An audio-based drone detection and classification system that identifies drones by their acoustic signature using mel-spectrograms and a CNN. Built as a defense/security ML project — the idea is that drones produce distinct harmonic patterns from their propellers, and a model can learn to spot these even in noisy environments.
- Detects whether a drone is present in a 1-second audio clip
- Classifies between two drone types (Bebop, Mambo) and non-drone sounds
- Tested under different noise levels (SNR 0-20dB) to simulate real-world conditions
| Condition | Accuracy | Drone Recall |
|---|---|---|
| Clean | 97.1% | 97.0% |
| 20dB SNR | 93.4% | 95.5% |
| 10dB SNR | 88.9% | 91.5% |
| 5dB SNR | 84.9% | 87.0% |
| 0dB SNR | 79.1% | 81.5% |
- Open
Drone_Detection.ipynbin Google Colab - Select Runtime → Change runtime type → T4 GPU
- Run all cells — dataset downloads automatically (~200 MB)
- Python, PyTorch, librosa
- Mel-spectrogram feature extraction
- CNN with 3 conv layers, early stopping, class weights
- SNR-based robustness evaluation
DroneAudioDataset by Sara Al-Emadi (IWCMC 2019) — 11,704 audio samples across 3 classes (Bebop drone, Mambo drone, non-drone environmental sounds).
- Scale to DADS dataset (180k samples, 15+ drone models)
- Test with real-world noise (wind, traffic) instead of white noise
- Try Audio Spectrogram Transformer architecture