Introduction

As ecologists, we aim to understand the dynamics of organisms and their relationships with their surroundings. We also bear a crucial responsibility to understand human impacts on, and predict future trajectories of, global biodiversity. In particular, given that the loss of biodiversity and ecosystem services is an ongoing crisis throughout the planet, the efficiency of the approaches we take matters; we need to understand what is going on with global biodiversity and take appropriate measures as soon as possible. Thus, the problem is how we can promote our understanding of ecological systems and project the future of biodiversity in an efficient, yet effective way (Starfield 1997; Evans 2012).

Modelling should be able to provide a solution to tackling this problem (Evans et al. 2012). Modelling has now become an important tool in the study of ecological systems for developing hypotheses, explaining existing data, conducting experiments, formulating predictions and consequently guiding research (Levin et al. 1997; Starfield 1997; Green et al. 2005). Typical examples include species distribution models, which aim to describe the dynamics of species spatial distribution, and population dynamics models, which are targeted to the temporal dynamics of population sizes. The critical importance of ecological modelling is well illustrated in, for example, studies on ecological phenomena at a large spatial and temporal scale. Widespread concerns about the impact of global environmental changes, including climate change, on biodiversity and ecosystem services have led to an urgent need to predict the future state of biodiversity at a global spatial scale and a temporal scale of decades to centuries for establishing effective policies (Green et al. 2005; Kerr et al. 2007). Large-scale ecological studies, however, are logistically more difficult to conduct than are small-scale studies (Root and Schneider 1995). Consequently, at a large scale, available information is usually spatially and temporally heterogeneous, often containing unknown levels of bias and large errors (Graham et al. 2004; Guisan et al. 2006). Modelling currently provides the most comprehensive and flexible approach to understanding and projecting the dynamics of organisms at a broad spatial scale while dealing with such problems that are inherent in large-scale ecological studies (Guisan and Thuiller 2005; McMahon et al. 2011).

Ecological systems, however, comprise complex networks of individuals interacting with each other and with their environment at multiple scales and, thus, can hardly be explained accurately in every aspect of time and space from a single, albeit complex, model (Guisan and Zimmermann 2000; Porté and Bartelink 2002). For example, the dynamics of organisms are affected by a wide range of environmental factors at hierarchical scales (Wiens 1989; Levin 1992), such as patch composition and configuration at a landscape scale (Dunning et al. 1992; Wiens et al. 1993) and climate and topography at a macro scale (Hawkins et al. 2003; Davies et al. 2007; Yamaura et al. 2011). The relative importance of different processes in governing ecological dynamics also differs among scales (Root and Schneider 1995; Vellend 2010). For instance, at small spatial and temporal scales, the spatial distribution of animals is driven by patch use and within-patch movements of individuals (Bernstein et al. 1988; Mueller and Fagan 2008). At a landscape scale, resource utilisation, dispersal, colonisation and population extinction compose spatial population dynamics (Turner 1989; Dunning et al. 1992; Hanski 1999) while niche shifts and conservatism, speciation and species extinction are important processes that govern species dynamics at broad spatial and temporal scales (Willig et al. 2003; Wiens and Donoghue 2004). Our survey approaches, and consequently available data, are also restricted by the scale of focal processes. Detailed information on organisms’ dynamics, such as individuals’ behaviour or life-history events, can usually be obtained at small spatial and temporal scales through direct observation or experiments (Sutherland et al. 2004). On the other hand, in macroecological studies, which are usually conducted at large spatial and temporal scales, available information is limited to readily measurable variables, for example, body size, abundance or geographic range for well-studied groups of organisms, such as terrestrial birds, mammals and plants (Brown 1999).

