DroneDetect V2 - RF Classification Pipeline

Machine learning pipeline for drone RF signal classification using the DroneDetect V2 dataset.

References

Paper	Contribution
Swinney & Woods, 2021	DroneDetect dataset, PSD vs spectrogram comparison
RF-UAVNet (IEEE, 2022)	1D CNN architecture for drone classification (raw IQ signals)
Kiliç et al., 2021	RF signal classification with machine learning
Swinney et al., 2021	Flying mode classification via ResNet50

Dataset

DroneDetect V2 (Swinney & Woods, 2021)

• 60 MHz sampling rate
• 2.43 GHz center frequency
• 7 drone types: DJI Air 2S, Parrot Disco, DJI Inspire 2, DJI Mavic Mini, DJI Mavic Pro 1/2, DJI Phantom 4
• 3 states: ON, Hovering (HO), Flying (FY)
• 4 interference conditions: CLEAN, Bluetooth, WiFi, Both

Installation

# Install uv (if not already installed)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Clone and install with PyTorch CPU
git clone <repo-url> && cd mldrone
uv sync --extra torch-cpu

# OR with PyTorch GPU (CUDA 12.1)
uv sync --extra torch-gpu

# Optional: add visualization tools (torchview, torchinfo)
uv sync --extra torch-cpu --extra viz

# Extract dataset
unzip /path/to/DroneDetect_V2.zip -d ./data/

# Test installation
uv run python -c "import torch; print(torch.__version__)"

Configuration

cp .env.example .env

Variable	Default	Description
DRONEDETECT_DATA_DIR	./data/raw	Raw dataset directory
DRONEDETECT_FEATURES_DIR	./data/features	Extracted features
DRONEDETECT_MODELS_DIR	./models	Trained models
DRONEDETECT_METADATA_CACHE	./data/metadata_cache.parquet	Metadata cache

Usage

1. Explore Dataset

Open notebooks in your IDE (VSCode, PyCharm, etc.) with the .venv Python interpreter. (check out figures folder and related pdf for output)

2. Extract Features

Run preprocessing notebooks to extract:

• PSD features (for SVM)
• Spectrograms (224x224 for CNNs)
• Downsampled IQ (2x10000 for RF-UAV-Net)

Features saved to ./data/features/*.npz

3. Train Models

• 021_training_svm.ipynb: SVM on PSD features
• 022_training_cnn.ipynb: VGG16 + ResNet50 on spectrograms
• 023_training_rfuavnet.ipynb: 1D CNN on raw IQ

Models saved to ./models/*.pth or *.pkl

4. Run Inference

uv run streamlit run interface/app.py

See inference page

Models

Model	Input
SVM (RBF)	PSD
VGG16-FC	Spectrogram
ResNet50-FC	Spectrogram
RF-UAV-Net	Raw IQ

License

MIT