Can a model trained on satellite data really find brambles on the ground?