Executive Summary

Key Figures

Source: Guo et al., MAD-LEO: A Maneuver-Annotated Orbital Dataset for LEO Satellites with Tiered Multi-Source Evidence, arXiv:2609.08556 (2026), Data Records and Technical Validation

Low Earth orbit is filling up quickly. By the count the paper cites, SpaceX had launched more than 9,000 Starlink satellites by the end of 2025, and that company alone performs tens of thousands of avoidance maneuvers during each six-month reporting period. At this scale, orbit maintenance and collision avoidance become daily work, and a single unreported maneuver invalidates the predicted trajectory that conjunction screening runs on.

Research into finding maneuvers automatically goes back decades, yet the material for training and scoring such a method is thin. Two-line element sets, the most widely available public orbital data, have low temporal resolution and carry no covariance information, which limits their use for precise maneuver analysis. Centimeter-level precise orbit products exist for only a handful of geodetic and altimetry missions. Many studies fall back on simulated trajectories, and a simulation cannot represent the full range of real operating conditions.

The labels that remain are in no better shape. The diagnosis the authors set out in their background section is brief. Most published maneuver labels either lack independent evidence or are the output of a detection method, and a label of that kind cannot act as ground truth in an evaluation. Even the closest prior resource pairs long TLE histories for fifteen satellites with maneuver timestamps, without recording which independent observation supports any given timestamp.

The labels in MAD-LEO are not detector output. They came out of what the operating agencies themselves report. The team parsed the mission maneuver histories published by the International DORIS Service and pulled 1,134 events from them, covering eleven geodetic and altimetry satellites that carry DORIS receivers. The span opens with the first TOPEX/Poseidon maneuver in August 1992 and runs through Jason-1, Jason-2 and Jason-3, CryoSat-2, HY-2A, SARAL/AltiKa, Sentinel-3A and 3B and Sentinel-6A, and on to SWOT, the newest of the eleven; the last record in the table is from July 2026. The annotation table carries status fields such as truth_status, which record that a label is mission-reported rather than generated by a detection algorithm.

Each maneuver gets a 30-hour analysis window, running from 6 hours before the reported epoch to 24 hours after it. The two sides differ in length because catalog records can pick up a change well after the reported time. The table also keeps the epoch the window was hung on. Of the 1,134 events, 1,089 are anchored at the reported first-impulse time, and the other 45 use the reported operation end time, the only epoch those records supply. Response, here and below, is how far the orbit's semi-major axis moved between the two ends of a window, measured in meters. Those 45 show a median response of 3.3 m against 20.7 m for the first-impulse events, since most of them are early TOPEX/Poseidon records logging an interval, not an instant. The public label vocabulary holds three values, event, no_event and ignore, and this release contains no ignore at all. That value is reserved for records with no assignable UTC epoch or with an unresolved conflict against their source, and the paper attaches one sentence to it: "Insufficient evidence coverage is never grounds for ignoring a mission-reported event."

Quiet stretches were built as well. For each satellite the team laid a non-overlapping 30-hour grid and picked the stretches clear of any maneuver window, which produced 1,139 no-event controls, and of those it recommends 274 whose evidence coverage matches that of the events for balanced evaluation. Fourteen of the 1,139 control windows stayed in the release flagged as suspected unreported orbit change, because their TLE-derived response exceeds 20 m. Seven of the fourteen sit in the early-2023 period when SWOT was raising its orbit, which the DORIS maneuver histories do not cover. Rather than delete them, the team flagged them and examined all fourteen one by one in the validation section.

2.1For 6,785 Starlink Satellites, the Label Column Says Nothing

The second half of the dataset is Starlink. Operator-published predicted ephemerides were paired with catalog TLE records over the same period for 6,785 satellites, covering 107 continuous hours from 06:00 UTC on 26 November to 17:00 UTC on 30 November 2024. There are more than 43.3 million predicted states in it at a 60-second cadence. Next to the 1,134 events of the first half, the scale is of a different order.

This half has no labels. An operator ephemeris is a predicted trajectory rather than a real-time observation and may differ from the actual motion, and validated maneuver records for individual satellites are not published. The paper states flatly that this subset must not be used as maneuver ground truth, and sets it aside for testing at real operational scale the methods developed on the first subset. The absence of labels became one of the things the dataset publishes about itself.

