01Light and shadow swamp the signal
On a carved facade, the difference between morning and afternoon is larger than the difference between this year and last. Generic change detection flags the time of day.
What we do. The detector is evolved per site on image pairs where the true change is known, so it learns to ignore that site’s light and keep that site’s cracks.
02Every site is a different material
Sandstone, limestone, mud brick, and basalt weather differently and photograph differently. One model does not fit the register.
What we do. A model family evolved per material and per site, starting from the nearest neighbor in the portfolio rather than from scratch.
03Real change is rare, so there is little to learn from
Most image pairs show no structural change. Models trained on that learn to say nothing.
What we do. Detection of every known past change event at the site is a hard floor, and the loop weights the rare true changes so the detector commits on them.
04A flag is not a plan
A list of ten thousand possible changes is not useful to a team of five geologists.
What we do. Flags are ranked by the site’s existing risk assessment and the size of the change, into a survey schedule the team can actually fly.