This scale-dependent nature of ecological systems and survey approaches means that it is crucial to adopt a set of modelling approaches suited to ecological processes at hierarchical spatial and temporal scales. The importance of integrative approaches in ecological studies is not a novel idea but has been argued repeatedly by earlier studies (Lawton 1999; Simberloff 2004), including those focussing on ecological models (Levins 1966; Guisan and Zimmermann 2000). However, few reviews to date have focussed on integrative approaches to modelling ecological dynamics in the field of applied ecology. Given the urgent necessity to promote the understanding of biodiversity status and human impacts on it at the global, regional and national level (Balmford et al. 2005), such integrative approaches would achieve the maximum effect in tackling applied ecological problems. Although Evans (2012) is a rare exception in that it reviews existing approaches to ecological modelling for understanding the ecological impact of environmental change, the review does not explicitly focus on the difference in dynamics, processes and available information among scales, and the consequent advantages and disadvantages of different modelling approaches.

Thus, this review focusses on how modelling approaches can be used to understand the dynamics of organisms for the purpose of tackling applied ecological problems. The aim of this review is not to cover a whole range of ecological models but to introduce applications at opposite ends of the spectrum: mechanistic models targeted to local-scale dynamics of organisms, and statistical models for macro-scale dynamics. Reviewing the pros and cons of these two extremes would highlight the necessity of integrative modelling approaches to tackling ecological processes at hierarchical spatial and temporal scales.

Mechanistic models for local-scale dynamics

Mechanistic models explicitly capture hypothetical ecological processes (Guisan and Zimmermann 2000) and, thus, are likely to be robust under new environmental combinations in new locations, but are usually limited by the availability of data for model parameterisation (Porté and Bartelink 2002; Jongejans et al. 2008; Kearney and Porter 2009). Ecological studies at a local scale have a long history (Wiens 1989; Wu and Loucks 1995) and consequently, a wide variety of survey techniques have been established for obtaining detailed information on ecological processes at a local scale. Therefore, mechanistic models provide a powerful approach to tackling ecological dynamics at a local scale. Examples of mechanistic models in the field of applied ecology include matrix population models (Akçakaya et al. 1999), metapopulation models (Hanski 1999) and individual-based models (Grimm and Railsback 2005). The following section introduces actual applications of such mechanistic models, focusing on models based on behavioural decisions by individuals.

Behaviour-based models as an example of mechanistic models

Ecological dynamics at a local spatial scale are usually governed by processes that occur at a short temporal scale, and for animals, behavioural decisions by individuals play a fundamental role in shaping spatial population dynamics at a relatively small spatial scale (Dunning et al. 1995; Sutherland 1996). One advantage of understanding the behavioural decisions made by individuals is that it is possible to predict their behaviour and consequent spatial population dynamics in novel environments such as those resulting from environmental change (Sutherland 1998). There are also unexpected but critical findings of management consequences that cannot be derived without considering detailed behavioural processes in individuals (e.g., Goss-Custard et al. 2004). Thus, models based on behavioural decisions are a particularly powerful approach for a wide variety of applied problems in ecology (e.g., projecting spatial distribution: Railsback and Harvey 2002; Bar-David et al. 2005, assessing extinction risks: Schiegg et al. 2005; Rossmanith et al. 2006; Revilla and Wiegand 2008, measuring landscape connectivity: Kramer-Schadt et al. 2004; Revilla et al. 2004, predicting home ranges: Mitchell and Powell 2004, 2007, predicting the impact of deforestation: Satake and Rudel 2007).

