The 2020 MIT Lincoln Laboratory study that trained the first low-altitude aircraft encounter models from crowdsourced OpenSky Network ADS-B observations (Weinert, Underhill, Gill and Wicks, IEEE HPEC 2020) reported an explicit set of limitations and open issues throughout its text. The still-open problem concerns the evidence base itself: the deployed generative encounter models do not represent where or when aircraft actually fly (no geospatial or temporal structure), and aircraft-type identification rests on only four national registries with inconsistent manufacturer and model strings; the roughly 50 percent drop in observed traffic between 2019 and 2020 reported by the paper additionally raises the question about whether those distributions are stationary over multi-year periods. The problem is to extend the OpenSky-based low-altitude encounter model family so that it represents the joint distribution of aircraft type, location, and time, with a defensible treatment of crowdsourced sensor coverage.

Background

An aircraft encounter model is a generative statistical model — in this lineage, a dynamic Bayesian network, i.e. a graphical model whose random variables evolve over discrete time steps — that represents how aircraft behave during close encounters; Monte Carlo simulation of such models is the standard way to estimate midair-collision risk for certification of collision-avoidance systems. Since 2008, MIT Lincoln Laboratory has built encounter models of aircraft in the United States National Airspace System (NAS) for exactly this purpose (Kochenderfer, Edwards, Espindle, Kuchar and Griffith, Journal of Guidance, Control, and Dynamics, 2010). The classic models were trained from radar observations of the 84th Radar Evaluation Squadron (RADES) ground-based secondary-surveillance-radar network, whose Mode 3A/C transponder replies carry no aircraft-type information, and whose position accuracy forced a 500-foot floor on modeled altitudes. The integration of small unmanned aircraft systems (UAS) into the NAS — and detect-and-avoid (DAA) performance standards for them, such as ASTM F3442 — requires models of manned-aircraft behavior at much lower altitudes, where UAS actually operate. To that end, Weinert, Underhill and Wicks (IEEE Aerospace Conference, 2019) showed that the OpenSky Network, a crowdsourced network of ground-based receivers that observe aircraft broadcasting Automatic Dependent Surveillance–Broadcast (ADS-B, the satellite-based position and identity broadcast now standard on manned aircraft), could supply suitable training data. Weinert, Underhill, Gill and Wicks (IEEE HPEC 2020) then executed the full pipeline on the Lincoln Laboratory Supercomputing Center: 85 Mondays of raw observations (about 3.9 billion), organized into a directory hierarchy keyed by aircraft type (identified by parsing the national aircraft registries of the United States, Canada, the Netherlands, and Ireland, matched through the International Civil Aviation Organization (ICAO) 24-bit transponder address, the unique identifier each aircraft broadcasts) and archived with LLMapReduce, a multi-level MapReduce-style tool. They trained three type-specific Bayesian-network encounter models — fixed-wing multi-engine, fixed-wing single-engine, and rotorcraft — for aircraft between 50 and 5,000 feet above ground level (AGL). The paper's limitations and unresolved issues appear throughout its Sections III–IV. First, it had not been assessed whether these models can serve as surrogates for aircraft types or operations not explicitly modeled, and they do not fully supersede the RADES-based models, which represent different sensor characteristics and regions; a 2021 follow-up (Underhill and Weinert, Journal of Air Transportation) quantified model similarities and issued guidance on surrogate use for cooperative and noncooperative aircraft, but only for the three modeled types. Second, like their RADES predecessors, the models do not represent the geospatial or temporal distribution of the training data — they cannot say whether more aircraft were observed over New York City than Los Angeles — which matters because safety simulations must weight traffic by where and when UAS fly. Third, type identification depends on four registries, and manufacturer/model strings are inconsistent within and across registries (for example 'Cessna 172', 'Textron Cessna 172', and 'Textron C172' denote the same aircraft); the paper proposes fuzzy string matching as future work for identifying similar aircraft manufacturers and models in the traffic-composition analysis. Fourth, the archiving step was run without load-balancing optimization — since resolved by a 2021 HPEC follow-up using self-scheduling and triples-mode parallelization, which cut execution time from weeks to days. Fifth, the roughly 50 percent decrease in raw observations per hour from 2019 to 2020, attributable to the COVID-19 pandemic's curtailment of flying; the paper reports this drop without assessing whether the trained models are stationary over multi-year periods. Sixth, the per-type traffic distributions are not representative of all United States operations, since not all aircraft carry ADS-B; a companion OpenSky Symposium paper (Weinert, Brittain and Guendel, 2020) quantified ADS-B equipage rates but the representativeness caveat stands. A further structural difficulty, acknowledged in the 2021 follow-up, is that OpenSky is a crowdsourced receiver network whose coverage is geographically uneven, so observed traffic density conflates aircraft presence with receiver coverage — precisely the missing 'traffic density and surveillance coverage information' that makes weighting one model over another unjustifiable today.

