What are specific drawbacks of topographic heterogeneity that concern you? And what variables are you considering to be predictors of your models?

I agree with Marcelo. You should include topographic layers as predictors to account to topographic heterogeneity. This solves the problem most of the times.

Thank you for the answers.

Bioclim variables such as temperature, precipitation may not be reliable, because of topographic heterogeneity. Climatic variables may change by different altitudes. Narrow micro-climate areas may be ignored on low resolution layers. I think it is possible to eliminate this using high resolution layers, however it may not be feasible for a broad area such as Anatolia.

I am sorry, I don't know your region. But, I believe that if you use the layer altitude from Wordclim, this would account for topographic heterogeneity considering that your region covers a large area.

I agree that you should use topographic variables if they are available. You should transform aspect to make it meaningful. A cosine transformation works well for some areas (-1=south, 0 = east and west,1 = north).

In spatial models, you should not delete any drawbacks of topographic heterogeneity. You should instead add it to the modelisation procedure, all these describing the spatial discontinuity variables as predictors.

I agree with Federico. Topography may have predictive value as well so you should add it to your model. Why not create a triangulated irregular network (TIN). From a Digital Elevation Model (DEM), a contour image must be made which connects connecting locations with the same elevation (isoclines) which is a interpolation process. The interval between successive contours is user-defined.

See Topic 11 at http://www.innovativegis.com/basis/MapAnalysis/Default.htm

and for example:

Rashid H (2010). 3D Surface-area computation of the state of Jammu & Kashmir using Shuttle Radar Topographic Mission (SRTM) data in Geographical Information System (GIS). Journal of Geomatics 4 (2): 77-82

You could also check for a spatial structure (apart from topography) in the distribution of species by using the PCNM method (principal coordinates of neighbour matrices; search for Borcard and Legendre 2002; Borcard et al. 2004), now called called MEM (Moran’s eigenvector maps; Dray et al. 2006). You can model structures at scales ranging from the broadest, encompassing the whole of Anatolia, down to the finest, whose sizes are of the order of magnitude of the sampling interval. I used R with the manual ' Numerical Ecology with R. Borcard.

ISBN 978-1-4419-7975-9

DOI 10.1007/978-1-4419-7976-6

Bioclim Variables are generated with use of topography and climate and they reflect differences in climate based on elevation and major topographic features. However, the resolution of bioclim layers is only about 1x1 km. This means that your predictive maps are going to be only 1 km precise (Which is very gross, especially for some plants and plant communities that linked to specific substrates.) You may wish to consider developing a country or region-wide model at that large scale (1x1 km), and then develop additional local scale models (for discreet watersheds, or protected areas, or may be even larger units) which count more on information derived from more detailed topographic maps and satellite images, combined. Ronny brough up interesting references, certainly worth to check. One question would be whether you would like to use presence only or presence absence analytic approaches. What species do you study?

Two additional variables that can be of interest in areas with a lot of topographic relief are insolation (solar radiation) and topographic position. ArcGIS 10.1 has a decent built-in solar radiation toolset (see http://resources.arcgis.com/en/help/main/10.1/index.html#/How_solar_radiation_is_calculated/009z000000tm000000/ ) and even if you don't have access to ArcGIS, the links have some decent references to help you if you want to explore other tools. Topographic position can describe whether a particular location is on a ridgetop or summit, in a valley bottom, or mid-slope, and can be an excellent additional topographic variable not captured by other means; it is analyzed as an index of concavity. See the poster by A.D. Weiss, 2001, Topographic Position and Landform Analysis, for a often-cited method using focal statistics.

Thank you for a very interesting discussion (especially for helpful readings, thanks Ronny). I will take into consideration all suggested variables and methods.

Tigran, I have used Maxent with presence data. I had failed to modeling some narrow distributed endemic plants such as Pyrus serikensis and Fritillaria kittaniae, so I have decided to choose wider distributed species for modelling. Still I would be very happy, if I can find an analytic approach that is useful for narrow distributed species. 

Many good suggestions already, especially the DEM derivatives: elevation, slope, aspect, solar radiation and Topographic Position Index); attached articles covering most of these issues, including the question on spatial modelling of plant species with narrow ranges.

A variable that showed to be predictive is our case of narrowly distributed plant species in the mountainous Mediterranean is plant substrate (soil; lithology); from your topographic heterogeneity combined with my experience in Turkey, I hypothesize that your area of interest may contain lithological heterogeneity as the driver of plant species distribution. 

Article Fine resolution distribution modelling of endemics in Majell...

Article Sharing natural resources: Mountain gorillas and people in t...

An index of terrain roughness can also be a good predictor. See for example

http://gis4geomorphology.com/roughness-topographic-position/

The 3rd index (slope variability) worked well to predict badger distribution. I haven't tried to use it for plants, but maybe you can test it...