Although models based on behavioural decisions by individuals have been adopted for a wide range of taxa and purposes, an outstandingly successful example of such models is the application to understanding anthropogenic impacts on the populations of bird species, particularly in coastal and farmland habitats (Stephens et al. 2003; Stillman and Goss-Custard 2010). These models are often spatially explicit individual-based models that project changes in the spatial distribution of target species and population consequences (Fig. 1). Although bahavioural processes can be incorporated in individual-based models either as empirically derived decision rules or as optimality rules (Feró et al. 2008), models of the latter category, where behaviour is predicted on the basis of short-term proxies for fitness or long-term fitness considerations, have usually been adopted by studies of birds in coastal and farmland habitats (hereafter, behaviour-based models, Goss-Custard and Sutherland 1997; Sutherland 2006). Model structures and parameters are based on knowledge acquired from intensive studies about the behavioural ecology of the target species, such as the functional response showing the relationship between food density and intake rate (Goss-Custard et al. 2006b; Stillman and Simmons 2006; Smart et al. 2008), diet and patch selection (Vickery et al. 1995; Gill 1996; Nolet et al. 2002; Amano et al. 2006a), resource dynamics (Nolet et al. 2001), interference competition (Caldow et al. 1999; Triplet et al. 1999), exploitative competition (Rowcliffe et al. 2001; Nolet et al. 2006b), behavioural responses to human disturbance (Gill et al. 2001; Stillman and Goss-Custard 2002; Amano et al. 2004), migratory behaviour (Bauer et al. 2006, 2008; Moriguchi et al. 2010) and theoretical frameworks that integrate the revealed processes (Sutherland and Anderson 1993; Sutherland and Dolman 1994). The developed models have been applied for making a wide range of predictions such as regarding carrying capacities (Sutherland and Allport 1994; Stillman et al. 2000; Nolet et al. 2006a), and the impact of habitat loss (Pettifor et al. 2000; Goss-Custard et al. 2006a), agricultural practices (Johst et al. 2001; Amano et al. 2007; Butler et al. 2010; Catry et al. 2012), human disturbance (West et al. 2002; Klaassen et al. 2006), fisheries (Stillman et al. 2001) and climate change (Durell et al. 2006; Bauer et al. 2008).

Fig. 1
figure 1

Simplified sketch of processes, factors and their interrelationships included in typical behaviour-based models. Note that this schematic diagram does not necessarily represent all potential factors and processes

Full size image

Limitations of and challenges for mechanistic models

Despite the success in the examples above, the relatively limited variety of species to which behaviour-based models have been applied to date clearly points to a drawback of mechanistic models. Typical behavioural-based models require detailed information on interactions both between individuals and between an individual and its environment, as well as the dynamics of prey populations, which may limit quick and widespread implementation (Bradbury et al. 2001; Feró et al. 2008). Therefore, in order to make the best use of mechanistic models like behaviour-based models, two issues—model complexity and parameterisation—need to be explored further in future studies, as highlighted by Bradbury et al. (2001).

First, the level of model complexity should be explored carefully. The strength of inferences from mechanistic models depends on the identification of key limiting processes (Elith et al. 2010; Kearney et al. 2010); incorporating unnecessarily detailed processes would make it difficult to parameterise models and interpret outputs (Van Nes and Scheffer 2005). Comparing the performance of models with different degrees of complexity, as has been done in some studies (Stephens et al. 2002; Goss-Custard et al. 2003), should be encouraged further (Orzack 2012). For example, Amano et al. (2006b) developed four different behaviour-based models for white-fronted geese Anser albifrons with and without the assumptions of (1) individuals’ complete knowledge of foraging patch quality, and (2) benefits of group foraging, and tested the ability of these models to reproduce the observed patterns in spatial distribution and fat deposition parameters, concluding that both the assumptions are necessary to predict the spatial and temporal dynamics of foraging goose populations accurately. For applied problems, the choice of the most appropriate model also depends on the management objectives (Jongejans et al. 2008). For instance, to investigate the effect of the type of agricultural land-use on the breeding success of lesser kestrels Falco naumanni, Rodríguez et al. (2006) used individual-based models that assume that individuals exploit only one patch type. Catry et al. (2012) expanded this model to include the dynamics of multiple land-use types and optimality-based patch selection by individuals, successfully evaluating the impact of spatial and temporal changes in agricultural practices in Portugal. Model complexity translates to cost in terms of computing resources and increased error propagation and, thus, the construction of unnecessarily complex models should be avoided (Clark and Gelfand 2006; McMahon et al. 2011).

