01 / THE QUESTION
The problem
Remote-sensing projects depend on the quality of their imagery and labels. Duplicates, unusable scenes, and inconsistent annotation can undermine a model before training begins.
02 / THE MAKING
My role
- I worked on turning a specialist exploration problem into a software and AI product direction.
- The team built a corpus from Landsat imagery, checking source quality and eliminating duplicate outputs.
- The team established and tested a two-annotator review process with arbitration for disagreements.
03 / THE REALITY
Where it stands
The documented pipeline evaluated 10,348 source scenes and produced 9,732 unique clean images. A 150-image annotation pilot measured 82.7% agreement.
A note on scope. Full annotation is pending and there is no trained model or mineral-discovery claim. The figures describe the documented corpus and pilot, not exploration accuracy.