Ecological Modelling 312 (2015) 335–346 Contents lists available at ScienceDirect Ecological Modelling journa l h om epa ge: www.elsevier.com/locate/ecolmodel Living on the edge in species distribution models: The unexpected presence of three species of butterflies in a protected area in southern Spain Pilar Fernández ∗ , Diego Jordano, Juan Fernández Haeger Department of Botany, Ecology and Plant Physiology, Universidad de Córdoba, Campus de Rabanales, Edificio C-4, 14071 Córdoba, Spain a r t i c l e i n f o Article history: Received 8 January 2015 Received in revised form 26 May 2015 Accepted 28 May 2015 Keywords: MaxEnt Transferability Species distribution models Butterflies Marginal populations Plebejus argus a b s t r a c t MaxEnt (Maximum Entropy) modelling method is probably the most popular technique to model species distributions based only on the presence records across broad spatial scales. Although it is widely used, there is much controversy about the transferability of models between different geographical areas. Transferability might be more questionable when it comes to predict the distribution of peripheral popu- lations at the margin of the species geographical range, where they may be affected by and adapted to environmental conditions different from those of core populations. To explore transferability of MaxEnt models among sectors of the geographic range, we selected three butterfly species with wide distribu- tions and peripheral populations at their southernmost margin in the Iberian Peninsula, namely Plebejus argus, Cyaniris semiargus and Pyronia tithonus. Using data from the Atlas of the butterflies of the Iberian Peninsula and Balearic Islands as well as both climate and land use data, we modelled their potential distribution ranges in Spain. In addition, we also independently modelled their distributions separately in three concentric sectors of their range. We then investigated the transferability of the models between sectors and the effect of varying the regularization parameter. Our results show that when developing species distribution models the quality of occurrence data should be carefully checked, paying special attention to both their number and spatial distribution and avoiding possible significant biases. The transferability of the models tends to decrease when data from increasingly distant sectors are used as test data. More precisely, and independently of the regularization parameter value, models built using occurrence data either from the core or the intermediate sectors failed to adequately predict the distribution of the three butterfly species in the peripheral sector, especially in Do˜ nana National Park. © 2015 Elsevier B.V. All rights reserved. 1. Introduction Species distribution models (SDMs) are specific tools developed for predicting the habitat and potential distribution of plant and animal species given a set of occurrence data, albeit incomplete, and environmental data. SDMs have acquired increasing impor- tance in biodiversity conservation (Fielding and Bell, 1997; Araújo and Luoto, 2007; Mateo et al., 2011), and to exploring the potential effects of global climate change on biodiversity loss and on shifts in species distributions (Pearson and Dawson, 2003; Elith et al., 2010; Romo et al., 2013.) ∗ Corresponding author. Tel.: +34 957218596. E-mail address: bv2ferop@uco.es (P. Fernández). A variety of modelling techniques ranging from classic logistic regression models coupled with GIS, generalized additive mod- els (GAM), GARP (a genetic algorithm approach) and others, are available (Virkkala et al., 2005; Wisz et al., 2008; Titeux et al., 2009). Some authors focus on modelling the environmental con- ditions that meet a species’ ecological requirements and predict the relative suitability of habitat, aiming to produce the so called environmental niche models (ENMs) (Warren and Seifert, 2011). In practice, when modelling across large geographic areas there is usually a lack of data concerning important niche dimensions linked to biotic factors, while detailed data of climatic variables, altitude, slope, or aspect are more easily available. In any case, the model results are often flawed by problems like small sample sizes, biased data or unrepresentative sam- ples (Dennis and Thomas, 2000; Stockwell and Peterson, 2002; Romo et al., 2006; Pearson et al., 2007; Papes ¸ and Gaubert, 2007). http://dx.doi.org/10.1016/j.ecolmodel.2015.05.032 0304-3800/© 2015 Elsevier B.V. All rights reserved.