Methodology
How the numbers are made
Data sources
- Orbital element sets come from CelesTrak's current GP data interface (
gp.phpqueries in the documented TLE output format), a free public service maintained by T.S. Kelso. Detail, pass, and search views use bounded curated groups (crewed stations, GNSS constellations, scientific satellites) plus single-object lookups. The whole-sky globe is the one exception: it needs every active object (about 16,000 element sets), fetched in a single request at most once every 6 hours and shared by all visitors. - Each deployment also carries its own copy of the whole active catalog, fetched when the site is built (and refreshed every two days by a scheduled job). A server that starts cold answers from it at once instead of depending on CelesTrak in that moment; CelesTrak has refused the full-catalog request from our hosting provider's servers while answering other networks. The copy always shows its real fetch time, is refreshed from CelesTrak in the background when it is more than 6 hours old, and is never served once it is more than a week old.
- The whole-sky copy is also kept on the server's disk with its original fetch time, so a restart reuses it instead of downloading it again; it is served as “stale” for at most 48 hours after that fetch, then dropped. After a failed fetch the server waits 10 minutes before asking CelesTrak again rather than retrying on every page view. A feed far smaller than the active population is refused as truncated, never drawn as the sky.
- Requests are cached server-side for 2 hours. A record older than that but less than a day old is answered at once, labelled “cached”, while a fresh copy is fetched in the background for the next request, so an unreachable provider never holds the page up for its 15-second timeout. Older records wait for the provider; when it is unreachable, previously fetched data is served and clearly labelled “cached” or “stale: provider unavailable”. The application never fabricates a response when the provider is down.
- CelesTrak states its data is free for public use with attribution and asks bulk consumers to cache politely; our client throttles its own outbound requests (≥ 2 s spacing) and times out at 15 s. Review CelesTrak's current terms before altering request patterns.
- The NORAD satellite catalog identifier (column 3–7 of the TLE) is the join key across all features. Names, COSPAR designators, and classifications accompany each element set from the provider response itself.
The atlas: constellations and orbit shapes
- The atlas uses the same whole-sky copy as the globe. Inclination, eccentricity and mean motion are read from each TLE's fixed columns; perigee and apogee altitudes are Keplerian close forms (a = (μ/n²)1/3, WGS-72 μ, minus the equatorial radius). Each dot sits at the mean of the two. These are approximations of the mean orbit, not positions: they ignore the oblateness and drag terms SGP4 models.
- Constellation membership comes from the provider's object names, by whole-word prefix: STARLINK, ONEWEB, KUIPER (Amazon), QIANFAN, HULIANWANG (Guowang), IRIDIUM; navigation is NAVSTAR/GPS, GSATnnnn (GALILEO …), BEIDOU, any name containing GLONASS, QZS and IRNSS. India's GSAT-n communications satellites and COSMOS satellites that aren't GLONASS stay in “everything else”. Name rules are transparent but not authoritative: anything they don't match is never guessed into a constellation.
- Launch years come from each element set's international designator (line 1, columns 10–11: 57–99 means 1957–1999, 00–56 means 2000–2056). The launch chart counts satellites that are still active, by the year they launched. It is not a count of launches per year: retired, failed and re-entered satellites aren't in the active catalog.
- Orbit regimes match the globe's colouring: eccentricity above 0.25 is highly elliptical; otherwise a period under 129 minutes is low Earth orbit and 22–26 hours is the geostationary belt; the rest is medium orbit.
Close-approach screening
- Close approaches runs in a Web Worker in your browser against the same whole-sky copy, following the classic filter chain of Hoots, Crawford & Roehrich (1984). Radius filter: objects whose perigee–apogee range can't come within the screening distance of the target's (plus a 30 km pad for the gap between mean-element and SGP4 radii) are dropped, typically 95–98% of the catalog.
- Sweep: survivors and the target are propagated with SGP4 every 60 s. A local minimum of the sampled separation brackets each approach. Between samples two objects close by at most their relative speed (bounded by 16 km/s, a head-on meeting in low orbit) times the step, so minima sampled further out than the screening distance plus that bound are provably not approaches and are skipped. The same bound lets the sweep jump ahead: from a separation d, nothing can come within the screening distance for (d − screening distance) ÷ 16 km/s, so those samples are never computed. A unit test checks that this finds exactly the approaches the full sweep finds.
- Refinement: golden-section search narrows each bracket to under a millisecond to find the time of closest approach. The miss vector is resolved into the target's radial / in-track / cross-track frame. Unit tests check this against exact two-body orbits built to meet at a known instant and distance, and against real station element sets.
- Pairs closing slower than 50 m/s, or staying inside the screening distance for the whole window, are co-orbiting (docked modules share one element set; visiting vehicles fly with the station) and are listed apart from approaches.