Second, a novel approach to model parameterisation has recently opened the door to a quick and widespread implementation of mechanistic models. Although traditional mechanistic models have used parameters that are estimated statistically from different studies or sources (Fig. 2a), the novel approach integrates the traditionally disparate treatment of ‘mechanistic understanding’ and ‘statistical parameter estimation’ into a single process, making it possible to account for uncertainties in models and parameters (Fig. 2b, Clark 2005; Clark and Gelfand 2006). Typically, Bayesian methods and maximum likelihood methods (Patterson et al. 2008; Schick et al. 2008), but also other methods such as artificial neural networks (Dalziel et al. 2008) and signal processing (Boettiger et al. 2011), have been applied to parameterise mechanistic models incorporating various behavioural parameters from readily obtained distributional and trajectory data without direct behavioural observations (Table 1). Although many of these statistical approaches are based on the direct calculation of likelihoods, our ability to work out the likelihood functions is sometimes severely constrained by mathematical difficulties, particularly in models of complex stochastic systems, such as individual-based models with many hidden states (Beaumont 2010; Hartig et al. 2011). In such a case, a technique using ‘Stochastic Simulation Models’ is a powerful alternative approach to parameterising mechanistic models from data (see Fig. 2 in Hartig et al. 2011). Instead of calculating likelihoods directly, Stochastic Simulation Models usually take the following three steps: (1) calculate summary statistics of observed and simulated data, (2) approximate the likelihood of obtaining the observed data from the model with parameters based on the calculated summary statistics, and (3) estimate the shape of the approximate likelihood as a function of the model parameters using computationally intensive techniques, such as Approximate Bayesian Computing (Beaumont 2010; Csilléry et al. 2010) or Bayesian calibration (van Oijen et al. 2005). Pattern-Oriented Modelling (Wiegand et al. 2003; Grimm et al. 2005) does not explicitly approximate likelihoods but also applies the same concept. For example, Martínez et al. (2011) developed an individual-based model that considers basic demographic processes and interactions such as competition and facilitation in alpine tree-line ecotones, and successfully parameterised the model using the Stochastic Simulation Model approach based on Bayesian methods. Although uncertainties in model parameters, error propagation and ad hoc methods of model selection have been the major points of criticism against complex mechanistic models like individual-based models (Wiegand et al. 2003; Grimm and Railsback 2005; Grimm et al. 2005), the implementation of Stochastic Simulation Models provides a robust framework that allows such statistical inference even for complex mechanistic models (Hartig et al. 2011; Martínez et al. 2011).

Fig. 2
figure 2

a The approach in traditional mechanistic modelling, where parameters are estimated statistically from different studies or sources (shaded boxes) and then used in mechanistic models as inputs. b The novel approach with mechanistic models, which integrates the traditionally disparate treatment of ‘mechanistic understanding’ and ‘parameter estimation’ into a single process, making it possible to account for uncertainties in models and parameters simultaneously. Models are usually parameterised based on either the likelihood (maximum likelihood/Bayesian methods) or approximated likelihood (Stochastic Simulation Models: see main text)

Full size image
Table 1 Examples of studies where behavioural processes were estimated by directly fitting mechanistic models to data
Full size table

Statistical models for macro-scale dynamics

Widespread concerns over the possible impacts of global environmental changes on biodiversity have led to the increasing importance of understanding spatial and temporal dynamics of organisms at broad (typically from country to global) scales (Kerr et al. 2007; Kühn et al. 2008). This is also well reflected in recent global efforts to assess the status of global biodiversity and ecosystem services, such as the Convention on Biological Diversity’s 2010 target (UNEP 2002) and the Millennium Ecosystem Assessment (2005). Macroecology, which searches for ecological patterns in the spatial and temporal dynamics of organisms at broad spatial scales and develops theoretical explanations for these patterns, is a powerful approach to tackling applied problems at a broad spatial scale (Brown and Maurer 1989; Lawton 1999). Macroecological studies often rely on readily obtainable data, such as abundance, distribution or phenology of common species (Graham et al. 2004; Dickinson et al. 2010). Data-hungry mechanistic models are not necessarily easy to develop with such data (Urban 2005; Wiens et al. 2009). More importantly, such data are usually based on citizen science or natural history collections and suffer from sampling bias and errors (Dickinson et al. 2010; Snäll et al. 2011), which should be addressed to identify ecological patterns accurately. Statistical models thus play an important role in dealing with problems that are often inherent in macroecological data.

