Custom Satellite Imagery Datasets
Physics-accurate, mission-specific training data for on-board AI built to your sensor, orbit, and target constellation.
Design Your Dataset ↓Mission-specific training data, built to your specifications
Getting high-quality, mission-specific training data for on-board AI is hard. Public benchmarks like SPEED, SPARK 2022, and DLVS³ are anchored to a handful of fixed spacecraft, wrong targets, wrong sensors, wrong orbits for your mission.
SatScenes delivers custom photorealistic labeled datasets built to your requirements: your sensor geometry, your orbital scenario, your target constellation. Every dataset is generated with a physics-accurate rendering engine producing imagery that is photometrically faithful to real space cameras, not just visually plausible.
Each frame independently samples the full scene parameter space: target scale, in-frame position, attitude, lighting condition, background type, and focus, ensuring no two frames are alike and your model trains on genuinely diverse data. Class composition is controlled by probability weights, so you define exactly which satellite models appear and in what proportion.
The model library covers spacecraft from 3U CubeSats to the International Space Station and the most common rocket bodies, across low, geostationary and high Earth orbits, each with physical dimensions, body axes, per-face component labels and calibrated surface materials (MLI, solar cells, radiators, paints, bare metals). Every model states how accurate it is, so a dataset can keep only the most faithful ones. Targets can be selected by nation, operator, orbit class or object type, and your own models can be added.
Output is RGB frames with per-frame JSON metadata: bounding box annotations, camera intrinsics, satellite model, position and attitude, ready to convert into YOLO, COCO, or any annotation format your pipeline expects. Per-pixel masks of components, surfaces and instances can be added for segmentation and part-level tasks.
Train satellite detection and classification models across LEO, MEO, and GEO orbits. Datasets cover diverse lighting conditions, background types, and target attitudes, giving your model the breadth to perform in real operational environments.
Enable autonomous proximity operations, rendezvous, and satellite inspection with models trained on photorealistic imagery matched to your camera geometry and mission profile.
Don't train against irrelevant targets. Define exactly which satellites your system will encounter and generate a dataset focused on your catalogue, improving model precision and reducing false positives in deployment.
Future dataset types will support position, velocity, and pose estimation, extending the service to academic research and next-generation on-board capabilities. New satellite models can be added without pipeline changes.
Powered by the SatScenes rendering pipeline
Custom datasets are generated through SatScenes, a physics-accurate rendering pipeline built specifically for this purpose. From a single JSON configuration file to a fully labeled dataset, the entire process is automated: no manual scene setup, no hand-labeling.
The sensor model goes beyond geometry: it simulates quantum efficiency per color channel, 12-bit ADC, readout noise, dark current, pixel response non-uniformity, and Gaussian PSF, producing imagery that faithfully replicates the photometric characteristics of real space cameras. Support for lens distortion coefficients (radial k1/k2/k3, tangential p1/p2) is in active development and will enable faithful replication of any specific camera geometry.
Define target models, class weights, orbit altitude range, sensor profile and scene diversity parameters via JSON config.
The rendering engine executes the full scene graph (Earth, target satellite, lighting, camera), one frame per parameter sample, fully automated.
RGB frames with per-frame JSON metadata, including bounding box coordinates, camera intrinsics, and full scene parameters, ready to convert into YOLO, COCO, or any custom annotation format. Optional per-pixel component, surface and instance masks.
Examples of deliverable imagery
Real photos of satellites in orbit, reproduced with the same camera
Render-ready models with component labels, as of October 2026
#1 for number of supported spacecraft models
176 distinct satellite and rocket-body models
More than any public spacecraft imagery dataset: the largest counts 136 spacecraft (SpaceSense-Bench, 2026). Real satellites alone: 74.
8,901 operational satellites represented
52 % of all satellites operating in orbit today
Out of 17,270 operational satellites, each covered by a model of the same satellite or of its series.
