Training and intercalibration reduce observer-induced variability in forest vegetation surveys