Source-linked AI summary
MAD-LEO: A Maneuver-Annotated Orbital Dataset for LEO Satellites with Tiered Multi-Source Evidence
Zhixin Guo, Qi Shi, Xiaofan Xu, Linqiang Ge, Hua Zhu, Liyan Ben, Bendian Nie, Yuanrui Zhao, Xiaohan Li
TL;DR
Public maneuver labels remain scarce, and widely available TLEs have temporal-resolution and covariance limits. MAD-LEO assigns evidence tiers to mission-reported maneuver windows and compares source-specific orbital responses. TLE- and precise-orbit-derived responses correlate at 0.98 across 556 dual-computable windows.
Problem
Public maneuver labels are scarce, while widely available TLE data have limited temporal resolution and lack covariance information, and many available labels lack independent evidence.
Source: Background & Summary · pp. 1–2Method
MAD-LEO assigns mission-reported maneuver windows tiers according to the availability of independent TLE, precise-orbit, and SLR evidence.
Source: Methods · p. 4Results
0.98 Pearson correlation between TLE- and orbit-derived semi-major-axis responses across 556 dual-computable windows.
Source: Technical Validation · p. 10Takeaways & Limitations
The confidence tier describes the extent of multi-source observational support, not the reliability of the mission-reported label itself.
Source: Methods · p. 4The labels cover eleven geodetic and altimetry missions, and transfer to large impulsive or avoidance-class maneuvers has not been validated; semi-major-axis responses are insensitive to pure out-of-plane maneuvers.
Source: Usage Notes · p. 16
Abstract
from arXiv · showhide
With the rapid development of aerospace technology and the large-scale deployment of low Earth orbit (LEO) constellations, the risk of orbital collisions has increased, creating a growing demand for reliable observations of satellite maneuvers. However, public datasets containing real maneuver records remain scarce. We present MAD-LEO, a Maneuver-Annotated orbital Dataset for LEO satellites. The mission-reported subset contains 1,134 maneuver events from eleven geodetic and altimetry satellites spanning 1992 to 2026, with labels taken directly from mission-published maneuver histories. Each event is checked against two-line element (TLE) data, precise orbit products, and satellite laser ranging (SLR) observations, with evidence tiers assigned according to data availability. The operational subset pairs operator-published ephemerides for 6,785 Starlink satellites with cataloged TLE records over a continuous 107-hour period. Technical validation across seven machine-readable experiment suites confirms the cross-source consistency of the labels and the evidence products.
Background & Summary
Public datasets with real maneuver labels remain scarce, and existing orbital data and methods leave gaps for maneuver detection. MAD-LEO combines mission-reported events with multi-source evidence and a separate operational Starlink subset.
Background & Summary
- Reliable maneuver information matters for collision-risk assessment because one unreported maneuver can invalidate a trajectory used in conjunction screening.The LEO population is growing, with large constellations requiring frequent orbit maintenance and collision avoidance.Source: Background & Summary · p. 1
- Existing data-driven studies mostly target orbit prediction, while the few maneuver-detection approaches rely on TLE anomaly scoring or sparse-radar filtering and reachability analysis.Classical orbit-determination methods commonly assume a known dynamical model and are often evaluated on simulated trajectories with predefined maneuvers.Source: Background & Summary · p. 1
- Public orbital maneuver labels remain scarce, and many existing labels lack independent evidence or come from detection methods, making them unsuitable as ground truth.Public TLEs have limited temporal resolution and no covariance information, while higher-accuracy orbit products cover only a limited set of geodetic and altimetry missions.Source: Background & Summary · pp. 1–2
- MAD-LEO pairs mission-reported maneuver histories with TLE, precise-orbit, and ranging evidence, while its operational subset pairs Starlink ephemerides with TLE records.The operational records cover 6,785 satellites over 107 continuous hours and remain unlabeled because validated public maneuver annotations are unavailable.Source: Background & Summary · p. 2, Figure 1 · p. 2
Methods
MAD-LEO combines mission histories and orbital observations through shared parsing and standardization, then aligns each subset using procedures suited to its records. Mission-reported events receive evidence tiers and are paired with sampled no-event windows; the Starlink subset remains unlabeled.
