Aim: The unsupervised nature of traditional numerical methods used to classify vegetation hinders the development of comprehensive vegetation classification systems. Each new unsupervised classification yields partitions that are partly inconsistent with previous classifications and change group membership for some sites. In contrast, supervised methods account for previously established vegetation units, but cannot define new ones. Therefore, we introduce the concept of semi-supervised classification to community ecology and vegetation science. Semi-supervised classification formally reproduces the existing units in a supervised mode and simultaneously identifies new units among unassigned sites in an unsupervised mode. We discuss the concept of semi-supervised clustering, introduce semi-supervised variants of two clustering algorithms that produce groups with crisp boundaries, k-means and partitioning around medoids (PAM), provide a free software tool to perform these classifications and demonstrate the advantages using example data sets of vegetation plots.
Methods: Semi-supervisedmethods use a priori information about groupmembership for some sites to define centroids (k-means) or medoids (PAM) of site
groups that represent previously established vegetation units. They identify these groups in a species hyperspace and assign new sites to them. At the same time, they search for a user-defined number of new groups. We compared the unsupervised, supervised and semi-supervised methods using an example of a forest vegetation data set that was previously classified using expert knowledge, and assessed how well these methods reproduced vegetation units defined by experts. Then we compared supervised and semi-supervised methods in a task when a grassland vegetation classification established in one country was extended to two neighbouring countries.
Results and conclusions: Example analyses of vegetation plot data sets demonstrated that semi-supervised variants of k-means and PAM are extremely
valuable tools for extending existing vegetation classifications while preserving previously defined vegetation units. They can be used both for identifying so far unrecognized vegetation types in the regions where a vegetation classification already exists and for extending a vegetation classification from a particular region to neighbouring regions with partly identical but partly different vegetation types. Both k-means and PAM provide site groups with crisp boundaries, whichmakes thema simpler alternative to fuzzy clustering methods.