Statistical approaches to assessing the status of populations and species

For example, assessing the status of populations and species at a macro scale typically comprises two steps: quantifying population trends (changes in population sizes), and identifying drivers that are responsible for the revealed trends (Fig. 3). Population trends of species at a broad spatial scale are quantified using long-term, large-scale count data, which usually include several problems that have recently been recognised increasingly, such as spatial autocorrelation (Kissling and Carl 2008), incomplete samples (ter Braak et al. 1994), differences in observers’ abilities (Sauer et al. 1994) and imperfect detection of species (Royle et al. 2005). Consequently, a wide range of statistical models have been applied to quantifying population trends while dealing with those problems. Conventional approaches have used either generalised linear models or generalised additive models to estimate population trends from data with missing values (ter Braak et al. 1994; Fewster et al. 2000). More recently, hierarchical models have had a wide application for overcoming a number of problems that had not been addressed in the conventional approaches when estimating population trends (Table 2). These statistical models have contributed successfully to assessing the status of populations and species at national (Conrad et al. 2004; Battersby and Partnership 2005; Van Dyck et al. 2009; Amano et al. 2010; Baillie et al. 2010; Kasahara and Koyama 2010; Sauer and Link 2011), regional (Van Strien et al. 2001; Gregory et al. 2005) and global scales (Collen et al. 2009; Butchart et al. 2010).

Fig. 3
figure 3

A statistical approach to assessing the status of populations and species at a macro scale, which typically comprises two steps: a quantifying population trends, and b identifying the drivers responsible for the revealed trends

Full size image
Table 2 Problems that need to be dealt with when quantifying population trends using macro-scale data, and example studies that have tackled each challenge
Full size table

Once population trends are quantified, the underlying drivers can be explored effectively by comparative methods of assessing extinction risk, which aim to identify factors associated with species showing serious population declines (Fisher and Owens 2004). Factors related to the risk of decline and extinction include species life-history and ecological traits, such as body mass or geographical range (Bennett and Owens 1997; Reynolds 2003), evolutionary history (Purvis et al. 2000; Purvis 2008), external anthropogenic threats (Cardillo et al. 2004, 2005) and large-scale climate variability (Brander 2007; Waite et al. 2007). Statistical models again play an important role in untangling the effects of these multiple drivers of population decline. For example, since related species cannot be assumed to be independent data points, phylogenetic generalised least-squares models (Grafen 1989; Martins and Hansen 1997) have been adopted in comparative methods to identify correlates of population decline while dealing with phylogenetic non-independence among species (Shultz et al. 2005; Amano and Yamaura 2007; Purvis 2008).

Recently, comparative methods of extinction risk have shown further development in two aspects. First, an increasing number of studies has found that there are non-linear effects of species characteristics (Cardillo et al. 2005) and interaction effects of intrinsic and external factors (Murray et al. 2010) on species extinction risk. Such non-linearities and interactions of potential drivers can be explored effectively by machine learning methods (Olden et al. 2008). Tree-based models such as Decision Trees (De’ath and Fabricius 2000) and Random Forests (Liaw and Wiener 2002) have been applied in particular to a wide range of ecological topics (Olden et al. 2008). In fact, recent studies have applied tree-based approaches successfully to model population declines (e.g. Jones et al. 2006b; Sullivan et al. 2006; Davidson et al. 2009; Murray et al. 2010). However, although for practical purposes it is an advantage that tree-based models do not require phylogenetic information, which is often difficult to obtain (Davidson et al. 2009), the inability to account for phylogenetic non-independence also means that tree-based models cannot reduce the influence of clade-specific relationships in data (Bielby et al. 2010). Thus, tree-based models may be used effectively as a first step in identifying non-linearities and interaction terms of drivers for inclusion in phylogenetic generalised least-squares models in future efforts to assess species susceptibility to extinction (Bielby et al. 2010).

