Turn a map into usable evidence

Two images must refer to the same ground locations before their differences are interpreted. In an invented example, a road’s edge is shifted by one 10-metre grid cell between maps because of a positioning error. Subtracting the images can produce apparent change along the edge even if nothing on the ground changed. Check the coordinate system and alignment against suitable reference features.

Resampling places data onto a chosen grid. It can make layers convenient to compare, but it does not create new independent measurements or remove every positioning error. Keep the original acquisition time, sampling and processing history. If one layer is more uncertain than the other, the final map should not imply that all its boundaries are equally precise.

A fictional validation exercise labels 100 checked locations as water. Independent reference observations show that 80 are water and 20 are not. The fraction of water labels that are correct is 80 percent. Now suppose the reference set contains 120 water locations in total. The map found 80 of them and missed 40, so it detected about two thirds of the reference water locations.

These two questions measure different things: how trustworthy the positive labels are, and how much of the real water was found. A map can perform well on one and poorly on the other. Reference observations need appropriate dates, locations and independence from the classification being tested. Validating with the same labels used to construct a rule can give a misleadingly optimistic result.

Two validation questions use different denominators

Intersecting a flood-extent layer with a settlement or road layer can identify places for closer investigation. It does not by itself measure water depth, structural damage, whether a route is passable or how many people were present. A feature may also fall near an uncertain boundary. Useful reporting distinguishes mapped exposure, independently confirmed impact and unresolved questions.

For a final invented example, 500 ideal square water pixels at 5-metre spacing cover 12,500 square metres, or 1.25 hectares. That is the area represented by the classification under the stated assumptions. It is not automatically an exact damaged-crop area. A useful product includes its date, input sources, validation limits and a clear explanation of what the mapped quantity means.