Problem Statement

The problem is to extend the OpenSky Network-based low-altitude encounter model family so that it represents the joint distribution of what, where, and when aircraft fly. Concretely: (1) train generative encounter models of manned-aircraft behavior in the source's 50–5,000 feet AGL band that are conditioned on geospatial location and on time — operationalizing the source's 'temporal distribution' as diurnal and seasonal structure — in a way that defensibly separates true traffic density from the geographically uneven coverage of the crowdsourced OpenSky receiver network; and (2) extend aircraft-type identification beyond the four-country registry parsing so that type-conditional modeling covers more of the observed fleet and the surrogacy question for currently unmodeled types can be assessed. The overarching objective — the single scientific aim linking these components — is a defensible quantitative evidence base for low-altitude collision-risk simulation, so that safety analyses for UAS detect-and-avoid systems can weight and place encounter models according to where, when, and which aircraft actually operate, rather than sampling them uniformly as current practice does. That evidence base is one in which, on held-out OpenSky data, sampled encounters reproduce the observed joint spatial, temporal, and type distribution of traffic — with aircraft types resolved by the extended identification — and in which the sensor-coverage confound is separated from true traffic density. Whether the resulting distributions are stationary across multi-year periods, given the roughly 50 percent traffic drop between 2019 and 2020 reported by the source, is a question flagged among the difficulties above rather than a required deliverable of this problem.

References

  1. Andrew Weinert, Ngaire Underhill, Bilal Gill, Ashley Wicks, 'Processing of Crowdsourced Observations of Aircraft in a High Performance Computing Environment', 2020 IEEE High Performance Extreme Computing Conference (HPEC), pp. 1–6, 2020. DOI: 10.1109/HPEC43674.2020.9286229; arXiv:2008.00861; https://doi.org/10.1109/HPEC43674.2020.9286229
  2. Ngaire Underhill, Andrew Weinert, 'Applicability and Surrogacy of Uncorrelated Airspace Encounter Models at Low Altitudes', Journal of Air Transportation, vol. 29, no. 3, pp. 137–141, 2021. DOI: 10.2514/1.D0254; arXiv:2103.04753; https://doi.org/10.2514/1.D0254
  3. Andrew Weinert, Marc Brittain, Ngaire Underhill, Christine Serres, 'Benchmarking the Processing of Aircraft Tracks with Triples Mode and Self-Scheduling', 2021 IEEE High Performance Extreme Computing Conference (HPEC), pp. 1–8, 2021. DOI: 10.1109/HPEC49654.2021.9622862; arXiv:2110.12858; https://doi.org/10.1109/HPEC49654.2021.9622862
  4. Andrew Weinert, Ngaire Underhill, Ashley Wicks, 'Developing a Low Altitude Manned Encounter Model Using ADS-B Observations', 2019 IEEE Aerospace Conference, pp. 1–8, 2019. DOI: 10.1109/AERO.2019.8741848; https://doi.org/10.1109/AERO.2019.8741848
  5. Mykel J. Kochenderfer, Matthew W. M. Edwards, Leo P. Espindle, James K. Kuchar, J. Daniel Griffith, 'Airspace Encounter Models for Estimating Collision Risk', Journal of Guidance, Control, and Dynamics, vol. 33, no. 2, pp. 487–499, 2010. DOI: 10.2514/1.44867; https://doi.org/10.2514/1.44867
  6. Matthias Schäfer, Martin Strohmeier, Vincent Lenders, Ivan Martinovic, Matthias Wilhelm, 'Bringing Up OpenSky: A Large-scale ADS-B Sensor Network for Research', IPSN-14 Proceedings of the 13th International Symposium on Information Processing in Sensor Networks, pp. 83–94, 2014. DOI: 10.1109/IPSN.2014.6846743; https://doi.org/10.1109/IPSN.2014.6846743
  7. Andrew Weinert, Marc Brittain, Randal Guendel, 'Frequency of ADS-B Equipped Manned Aircraft Observed by the OpenSky Network', Proceedings of the 8th OpenSky Symposium, MDPI, 2020. DOI: 10.3390/proceedings2020059015; https://doi.org/10.3390/proceedings2020059015
  8. Lily Lee, Julian Fontes, Andrew Weinert, Laura Schomacker, Daniel Stabile, Jonathan Hou, 'TAACKIT: Track Annotation and Analytics with Continuous Knowledge Integration Tool', 2024 International Conference on AI x Data and Knowledge Engineering (AIxDKE), pp. 112–116, 2024. DOI: 10.1109/AIxDKE63520.2024.00028; arXiv:2412.16228; https://doi.org/10.1109/AIxDKE63520.2024.00028
  9. Ngaire Underhill, Evan Maki, Bilal Gill, Andrew Weinert, 'Estimating See and Be Seen Performance with an Airborne Visual Acquisition Model', 2023. arXiv:2307.05502; https://arxiv.org/abs/2307.05502