STARDUST

Synthetic data for aerial autonomy and ISR

EO, IR, and SAR imagery for airborne ISR, from altitude and slant range. Generate the rare targets, hard look-angles, and sensor conditions a flight campaign would take a year to capture, labeled to the pixel.

BOOK A DEMO READ THE DOCS

One dataset in a day, not eleven weeks.

You cannot fly your way to enough data

Airborne perception is throttled by flight hours. A single real ISR dataset takes around eleven weeks to deploy, collect, and label, and the targets that matter most are rarely in frame. Each new altitude, slant range, and look-angle multiplies the data you would need, and small-object labeling at scale stays manual, slow, and subjective.

Stardust generates sensor-true EO, IR, and SAR imagery across the full geometry of flight, with pixel-perfect labels and fine-grained classes. Rare targets that have no real dataset, low-profile vessels or specific aircraft types, become trainable on demand. Teams reach real-data accuracy around 12x faster, and produce in a day what a flight campaign produces in a quarter.

Why real data falls short

// collecting from the air today

  • ✕Flight time is expensive and the rare targets you care about are rarely in frame
  • ✕Slant range, altitude, and look-angle variation explode the data you need
  • ✕Annotating small objects from altitude by hand is slow, subjective, and does not scale
  • ✕EO alone breaks at night and in weather, and real SAR training data is scarce and hard to label

Sensor-true data, perfectly labeled

Every altitude and look angleRender targets across the full range of altitudes, off-nadir angles, slant ranges, and standoff distances.
EO, IR, and SAR, registeredDay, night, thermal, and radar in registration for fusion and all-weather ATR.
Small-object ground truthPixel-perfect boxes and masks, plus fine-grained classes, on targets too small to label reliably by hand.
Rare targets on demandGenerate classes that have no real dataset, from low-profile vessels to specific aircraft types.
SENSOR COVERAGE
EOMWIRLWIRSWIRSARFMV

EO/RGB, thermal MWIR and LWIR, SWIR, and physics-based SAR, configurable to match a specific sensor.

Inside the aerial data

Wide-area EO surveillance Airborne detection over open water
SWIR vs RGB See targets the visible band misses
Simulated SAR Coherent radar imagery for SAR ATR
Counter-UAS tracking Persistent track through maneuver
Pixel-perfect labels Automatic segmentation from altitude

What teams build with it

ISRWide-area detection and monitoring from airborne platforms.
EO/IR ATRAutomatic target recognition across day, night, and thermal.
Full-motion video analysisTrain detection and track on FMV from UAV platforms.
Fine-grained classificationTell aircraft and vessel types apart, not just detect them.
SAR ATR and ship detectionDetect and classify vessels in coherent simulated SAR.
Counter-UASDetect and track small drones against cluttered backgrounds.
Change detectionSurface what moved or appeared between passes.
Slant-range and oblique detectionDetect and track at standoff and off-nadir angles.
11 wks → 1 dayto build a labeled aerial dataset
12×faster to real-data accuracy
95%+accuracy detecting low-profile vessels from EO and IR
+25%F1 on fine-grained aircraft classification

Speaks your domain

The vocabulary, sensors, and benchmarks aerial teams actually use.

ISRATREO/IRMWIRLWIRSWIRSARSAR ATRFMVWAMIGEOINToff-nadirslant rangelook anglestandoffGSDNIIRSchange detectioncounter-UASfine-grained classification

Built with airborne ISR and geospatial programs.

NTT DataST Engineering

Questions teams ask

How do you train EO/IR ATR with synthetic data?

Generate sensor-true EO and IR imagery of each target class across look angles, ranges, and weather, with pixel-perfect labels, then train and test against held-out real data.

Can you simulate SAR imagery?

Yes. Stardust simulates SAR by ray tracing coherent complex radar returns, supports ascending and descending passes, and can composite synthetic vessels onto real SAR backplates. We build SAR ship-detection data with NTT Data.

How much faster is synthetic data than real flight collection?

A real aerial dataset takes around eleven weeks. Stardust produces one in a day and reaches real-data accuracy about 12x faster.

Can synthetic data detect rare aerial targets?

Yes. Classes with no real dataset, from low-profile vessels to specific aircraft types, are generated on demand. One program reached over 95% detection accuracy this way.

Does synthetic data work for counter-UAS?

Yes. A synthetic drone dataset beat a 13,000-image real set, 95.9% F1 versus 48.7%, built in hours instead of months.

Unlock unlimited aerial datasets

Tell us what you are building and the scenarios you need. We will get you access.

BOOK A DEMO EXPLORE STARDUST