Prediction models of vegetation communities perform better when including transition zones

VERMEERSCH S.

Plant Science and Nature Management, Department of Biology, Vrije Universiteit Brussel, Pleinlaan 2, 1050 Brussel
E-mail: sophie.vermeersch@instnat.be

It has often been demonstrated that communities can be useful indicators for environmental heterogeneity. However, in the case of complex spatial variability such as for soil factors, it is not obvious to filter out the information of indicator communities. The correlation of various soil factors to the constituting species of the communities is not obvious either, since most of the species present in the vegetations are neither characteristic nor differentiating species of communities. Several species have wide amplitudes in relation to environmental variables. Moreover spatial complexity often leads to the presence of two or more communities on a site. This species, environmental and spatial complexity, has particular consequences on the prediction efficiency of communities.

Prediction models based on generalised linear modelling use the relation of environmental variables to determine the presence or absence of species or of communities, where the information from the transition areas is often not taken into account. To optimally utilise the information enclosed in these areas, a weight was assigned to each of the relevés: “1” on sites where a single community is present, “<1” on sites where different communities are present. These weights are determined through CCA-analysis and are proportional to the relation of the communities to environmental factors at that site. The prevailing environmental variables regulating the presence of single communities and the species composition at those sites are used as reference systems to determine which share of the transition area can be attributed to each of the communities.

This approach has led to prediction improvement, since the validation of two types of models namely unweighted and weighted, on an independent subset revealed a better fit for the weighted models.