# AURA offline research baseline

An actual CPU logistic-regression classifier was fitted on the training split. No pretrained encoder, synthetic augmentation, live sensor test, or operational evaluation was used.

**These scores are source-confounded and come from one positive campaign. They are not evidence of field detection performance.**

| Split | Files | Source groups | Windows | Balanced accuracy | Drone precision | Drone recall | Drone F1 |
|---|---:|---:|---:|---:|---:|---:|---:|
| train | 78 | 50 | 98 | 0.750 | 1.000 | 0.500 | 0.667 |
| validation | 28 | 20 | 42 | 0.958 | 0.667 | 1.000 | 0.800 |
| test | 28 | 21 | 42 | 0.750 | 1.000 | 0.500 | 0.667 |

Metrics above are at file level, with each file score the mean of its non-overlapping five-second window scores. The JSON also reports means aggregated by original source recording, so multiple ESC excerpt takes are not mistaken for independent sources.

Threshold: `0.9249967798`, selected on validation only. Test confusion matrix, rows=true and columns=predicted, order background/drone: `[[24, 0], [2, 2]]`.

## Data and permissions

14 Salford M1 flyover event WAVs: 62,001,120 bytes; 120 CC-BY ESC-10 clips (40 each rain/helicopter/chainsaw): 52,925,280 bytes. Total audio: 114,926,400 bytes. All files have SHA-256 provenance and upstream integrity checks in `data/manifest.json`.

Salford v3: https://salford.figshare.com/articles/dataset/DroneNoise_Database/22133411/3 — CC BY 4.0. ESC-10-only license and individual source credits: https://github.com/karolpiczak/ESC-50/blob/33c8ce9eb2cf0b1c2f8bcf322eb349b6be34dbb6/LICENSE — CC BY 3.0 for the subset, with original CC0/CC-BY credits retained. Both licenses permit commercial reuse and redistribution subject to their terms. The full ESC-50 dataset is noncommercial and was not selected.

## Reproduction and inference

```sh
python3 scripts/download_data.py
python3 scripts/train_baseline.py
python3 -B -m unittest discover -s tests -v
python3 scripts/predict.py --audio data/raw/salford/Ed_3p_10_F15_N_W_dw_ev1_M1.wav --model models/baseline.joblib
```

Dependencies are pinned in `requirements.txt`. The model contains only the training-fitted scaler and classifier; validation selected the threshold. Audio normalization is deterministic per window. Splits precede windowing; all source takes, microphone counterparts, or future augmentations must remain in the same original-recording group.

Model SHA-256: `dee018de0c1a950a89e5161395708ca70825cd0ec6df8e2f2558b2f84678e189`. Machine-readable results and environment/source hashes are in `reports/metrics.json`. Only load trusted local joblib models.

## Limitations

- All positives are from Salford and all negatives from ESC-10. Source, recorder, processing, and environment are confounded with class; a good score can reflect collection recognition.
- Salford is one Edzell campaign on 2022-08-17. Events are split disjointly, but sessions/campaigns are not independent. Continuous master-recording IDs are unavailable; filename event groups are the strongest released identifiers used here.
- Only 14 positive events and four publisher drone codes are present, with each code seen in training. This is not an unseen-aircraft, unseen-site, or unseen-campaign evaluation.
- Salford provides event-level flyover labels, not temporal drone onset/offset annotations. ESC-10 categories are negative proxies, not independently certified drone-absence recordings. No exhaustive listening audit was performed.
- The negative subset contains rain, helicopters, and chainsaws only; it does not represent all environmental backgrounds or the deployment prevalence of either class.
- The fitted logistic-regression output is an uncalibrated model score. The validation-selected threshold is a research setting, not an operational alert threshold.
- These are passive microphone recordings. No emitted probe, active sensing, live physical-sensor/drone test, localization, range measurement, or operational validation was performed. Speaker playback would test a replay chain, not live-drone detection.
- Results describe this recorded-data split only. Raw files and model artifacts remain local and ignored. Source copyright licenses do not themselves certify downstream model clearance or endorse this work.