Second, disentangling the effects of species traits, phylogeny and spatial context on population decline and species extinction would be a challenge to be addressed in comparative methods (Purvis 2008). Species traits, phylogeny and space are related closely to each other through phylogenetic trait conservatism (Freckleton et al. 2002; Maherali and Klironomos 2007) and phylogenetic niche conservatism (Peterson et al. 1999; Wiens 2004), making it difficult to evaluate the independent effect of each factor. Recent advances in statistical models for quantifying the relative effect of traits and phylogeny (Desdevises et al. 2003; Diniz-Filho and Bini 2008) and space and phylogeny (Freckleton and Jetz 2009) could be a breakthrough to solve the problem, although only a few studies to date have adopted these modelling techniques to quantify the relative contribution of these factors to species extinction risk (e.g. Safi and Pettorelli 2010).

The examples above focus on the status of populations and species, but macroecological approaches described in this section also play an important role in assessing the loss of functional diversity (Şekercioğlu et al. 2004; Flynn et al. 2009) and consequent decline in ecosystem services (Butchart et al. 2010; Garibaldi et al. 2010; Keesing et al. 2010).

Limitations of and challenges for statistical models in macroecological studies

Obviously, one big disadvantage of correlation-based statistical models is the lack of mechanistic structures, making it difficult to extrapolate to novel conditions (Sutherland 2006). This is particularly critical in applied ecological problems, which often require predictions under novel conditions (e.g. impact of climate change: Thomas et al. 2004; spread of invasive species: Bradley et al. 2010). For instance, climate envelope models have typically been used in efforts to assess biodiversity consequences of climate change (Heikkinen et al. 2006). However, recent studies have pointed out that the reliance on such correlation-based statistical models can lead to inaccurate projections of changes in species spatial distribution (Beale et al. 2008; Duncan et al. 2009). Consequently, there is an increasing awareness that, for the accurate understanding of the impact of climate change, it is necessary to incorporate biotic interactions (Davis et al. 1998; Araújo and Luoto 2007), physiology (Kearney and Porter 2009), dispersal limitations (Svenning and Skov 2007) and plasticity and evolution of niches (Pearman et al. 2008; Wiens et al. 2010) in predictive models.

Incorporating such mechanistic structures in ecological models at a macro scale is not at all an easy task, considering the general lack of information at this spatial scale. Nevertheless, recent studies have successfully enhanced predictions about the spatial population dynamics of organisms by incorporating information derived from mechanistic models in statistical models (Amano et al. 2008; Kearney and Porter 2009; Morin and Thuiller 2009; Elith et al. 2010). The use of meta models, such as graph models, which extract key processes from but are much simpler than detailed mechanistic models, will also be a promising approach to applying a mechanistic understanding over large spatial scales (Urban 2005). Novel statistical approaches to model parameterisation, such as Bayesian methods and Stochastic Simulation Models as described earlier in this paper, also offer an alternative and promising modelling method that allows the development of mechanistic models with macroecological data (Hartig et al. 2011; McMahon et al. 2011). Bled et al. (2011) represents one excellent example of such a novel modelling approach for assessing processes underlying the spatial population dynamics of organisms at a macro scale. Their hierarchical Bayesian model explicitly incorporates the invasion process of Eurasian collared doves Streptopelia decaocto through the estimation of site-persistence probability, initial colonisation and recolonisation, and their relationship with local population density, while accounting for the detection probability of the species. The model was fitted, using a Bayesian method, to distribution data derived from the North American Breeding Bird Survey, and the density-dependent nature of the invasion process was successfully inferred from the estimated parameters (Bled et al. 2011). Kadoya and Washitani (2010) also adopted a Bayesian method to estimate parameters of the immigration and establishment processes of alien bumblebees Bombus terrestris in Japan from spatio-temporal presence/absence data.