- Limits: the active feed has no debris or rocket bodies; TLE positions are uncertain at the kilometre level and carry no covariance, so there is no collision probability here. Operators screen with precision ephemerides and covariance (conjunction data messages), and that is what drives real avoidance manoeuvres.
Orbital elements and their epoch
An element set (TLE) is a snapshot of a satellite's orbit at its epoch, the TLE's instant (line 1, columns 19–32, with the fractional day-of-year). Every derived quantity in this application (position, ground track, pass times) inherits the epoch, and we display it alongside every result. Each fetch records the retrieval time so “how old is this data” always has an exact answer. Elements older than 36 hours are flagged in the UI: a conservative operational-difference warning from the SGP4 maintenance model, not an asserted universal accuracy cutoff.
SGP4 propagation
Positions are propagated with SGP4, the Simplified General Perturbations model (coded here via the maintained satellite.js implementation). SGP4 models secular and periodic perturbations (J2/J4, atmospheric drag via B*, resonant terms) near Earth, and SDP4-style deep-space corrections for orbits beyond ~225 minutes of period. It is the standard model trained to consume NORAD element sets; results are in the TEME frame in kilometres and km/s, exactly what the element sets were fit to. This application never substitutes a simplified Kepler orbit for SGP4.
Coordinate frames and time
- All timestamps are UTC; the application performs no local-time arithmetic internally and labels displayed timestamps with UTC explicitly.
- SGP4 output (TEME) is converted to earth-fixed (ECEF) using the Greenwich sidereal time (GMST) for each sample, then to WGS-72 geodetic coordinates (the same ellipsoid the element set model uses internally). Differences between WGS-72 and WGS-84 geodetic coordinates are bounded by ~180 m (smaller than the prediction error budget) and documented rather than silently mixed.
- Pass calculations use topocentric azimuth/elevation from the observer's WGS-72 geodetic location (height treated as above-mean-sea-level; the geoid-vs-ellipsoid difference of up to ~100 m shifts pass timing by well under a second at LEO rates).
Ground tracks and pass predictions
- Ground tracks sample SGP4 positions at bounded intervals (default 240 samples over 24 h). Samples are split into segments at the ±180° antimeridian so plots never draw a false line across the map; the split edge point is interpolated where a track crosses it.
- Pass predictions probe the elevation angle every 30°·s of time (30 s steps), find above-threshold crossings, refine each rise/set crossing with bisection to < 1 s, and refine the maximum-elevation instant with a ternary search. Passes that are already above threshold when the window opens (or never set inside it) are reported with an explicit truncated marker rather than invented boundary times.
- The search window is validated server-side (≤ 30 days) and bounded to at most 25 satellites per request: one SGP4 sweep per satellite per request, ~2,880 evaluations for 24 h; cheap enough for the free tier while remaining bounded.
Prediction limitations (read this before mission planning)
- Predictions, not measurements. Everything labelled “position” here is an SGP4 propagation from a published element set, never a live telemetry feed from the satellite.
- Accuracy decays from the epoch. SGP4+TLE is historically good to very roughly ~1 km at epoch for well-tracked LEO objects, growing with prediction duration; GEO and highly-dragged objects differ. Exact figures depend on the specific object, the fit quality, and weather in the upper atmosphere. We show element age and let you judge against the guidelines linked below.
- Passes ≠ visibility. A predicted pass is orbital geometry relative to your coordinates. Visual-visibility judgements model sunlight geometry: the satellite must be lit (cylindrical shadow test) and the observer's sky must be past civil twilight (sun below −6°). Where a published value exists, it also estimates apparent magnitude, corrected for phase. Light pollution and eyesight are not modeled, so this still is not an observing forecast.
- Clouds are forecast, shown beside the verdict, never folded into it. Each visible pass on the home page carries the forecast total cloud cover for its midpoint, from Open-Meteo (CC BY 4.0). Open-Meteo blends global models (ECMWF IFS, NOAA GFS, DWD ICON) with higher-resolution regional ones where they exist. Cover is forecast at whole hours, so a pass's value is interpolated between the two hours around it, and a pass outside the forecast reads “beyond the cloud forecast”, never a guess. The words use 20 / 50 / 80 % cut-offs (clear / partly / mostly cloudy / overcast), loosely the meteorological okta bands. Only overcast is called “likely hidden”: a satellite is a point of light and can shine through a gap. This is a model grid cell, not your patch of sky. The pass stays listed whatever the forecast says, because a forecast is the thing most likely to be wrong here. Your browser asks Open-Meteo directly, sending your spot rounded to about 1 km, which is still finer than any model grid cell.
