sgp4-predict — Python bindings

Python bindings for the sgp4-predict Rust library: satellite pass prediction from TLE or OMM data, with type stubs and native datetime support.

pip install sgp4-predict

Quick start

Find the passes over a ground station and sample each one:

from datetime import datetime, timedelta, timezone
from sgp4_predict import GroundObserver, Interval, Predictor, Tle

tle = Tle(
    "SENTINEL-2C",
    "1 60989U 24157A   25356.66913557  .00000141  00000+0  70244-4 0  9990",
    "2 60989  98.5671  69.0082 0001197  95.1447 264.9872 14.30821394 67740",
)
predictor = Predictor.from_tle(tle)

glasgow = GroundObserver(latitude_deg=55.86, longitude_deg=-4.25, altitude=40.0)

start = datetime(2025, 12, 22, tzinfo=timezone.utc)
window = Interval(start=start, end=start + timedelta(days=1))

for transit in predictor.transits_iter(glasgow, window, min_elevation_deg=5.0):
    print(f"AoS {transit.start} → LoS {transit.end} ({transit.duration_seconds:.0f}s)")

    # A Transit is itself an interval, so it can be passed straight back in.
    for t, obs in predictor.observation_iter(glasgow, transit, timedelta(seconds=10)):
        print(f"  {t}  az={obs.azimuth_deg:.1f}°  el={obs.elevation_deg:.1f}°")

Predictor

Construct from a Tle, or from Elements parsed from OMM JSON:

from sgp4_predict import Elements

predictor = Predictor.from_tle(tle)                  # raises ValueError if malformed
predictor = Predictor(Elements.from_json(omm_json))  # CCSDS OMM (Celestrak, Space-Track)

SGP4 accuracy degrades with element age — check predictor.tle_age_seconds(now) and treat LEO TLEs older than 3–7 days with caution. Fresh TLEs are available from CelesTrak.

Besides transits_iter, Predictor yields apsides, sunlit and eclipse windows, ground-track points, area overpasses and raw state vectors, and answers point queries such as propagate, observe_at and sub_point. The package ships type stubs, so an IDE or help(Predictor) lists them with their signatures and units.

Every method taking a time range accepts any object with .start and .end datetime properties — an Interval, a Transit, an Illumination, an AoiWindow, or your own type. All iterators are lazy.

Predictor.with_refinement(Refinement(...)) returns a copy with a different root-finder configuration for event times.

Areas of interest

An area is a region on the ground; aoi_iter yields the windows in which it is within the payload's reach. Points are LatLon objects, Geodetic objects whose altitude is ignored — so a sub_point result can be passed straight in — or plain (latitude_deg, longitude_deg) tuples.

from sgp4_predict import Circle, LatLon, Polygon, Rectangle

# An arbitrary ring. Concave and self-intersecting rings are both fine, the ring
# closes itself, and vertex order does not matter.
scotland = Polygon([(54.0, -8.0), (54.0, -1.0), (60.0, -1.0), LatLon(60.0, -8.0)])

# A latitude/longitude box, whose north and south edges follow their parallels
# exactly. Runs eastward from the south-west corner, so this one wraps the
# antimeridian.
pacific = Rectangle((-20.0, 160.0), (20.0, -160.0))
arctic = Rectangle.latitude_band(66.5, 90.0)

# A circular area 500 km across. The radius is angular — a degree of arc is about
# 111.2 km on the ground.
cape_town = Circle((-33.9, 18.4), radius_deg=2.25)

for overpass in predictor.aoi_iter(scotland, window):
    print(overpass.start, overpass.end)

A malformed area raises ValueError — fewer than three distinct vertices, a latitude outside [-90, 90], a nan or infinite coordinate, a polygon larger than a hemisphere, an empty box, or a circle radius outside (0, 90).

By default a window opens when the ground track itself crosses into the area. Pass max_off_nadir_deg — the half-angle of the satellite's field of regard, the largest nadir angle the payload can be slewed to — and the window instead covers whenever the area is within reach. coverage chooses whether any part of the area or all of it must be in reach:

from sgp4_predict import Coverage

for window in predictor.aoi_iter(cape_town, interval, max_off_nadir_deg=30.0):
    print(window.start, window.end)

# Every part of the area reachable at once, rather than any part of it. Needs a
# field of regard wider than the area, or no window ever opens.
predictor.aoi_iter(cape_town, interval, max_off_nadir_deg=30.0, coverage=Coverage.Full)

A window longer than max_window_duration — one hour by default — raises RuntimeError rather than being yielded, since a window that long usually means the area is bigger than intended. Raise the cap for an area that really is near-global:

band = Rectangle.latitude_band(-90.0, 60.0)   # ~85 min of each 100-minute orbit
predictor.aoi_iter(band, window, max_window_duration=timedelta(hours=2))

An area the ground track never leaves at all — a whole-Earth box, or a band wider than the orbit's inclination reaches — has no window end to find, so it raises whatever the cap is set to.

min_step is the shortest crossing the scan is guaranteed to see. Lower it below the default second for an area the ground track crosses faster than that. It also raises the scan's ten-minute upper bound wherever it exceeds it, so a min_step above that pins every step there and a small area is passed straight over.

Every detection method takes its tuning the same way, as keyword-only arguments left at the library's defaults unless passed.

Polygon edges are great-circle arcs, so two vertices at the same latitude are not joined along the parallel: the arc bows toward the nearer pole, by about 0.05° for the 7° ring above and roughly 8° for vertices a quarter of the globe apart. Densify edges that long, or use Rectangle when the region really is "these latitudes by these longitudes". A polygon must also fit inside a hemisphere — polar caps, equator-spanning and antimeridian-spanning areas are all fine, a region larger than half the globe is not.

Coordinate frames

propagate() returns a StateVectorTeme. The frame chain is available step by step if you need an intermediate:

sv_teme = predictor.propagate(t)   # StateVectorTeme
sv_ecef = sv_teme.to_ecef(t)       # StateVectorEcef  (GMST rotation)
sv_enu  = sv_ecef.to_enu(gs)       # StateVectorEnu   (geodetic to local ENU)
obs     = sv_enu.to_observation()  # equivalent to predictor.observe_at(t, gs)

All three expose .position and .velocity as Vec3(x, y, z).

Units

SI throughout: positions in metres, velocities and range rate in m/s, apsis altitude in metres above the WGS-84 equatorial radius. Angles are plain floats in degrees — every angular name is suffixed _deg.

Azimuth is measured clockwise from north over (-180, 180], so a southwesterly bearing is negative rather than the [0, 360) most tracking software reports.

Development

cd sgp4-predict-py/
uv sync --extra dev   # create .venv and install dev dependencies
make dev              # compile the Rust extension in-place (maturin develop)
make test             # compile + run pytest
make lint             # ruff check --fix + ruff format

make targets use uv run, so no venv activation is needed.

License

Licensed under either of

• Apache License, Version 2.0 (LICENSE-APACHE or https://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or https://opensource.org/licenses/MIT)

at your option.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this crate by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.