Methods
- The mission-reported subset combines maneuver histories for eleven geodetic and altimetry satellites with TLE, precise-orbit, and SLR data; the Starlink subset combines predicted ephemerides with TLE records.Mission histories come from the International DORIS Service, and the Starlink records span 107 hours in November 2024.Source: Methods · pp. 2–3
- Sources are parsed, validated, standardized, and grouped by satellite, with UTC times, consistent units and reference frames, and provenance retained for traceability.TLE records are checksum- and catalog-checked; operational rolling ephemerides are deduplicated deterministically by filename order and checked for temporal overlap with TLEs.Source: Methods · pp. 3–4
- Mission-reported records are aligned to asymmetric 30-hour windows, extending from six hours before to 24 hours after the representative maneuver epoch.TLE coverage requires catalog epochs bracketing the window, precise-orbit coverage requires temporal overlap, and SLR coverage is assessed over an interval extended by a day at each end.Source: Methods · p. 4
- Each mission-reported maneuver window receives a tier based on independent TLE, precise-orbit, and SLR availability; tiers describe observational support, not label reliability.Tier A requires all three sources, Tier B requires TLE and precise-orbit data, and Tier C indicates that one or more principal sources are unavailable.Source: Methods · p. 4
- Mission-reported maneuvers define positive events, while orbital changes are not used to create labels; insufficient evidence leaves an event retained with its coverage and tier recorded.No parsed record in this release carries the ignore label; 30-hour no-event windows are sampled on a non-overlapping grid and matched by year to reported maneuvers.Source: Methods · pp. 4–5
- The matched no-event subset contains 274 windows, resampled by satellite to reproduce the Tier A share of maneuver events where eligible pools permit.The operational Starlink subset has no independently validated maneuver annotations and therefore no maneuver ground-truth labels.Source: Methods · p. 5
Data Records
The release provides annotation and window tables alongside per-satellite orbital evidence and a large Starlink ephemeris record. It preserves source traceability and flags selected data-quality conditions rather than silently removing affected records.
Data Records
- The 40-file, approximately 4.3-GB release uses CSV annotation tables and Zstandard-compressed Parquet orbital records, with double underscores encoding logical paths in deposited filenames.The metadata document provides complete column dictionaries.Source: Data Records · p. 5
- The annotation layer contains 1,134 mission-reported events and 1,139 no-event windows, including a recommended matched subset of 274 controls.The event-window table records per-source coverage and confidence tier; fourteen stable windows are flagged for a potentially unreported maneuver and retained for examination.Source: Mission-reported subset · p. 6
- The orbital evidence comprises TLE files for eleven satellites, precise-orbit files for ten, and SLR files for eleven, with about 2.1 million normal points and 24 million full-rate observations.Some normal-point sigma values are source-side placeholders, and 20 rows with implausible ranges or cross-target contamination are flagged rather than removed.Source: Mission-reported subset · p. 6
- The Starlink ephemeris file contains 43,361,358 predicted states for 6,785 catalog objects over 107 hours at a nominal 60-second cadence.The published predictions for three deorbiting objects include 5,861 retained below-surface rows flagged for filtering in altitude statistics.Source: Operational subset · p. 7
Technical Validation
Seven machine-readable validation suites test annotation traceability, cross-source response consistency, evidence quality, and the operational-subset boundary. The checks report strong agreement between response estimates while documenting limits tied to coverage, mission era, and the largest maneuvers.
Annotation traceability
- 687 of 742 benchmark events matched MAD-LEO (92.6%), and 687 of 742 in-span MAD-LEO events matched the benchmark; the +1-day offset reflects a benchmark date-parsing difference.The benchmark ends in October 2022, before SWOT’s launch.Source: Technical Validation · p. 8, Figure 3 · p. 10
Response robustness
- The TLE conversion differed from the SGP4 Brouwer-path response by a median 2.2 cm, but 5.6% of events exceeded 1 m, comparable to the smallest responses.Recomputing with a symmetric ±24-hour window changed median absolute response by 0.3 m (1.5%).Source: Technical Validation · pp. 10–11
Specificity and SLR audit
- Stable controls had a 95th-percentile TLE response of 8.0 m, while SLR-audited windows showed a median before-to-after residual shift of 0.23 m.The SLR audit covered 296 windows with at least three observations on each side; four windows exceeded 1 km RMS and were disclosed.Source: Technical Validation · p. 11, Figure 4 · p. 12
Cross-source response consistency
- For 556 dual-computable windows, TLE and orbit responses had Pearson correlation 0.98; median TLE response was 20.3 m for events versus 1.1 m for stable controls.The paired estimates had medians of 23.8 m (TLE) and 26.5 m (orbit).Source: Technical Validation · pp. 9–10, Figure 4 · p. 12
Operational subset boundary
- For the Starlink slice, median TLE–ephemeris residuals ranged from 0.58 to 13.0 km at five-to-seven-hour median propagation ages, but the age cap makes these medians optimistic.The slice documents orbit-change activity but supplies no validated maneuver labels; a 4.3 km semi-major-axis decrease identifies one maneuver crossing the propagation interval.Source: Technical Validation · pp. 12–14
Tier stratification
- Tier-A and tier-B windows had similar three-sigma coverage (91.8% versus 91.7%), while tier C had a heavier TLE-response tail, with KS distance D = 0.30 against tier A.The tier differences track mission-era coverage; tier-C large responses concentrate in commissioning and disposal periods.Source: Technical Validation · p. 14, Figure 8 · p. 16
Harmonized state estimation
- SGP4 residuals rose from 0.48 km below six hours to 0.86 km at 24–48 hours, and three-sigma coverage fell to 84.2% on historical evaluation targets.The uncertainty model is a consistency bound, not a calibrated confidence interval; its linear form is a local approximation valid to 48 hours.Source: Technical Validation · p. 14, Figure 7 · p. 15
Usage Notes
The usage notes distinguish the intended roles and evidence limits of the two subsets, and specify how to select, report, and reproduce analyses. They also identify scope boundaries and caveats that affect interpretation of the released data.
