AeroScope began with a plain question, "what is flying over me, and can I trust it?", and a conviction that the answer should be checkable. It is built by one independent researcher, runs entirely on open data, and refuses to dress up rules as magic.
Every airliner, business jet, helicopter and many larger drones announce themselves roughly once a second on 1090 MHz. Dozens of volunteer networks already collect those broadcasts and publish them for free. The data is genuinely open, which is remarkable. The problem is that it arrives as undifferentiated dots.
Conventional flight trackers are built around a different question. They answer "where is flight BA117", which is exactly right if you are meeting someone at an airport. It is the wrong shape of question if you are standing in your garden listening to something circle, or if you are an analyst asking whether the traffic over a particular facility looks normal this week.
AeroScope inverts it. You fix a location and the platform continuously evaluates everything above it: what each aircraft is, how it is behaving, whether its own signal is internally consistent, and which of the dozens overhead actually deserve a second look. That is what "observer-relative" means, and it is the single design decision the whole project follows from.
Each cycle runs raw positions through an enrichment and detection pipeline before anything reaches the screen. The stages are deliberately boring and inspectable:
| Stage | What happens |
|---|---|
| Fusion | More than 60 public ADS-B feeds are polled and merged. Reports of the same airframe from several networks are de-duplicated on the ICAO 24-bit address, so one aircraft is one track rather than three dots. See data sources. |
| Enrichment | Registry lookup attaches type designator, operator, engine count and wake category where a record exists. Fields are left blank rather than guessed. |
| Attention scoring | Eight weighted, rule-based factors produce a transparent 0 to 100 score: emergency squawk, military hex allocation, proximity to the observer, altitude and speed anomalies, loitering or orbit patterns, integrity-check failures, unregistered identity and drone-like kinematics. Every score shows its contributing factors. Read the scoring method. |
| Integrity | A DO-260B style check on NIC, NACp, NACv and SIL, a Kalman normalised-innovation test on the trajectory, and self-consistency residuals such as geometric versus barometric altitude. See spoofing detection. |
| Anomaly consensus | Six independent detectors vote. A flag is raised only when at least two agree, which suppresses the false alarms a single detector produces. See behavioural baselines. |
| Delivery | Positions refresh on a fast loop of about seven seconds, with client-side dead-reckoning between fixes so motion is continuous. The heavier analysis runs on its own cadence of roughly twelve seconds so it never blocks the map. |
The detection stack is deliberately torch-free. It is built on River, scikit-learn, pykalman, OpenAP, stumpy and PyOD, chosen because every flag they raise can be traced back to a rule or a statistic. There is no deep learning and no reinforcement learning. An earlier RL prototype was built, evaluated, found not to justify its opacity, and retired. References to it were removed from this site rather than left to quietly inflate the capability list.
The AeroScope ADS-B Anomaly Benchmark is published under CC-BY 4.0 with 38 documented columns, a reproducible IsolationForest baseline and its limitations stated up front. Publishing the dataset means the detection claims can be checked by someone else rather than simply believed. See research and dataset.
AeroScope ranks attention. It does not judge intent. It sees only what aircraft choose to broadcast, so silent aircraft, most small consumer drones and anything operating without a transponder are invisible to it. Where a capability has an edge, the platform says so in plain language on the limitations page instead of hiding it in a footnote.
Being useful to serious users means being precise about the boundary of the tool. AeroScope is an awareness, research and analysis platform. It is explicitly not the following things:
AeroScope is designed, written, deployed and maintained by one independent researcher in the Department of Computer Science at COMSATS University Islamabad, Islamabad, Pakistan. Every design decision described on this site, from the choice of a consensus rule over a single classifier to the decision to publish the benchmark openly, traces back to one accountable person rather than an anonymous team.
That is a real limitation on capacity and it is stated openly on the status page. It is also the reason the project can afford to retract a claim the moment it stops being true, which a commercial roadmap usually cannot.
Identity and provenance are published so the work can be attributed and checked: ORCID 0009-0002-4104-2680, the Wikidata entity for the platform, and the open dataset record on Hugging Face.
| If you are | Start with |
|---|---|
| Curious about your own sky | What is flying over my house, then open the live map and set your location. |
| A researcher | The open benchmark dataset and the detector design. The data is CC-BY 4.0 and citable. |
| A receiver-network operator | Data sources for how your feed is used and attributed, and receiver setup if you want to contribute one. |
| A security or OSINT analyst | ADS-B security and spoof detection, plus the REST and WebSocket API. |
| Building with AI assistants | The public MCP endpoint, which lets an assistant query live airspace and this documentation directly. |
If you are a researcher, a receiver-network operator, a journalist or an organisation that cares about ADS-B integrity and open aviation data, I would genuinely like to hear from you. Get in touch.