Why this method
Remote measurement has a credibility problem it earned. The field’s history is full of maps that looked confident and meant little, indices renamed as insights, and accuracy figures quoted far from the ground they were measured on. Anyone who has paid for one of those maps carries the scepticism into every conversation that follows. We build for that reader.
Our answer is a method rather than a promise. Fit physics against public reference standards so anyone can check the chain. Refuse the pixels that cannot be measured instead of averaging them away. Publish what a score means and what it does not. None of this is novel science; it is laboratory discipline carried into orbit, which is exactly why it works.
Honesty here is not a virtue we advertise, it is the only approach that survives contact with a technical buyer. A geologist, a reservoir engineer or a forester will test a layer against ground they already know. A method that admits its limits passes that test and earns the next job. A method that overclaims fails it once, permanently.
The principles
Seven rules, enforced in code, applied to every layer the platform produces.
Provenance travels with the result
Each layer records the scene it came from, the acquisition time, the processing time, and the version of the model that ran. The record ships with the export. A map opened six months from now still traces back to its inputs.
Public reference standards
Every fitted product is matched against published laboratory reference standards rather than a proprietary training set. The physics is public, and the chain from measurement to answer can be checked by anyone who cares to.
Masked ground is counted, not guessed
Cloud, shadow and unreadable pixels are marked as no data, and the masked share is reported with the result. On a live mineral scene in the demo workspace that share is 16.1%, stated plainly rather than folded into the nearest class.
Confidence ships with every answer
A classification is never delivered alone. Each one carries a per-pixel fit quality layer scoring how well the ground matches the class it was given, so a reader can see exactly where the map is strong and where it is thin.
Statistics are published as measured
A scene statistic is the mean or range of the measured value over that scene on that date, and it is labeled that way. A mean detection score of -0.07 is reported as -0.07.
Frozen, versioned models
A model that scores ground is pinned before it ships and named on every layer it produces. Results do not drift under a moving model, and two maps made a month apart can be compared because the same versioned method made both.
Independent cross-checks
The mineral lane is run end to end against an independent expert-system classifier over the same ground, a check we do not control and cannot tune. Agreement is earned, not asserted.
Where the numbers stop
Every figure the platform reports is a scene-level measurement from a single acquisition. It describes surface response in one place on one date, and it is dated so you can tell.
Mineral results describe surface alteration, never concentration, grade or reserves. Methane and vegetation statistics are screening values: they rank ground and direct attention, and field measurement settles what they raise.