End-to-end satellite imagery processing, proven at scale

One real order

Fourteen captures in. One image out.

The Burdekin River delivery, replayed from its own files. Every layer is a real render, every polygon is the delivery's own metadata, and the registration numbers were measured on this order's captures.

The pipeline

How we process every order

Every step below runs on this real order as you scroll — the same automated chain behind every delivery. Nothing manual, nothing skipped.

Captures come off the downlink as Level-1 products — 11-bit pixels in sensor geometry, with RPCs, ephemeris and calibration coefficients intact. Nothing is resampled until the geometry is solved.

scenes in block

1 / 3
gsd

0.30 m
bands

PAN + 4 × MS
bit depth

11-bit
Tie points are matched across every overlap, then a bundle block adjustment solves all captures into one relative frame. Scene-to-scene drift collapses to sub-pixel before anything else happens.

tie points

14,202
relative rmse

0.31 px
solution

bundle block
Refined RPCs are anchored to ground control and a DEM, and every pixel is map-projected. Supply GCPs or a base layer and the block adjusts to it — your survey is the reference, not the satellite's.

control residual

0.24 m
reference

your GCPs
dem

1 m DSM
crs

EPSG:32633
The panchromatic band carries spatial detail; the multispectral bands carry the radiometry. Fusion injects PAN detail into each MS band while preserving band ratios — downstream indices stay valid.

pan

0.30 m
ms

1.2 m · 4-band
fused

0.30 m · 4-band
spectral ratios

preserved
Top-of-atmosphere radiance becomes surface reflectance: aerosol optical depth is estimated per scene, then path radiance, haze and adjacency effects are modelled out. Values become physically meaningful — comparable across dates and sensors.

input

TOA radiance
output

BOA reflectance
aod @ 550 nm

0.18
sun elevation

52.3°
Clouds, cirrus and their projected shadows are detected per pixel. The masks do more than flag bad data — they tell the mosaic engine which scene can contribute clean pixels where.

cloud

4.6 %
shadow

1.9 %
clean coverage across block

100 %
Cutlines are solved as least-cost paths — through low-contrast terrain, around masked clouds, never across structures. Watch each line detour: the area it cuts off is re-sourced from the scene that is clean there.

cost terms

gradient · cloud mask
cloud crossings

0
route

least-cost
Scene radiometry is normalised across the block, seams feathered, and the result written as one continuous surface with no visible joins. That is Legato: every stage above tied into one automated flow — nothing manual between them, nothing skipped — and the seamless, join-free mosaic is the proof the whole chain ran. Automated QC signs off geometry, radiometry and coverage before it ships.

seam Δe

0.4
qc · geometry

pass
qc · radiometry

pass
process

Legato · gapless
delivery

COG · tiles · STAC

01 / 08

Source capture & ingest

Captures come off the downlink as Level-1 products — 11-bit pixels in sensor geometry, with RPCs, ephemeris and calibration coefficients intact. Nothing is resampled until the geometry is solved.

scenes in block: 1 / 3
gsd: 0.30 m
bands: PAN + 4 × MS
bit depth: 11-bit

End-to-end satellite imagery processing with Legato, proven at scale

Fourteen captures in. One image out.

Source capture & ingest