Future challenge for applied ecological modelling

This review has briefly introduced the application of two types of ecological models to applied problems. As indicated above, no single type of model is sufficient to understand and predict ecological dynamics at hierarchical scales, given the difference in processes operating and information available among scales. Thus, the complementary use of mechanistic models and statistical models according to the scale, available information and skills would be an effective approach to tackling the dynamics of organisms in a changing world (Bradbury et al. 2001; McMahon et al. 2011, Fig. 4). Considering that all models, including the mechanistic and statistical models introduced here, inevitably contain a black box, which hides the underlying details, no type of model should be dismissed automatically as inferior just because it does not include all of the mechanistic details (Orzack 2012). What truly matters is where to use such black boxes in model structures, and the consequent predictive performance of the models (Orzack 2012), both of which should be taken into account more explicitly in future modelling studies. At the same time, efforts to overcome the drawbacks of each type of model represent a major challenge for the next decade. Novel approaches to model parameterisation, such as Bayesian methods and Stochastic Simulation Models, improve the chances of a quick and widespread implementation of mechanistic models even for ecological dynamics at large spatial and temporal scales. In fact, these approaches break down the boundaries between statistical models and traditional mechanistic models, providing a comprehensive framework for the field of ecological modelling (Fig. 4).

Fig. 4
figure 4

A diagram of the amount of available information, statistical models, traditional mechanistic models and novel mechanistic models along the gradient of temporal/spatial scales and skills required for modelling. The amount of available information is generally large at small spatial and temporal scales. Statistical models, including the comparative methods, can be applied to a wide range of scales but are particularly important for ecological dynamics at large spatial and temporal scales, where it is usually difficult to implement traditional mechanistic models due to the lack of necessary information. Novel statistical approaches to model parameterisation allow a quick implementation of mechanistic models even with information available at large spatial and temporal scales, consequently providing a comprehensive framework for the field of ecological modelling

Full size image

One of the largest challenges faced by ecological modelling and, more broadly, conservation science, are the reported gaps and heterogeneity in the geographical and taxonomical coverage of existing information (Collen et al. 2009; Pennisi 2010). For example, 18% (7,976/45,301 species) of animal species that have been evaluated to date by the IUCN have so little information available for assessing population status that they can be judged only as Data Deficient (IUCN 2011). Thus, the challenge is how to effectively understand and predict the dynamics of organisms with ecological modelling under such uncertain and temporally and financially constrained conditions. Although many studies (e.g. Ludwig 1999; Brook et al. 2000; Coulson et al. 2001) have explored the validity of ecological models for predicting future status of organisms, surprisingly few studies to date have tested if ecological models that have been developed for a particular species in a particular region can be used to effectively predict the dynamics of other species or in other regions (Amano et al. 2011). In this regard, efforts to test the predictability of species’ extinction risk using trait-based comparative methods (e.g. Cardillo et al. 2008; Pocock 2011) and the spatial transferability of predictive models (Broennimann et al. 2007; Whittingham et al. 2007), though still rare, should be encouraged further.

Finally, special emphasis should be placed on the importance of improving biodiversity information across the globe. As discussed above, the possibilities and applicabilities of ecological modelling are restricted largely by the amount of information available for a target system. Systematic biodiversity monitoring (Pereira and Cooper 2006; Lindenmayer and Likens 2009), citizen science (Dickinson et al. 2010) and natural history collections (Graham et al. 2004) all have the potential to provide invaluable information for developing effective ecological models. Improving the accessibility of existing but isolated empirical data by overcoming technological and cultural challenges (Jones et al. 2006a; Reichman et al. 2011) is also an effective strategy.

In conclusion, the complementary use and improvement of multiple model types, increased use of novel model parameterisation, examination of the applicability of models to species and regions with little information, and the achievement of wider biodiversity information availability are important challenges for the effective use of modelling in applied ecological problems. Enhancing collaborative partnerships among empirical ecologists, theoretical ecologists and ecological modellers would promote the exchange of information and ideas for these purposes (Green et al. 2005; Lindenmayer and Likens 2011). However, considering recent declines in fundamental fields such as taxonomy and natural history (Noss 1996; Hopkins and Freckleton 2002; Uniyal 2011), encouraging inter-field collaborations may not be enough. To borrow Weiner (1995)’s phrase, individual ecologists should aim to be both modellers and empiricists. Modellers should learn as much as possible about the natural history of the systems they are trying to model, and empiricists should learn as much as possible about models that may be relevant to their research (Weiner 1995). Such pluralism, not only within ecological modelling but across the field of ecology, might be the best strategy for unraveling the dynamics of organisms in a rapidly changing world.