Plus 3 further institutional satellites. More invented satellites can be designed on demand.
| Nation | Models | Share of its operational satellites represented |
|---|---|---|
| United States | 55 | 62.9 % |
| United Kingdom | 1 | 93.1 % |
| Finland | 1 | 50.0 % |
| Italy | 2 | 26.2 % |
| France | 3 | 14.5 % |
| ESA | 3 | 3.8 % |
| Japan | 4 | 1.3 % |
| Russia | 1 | 0.3 % |
| Argentina | 2 | 0 % |
| Canada | 1 | 0 % |
| International (ISS) | 1 | — |
A model represents an operational satellite when it is that satellite or the same series and version (for Starlink, the same GCAT bus family); the represented satellites also include ones operated by the Netherlands, Poland, the UAE, Greece, Portugal and Sweden. Operational satellites from the CelesTrak catalogue, bus versions from GCAT (J. McDowell, planet4589.org), October 2026.
I'm Lorenzo Voli, a Space Engineer and AI researcher with 10+ years of software experience in particular for spacecraft avionics and spacecraft HIL framework development, currently serving as Institutional & Defense BU Tech Strategist at D-Orbit S.p.A.
My work spans the full stack of spacecraft software: from RTOS embedded systems and CAN bus protocols on ION spacecraft, to leading avionics SW development for a SAR satellite of the Iride constellation, including payload data handling, FDIR, and symmetric encryption subsystems.
I hold an MSc in Space Engineering from Politecnico di Milano, where my thesis research on adaptive hybrid control for robotic manipulators was published in the Journal of Intelligent & Robotic Systems (Springer, 2018).
SatScenes is a personal research initiative born from the need for flexible, physics-accurate training data for on-board AI, a gap I encountered firsthand in operational satellite programs.
Journal of Intelligent & Robotic Systems, Springer, 2018
Read on Springer ↗Interested in a custom dataset, a collaboration, or just curious about what I'm building? I'd love to hear from you.
If you are a university professor or student working on AI applied to space technology (detection, navigation, pose estimation), I'd be glad to collaborate on research datasets and joint exploration.
If you're a space and AI enthusiast looking for adventure, let's connect, share dreams, and see where curiosity takes us.
Message received. I'll get back to you soon.
SatScenes ranks #1 for number of supported spacecraft models: 176 distinct satellite and rocket-body models as of October 2026, 74 of them real satellites. The largest public spacecraft imagery dataset, SpaceSense-Bench (2026), counts 136 spacecraft.
Yes. Public benchmarks such as SPEED, SPARK 2022 and DLVS3 are anchored to a handful of fixed spacecraft. SatScenes generates custom datasets built to your sensor geometry, orbital scenario and target constellation, choosing the satellites from a library of 176 models.
8,901 of the 17,270 satellites operating in orbit today, 52 %, are covered by a model of the same satellite or of its series (CelesTrak catalogue + GCAT bus versions, October 2026).
Yes. You define exactly which satellites appear and in what proportion through class weights; targets can be selected by nation, operator, orbit class or object type. If a satellite is missing you can provide its 3D model (.glb, .gltf or .fbx), and new models can be added without pipeline changes.
The sensor model simulates quantum efficiency per colour channel, a 12-bit ADC, readout noise, dark current, pixel response non-uniformity and a Gaussian point spread function. Lens distortion (radial k1/k2/k3, tangential p1/p2) is in active development.
RGB frames with per-frame JSON metadata: bounding boxes, camera intrinsics, satellite model, position and attitude, ready to convert into YOLO, COCO or any annotation format. Per-pixel masks of components, surfaces and instances can be added for segmentation and part-level tasks.
Yes. The library includes the 13 most common rocket bodies, about two thirds of the rocket bodies in orbit, alongside satellites in low, Sun-synchronous, medium, highly elliptical and geostationary orbits.
Images come from a physics-accurate rendering engine that is photometrically faithful to real space cameras. Real photographs of spacecraft in orbit, such as the ISS seen from Crew Dragon, have been reproduced with the same camera to validate geometry, models and materials. Every model states how accurate it is, so a dataset can keep only the most faithful ones.
Universities and researchers working on AI for space (detection, navigation, pose estimation) interested in research datasets and collaborations, and space and AI enthusiasts. Get in touch through the contact form.