I think, that future could be in integration of principles of http://linkeddata.org/ with GIS data. This is probably way, how to go to more complex model. This is only in first stage, but now it becomes more and more popular in GIS research community.

I agree that spatial data quality is a key problem. However, increasingly we are not having to make a trade off between the extent and the grain of landcover and landuse data. Certainly in cities there seems to be an interest in generating high resolution datasets for Lidar, near infrared and photographic data (allowing tree cover to be mapped). In addition, there is much data that captures social systems and features that may have a strong influence on ecological performance. For example, census data can be available at a very fine grain, the locations and nature of businesses are often available from local authorities and pollutants such as noise, particulates and artificial lighting can act as proxies for certain types of anthropogenic disturbance http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0033300

Another issue that may be worth more consideration is the modeling of available/accessible habitat. In many cases, species presence may not simply relate to the habitats within say 500m of a location, but to a subset of habitats that are adjacent to a network of features that may provide safety or additional resources.

http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0033300

Whilst some species are able to range over the landscape selecting desirable habitats wherever they are, it is clear that for others, much of the landscape may be functionally unavailable. It is relatively simple to mask out parts of the landscape within a GIS, as long as data is available from fieldwork on barriers and typical movement thresholds.

Spatial data quality is a key problem, but also the selection of best spatial scale to modelling SDM (species distribution models) is key problem. The modelling process uses a number of parameters to predict others, such as the occurrence of determinate species, population size, habitat suitability or biodiversity. For example is well recognized as the heterogeneity of landscapes can lead to changes in species’ abundance and biodiversity. However, landscape metrics depend on maps and spatial scales when it comes to undertaking a GIS analysis. For this reason also the the multi-spatial scale approach could be very useful for assessing the accuracy of surrogates of response variables and in order to optimize spatial model procedures for predictive mapping, mainly over extensive areas.

http://onlinelibrary.wiley.com/doi/10.1111/j.2041-210x.2012.00253.x/abstract

http://mammal-research.org/pdf/Loe2011.pdf

Look also the principles of down-scaling analysis...

I think another component that needs to be considered in habitat models is the animal's/organism's dispersal capability. Too often we use prediction maps from habitat distribution models (HDM) and expect to find organisms there based on species-habitat relationships. However, suitable habitat means little if the organism cannot disperse to it in the first place (due to barriers, physical or otherwise, in between). Combining HDMs with dispersal-potential models I think is an important aim, especially when working on intermediate spatial scales.

I think Federico Morelli observation is very important, define the resolution and scale for the input variables and subsequently the models are crucial. Nowadays many HDM models use Global data sets in particular for climate data. Providing the coarse resolution of these data, they cannot be used to produce good models at local or regional scales. In some regions of the world there is a lack of more accurate environmental data, and then the use of these databases may be acceptable nevertheless in such cases the number of variable need to be keep to a minimum avoiding redundancy so not to carry on with the intrinsic errors produced by the lack of accuracy of such variables. In mountain regions and complex landscape this is a delicate issue and decisions in terms of data quality are crucial as this change the whole outputs of the models. I advise also the use of platforms like BIOMOD that will help to analyze the performance of different models according to the variables used and the object to be modeled. We found very different results in terms of the different model performances for animals vs. plants. It is clear that adding movement trajectories to the animal models provide much better results but obtaining the data is also more expensive.

To look for something really new in quantification of landscape structure across multiple scales to derive more accurate GIS-based habitat models see "Landscape patterns from mathematical morphology on maps with contagion" from K.H. Riitters et al. Landscape Ecol, Vol. 24: 699-709.

First an conceptual consideration, then an empirical one.

Maximum entropy algorithms are applied from astronomy to image interpretation in surveillance cameras and medical scans. As such the output it is not about biological species, geography, ecology or suitability in particular. These terms represent relabelling of the output. It is about 2-D (spatial) patterns of a selected object related/associated with fields of (proxy) variables. Maxent provides a probability that the selected object can be predicted at unsampled sites/spots by the selected fields of variables given the selected area extent. Not included is the temporal dimension. If the object on inspection is repeatedly not found at sites with a calculated high probability of occurrence, or often found at sites with a low probability of occurrence, the model needs to be revisited in all its aspects, particularly the relevance and the spatial accuracy of the predictor layers. In my experience, the absence of historical legacies and land tenure as predictor variables are common culprits in unrealistic predictions.

Now the empirical case. Unhappy with the prediction of the cheetah in Namibia by the classical biophysical environmental predictors (climate, substrate, human population density, land use/cover), we entered prey and potentially competing predators. What do you expect to be the outcome in terms of significant predictors of cheetah distribution (Unpublished, but the data set is available for demonstration or testing).