- Brightness: only from published standard magnitudes. The standard magnitude is an object's visual magnitude at 1 000 km range and 90° phase angle (“half illuminated”), with phase adjustment 2.5·log₁₀(0.5 + 0.5·cos φ) (Mallama 2020, §2). We estimate m ≈ mstd − 15 + 5·log₁₀(range km) − 2.5·log₁₀((1 + cos φ)/2) − 0.753, which equals mstd at 1 000 km and 90°. Real satellites are not spheres: attitude and solar-array orientation move their brightness around this model, so treat estimates as rough. Objects without a published value get no magnitude at all (“brightness unknown”): a size-blind default would rate a cubesat as bright as a space station. The values in use, with their sources and known weaknesses:
- NORAD 25544: -1.3, T. Molczan, SeeSat-L, 2005-08-08 (satobs.org/seesat/Aug-2005/0114.html): 47 visual observations by R. Eberst, Dec 2002–Jul 2005, normalized to 1 000 km and 90° phase. Caveat: Derived before the station's truss and solar arrays were completed (2009); today's larger ISS is likely brighter, so estimates may read faint.
- NORAD 48274: 0.0, Heavens-Above satellite information for 48274 (heavens-above.com/SatInfo.aspx?satid=48274: "intrinsic brightness 0.0 at 1000 km distance, 50% illuminated"), as quoted by a web search on 2026-10-16; the page itself refused our connection, so the value is not yet confirmed first-hand. Caveat: Single published value with no stated observation basis or date; the station grew from one module to three in 2022 (Wentian, Mengtian), and it is unclear which configuration the value describes.
- The visible window. Each pass is sampled 24 times between rise and set; the first and last visible samples bracket when it can actually be seen, and each edge is refined by bisection to about a second (finer would be false precision; the shadow model itself is good to seconds). The window is assumed contiguous: within one pass the sky only darkens or brightens, and the satellite crosses the shadow boundary at most once. An edge in mid-sky has only two possible causes under this model, and the home page names which: Earth's shadow (“fades into Earth's shadow 25° up in the E”) or twilight (“lost in the brightening sky”). The home page lists only passes with a visible window still ahead; every geometric pass remains one click away.
- Sky charts. Each pass is drawn on a chart of your sky: the zenith at the centre, the horizon on the rim, elevation falling linearly outward (rings at 30° and 60°), and east on the left, a star chart's orientation, the sky seen from below, so held overhead with N toward north every point sits where it will appear. The dashed arc is the same 24 rise-to-set samples the visibility test uses; the bright line runs between the refined window edges above, not between coarse samples, so it starts and ends where the text says. A ring marks a mid-sky entry into Earth's shadow. The peak is also given in fists at arm's length, about 10° each and rounded to half a fist (hands vary, so finer would be false precision), and from 75° up simply as “nearly straight overhead”.
- Calendar events. “Add to calendar” builds an iCalendar (RFC 5545) file in your browser (nothing is sent anywhere), spanning only the visible stretch, with a 10-minute alarm. A calendar file cannot update itself, and predictions move as new element sets arrive (a station reboost moves them most), so the event links back to the live prediction.
- Your spot. Device location, a typed place, or typed coordinates. Coordinates are read in your browser. A place name is looked up by Open-Meteo's geocoder (GeoNames places, CC BY 4.0): your browser sends it the text you typed, and nothing else. It is called from your browser rather than our server because its free tier is rate-limited per IP address, so one shared server address would run out for everyone at once. A place is a point, its centre, which is fine here: for the ISS, 400 km up and moving 7.7 km/s, standing 10 km away shifts a pass by a second or two and its directions by at most about a degree. The spot you pick is remembered in this browser only (local storage) so a return visit skips the question; “change spot” forgets it.
- Times are the observer's. Pass times are shown on the wall clock of the coordinates you entered, not your device's: the server resolves the time zone offline from the coordinates with tz-lookup, which uses simplified zone boundaries. Its authors measure that about 5% of random points on Earth (mostly remote areas and open ocean) resolve to a zone with a different UTC offset; populated places resolve reliably. Open ocean uses nautical UTC±N zones. When the zone differs from your device's, the page says so, and the technical pass table also keeps UTC.
- Light-model approximations. The sun position is the low-precision Astronomical Almanac approximation (~0.01°, far coarser than we need for the lit/unlit decisions); shadow timing ignores penumbra and atmospheric refraction (seconds of difference); the phase term is the illuminated fraction of a sphere, the convention standard magnitudes are defined with, rather than a diffuse (Lambertian) sphere's phase function. Standards for a plan-outside tool, not for photometry.
