Species data
The list of species of microchiroptera present in Australia and Papua New Guinea was compiled from several sources including:
- Van Dyck & Strahan (2008)
- Flannery (1995a, b)
- Bonaccorso (1998)
- Churchill (2008)
- Parnaby (2009) for clarification of large-bodied Nyctophilus species
- Reardon et al. (2008) for their taxonomic work resolving a part of the Australian Mormopterus enigma
Species occurrence records were downloaded from two web portals, Atlas of Living Australia (ALA; www.ala.org.au) and the Global Biodiversity Facility (GBIF; www.gbif.org)
Records for large-bodied Nyctophilus were revised using the extensive list of re-identified museum specimens published by Parnaby (2009). Likewise, Reardon et alet al. (2007) was used to identify museum records assigned to the new species Mormopterus eleryi.
Data were cleaned to remove any records with missing latitude and longitude, or whose coordinates were clearly in error (e.g. plotting in the sea using a GIS program).
The complete species list is available here and includes links to the model output for each species. Alternatively, see the links to family accounts at the bottom of this page.
Climate data
For current or baseline climate I used the WorldClim data at 5 arc minute resolution (Hijmans et al. 2005; www.worldclim.org). These data are the 19 basic bioclim variables based on climate average data for the period 1960-1990.
I produced future climate using IPCC Fourth Assessment Report (AR4) General Circulation Models (GCMs). Data for a single run from each of 14 GCMs was downloaded from the Climate Model Inter-comparison Project (CMIP) website (www-pcmdi.llnl.gov/projects/cmip/index.php). The anomoly or “Delta” method outlined in Hijmans et al. (2005) was used to adjust observed baseline climate represented by the WorldClim data to provide estimates of mean monthly maximum and minimum temperatures and mean monthly precipitation averaged over the period 2046 to 2055 giving a decadal climate for the decade centred on 2050. The monthly data was then used to calculate the standard 19 bioclimatic variables defined by Nix (1986) and Busby (1991).
An important limitation of the method I have developed is that it relies on the availability of data for minimum and maximum temperature, and precipitation from the GCM runs. Only 14 of the 23 AR4 GCMs provided these data, and all but four only provided data for the decade centred on 2050. An alternative approach to downscaling first generates statistical estimates of GCM data in years not supplied from the CMIP repository and for greenhouse gas scenarios not required to be provided by the modellers. This method, known as “pattern scaling", seems to be able to produce useful data for species distribution modelling but this wasn’t clear to me at the time I began developing my own future climate data in 2008. I was frankly rather sceptical of using a statistical model of a complex stochastic process model like a GCM - a rough model of a rough model! - and my experiments with using the then available version of the best software tool for pattern scaling were not productive of sound data. However, after Damian Fordham’s excellent examination of an updated version of the most commonly used software for pattern scaling (Fordham et al. 2011, 2012), it may be worth pursuing.
The core set of future climate data I use were developed about four years ago, and recently I expanded coverage from 4 GCMs with some skill in the Australian region (Suppiah et al. 2007; but see Perkins et al. (2007a, b) for slightly different outcomes) to 14 GCMs. I continue to use my own future climate data. Free sources of downscaled future climate data are available at the CCAFS website and CliMond website.
A list of the 14 GCMs I have used is available here.
Species models
Cleaned species occurrence data were used to build species distribution models using a number of modelling tools. Each modelling method may introduce bias in fitted models due to various aspects of the modelling algorithm and how well the data match the assumptions of a particular modelling method. It is widely understood that an ensemble of model results provides the best indication of the most likely relationship between species occurrence and environmental conditions (Araújo and New (2007)).
The models I present on this website only use climate data, and they are more accurately described as “climate envelope models”, or “climatic niche models”, although current terminology in the ecological modelling community often uses the generic term “species distribution models” or SDMs. The term “climatic niche model” is informative because the models I present here are attempts to model the locations without occurrence records that fit within the known climate component of the species’ realised niche.