Did you try with Maxent in R adding topographic layers???

Many good suggestions. There are plenty of examples in the literature about how best to apply elevation in your modeling. The bottom line is, topographic heterogeneity within your study area is a GOOD thing. Species are most likely responding to such gradients (or some underlying factor for which elevation is a surrogate) and they could certainly prove useful in identifying distributional patterns. More concerning is a lack of gradient. All statistical models are searching for divisions within inputs in order to rank or categorize relationships. Without obvious differences, patterns are more difficult to discern. Given that, I would have to disagree with Daniel's disagreement regarding elevation as an appropriate environmental variable. One would be hard pressed to find variables that are not conflated in some way or another - especially climatic derivatives. The only solution would be to run models with only one variable so as not to incorporate overlapping factors. Elevation is often times one of the most accurate and spatially resolute variables available and is easily gathered, processed, and incorporated in SDM. In the end, your interpretations of model results and application (as well as an understanding of all incumbent caveats ) are what matters the most. If models suggest certain environmental variables are driving results but those variables don't seem intuitively and biologically important, they most likely should be viewed with suspicion. All variables are actually surrogates for underlying biological processes.

Whitebox is a very good DEM processing tool that is free from John Lindsay at the University of Guelph. You can generate any number of DEM-derived topographic derivatives for SDM modeling.

Daniel (above) has a point: it is crucial to avoid variables that are collinear with other variables as predictors. However, it is worth building on one hand models that use more "proximate" variables (radiation, temperature, substrate...) and on the other simpler "ultimate" variables, or "compound" variables (altitude, topographic position). The former are captured in the latter, but there is always some aspects of altitude that we don't understand. And because they don't vary at the same scale, it is possible that a model with only altitude and topography will perform better than a model with 20 proximate variables. But, again, it is critical to avoid any collinearity between your predictors, or you model won't mean anything (something that everyone agrees on, but many ignore). Good luck.

I live in British Columbia, Canada, an area with high topographic diversity.  Not knowing what base layers you can rely on, my answer may be of limited use.  My approach would be to map polygons based on soil, terrain, and existing vegetation cover as well as aspect/elevation which may be pushed into a site severity index together so that you have a model that captures these important variables for plant cover.  Do a google for  BC BIogeoclimatic Mapping and you will find a good body of literature for such a system.

Where I find many models fall apart is through us not necessarily knowing the plants that we are modelling well enough. An example is Townsendia parryi, a small and beautiful odball endemic to my greater region.  Our literature review and field investigations suggested it is nearly exclusively a plant of exposed high elevation ridges.  Last year I found this plantt growing in the valley bottom on a south facing, dry ravelling slope thus completely ruining our assumptions.

Wishing I could go botanise in Turkey.

We have some fine-scale modeling 30-50m scale) of species distributions using Canonical Correspondence Analysis.  Inputs were elevation, topographic position, insolation, slope, and geologic substrate.  We did not have fine scale temperature and precipitation, so used elevation as a surrogate (works within a single mountain range).  Two papers are:

https://www.researchgate.net/publication/237262592_Plant_Species_Distributions_under_Present_Conditions_and_Forecasted_for_Warmer_Climates_in_an_Arid_Mountain_Range?ev=prf_pub

https://www.researchgate.net/publication/227017019_GLM_versus_CCA_spatial_modeling_of_plant_species_distribution?ev=prf_pub

Article Plant Species Distributions under Present Conditions and For...

Article GLM versus CCA Spatial Modeling of Plant Species Distribution

I just decided to add few pages with references from my list that may help you.

http://onlinelibrary.wiley.com/doi/10.1111/j.1600-0587.2013.07872.x/pdf

http://www2.unil.ch/biomapper/

You may also check an unpublished manuscript on the local scale habitat suitability modeling for Persian Toad-Headed Agama on my research gate page. I am actually currently advancing that study with other approaches, not covered in that manuscript, but it will give you an idea about some aspects of local scale modeling. I am also interested in modeling ranges of some plants that may be threatened in Anatolia, including Calligonum polygonoides, Rhynopethalum gibbosum, etc. I would be interested to learn more about your hypotheses and approaches.

If you do have access to ArcGIS, Jeff Jenness has incorporated the methods in A.D. Weiss into a toolset that can be downloaded from http://www.jennessent.com/arcgis/land_facets.htm.

http://www.jennessent.com/arcgis/land_facets.htm

Thank you very much for all answers. 

Turkey has an excellent species diversity. There are a number of revision study that may help anyone who is interested in the flora of Turkey (Although site is in Turkish I hope it is helpful because of paper links):

http://www.flora.org.tr/revizyonlistesi.html