engleski

Modelling the number of invasive plants in Croatia according to the habitats and bioclimatic factors: importance of the quality of the data

Plant species, defined as invasive, at the territory of Croatia (Boršić et al., 2008 ; Mitić et al., 2008 ; Nikolić, 2013)are in research focus due their potential and real threat to the overall biodiversity. Flora Croatica Database (FCD http://hirc.botanic.hr/fcd) is a comprehensive database including information about Croatian flora from diverse studies and research projects with various focuses, research intensity, spatial extent and temporal coverage. Database contains information extracted from published literature, unpublished field observations, and form herbarium collections. Across the country, there is substation difference in research focus. For example, attractive areas on the coast are more researched in contrast to the eastern part of the country, and also, some protected areas are far more researched than average. The purpose of this analysis was making predictions on number of invasive plant species across the whole territory of the Republic of Croatia using available information on invasive plant species occurrences. This analysis is focused on main question: How the result of predicting number of invasive plants differs in relation to quality of data entered into model? Is it possible to detect threshold in research intensity after which data entered make model less precise? Information of invasive plants was extracted from FCD. Analysis was made in R using diverse packages for data manipulation and aggregation and SAGA GIS ver. 2.0.8. functionalities run via RSAGA package. Point information on presence of some invasive plant species was transformed into variable „invasion“ – number of unique plant species detected at regarded spatial resolution. As surrogates for research effort we used information on total number of records in CFD per spatial units. Variable named gower is the result of the gower clustering algorithm at 10 km resolution prepared classifying whole territory of Croatia into previously defined number of classes according to the percentage of the habitats and bioclimatic variables. We prepared five datasets to test model performance. First dataset was those quadrants in each gower class that are farely good researched (effort per class between 0.80-0.95 percentile). Similarely, other data sets included more quadrants since we broadered quality ranges 0.65- 0.95 for second ; third from 0.50-0.95 ; fourth 0.35-0.95 quantile and fifth dataset ranging from 0.15-0.95 quantile effort per gower class. For model validation we took only the best researched grid cells in whole territory of Croatia (>95quantile total effort). Predictions were obtained fitting generalized linear model defining Poisson distribution as error distribution and log link via glm function(stats package). The structure of spatial autocorrelation in the data was modelled with fitted variogram model of covariance structure in the data (gstat package). Best model (according to root mean squared error (RMSE)) was the one where model fitted using quadrants with sampling effort above the mean in each gower class. Worst model was one with small number of quadrants of highest research effort (0.80-0.95). The greatest relative difference in predictions of best and worst model was detected in areas with lowest research effort.

modelling; invasive plant; biological databases

nije evidentirano