Models trained on someone else's ground

Segmentation models carry the assumptions of the imagery they were trained on — field shapes, road patterns, roof materials, vegetation, the texture of disturbed ground.

Move that model to a different continent and accuracy drops in ways that are hard to see until someone drives to a flagged site and finds nothing there.

The fix is unglamorous: engineer your own datasets on the terrain you actually operate over. It is slow, it is the least fashionable part of the work, and it is the part that makes the output trustworthy enough for an enforcement decision.