Why geospatial AI needs human review

AI as an analytical assistant

Models for segmentation, classification, and change detection can scan large areas quickly and consistently. Their real value is prioritization: showing people where to look first.

Treating model output as a final answer skips the step where context, local knowledge, and judgement come in.

Where models struggle

Satellite imagery varies enormously across the planet, and models trained on one setting can falter in another.

• Regions, landscapes, and building styles under-represented in training data
• Seasons and illumination conditions the model has rarely seen
• Thin cloud, haze, and shadows that mimic surface change
• Changes that look alike from orbit but have different causes

Confidence only helps when it is calibrated

A score of 0.9 is only meaningful if, across many similar cases, about nine in ten are correct. Uncalibrated scores can create false certainty.

Thresholds and confidence indicators should be shown only when they have been validated for the setting where they are used.

A practical review loop

Review works best when it is part of the workflow rather than an afterthought.

• Show each finding with the observations, dates, and quality notes behind it
• Let reviewers confirm, correct, or dismiss, and record who did what
• Use corrections to evaluate and improve models over time
• Keep a clear status on every finding: detected, flagged, confirmed, or dismissed

Document intended use and limits

Responsible geospatial AI documents what each model is for, where it was evaluated, and where it is known to be weak. Evaluation on geographically diverse imagery and across seasons is part of that record.

Risk indicators and change findings are decision support. They do not replace engineering, regulatory, emergency-management, or other professional assessments.