Geospatial and remote-sensing AI
Imagery corpora, annotation protocols and sensor selection for teams who need remote-sensing work that will survive a technical review.
The model is not where the difficulty is
Most remote-sensing projects that disappoint did not fail at the modelling step. They failed earlier, in ways nobody looked for, and the modelling step simply inherited the problem and reported a number.
Two failures account for most of it.
The corpus is worse than it looks. We were once handed a prepared set of 6,095 RGB images. Hashing the contents showed 1,202 unique files — around eighty per cent were byte-identical duplicates. Annotating that set would have produced a corpus that looked four times larger than it was, an accuracy figure inflated by test images that had also been trained on, and no way to discover either. The check took minutes and ran before any annotator was paid.
The sensor already decided what is knowable. An instrument without a blue band cannot produce natural-colour composites, which means a whole class of visual judgement is off the table no matter what happens downstream. Multispectral imagery does not become hyperspectral because a model was applied to it. Choosing the instrument is the first modelling decision, and it is usually made by whoever had an archive lying around.
"What was that accuracy figure measured on?"
If the answer is not a specific, named set with a known labelling protocol, the figure is decoration.
How we approach it
Build the composites from raw bands. Ready-made products hide their processing history. Compositing from per-band scenes means the rejection criteria are ours, they are written down, and the rejection log is part of the deliverable — how many scenes were dropped for cloud or no-data coverage, how many were edge slivers, and what remained.
Deduplicate before annotating. Content hashing is the cheapest quality control in this field and the most frequently skipped one.
Design the taxonomy for the annotator, not for the paper. A small number of classes with unambiguous decision rules, each reachable with a single keystroke. A taxonomy that requires deliberation produces disagreement that looks like signal.
Make agreement measurable, then measure it. Every task is seen by two annotators, with a third arbitrating disagreements — selected deterministically by task id, so arbitration is idempotent and the load spreads evenly, and re-running the pipeline cannot silently change who adjudicated what. Then agreement is measured and reported for exactly the set it was measured on.
What you get
A sensor decision with its reasoning. An imagery corpus with provenance and a rejection log. A deduplicated, content-addressed set. A taxonomy and an annotation protocol reproducible from committed configuration. A measured agreement figure. And a plain statement of what your imagery cannot tell you.
When this is the wrong choice
- Satellite analysis narrows search areas. It does not find deposits. Vegetation, surface cover and weathering all degrade what any orbital instrument can report, and no amount of modelling removes that.
- If you need hyperspectral mineral identification, multispectral archive imagery is the wrong instrument and we will say so rather than build on it.
- If you need sub-metre detail or frequent revisits, a decades-old archive at moderate resolution is not the right source, whatever its price.
- If no ground truth exists and nobody can produce labels, no annotation protocol saves the project. That is a data-collection problem wearing a machine-learning costume.
- Classification accuracy is not exploration success, and the two are not interchangeable in a sentence, a slide, or a contract.
- Our own current deliverable is the corpus and the protocol. No trained model or model artifacts exist yet from this work, and anyone describing that stage differently is selling something.