1. The console in the sky

William Gibson’s Neuromancer is full of operators pulling data out of raw signal and turning it into something actionable. The desktop’s compute attachment is doing the same thing, except the console is moving at 7 kilometres per second and the data is photons that have already bounced off a field or a roof.

Entry 233 sized the sensor’s power and thermal impact. This entry decides what the compute attachment does with the raw images before they go down to customers.

2. Why process on orbit at all

The simplest pipeline is to dump every raw frame to the ground and let ground stations do the rest. That avoids putting heavy software on the desktop and lets operators use mature tools. But it also wastes downlink bandwidth on clouds, dark frames, and redundant overlap.

The desktop has compute capacity, local storage, and an incentive to reduce bandwidth costs. It makes sense to do enough processing on orbit to:

  • Discard or deprioritise unusable frames.
  • Compress the useful ones efficiently.
  • Attach geometry and metadata so ground systems can ingest the product directly.

The goal is not to replace the ground pipeline. The goal is to make the downlink payload smaller and more valuable.

3. Level-0 to level-1A

Level-0 is the raw detector readout with telemetry attached. The first processing step reconstructs the image frames, applies bad-pixel maps, and subtracts dark current and bias. This produces a level-1A product: a radiometrically corrected image in detector geometry.

This step is deterministic and well suited to a small onboard computer with a GPU or FPGA. It needs calibration tables uploaded from the ground and updated periodically as radiation damage accumulates.

4. Level-1B to level-1C

Level-1B adds geometric correction. Using spacecraft position from GPS, attitude from star trackers, and the bracket angle from encoders, the desktop maps each pixel to a location on the ground. Level-1C resamples the image to a standard map projection, typically UTM, with a consistent pixel grid.

For a 3–5 m imager, full orthorectification that accounts for terrain is overkill for many customers. A simple terrain-corrected product using a public digital elevation model is enough for agriculture and forestry. Customers who need survey-grade geometry can reprocess the level-1A data on the ground.

5. Cloud and quality screening

A cloud mask is the most valuable piece of onboard filtering. If an image is more than a threshold percentage of cloud cover, it can be compressed at lower priority or dropped from the downlink queue entirely.

Simple cloud detection uses the near-infrared and short-wave infrared bands with thresholds on reflectance. More advanced methods use machine learning classifiers trained on labelled data. For the first desktop, a threshold-based cloud mask is enough. A neural-network classifier can be added later once a training dataset exists.

Quality screening also flags images with excessive smear, saturation, or pointing error. These frames are logged but not downlinked.

6. Compression

Lossless compression of raw 11-bit imagery typically achieves a 1.5:1 to 2:1 reduction. Visually lossless compression can reach 4:1 to 8:1 without degrading most analytic products.

The desktop should store raw level-0 data for a limited time and downlink compressed level-1C products as the default. Customers who need lossless data can request it as a higher-priced product and trigger a dedicated downlink pass.

7. The object-store staging model

Each processing step writes its output to the object store described in entry 224. The pipeline becomes a set of jobs triggered by new arrivals:

  • ingest writes level-0.
  • l1a produces radiometrically corrected frames.
  • l1c produces map-projected images.
  • cloudmask attaches a quality band.
  • downlink-queue decides what gets transmitted.

This model reuses the same compute-storage-communications architecture that handles customer payloads. The desktop does not need a separate image-processing appliance.

8. Split between orbit and ground

Not everything should be done in orbit.

On orbit: radiometric correction, geometric tagging, cloud screening, compression, queue management.

On ground: atmospheric correction to surface reflectance, change detection over time series, machine-learning feature extraction, customer-specific analytics, archiving.

The ground pipeline is where the desktop’s business model lives. The onboard pipeline exists to feed it efficiently.

What this changes

  • The desktop runs a standard L0-to-L1C pipeline on board for its baseline imager.
  • Cloud and quality screening reduce wasted downlink bandwidth.
  • Compression is tuned to the product tier: lossy for visual products, lossless available on request.
  • The pipeline uses the existing object-store and job model, not a separate subsystem.
  • The next entry can define the data products the desktop sells and their pricing logic.