- The globe's satellite and 3D model. The followed object is placed from SGP4 samples every ~5 minutes over a 6-hour window, interpolated between them; the time slider, “live”, and 60× fast-forward all move the same simulation clock, so the satellite and every whole-sky dot always show the same instant. The ISS is drawn with NASA's “International Space Station (ISS) (B)” model (NASA 3D Resources, credit NASA/Michael D. Carbajal; NASA states its 3D resources are free and without copyright). Two things about it are not measurements: it is drawn enlarged so it stays visible (the real station is about 109 m across), and its attitude is approximated as pointing along its direction of travel, close to how the ISS usually flies, because attitude is not part of an element set. Other objects are shown as dots: we only use models we can source.
- Whole-sky dots between SGP4 runs. Your browser runs SGP4 for every whole-sky object at once (every 5 seconds live, every quarter-second at 60×), and between those runs moves each dot in a straight line along its computed velocity. Over that span a straight line strays from the curved orbit by at most about a kilometre (checked against SGP4 itself in the tests), well under one pixel at globe scale. To save work, a dot is redrawn only when it would move at least half a pixel on screen, so on the far side of the Earth from the camera dots can lag their computed spot by that much.
- Manoeuvring objects (spacecraft with thrusters, de-orbit schedules, dodging debris) can depart from published elements quickly after manoeuvre; predictions for such objects deserve extra caution.
- Not for collision avoidance, safety-of-life, or operational mission planning. This is an educational/analysis product. Use dedicated professional services for those cases.
Identifying a sighting (“What did I just see?”)
- What is searched. Every object in the active catalog the globe uses, propagated with SGP4 in your browser (a Web Worker) every 10 s for 4 minutes either side of the time you give: about 49 instants × 16,000 objects, roughly 780,000 propagations. Ten seconds matters: an object passing overhead crosses about 1° a second, so coarser steps could jump straight over the patch of sky you describe.
- What counts as a match. At some sampled instant the object was at least 10° up, in direct sunlight (the same cylindrical shadow test as for passes), and your sun was at least 6° below the horizon. It was crossing the sky at least 1° a minute (you saw it move; this leaves out geostationary and navigation satellites, which sit 20,000 to 36,000 km away and aren't naked-eye objects in normal conditions). And it came within 25° of what you reported, about one compass point either side.
- Distance from your report. “Looking NW, halfway up” is taken as the arc of sky due NW between 30° and 60° up (10–30° for “low”, 60–90° for “nearly overhead”, horizon to zenith when unsure). A candidate's distance is the angle to the nearest point of that arc, solved in closed form, so the zenith counts as on every line of sight.
- Which way it moved (optional). An object's heading is the compass direction of its 10-second step across the sky, taken from the horizontal part of that step. Overhead, that is the direction it travels; low in the west and climbing, it reads “toward the E”, which is how people describe it. A remembered direction keeps candidates heading within 45° of it. Around dusk, hundreds of Starlinks can be sunlit in one quarter of the sky, and this is the filter that narrows them most.
- Ranking. Closest to your report first, in 5° steps (finer would claim more precision than “it was over there” contains). Within a step, objects with a published standard magnitude come first, brightest first, then nearer before farther. A “satellite train” is folded into one entry: at least four satellites from one launch within 10° of each other on the sky at the same instant. Sharing a launch is not enough. Thousands of Starlinks from years-old launches are spread around their orbits, so a few launch-mates crossing the same quarter of the sky is ordinary and stays listed one by one.
- Checked against the pass predictor. The search and the pass predictor are separate code paths. A test asks the search about every moment the pass predictor calls a test satellite visible, from five places on Earth, and requires the satellite named within 1°. Its look-angle maths is also checked against satellite.js to nine decimal places.
- What it can't know. Classified objects aren't in the public catalog, and new launches take a day or two to appear. Aircraft, meteors and planets are described, not computed. Times are your device's clock.
Automated validation
The SGP4 integration is validated against the official SGP4 verification suite published with Vallado's reference implementation (`SGP4-VER.TLE` + `tcppver.out`, mirrored by the python-sgp4 project). Our test regenerates the expected 700 output rows and compares the first seven numeric fields (TEME km, km/s) against the file at a tolerance of 2e-7, the same bound the python-sgp4 project uses, across all 33 element sets, including the known SGP4 failure cases.
Additional suites cover: TLE parsing/checksum rejection, coordinate conversions, ground-track antimeridian splitting, pass-finding bracket/bisection behaviour (including onset-at-window-start edge cases), the CelesTrak client against malformed/timeout/rate-limit fixtures, API validation and bounds, and snapshot persistence when configured. Run everything with npm test; the e2e flows live under tests/e2e (Playwright).
The code and its tests are on GitHub.