Technical Notes
Is it a change, or a different photograph?
When an image pair disagrees, check alignment, coverage and capture conditions before escalating the difference.
Reference from 2023: USGS: Guidelines for calibration of uncrewed aircraft systems imagery

A model flags the difference. The reviewer checks it.
A dark patch could be a new object or a shadow from an object already there. A shifted edge can come from registration error. Missing imagery may simply mean cloud or no overlap. Treat an automated result as a candidate. Record what would turn that candidate into an observation worth handing to an operator.
Use a baseline you can defend
Start with fixed features on both dates. Check the alignment and restrict review to usable overlap. Look for changes in view angle, illumination, season, moisture and resolution before comparing the ground. If the captures are not comparable in an area, say so. “Not assessable” means something different from “no change seen.” Keep the baseline identifier and your reason for choosing it.
Example: a mark beside a terminal gate
A dark shape appears near the lane. In the original frame, check whether the edge follows a crane shadow. Compare another view if one exists, then look at the lane boundary and the operating record. Material can be present without blocking access. A work order may explain why it is there, but the operator may still need to confirm whether vehicles can pass.
Count failures beyond the obvious flags
Review places the model did not flag as well as places it did. Split evaluation by site or time so neighboring crops do not make the test look easier than the next assignment. Track unsuitable coverage separately from a missed finding. Otherwise the team may try to fix a detector when the real problem is acquisition.
Practical checklist
- Check stable ground and valid overlap before interpreting differences.
- Compare the original frames, not just the overlay.
- Save the baseline, review outcome and any follow-up request.