1. Understanding large-scale distribution shifts is vital to predict species responses to changes in the environment. Such shifts occur as a consequence of habitat distribution, dispersal processes and the influence of environment factors.
2. Historical distribution records provide a long-term perspective on change. Here, we use presence/absence data on 1781 British plant species to examine distribution shifts in the 20th century. Our aim was to determine the importance of long- and short-distance colonisation in the spread of plant populations.
3. We consider three contrasting models of distribution change: random colonisation, where colonisation occurs independently of the established distribution, localised phalanx, where favourable local conditions and/or short-distance dispersal results in colonisation of neighbouring sites only, and phalanx-spread, which is the diffusion-like spread of a population through localised colonisations.
4. We fitted a set of four generalised linear models to contrast various forms of the three mechanisms of distribution change. Model selection was used to assess the relative fit of each.
5. Overall rates of extinction and colonisation were low, but highly variable. They were strongly linked with total occupancy, which reduced the probability of extinction and increased that of colonisation.
6. Comparison of the models indicated that the majority of distributions change through a phalanx-spread process. This indicates that local habitat distribution, as well as localised dispersal, are key driving processes. Long-distance colonisation is, in particular, very rare.
7. An additional set of three models was fitted using rainfall and temperature data. These models were of improved fit compared to the models of distribution change. A clear climate signal could be identified for c. 45% of species, while for the remainder there was a signal that could not be unequivocally attributed to rainfall or temperature.
8.Synthesis. Our analysis shows that the spread of most species is spatially restricted at a national scale. This is likely to be a consequence of habitat suitability and dispersal. Habitat changes, such as climate, occur (on a decadal scale) through relatively slow spread of changes. Dispersal is typically restricted, so that new habitat is usually colonised only if it is adjacent to already colonised patches. Our findings highlight the importance of prior knowledge of dispersal mechanisms as well as local habitat structure and availability in species distribution modelling research and species conservation initiatives.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
The majority of studies on distribution change use purely static data to predict future distributions, and only very few include a dynamic perspective (Guisan & Zimmerman 2000). This is due to the relative ease with which climate or habitat data can be correlated with a static representation of species occurrence. However, use of multiple time points, with information on the likely habitat and environmental changes, is more powerful for identifying drivers of change and should result in greater predictive ability (Willis et al. 2009).
Historical presence/absence distribution records are useful in developing general predictions of large-scale dynamics in plants as they provide a long-term perspective. Using distribution data at multiple discrete time points allows an assessment of how these environmental factors contribute to population change and should indicate drivers of change. Kery, Gardner & Monnerat (2010) showed that for checklist records across multiple sites, a particularly important issue is recorder intensity. Areas that are sampled more intensively naturally tend to yield more records. They show that recorder effort may not be randomly distributed with respect to environmental gradients and variation in recorder effort may confound attempts to model the effect of the environment on distribution change. As they point out, this is less of an issue for atlas data collected on large temporal or spatial scales, where detection is more probable. However, the issue of recorder intensity can nevertheless be a problem even in such data sets, for example, when comparing records from two periods (e.g. Telfer, Preston & Rothery 2002).
Here, we aim to characterise long-term, large-scale distribution changes for a large number of species and consider them in the light of proposed frameworks for large-scale population dynamics in plants. We use the species distribution data to highlight the importance of a dynamic perspective when predicting distribution change through time or as a response to climate or habitat change. Our results show the variability in distribution dynamics at the species level, but also identify underlying patterns in changes occurring for a large group of species. We examine the role of local habitat availability (spatial structure) and the relative importance of short- and long-distance colonisation.
