context_classification.md

# RFUAV Classification Scope

We will use RFUAV for the first narrow-band target segment: Mavic/Autel-class COTS drones. The challenge also mentions Supercam (~1.3 GHz), Zala/Lancet (~2.2 GHz), and Merlin/Orlan (~2.6 GHz), but RFUAV does not provide public labeled signatures for those military UAVs. Therefore, we avoid overclaiming and train only on classes that match the available public data and the Mavic/Autel requirement.

## Final Selected RFUAV Classes

- DJI MAVIC3 PRO
- DJI MINI3
- DJI MINI4 PRO
- DAUTEL EVO nano

## Operating Frequency Segment

These classes represent the Mavic/Autel operational segment, mainly:

- 2.4 GHz band: 2.400-2.4835 GHz
- 5.8 GHz band: 5.725-5.850 GHz
- Some newer DJI O4 systems may also support 5.170-5.250 GHz where regulations permit.

Exact RFUAV capture frequencies should be read from the dataset XML metadata for each class. The model scope is intentionally limited to the 2.4/5.8 GHz COTS drone segment.

## Why Only These Classes

This 4-class subset is the practical hackathon scope: about 18.6 GB compressed instead of the full RFUAV dataset. It focuses on Mavic/Autel-type targets, avoids unrelated RC controller classes, and leaves enough time to download, extract, inspect raw IQ data, preprocess windows or spectrograms, and train a baseline model.

We will use Hugging Face compute so the raw RFUAV archives do not need to be stored locally. The expected workflow is to download/extract the selected RFUAV archives remotely, create a smaller derived dataset, and train from that processed dataset.

## Modeling Approach

We will work from raw IQ windows, not only the processed image preview. An IQ window is a fixed slice of complex RF samples: I/Q pairs over time.

Planned input representations:

- FFT vector: raw IQ window -> log-power FFT vector. Use as a fast baseline with Random Forest, SVM, MLP, or a small 1D CNN.
- Raw IQ: I and Q as two time channels, shape [2, N]. Use a 1D CNN to preserve time-domain and phase information.
- Spectrogram: raw IQ -> FFT over time/STFT -> log-power time-frequency matrix. Use a 2D CNN such as ResNet18 or EfficientNet-B0 as the main classifier.

The spectrogram 2D CNN is likely the strongest demo model, while the FFT and raw-IQ models provide RF-native baselines and make the signal-processing pipeline defensible.