planequery-flights-algorithm

The world’s best ADS-B → flights algorithm.

See paper/Evaluating-and-Improving-Flight-Derivation-from-ADS-B-Data-Draft-Submission.pdf

Performance has been measured against OpenSky’s and ADSBExchange’s algorithms using SWIM SFDPS and Bureau of Transportation Statistics (BTS) flight data. It has also been tested against EUROCONTROL flight data. To assess the algorithm’s performance yourself, the repository includes flights generated from ADSB.lol ADS-B data from 2026-06-01 to 2026-07-01: data/flights/algorithm/v1/adsblol/year=2026/month=06.

The algorithm uses gradient-boosting ML for airport determination.

Test date: 2026-03-01 UTC

adsb_src	test_src	gold_src	true_positive	false_positive	false_negative	precision	recall	f1	takeoff_airport_ident_match_pct	landing_airport_ident_match_pct	airport_ident_match_pct
adsb.lol	algorithm	sfdps	6,322	38	52	0.9940	0.9918	0.9929	99.24%	98.54%	97.90%
adsb.lol	algorithm	bts	3,320	2	10	0.9994	0.9970	0.9982	99.76%	99.73%	99.49%
adsbx	algorithm	sfdps	6,357	61	17	0.9905	0.9973	0.9939	99.91%	99.80%	99.72%
adsbx	algorithm	bts	3,330	1	0	0.9997	1.0000	0.9998	100.00%	100.00%	100.00%
adsbx	adsbx	sfdps	5,991	60	378	0.9901	0.9407	0.9647	99.88%	99.92%	99.80%
adsbx	adsbx	bts	3,226	2	102	0.9994	0.9694	0.9841	100.00%	100.00%	100.00%
opensky	opensky	sfdps	5,820	219	553	0.9637	0.9132	0.9378	73.25%	67.47%	50.95%
opensky	opensky	bts	3,269	46	63	0.9861	0.9811	0.9836	83.94%	75.41%	63.32%

Notes:

• precision, recall, and f1 are flight-segmentation metrics.
• Airport metrics are calculated only for flights that were successfully segmented.

Eurocontrol data

test_date	adsb_src	test_src	gold_src	true_positive	false_positive	false_negative	precision	recall	f1	takeoff_airport_ident_match_pct	landing_airport_ident_match_pct	airport_ident_match_pct
2024-06-01	adsb.lol	algorithm	eurocontrol	6,154	87	182	0.9861	0.9713	0.9786	0.9594	0.9366	0.9007

Setup

This repository uses Git LFS for model and data files.

# First install Git LFS using your platform's package manager or installer
git lfs install
git clone https://github.com/PlaneQuery/planequery-flights-algorithm.git
pip install .

Evaluation

Downloads ~3GB from https://github.com/adsblol/globe_history_2026/releases and runs processing that takes ~5 minutes on M4 Macbook.

python src/flights/evaluation/evaluation.py \
  --adsb-src adsblol adsbx \
  --test-src algorithm adsbx opensky \
  --gold-src sfdps bts \
  --test-dates 2026-03-01

Note: ADSB.lol has reduced data coverage for 2026-06-11, which will result in worse performance for that day.

Train Model

Downloads and processes 8 days of ADS-B data from https://github.com/adsblol/globe_history_2026/releases, ~24GB.

python src/flights_ai_model_airport/train.py

Contributing:

All contributions welcome. Open a PR or issue with datasets or days you want to test algorithm on.

AI Disclaimer

The airport model and testing apparatus used minimal AI assistance.
The flight-segmentation algorithm used extensive AI assistance.