It is vitally important when you look at the maps produced by my models not to over-interpret them. They are most definitely NOT SPECIES DISTRIBUTION MODELS and the maps are NOT MAPS OF PROBABILITY OF OCCURRENCE. There are some very strict requirements for output maps to be regarded as estimating probability of occurrence, and they are most definitely not met by the museum and herbarium data used for my models. My models estimate the distribution of climate suitability.
At present, I only provide models fitted using MaxEnt (Phillips et al. (2006), Phillips and Dudik (2008)) but output from other machine learning methods such as Boosted Regression Trees, Random Forest, Bayesian Additive Regression Tree, Artificial Neural Network and Support Vector Machines will be included in the future.
MaxEnt models
MaxEnt models were built using version 3.3.3k of the software (www.cs.princeton.edu/~schapire/maxent/) with the following key settings:
- 10-fold cross-validation;
- Threshold and Hinge features turned off to avoid spurious discontinuities in the output maps; and,
- "writebackgroundpredictions" was set to "true" so that all necessary information was available for computing regional AUC values as described below.
Using and R-script, I generated SWD-format files of current climate data at the filtered occurrence points for each species. Producing SWD-format background point files for the species included in this study present a challenge. Unconstrained selection of background points may cause problems with model fit and interpretation and several schemes have been suggested to constrain the geographical area over which background points are selected (e.g. VanDerWal et al. (2009), Phillips et al. (2009), Barbet-Massin et al. (2012)). Conversely, fitting MaxEnt models with too few background points are of extremely poor quality (pers. obs., and see Phillips and Dudík (2008)). However, I was unable to apply any of these methods to the present data because a very large portion of the species in the study have distributions that mean that too few background points would be selected using one of these methods. This situation arises in the set of study species via two quite different circumstances. First, a number of the species endemic to the Australian mainland have distribtuions covering most fo the contientn so that there is little unoccupied space from which to select background points. Second, many species occur on islands and island chains comprised of small islands. Restricting background point selection to a defiend radius (e.g. after the method of VanDerWal et al. (2009)) results in almost no terrestrial grid cells being included in the background data.
A solution was found in the use of Köppen-Geiger climate zones to restrict background point selection (Webber et al. 2011). The SWD-generating R-script created a background SWD file of 10,000 points by assessing which Köppen-Geiger climate zones the occurrences fell, and randomly selecting background points only from the global extent of those climate zones. I had previously developed an R-scripts in 2010 to create several Köppen climate classification maps based on two new schemes (Peel et al. 2007; Stern et al. 2000) for use in climate envelop modelling of weed species (again my employer at the time would permit further development and publication of the approach I was developing!!!). The best performing one was a global map of Köppen-Geiger climate zones created using the updated rules by Peel et al. (2007) which I based on the WorldClim 5 arc minute climate data, and this was used as the source of Köppen-Geiger climate zone information. (If anyone is interested, other Köppen-Geiger climate zone data sets are available at the CliMond website and here).
Fitted models were projected onto climate data trimmed to the Australia-PNG region (defined as 110 to 160 degrees longitude by 0 to -45 degrees latitude) and all subsequent analyses were based on these regional maps of climate suitability.
Selection of variables for MaxEnt models is a matter of active consideration, and my own (as yet) unpublished research indicates that large differences in the spatial arrangement of climate suitability values is linked to the variables used to fit a model. Paradoxically, there is nothing to choose between sets of variables using AUC values - they are practically the same no matter which variable set is used to fit the model. Given that it is very difficult to pre-select variables based on a knowledge of each species’ traits and physiological tolerances, I opted to run with a naive set of 16 of the orginal 19 bioclim variables. I excluded Mean Temperature of the Wettest Quarter (bio8), Mean Temperature of the Driest Quarter (bio9), and Precipitation in the Warmest Quarter (bio18) because they have problematic jumps in values within the region that lead to patchy models. I adopted the view that these were artefacts of the definition of these variables and smooth gradients are required for bio-physical predictor variables.
Model performance or quality was assessed in a number of ways. First, the Area under the Receiver Operating Curve (AUC) statistic for the fitted model was extracted from the MaxEnt output files and tabulated. These values reflect the predictive skill of the fitted model across the gobal scope of the climate data used for predictor variables. It is, therefore, an overall index of model quality.
