Dimension reduction and data sharpening of high-dimensional vegetation data: An application to Swiss mire monitoring

In an era of the availability of very large databases, the problem of efficient methods to analyze such datasets remains. In large scale forest and landscape monitoring projects for instance, appropriate data mining techniques that can summarize the overall status of landscapes are necessary for planning and implementing follow-up management strategies. We consider a vegetation data set consisting of species data from more than 120 mires spread across Switzerland with a total of 20,134 plots on 2658 vascular and non-vascular plant species. Using species indicator values as proxy for site conditions, we propose some simple strategies for data mining involving multidimensional scaling and nonparametric probability density function estimation, both of which are known in the classical statistical literature. We show how commonly known techniques can be adapted in a novel and effective approach for the specific purpose of analyzing plant species occurrence in the plots to identify site conditions that influence species configurations and species diversity, thus providing important information concerning aspects of large scale vegetation structure. Our results indicate high variations among the mires with respect to site conditions that affect species assemblage and species diversity. While species indicator values continue to be popular as well as subject of much debate and research in the ecological community, our experience shows that careful methods of analysis at large landscape scales can reveal some powerful results, which can be taken up as the starting point for the next level of investigation.