Formalized reproduction of an expert-based phytosociological classification

- Formalized reproduction of an expert-based phytosociological classification -

603

Cocktail classification

The Cocktail method (Bruelheide 1995, 2000) was

designed to mimic the classification procedure of the

Braun-Blanquet approach. Cocktail classification is ba-

sically created by expert knowledge and not by an

unsupervised algorithm of a computer program. An

expert makes subjective choices during the classifica-

tion process while the program suggests possible solu-

tions and ensures that particular steps in the classifica-

tion process are applied consistently throughout the data

set. Delimitations of the resulting vegetation units are

explicitly formally described, which means that also

new relevés that were not present in the original data set

can be unequivocally classified as belonging or not

belonging to the particular vegetation unit.

The Cocktail procedure starts with defining groups

of species that tend to occur together in relevés of a large

database. Using a large database that covers a broad

spectrum of different habitats and a large geographical

area is important for obtaining species groups of more

general validity. Species of the same group usually have

similar habitat requirements and phytogeographical af-

finities. Cocktail species groups correspond to the con-

cept of sociological species groups (Doing 1969) and

often they are closely related to the groups of diagnostic

species for particular vegetation units as recognized in

phytosociological literature. Extraction of each group

starts with one or a few species preselected by the

researcher. Other species with the most similar distribu-

tion across the relevés of the database are added stepwise

to this starting species or species group. In our case, co-

occurrence tendency of species was measured by the phi

coefficient of association (Sokal & Rohlf 1995; Chytrý

et al. 2002).

Unlike Bruelheide (1995, 2000) who used a fully

automated process of species group optimization, we

allowed for more manual control with the aim to arrive

at ecologically more coherent species groups. After

selecting a starting group of two or three species, we

calculated the phi coefficient of association between

each species in the data set and the group of relevés that

contained the starting species group. Of the species not

belonging to the species group, we usually chose the one

with the highest 

Φ value and included it in the group as

its next member. In a few cases we included the species

with the second or third highest 

Φ  value, particularly if

the species with the highest 

Φ  value  was already in-

cluded in another species group or had several times

more or less occurrences in the data set than the species

already included in the species group. This solution was

chosen because groups of species with large differences

in occurrence frequency would not be ecologically co-

herent: their species might have roughly identical eco-

logical optima but much more frequent species would

usually have broader ecological ranges. After including

the new species in the species group, we re-defined the

group of relevés and recalculated the phi coefficient for

all species in the data set and the new group of relevés.

If the species group disintegrated after this step, i.e.

some of the species not included in the species group

had a higher 

Φ  value than some of the species included,

the group was rejected. By contrast, if the species be-

longing to the group had the highest 

Φ  values, the group

was either accepted or further optimized by including

additional species. The optimization process was stopped

if any of the candidate species for addition in the next

step either caused group disintegration or substantially

changed the ecological coherence of the group.

To consider a species group as being contained in a

relevé, not all the species of the group need to be

present. Bruelheide (1995, 2000) defined the minimally

required number of species of the group as the intersec-

tion of expected and observed cumulative distribution

functions for relevés having 0 to k species, k being the

number of all species included in the group. However,

our pilot studies showed that this criterion tends to yield

a low minimum number of species if the group consists

of species that are rare in the data set and a high minimum

number of species if the group mainly includes common

species. Our data set of 21 794 relevés included many

different vegetation types, which made subalpine tall-

forb species relatively rare; then the calculated mini-

mum number of species was two for most groups. By

contrast, in our data set of 718 relevés, which contained

only subalpine tall-forb vegetation, tall forbs were rela-

tively common and the calculated minimum number of

species increased for several groups. This indicates that

the minimum criterion derived from the cumulative

distribution functions is strongly dependent on the data

set structure, which complicates the transfer of species

groups between different data sets. Therefore we em-

ployed a simpler criterion, taking half of the species of

the group as the minimum number, e.g. at least two out

of four or three out of five.

After defining several species groups, the Cocktail

method creates definitions of vegetation units by combi-

nations of species groups using logical operators such as

AND, OR or AND NOT (Bruelheide 1997). For exam-

ple, a relevé is assigned to vegetation type X if it con-

tains species group A  and at least one of the species

groups B or C, while at the same time the species group

D is absent. As our aim was reproduction of an expert-

based classification of Kočí (2001), we combined spe-

cies groups in such a way as to arrive close to an

understanding of associations in that classification.

However, pure combinations of species groups were

not sufficient to reproduce most of the expert-based

---