SatScenes

SatScenes generates photorealistic labeled imagery of satellites and space debris in Earth Orbit, but easily adaptable to other celestial body orbits, purpose-built for training deep learning object detection models on-board spacecraft. Where existing benchmarks such as SPEED, SPARK 2022, and DLVS³ are anchored to few and limited target spacecrafts, SatScenes scales without bound: adding a new target requires only a 3D model import, with zero pipeline changes.

SatScenes is designed to support a virtually infinite number of satellites and space debris models. Up to now 6 satellite models are available: Starlink V1, Sentinel-3B, ION Mk2, Nimbus, RadarSat-1 and a generic Stylized Satellite.

The rendering engine is a physically-based path tracer, producing frames with fully configurable optics and resolution. Field of view and sensor dimensions are configurable via JSON config. The possibility to configure lens distortion coefficients, e.g. radial k1/k2/k3 and tangential p1/p2, enabling faithful replication of any real camera geometry, is being integrated.

Orbital altitude is sampled from a configurable range, defaulting to the 400–800 km LEO band. Each frame draws independently from a full scene parameter space: target scale (extreme far to close-up), in-frame position (centred to partially visible), attitude (frontal/lateral/oblique/rear/edge-on), lighting condition (fully lit/partially lit/grazing/eclipse), background type (Earth dominant/Earth partial/deep space/Moon visible), and focus (sharp/slight blur/motion blur). This combinatorial diversity means no two frames are alike.

Class composition is user-controlled via probability weights: assign any distribution across your satellite catalogue — for example 60% Sentinel-3B, 20% ION Mk2, 10% Nimbus, 10% RadarSat-1 — and regenerate a balanced or intentionally skewed dataset in a single command. Output is RGB frames paired with vast set of metadata in json format, e.g. bounding box annotations, satellite model, position, attitude.

The flexibility in dataset configuration aims at supporting different use cases, from basic satellite detection/determintation for SSA/SDA applications in a wide range of orbits to creation of very specialized models for more specific applications.

Next version of the pipeline will target generation of datasets for position, velocity and pose estimation. The configurability of the pipeline will still provide an advantage with respect to existing datasets.