Species distribution modeling includes a variety of methodological approaches and software tools. In
our research we used various models to test the influence of some environmental variables on the
geographical distribution of centipedes (Chilopoda). We collected the centipede presence data from the
CHILOBIO database that consists of information about centipede findings in the area of Slovenia. The data
we used were gathered with a common soil sampling method with the use of sampling cylinders. We used
data of 18 most numerous centipede species from 29 sampling locations in the area of Kočevje region to
perform a canonical correspondence analysis (CCA). Furtherly we preformed two generalized linear models
(GLM) and a model preformed in the framework of Bayesian statistic. We used presence data for centipede
species Sigibius anici from 73 sampling locations throughout Slovenia. The presence data was binary and
quantitative, where we took into account the sampling effort expressed as number of sampling units on a
particular location. We tested the environmental variables for possible correlations before we used them in the
models and we made a selection of the most influential variables with an automated selection protocols. We
used variables as: terrain elevation, soil texture, content of organic matter in the topmost soil horizont and
terrain insolation in summer solstice to execute CCA and only the first three variables listed to perform the
other three models. The CCA ordination diagrams showed outlying of some species optimums for the chosen
environmental gradients. However, the optimums for the most of the species were located at the center of the
diagram, which can be due to various reasons. The results of Sigibius anici presence data modeling showed
poor model fit. Among the three of the variables used in the GLM only the variable with, content of organic
matter in the topmost soil horizont showed statistical significance. The model prediction estimates (AUC)
were around 0,5, which means that the predictions were close to random. Unsuccessful model predictions
could be due to some sources of data noise. In our case the most potential noises were: unprecisely calculated
population densities, violation of predisposition about sampling randomness and inadequate resolution of
environmental variables that failed to precisely express the conditions on the sampling locations.
|