Every supervised model is a compression of its training data — and in geospatial AI, that data must be created: polygons around buildings, masks over crops, labels on change. Annotation looks like unskilled work and is anything but; consistency rules, edge-case policy and quality scoring decide whether a model learns signal or noise. Teams that treat labelling as a production discipline ship better models than teams with fancier architectures.
Specification before annotation
Good programmes begin with an annotation specification: what counts as a building (sheds? ruins? under construction?), where boundaries run (roof edge or footprint?), how occlusions and ambiguity are handled, minimum sizes, class definitions with visual examples. Without it, ten annotators produce ten datasets.
The spec evolves through pilot rounds where disagreements surface — cheap corrections at page stage, expensive ones after a hundred thousand labels.
Quality as a measured property
Professional pipelines measure labels like any product: inter-annotator agreement on overlap samples, review tiers for hard cases, gold-standard tasks seeded through the stream, error taxonomies feeding annotator coaching. Quality scores accompany the dataset, not just the promise of care.
Geospatial adds its own checks: geometric precision against imagery, topology of adjacent labels, alignment across image dates, and CRS bookkeeping so labels land on the right pixels forever.
Strategy: what to label
Naive random labelling wastes budget on easy redundancy. Active learning routes annotators to samples the current model finds uncertain; hard-negative mining targets its confident mistakes; geographic stratification buys the diversity that transfers. Synthetic augmentation stretches datasets but never replaces real edge cases.
And datasets are living assets: versioned, documented (source imagery, date, licence, spec), and extended as deployment reveals new failure modes.