Materials and methods
In the British Isles, the distribution of plant species has been well documented for hundreds of years, often with national coverage. These distribution records, collected from various sources including the Botanical Society for the British Isles’ (BSBI) vice-county recorder scheme, as well as dedicated surveys, have been published as atlases on two occasions: first in Perring & Walters (1962), with an updated version (Preston, Pearman & Dines 2002) following when the original was found to have become outdated (Rich & Woodruff 1996). These will be referred to throughout the text as the first atlas and the new atlas, respectively. These atlases have allowed an assessment of the changes in the British flora over the second half of the 20th century, and there is a considerable evidence of significant changes to many distributions over this time, such as establishment and naturalisation of alien species in the flora (Rich & Woodruff 1996; Smart et al. 2005; Braithwaite, Ellis & Preston 2006). A range of studies have modelled changes in the British flora examining ecological correlates of range structure in rare and scarce plants (Pocock et al. 2006), climate impacts on a single-species distribution (Foody 2008) and the recent spread of alien species (Hulme 2009).
The distribution data used here were obtained from the vascular plant data base, which is maintained online (Botanical Society of the British Isles). Records are held for 6669 higher plant taxa in the British Isles. These were collected between 1629 and 2006 and are standardised and presented in the form of presence/absence distribution maps at the 10 km × 10 km scale. The full species list was refined by discarding low-level taxonomic distinctions including subspecies, hybrids and variegates. The data for each species were sorted to include only the records from two time periods: 1930–1960 and 1987–1999. These correspond with the full data collection periods for each atlas. At this point, the data set was further refined by discarding records for species with a negligible presence in the British Isles, defined as those whose range covered <1% (26 10 km2 sites) of land area throughout the study. The remaining records included all data listed under the standard taxonomic name for each species. This process resulted in a data set of 1781 species from the British flora.
The biases in atlas data and the pitfalls in using such data as an indication of distribution change have been widely discussed (Telfer, Preston & Rothery 2002; Braithwaite, Ellis & Preston 2006; Rich 2006). In this case, the two study periods are of differing lengths and also represent different levels of recorder effort. The distribution data for each species were therefore modified to contain records corresponding to a grid of 2646 10 km × 10 km squares, which are a comparable set of coordinates to those used in the first atlas, but which do not include squares recorded only in the new atlas (for example some coastal squares). Additionally, the distribution records from Northern Ireland and the Isle of Man could not be included in this study as the new atlas used a different grid map for these areas allowing no direct comparison with the first atlas.
Unknown Recorder Effort
Unfortunately, there is no independent estimate of recorder effort by site or between years within a census period. It is likely the new atlas represents an overall increase in recorder effort. When accounting for difference in recorder effort between atlas census periods, Telfer, Preston & Rothery (2002) calculated a distribution change index. This uses a weighted linear regression of the logit-transformed proportional range sizes to measure the performance of each species against the mean of all species. In a similar vein, we include in our analysis the change in the number of species recorded in each grid cell as a proxy for recorder effort. The rationale is that an increase in the number of recorded species is indicative, in part, of more observers at that location. The number of species recorded in each cell increased in most sites, owing to a combination of increased recorder effort in the second census and the introduction and spread of alien species in the flora. Areas of widespread loss are likely to indicate a change in high-order habitat suitability rather than recorder effort; however, these are rare. A distribution map of change in species number and further information on modelled effect size of recorder effort is included in the online supplementary information (see Figs S1 and S2 in Supporting Information). In terms of the analyses we report, the main objective is to contrast spatial models of spread, which we believe should be relatively insensitive to changes in recorder effort as there are a range of patterns observed across species. For other approaches to dealing with unknown recorder effort in occurrence data see Hill (2011).
Another problem often encountered when modelling spatial data is autocorrelation (Lichstein et al. 2002). Typically, species distributions will show positive spatial autocorrelation, with sites near to each other being more similar than expected by chance. Here, we use density of populated cells surrounding a focal cell in an autoregressive method. To check fitted models, we used Moran’s I to test for autocorrelation in model residuals (see Table S1).