None of that amounts to a claim that the orbits held still. In the released TLE table, the semi-major axis jumps by more than 100 m between neighboring epochs 10,498 times across 2,909 satellites, with a median jump of 338 m. A caveat travels with the count: 57% of those steps come from the deployment and transfer population below 480 km, where drag alone produces changes of comparable size. For NORAD 45184 the records show the semi-major axis dropping 4.3 km at 22:00 UTC on 26 November 2024, roughly fifteen times the 0.28 km that satellite loses to drag in 8 hours of a quiet period. All of this sits in the text of the paper while the label column stays empty.

All 1,134 rows are labeled event, but the independent observation behind each window varies a great deal. The researchers decided, for each window, which of the three sources were actually available. TLE coverage requires at least one catalog epoch at or before the window start and one at or after its end, precise orbit coverage requires the orbit product to overlap the window in time, and satellite laser ranging is judged over an interval extended by one day on either side, because laser passes are sparse. The verdict goes into the confidence tier column as A, B or C.

Sentinel-6A got its own handling. No finished precise orbit product exists for it, so daily GNSS tracking files take the orbit-coverage role in the tier assignment. Its windows enter the label-level analyses and drop out of every metric that needs an orbit state. The material is partial, and it stays in the dataset anyway, with a window-by-window note on how far it can be used.

The release holds 380 windows short of all three sources. Every one of them carries a machine-readable marker in the released alignment table naming the source that is missing. Because the release keeps that detail instead of flattening it into a single letter, a user can filter again on whichever source their own analysis actually needs.

Once the tiers are in place, one question follows. Is a lower-tier window actually lower in quality? The sixth of the seven validation suites takes that question head on.

The two metrics referenced to the orbit need a precise orbit product to exist, so they can only be compared across tiers A and B. On that set the two tiers sit close together. Three-sigma coverage runs 91.8% for A against 91.7% for B, sign agreement 96.4% against 94.9%, and the median period-averaged orbit response 22.0 m against 28.0 m. The TLE response, which is computable for every window, moves more: its median runs 20.2 m for A, 15.5 m for B and 31.8 m for C, and the tier C distribution trails a heavy tail whose 75th percentile reaches 1.87 km. Its Kolmogorov-Smirnov distance from tier A is 0.30 with a p-value of 3.5×10⁻¹². Tiers A and B differ statistically too on that metric, at a distance of 0.13 and a p-value of 0.014, though the gap is narrow next to what separates tier C.

A reader could take the tier C tail for degraded quality, but the authors offer another account. Large maneuvers cluster around the start of a mission, when a satellite is being placed in its orbit, and around its end, when the satellite is brought down, and tracking infrastructure happens to be sparsest in exactly those periods. A window with a large response is therefore more likely to be missing a source. The sentence the paper gives as the conclusion of this suite reads: "The tiers therefore separate coverage eras, not data quality: where a metric is computable at all, tier-B windows behave like tier-A windows."

Because tier composition itself varies with mission era, every tier comparison was run within a single satellite. Pooling satellites would mix orbital regime with tier and leave no way to tell which produced a difference. Measured within a satellite, the larger gaps in orbital elements between tiers lined up with mission phases that are documented in the record, such as the 2018 relocation of Jason-2 and the 1992 commissioning of TOPEX. A difference test cannot establish that no difference exists, so the authors pre-registered engineering tolerances and ran an equivalence test on top. Inclination came out equivalent in 14 of 19 within-satellite comparisons, while eccentricity and altitude exceeded the tolerance in 24 of 38, with every exceedance falling on a documented mission phase. The exceedances amount to at most 0.35% of the semi-major axis.

Those measurements turned straight into usage guidance. An analysis that genuinely needs all three sources behind it should stay in tier A, tier B remains quantitatively reliable on the sources it has, and tier C is meant for conservative use or manual review. A further clause asks anyone reporting results to state which subset and which tier selection they chose.

A label coming from a mission report is no guarantee that the parsing was right. So the researchers held their own labels against a maneuver benchmark that had been built from entirely different material. Matched under a one-day tolerance, 687 of the 742 events in the benchmark have a counterpart in MAD-LEO. That is 92.6%. In the other direction, 687 of the 1,134 MAD-LEO events appear in the benchmark, and since the benchmark's records stop in October 2022, 392 events lie outside the comparable span to begin with. Within that span the reverse rate is 92.6% as well. Widening the tolerance to two days brings 709 matches, and narrowing it to half a day drops the count to 80. A fall that steep inside half a day means the timestamps in the two lists were offset from each other by roughly a day.

