Using summed individual species models and state-of-the-art modelling techniques to identify threatened plant species hotspots

Reliable identification of hotspot areas with high numbers of threatened plant species has a central role in conservation planning. We investigated the potentiality of identifying the distribution, richness and hot-spots of threatened plant species at a 25 ha resolution using eight state-of-the-art modelling techniques (GLM, GAM, MARS, ANN, CTA, GBM, MDA and RF) in a taiga landscape in north-eastern Finland. First, the individual species models developed based on occurrence records of 28 species in the 1677 grid squares and derived from different statistical techniques were extrapolated to the whole study area of 41 750 km(2). Second, the projected presence/absence maps were then combined to create species richness maps, and the top 5% of grid cells ranked by species richness were classified as hotspots. Finally, we created an overall summary map by combining the individual hotspot maps from all eight modelling techniques and identified areas where the individual hotspots maps overlapped most. There were distinguishing differences in projections of the geographic patterns of species richness and hotspots between the modelling techniques. Most of the modelling techniques predicted several hotspot locations sporadically around the study area. However, the overall summary map showed the highest predictive performance based on Kappa statistics, indicating that the locations where the hotspot maps from the eight models coincided most harboured highest observed species richness. Moreover, the summary map filtered out the patchy structures of individual hotspot maps. The results show that the choice of modelling technique may affect the accuracy and prediction of hotspot patterns. Such differences may hamper the development of useful biodiversity model applications for conservation planning, and thus it is beneficial if the conservation decision-making can be based on sets of alternative maps and overlaying of predictions from multiple models. (C) 2009 Elsevier Ltd. All rights reserved.