Analyzing floristic inventories with multiple maps

a r t i c l e i n f o Spatial observations of plant occurrences contain a wealth of information on relations among species and on the relation between species and environmental conditions. Typically, inventory data of this kind are large co-occurrence matrices, and hence, direct ecological interpretations based on expert knowledge are often very difficult. Hitherto, ordination approaches have been used to construct a virtual ordination space (represented as one or multiple scatter plots) in which species that often co-occur are situated close together, whereas species that hardly co-occur are found far apart. In this study, we investigate a recently proposed ordination approach, multiple maps t-SNE, that constructs multiple, independent ordination spaces in order to reveal and visualize complementary structure in the data. We compare multiple maps t-SNE to several conventional ordination approaches, exploring a large inventory of vascular plant occurrences (FLORKART). Our results reveal that multiple maps t-SNE is well suited for the analysis of floristic inventories. In particular, multiple maps t-SNE uncovers the major dependencies of species co-occurrences on climate and soil biogeo-chemical preconditions. The construction of large floristic and vegetation databases over the course of the last century opened novel possibilities for regional, national, and continental biogeographical assessments (Bekker et al. Recent efforts bring together various sources of observation and allow even for global analyses (e.g., Scholes et al., 2008, see also the open access " global biodiversity information facility " http://data.gbif.org). Data collections of this kind allow scientists to address plant ecological questions across ecosystems (Kühn et al., 2004; Schmidtlein, 2004). Extracting the underlying environmental patterns from such data collections requires computationally powerful explorative multivariate tools (Mahecha and Schmidtlein, 2008). Indeed, the application of such tools has a long history in vegetation sciences, and a series of influential textbooks were published more than a decade ago (e. Since then, the use of ordination techniques such as principal component analysis (PCA), detrended correspondence analysis (DCA), classical multi-dimensional scaling (CMDS), or non-metric multi-dimensional scaling (NMDS) has become a standard. Nonetheless, novel developments in the field of dimensionality reduction and machine learning in general are largely ignored or only timidly transferred from informatics to other fields (Mjolsness and DeCoste, 2001). As a consequence, scientific progress regarding data exploration tools in ecology and environmental sciences is partly on halt (Mahecha et al., 2007). For instance, an important open issue is how large-scale patterns of plant co-occurrences are related …