What it does

ZetaMine analyses geographic regions using satellite imagery and machine learning, to support decisions about where mineral exploration is worth pursuing. The output is a narrowed search area, not a discovery.

That distinction matters more than any feature list. Orbital sensing is good at telling you where not to spend a drilling budget. Treating it as a deposit-finder is how exploration programmes waste money.

The dataset is the engineering problem

Most of the difficulty in applied remote sensing is not the model. It is that the corpus you train on is quietly worse than it looks.

We were given a ready-made directory of RGB images to work from — 6,095 files. Content-hashing them before anyone started annotating showed 1,202 unique images. Roughly 80% were byte-identical duplicates. Had we skipped that check, the resulting corpus would have appeared four times larger than it was, agreement statistics would have been inflated by repeated images, and any model trained on it would have carried the distortion silently.

So we built the composites ourselves:

10,348 source scenes — LANDSAT GLS2005 Landsat-7 ETM+, stored one greyscale image per band
Band 30/20/10 → R/G/B — merged into natural-colour composites
491 scenes rejected — more than 50% no-data
125 edge slivers rejected — partial-coverage artifacts
9,732 clean images, 741 MiB — every one content-hash unique

ASTER was evaluated and rejected: it has no blue band, so natural-colour composites are not possible, and the visual judgements the annotation protocol depends on become unreliable.

The annotation protocol

Eight single-choice Persian classes, hotkeys one through eight, in Label Studio. Each image is labelled by two annotators. Where they disagree, a third arbitrates — and the arbiter is chosen deterministically by task id, which makes the assignment idempotent and spreads load evenly across the team. Re-running the pipeline cannot silently change who adjudicated what.

We verified this end to end on live data, including the negative case: an arbitration submitted by the wrong annotator is refused.

Where this stands

The corpus is built. The protocol is validated. Pilot agreement across 150 images was 82.7% — 124 in agreement, 26 arbitrated.

Full annotation is pending, and no trained model exists yet. We are publishing the honest state of the work rather than a rounder number, because the whole argument for using remote sensing in exploration rests on being straight about what it can and cannot resolve.

Technical decisions

Including what we did not choose, and why.

Built our own natural-colour composites from raw per-band scenes rather than using a supplied RGB directory.

Alternatives: Use the ready-made RGB image set

Content-hashing the supplied directory showed 6,095 files contained only 1,202 unique images — roughly 80% byte-identical duplicates. Annotating that set would have produced a corpus that looked four times larger than it was.

Hash comparison run before any annotator time was spent.

Rejected ASTER as the imagery source.

Alternatives: ASTER, Landsat-7 ETM+

ASTER has no blue band, which rules out natural-colour composites and the visual judgements our annotation protocol depends on.

Two-of-two overlap with deterministic third-annotator arbitration.

Alternatives: Single annotator, Majority vote across three

A single annotator gives no agreement signal at all. Deterministic arbitration — selecting the arbiter by task id — is idempotent and spreads load evenly, so a re-run cannot silently change who adjudicated what.