Usage Notes
- The mission-reported subset is intended for maneuver training and evaluation, while the operational subset supports constellation-scale transfer and domain-shift studies, not maneuver ground truth.TLE records support magnitude analysis for every event; precise-orbit products provide independent magnitude estimates for tiers A and B, while sparse SLR observations audit event signatures.Source: Usage Notes · p. 16
- Use tier A for three-source corroboration, tiers A and B for quantitatively reliable available sources, and tier C conservatively or for manual review.The recommended coverage-matched stable subset has 274 windows; the 1,139-window full stable set prioritizes complete mission histories over balanced coverage.Source: Usage Notes · p. 16
- Partition by satellite to avoid leakage from satellite-specific signatures or by time to test within-mission extrapolation, and state the chosen construction.The release provides no recommended train, validation, or test partition; users can derive partitions from per-satellite and per-window metadata.Source: Usage Notes · p. 16
- Labels cover eleven DORIS-equipped geodetic and altimetry missions, dominated by in-plane, centimetre-per-second-class station keeping; transfer to large impulsive or avoidance maneuvers is unvalidated.Semi-major-axis response measures are insensitive to pure out-of-plane maneuvers; Sentinel-6A has daily GNSS tracking rather than a processed orbit product, so exclude it from precise-orbit-state metrics.Source: Usage Notes · p. 16
- The TLE mean-element-change and period-averaged orbit-change estimators correlate at 0.98 on 556 dual-computable windows but differ systematically for large maneuvers.Report the estimator, subset, evidence sources, confidence-tier selection, partition construction, and each configuration separately; document preprocessing for reproducibility.Source: Usage Notes · p. 16
- Interpret released evidence with its stated frame, measurement, and label conventions, and consult sparse-coverage gap reasons before interpreting missing coverage.Precise-orbit velocities are Earth-fixed, SLR precision fields describe within-normal-point spread, and the harmonization sigma is a consistency bound rather than a calibrated confidence interval. Audit candidate rows are evidence artifacts, not labels; the label vocabulary is event, no_event, and ignore, while qc_status or below_surface rows should be filtered before statistics.Source: Usage Notes · p. 17
- Reproduction can start from archived raw inputs or released snapshots, with repository commands to regenerate normalized sources, the alignment audit, response tables, and figures.Source: Usage Notes · p. 17
Author contributions statement
The contribution statement assigns leadership across the paper’s conceptualization, methods, software, validation, analysis, supervision, and writing, alongside named contributions to curation, investigation, resources, and visualization.
Author contributions statement
- Zhixin Guo led conceptualization, methodology design, software development, data validation, data analysis, and the original draft.Guo also contributed to review and editing.Source: Author contributions statement · p. 20
- Xiaofan Xu led project administration and supervision and contributed to conceptualization, funding acquisition, and review and editing.Linqiang Ge and Hua Zhu contributed to project administration, supervision, and funding acquisition.Source: Author contributions statement · p. 20
- Qi Shi contributed to conceptualization, curation, and investigation, while other authors contributed to resources, curation, visualization, investigation, validation, and writing.All authors reviewed and approved the final manuscript.Source: Author contributions statement · p. 20