spresense microcontroller. The system captures environmental sounds, applies low-pass filtering and downsampling, and generates spectrograms for classification by a Convolutional Neural Network. Trained on a dataset of 8,129 fire-related and 5,000 non-fire audio samples, the model achieved a 99.66% F1-score. Initial testing revealed variability in detection, but retraining on an expanded dataset improved accuracy, consistently identifying over 70% of fire sounds. The final 136 KB model balances performance and memory constraints, providing a solution for early wildfire detection with minimal human involvement.