The offsets in the matched pairs all leaned one way. Of the 687 pairs, 582 cluster at exactly one day. A pile-up like that signals a systematic date-handling error on one side. The authors went back to the raw IDS/DORIS records and confirmed it as a day-of-year conversion difference in the dates the benchmark published, then left their own labels untouched. They followed the discrepancy to its cause instead of making it go away.

The 55 that did not match came with an account too. Nineteen have responses under 5 m, buried in the catalog noise floor; 27 are tier C windows sitting at archive boundaries; and seven fall in both groups. They are not errors scattered at random but a concentration in the places where the benchmark is least likely to be complete. The distribution of response magnitudes points the same way. The matched events have a median TLE response of 13.2 m, while the 55 unmatched in-span events sit at 33.4 m and the 392 out-of-span events at 29.4 m. That describes a benchmark whose recall falls off both at the noise floor and out in the relocation-class tail, not a set of MAD-LEO events with nothing behind them.

The team also checked the state of the measuring instrument itself. Across 542 windows with at least two laser normal points on each side, the median shift in normal-point precision between before and after is 2.6 mm, against a typical precision for this instrument of 4 to 14 mm. Measurement quality, in other words, did not wobble at the epochs where a maneuver was reported. Next they set the laser-measured ranges beside ranges computed from the precise orbit products. That comparison ran on the 296 windows holding at least three observations on either side, where the median residual moved 0.23 m across the event. Four windows whose residual RMS exceeded 1 km stayed in the release, each published with its cause: a Jason-1 window that looks like a time-tagging problem, a Jason-3 window matching a documented multi-day orbit degradation, and CryoSat-2 and TOPEX windows at archive boundaries. Response magnitude also tells a maneuver window from a quiet one. The median TLE response is 20.3 m for events and 1.1 m for the no-event controls.

That is where the paper ends. To anyone who works with data, a familiar gap comes into view here. A label usually survives as one of two values, right or wrong. How many reviewers looked at it, whether two separate instruments returned the same answer, or whether one person glanced at a screen once and moved on, all of that disappears somewhere in the labeling process. And a model then learns from the thickly confirmed label and the thinly confirmed one at the same weight.

MAD-LEO holds that information inside the schema instead. Three pieces of it travel well.

• Keep the label column and the evidence column apart. The tier is not a second verdict on the label. It notes only whether three independent checks stood behind that label or one. Two axes that stay unmixed can later be updated one at a time.
• Do not delete missing evidence; mark it. Each of the 380 windows short of full three-source coverage says in machine-readable form what is absent and why. If a row goes out because its evidence looked thin, the reason it was thin leaves with it.
• Ship the tier-by-tier usage rules with the data. The paper calls tier A alone strict, A and B together relaxed, and all of them audit, and asks that any result state which of the three was chosen. When the people who built the data name the options first, the people using it stop inventing their own each time.

There is a reason this design came out of orbital work. Re-confirming a past maneuver for one satellite calls for a laser ranging network and precise orbit determination, equipment that exists for only a few missions, and an epoch that has passed cannot be measured again. Where re-measurement is impossible, the choice narrows to discarding the label or preserving the thickness of its evidence. A team stands at the same fork over a vision inspection on a manufacturing line after the product has shipped, or over a medical image on which two readings disagree.

The limits of using this dataset as it stands are also written into the paper. The maneuvers in the labels are mostly centimeter-per-second station keeping, so whether a method trained here carries over to the large impulses of collision avoidance has not been validated in this release. Because the response measures are built on the semi-major axis, they are insensitive to a maneuver that only tilts the orbital plane, and the authors say so. Alongside a design that records evidence as a tier, they narrowed by hand the range their labels claim to cover.

Thank you for reading this far. The paper is at arXiv:2609.08556, and the dataset is on Figshare under CC BY 4.0, 40 files totaling about 4.3 GB. The construction pipeline is public at github.com/sjtugzx/madLEO. If your team already runs a column for the evidence beside the label, we would be glad to hear how you divided it.