602
Kočí, M. et al.
researchers often take a conservative approach and rather
stick to the units of the traditional expert-based classifi-
cation. An increasingly common practice seems to be
combining numerical classification with subsequent in-
terpretation of clusters selected subjectively across dif-
ferent hierarchical levels, merging selected clusters so
that they better correspond to traditional units, or even
manually re-assigning selected relevés among clusters
(Hennekens et al. 1995; Bergmeier 2002; Krestov &
Nakamura 2002; Willner 2002). This practice of post
hoc manual re-arrangement of the numerical classifi-
cation results indicates that formalization of the tradi-
tional expert-based classification by cluster analysis and
related unsupervised methods has mostly failed so far.
An alternative to the commonly used numerical clas-
sification algorithms is the Cocktail method proposed
by Bruelheide (1995, 2000). This method produces for-
malized definitions of vegetation units by providing
unequivocal criteria for assignment of relevés to these
units. At the same time it appears to be able to delimit
vegetation units in a similar way to the traditional ex-
pert-based classification (Bruelheide & Chytrý 2000),
but with the elimination of the latter’s inherent incon-
sistencies. An important difference between the Cock-
tail method and the numerical classification algorithms
is that the Cocktail method does not assign all the
relevés in the data set to vegetation units. It preferably
defines vegetation units in those parts of the vegetation
continuum, where several species with rather narrow
ecological or geographical ranges meet, while those
parts of the vegetation continuum which contain only
widespread generalist species are often not assigned to
any vegetation unit by the Cocktail method.
However, non-assignment of some relevés to the
vegetation units may become a problem in some appli-
cations of vegetation classification, notably in vegeta-
tion mapping. Therefore it seems to be advantageous to
apply the Cocktail method in combination with numeri-
cal procedures that assign relevés to vegetation units by
calculating similarity between the relevés and constancy
columns of vegetation tables (Hill 1989; Dodd et al.
1994). If these procedures are run in large phyto-
sociological data sets, diagnostic species of vegetation
units can be formally defined (Chytrý et al. 2002) and
performance of the similarity calculations can be en-
hanced by positive weighting of diagnostic species.
The aim of the present paper is to test the ability of
the Cocktail method, combined with a newly designed
procedure of similarity-based assignment of relevés to
vegetation units, to reproduce an expert-based vegeta-
tion classification in a formal way. As a test data set, we
use the subalpine tall-forb and deciduous scrub vegeta-
tion of the Czech Republic, previously classified at the
level of associations by expert knowledge.
Material and Methods
Material
We took the classification of the subalpine tall-forb
and deciduous scrub vegetation of the class Mulgedio-
Aconitetea in the Czech Republic (Kočí 2001) as an
example of the traditional expert-based classification.
The data set used for creating this classification con-
sisted of 718 relevés of subalpine tall-forb vegetation,
with species cover estimated on the Braun-Blanquet or
Domin scale (Westhoff & van der Maarel 1978). This
classification was largely based on expert knowledge,
being a compromise between the field experience of the
author and different local classifications published in
earlier literature. The classification was aided by nu-
merical divisive algorithm of the TWINSPAN program
(Hill 1979), which was used in several successive runs.
Several TWINSPAN end-groups were either merged or
further divided according to the subjective opinion of
the author. Assignment of each of the relevés to the end-
groups was checked manually and some relevés were
eventually moved to groups other than those suggested
by TWINSPAN. In the end, each of the 718 relevés was
assigned to one of 16 recognized associations.
The Cocktail classification, which was used to for-
mally reproduce the expert-based classification by Kočí
(2001), was performed with a data set of 21 794 relevés,
containing all vegetation types of the Czech Republic.
This data set was taken from the Czech National Phyto-
sociological Database (Chytrý & Rafajová 2003), us-
ing a geographically stratified selection that made it
possible to avoid a great influence of the over-sampled
areas on the results. In this selection, we took only one
relevé of each syntaxon per grid square of 1.25 longi-
tudinal
× 0.75 latitudinal minute (ca. 1.5 km × 1.4 km).
The assignment to syntaxa at the level of association
(or alliance) was according to the original assignments
by the relevé authors. If two or more relevés of the
same association were encountered in the same grid
square, selection priority was given to the relevés with
recorded cryptogams and to newer relevés. If this
selection still yielded more than one relevé in the grid
square, one of them was selected at random. Due to the
stratified selection, several relevés contained in the
above-mentioned data set of 718 relevés were not
included in this data set.