Second, the quality of the models to predict occurrences within the Australia-PNG region was assessed by computing AUC values for occurrences falling within that region (subsequently referred to as regional AUC values). The computation used the MaxEnt output for model predictions at occurrence and background points for each replicate model trimmed to exclude points falling outside the Australia-PNG region. An R-script was developed to perform these tasks and compute the regional AUC values using the ROCR package in R and functions provided by Hand (2009).
Finally, AUC has been criticised for weighting commission and ommission errors equally Lobo et al. (2008), and for not using the same measurement scale for each model so that its use to select the best model in an ensemble of models is misleading (Hand (2009)). Hand's approach has been criticised by Flach et al. (2011) as being restricted only to a special set of circumstances. However, it appears that Flach et al's critique is based on an unrealistic set of assumptions or premises, and that Hand's reasoning is sound (Parker, 2011). Therefore, I used R code by Hand and available from this URL (last accessed 9 September 2012) to compute H measures for the fitted models. (Note: Subsequent to my adaptation of Hand's original R-code, Hand and colleagues have produced an R-package, Hmeasure, that I will apply when I next update data and refit models.)
A grand average climate suitability map was made across the replicates for each of the 14 GCMs (i.e. up to 140 files per species) and this was done using an R-script. All subsequent analysis was based on the grand average files.
Data analysis
Unless otherwise stated, all the analyses described below were undertaken using scripts for version 2.15 of the R statistical system (R Development Core Team, 2012).
Change in overall climate suitability
An overall index of climate suitability for each species, termed "intensity" (see Wilson, 2011), was computed for the current or baseline map, and for the grand average 2050 map. The relative or percentage difference between these values measures the degree of change in overall climate suitability potentially faced by a species. A limitation of this measure (like all indices that reduce spatial patterns to a single number) is that it hides whether the differences are due to loss or gain of high or low climate suitability. I therefore computed a dissection of the intensity into three "bins" representing high, medium and low suitability. The high bin ranged from 1 to 0.65, medium from 0.65 to 0.35, and low from 0.35 to 0.
Spatial shift in climate suitability
Wilson (2011) also developed a method to produce ordinations and statistical tests for collections of maps produced by programs like MaxEnt. The technique computes a matrix of pair-wise Hellinger distances for the collection of maps. The distance-based comparison of pairs of maps incorporates all of the difference in spatial distribution so that a distance of zero means two maps are identical and an increasing distance value indicates increasing differences in the distribution of climate suitability. Using an R-script I computed a Hellinger distance matrix, and looked for clusterings or patterns in a principal coordinates analysis produced using the R package labdsv.
Several years ago I also began using plots of changes in the centre of mass of MaxEnt output maps to characterise the broad shifts in the distribtuion of climate suitability. My employers at the time didn't like it and so it was never published. I discovered that this approach has been developed several times (see Woillez et al. 2007). See also Yates et al. (2010) who applied it to plant species in south-west Australia. I adapted an old R-script to compute the shift in center of mass between baseline and 2050 MaxEnt maps.
Climate change impacts and current conservation status
The level of threat to a species may be assessed using many criteria and numerous ways of combining information about the species into an index of threat have been devised. Few, if any, of these assessment methods incorporate predictions of impacts due to climate change. The most widely used system for mammals is the IUCN Red List method (IUCN, 2012). The version of the IUCN method in operation during 1998-1999 was applied to the Australian bat fauna as part of the development of the Australian Bat Action Plan (Duncan et al., 1999). Allowing for some taxonomic changes, I collated the table of IUCN threat categories for species and compared the rankings of threat against the rankings of species based on change in intensity suing Spearman's rank correlation (Zar, 2010).
Results
Species models
Models could be fitted for 81 species of the 101 species in the study.
Change in overall climate suitability
The majority of modelled species (61 or 75.31%) were predicted to face a decrease in overall climate suitability, and 20 species (24.69%) were predicted to see an increase in overall climate suitability.
Spatial shift in climate suitability
Climate change impacts and current conservation status
No significant relationship existed between IUCN threat status (with Data Deficient adn unassessed species excluded) and predicted overall change in climate suitability (rS = -0.143, n = 75, p = 0.222).