To examine the relationship between patterns of distribution shift and environmental change, we use climate data on rainfall and temperature as predictors of species colonisation and extinction dynamics. The climate data were obtained from the Climate Research Unit (CRU) at the University of East Anglia (Mitchell et al. 2004). This data set consists of 1200 monthly grids for observed climate covering the European land surface at a 10’ resolution. Data on mean precipitation and temperature were assigned to each of the 10 km2 sites in the British Isles. A single mean value for each variable was generated for time t + 1. Although coarse, the value used is appropriate for the scale and resolution of the plant species data and clearly summarises variation in climate within the study. We recognise that rainfall and temperature data do not represent all environmental factors relevant to distribution change. However, in terms of environment, these are particularly significant variables, for example, sites that are warmer are often dryer, two phenomena with opposing effects on plant growth. Therefore, they are likely to reveal the underlying climate signal in the data when accounting for the role of site dynamics and spatial arrangement of sites.
Models of Distribution Change
Here, we consider three simple mechanisms of distribution change, which we translate into statistical models for distribution expansion. According to the first model, expansion can occur randomly via long-range colonisation. In this model, the random appearance of new suitable habitat, combined with sufficient long-range dispersal, means that any site can become occupied. This results in an apparently scattered distribution (Fig. 1a). We term this the random colonisation model, and it is characterised by chance colonisation of favourable sites beyond the immediately neighbouring sites. In this model, the extinction or colonisation of sites does not depend on the spatial pattern of occupancy (Fig. 1a).
In the second model, colonisations occur in an aggregated manner with species colonising the empty sites immediately surrounding currently populated ones in a localised phalanx model (Fig. 1b). This can occur because either suitable habitat is created only in areas immediately adjacent to existing habitat or as a consequence of highly restricted dispersal. In this model, the colonisation of new sites, or the extinction of occupied sites, is highly spatially dependent (Fig. 1b). Colonisation increases as the frequency of populated neighbouring sites increases, while extinction decreases.
In the final model, spread is a spatially continuous process, which we term the phalanx-spread model of distribution change (Fig. 1c). In this model, newly suitable habitat is created as a large-scale front, and/or rapid spatially restricted colonisation results in a wave-like advance of a species into new habitat. Or, in reverse, degradation of habitat might lead to the extinction and loss of a species from large contiguous areas.
Distribution changes were analysed using Generalized Linear Models and autologistic regression (R core development team version 2.9; Dormann et al. 2007; Betts et al. 2009). We fitted models to reflect the different models of distribution change described earlier.
For each species, presence at each 10 km2 site in the study area at the time of first census (t) was used as a predictor for presence at the time of second census (t + 1), assuming a binomial error distribution and a logit link (logistic regression). The change in number of species found at each site in the study area between censuses (Δ) was included in each model to account for recorder effort (see Figs S1 and S2).
The first model was designed to reflect the random colonisation model:
For site i, the response variable, Pi,t+1, gives the probability (between 0 and 1) of that site being populated at time t + 1. β0 is the average probability that a site will be occupied at time t + 1. Pi,t is the presence or absence at time t and allows for differential colonisation or extinction depending on whether the site is currently occupied or not and βP models this dependency. This model predicts that distribution change is independent of the spatial arrangement of neighbouring populated sites.
To fit the localised phalanx model, Model I was modified by adding an additional term measuring the effect of populated neighbouring sites on occupancy in time t + 1:
βN is the strength of the effect of Ni,t on occupancy, with Ni,t varying between 0 and 8 populated neighbour sites surrounding focal site i. We hypothesise that an increase in Ni,t will increase the probability of a site being colonised and will buffer established populations, through dispersal and local population expansion. Ni,t only contributes to predicting expansion on the periphery of the established distribution.
Where a distribution has undergone a wide-ranging expansion, colonisation of sites clearly occurs beyond those peripheral sites. Model III was designed to distinguish whether these colonisations occur by the phalanx-spread model:
In Model III, spatial dependency is modelled through the occupancy at time t + 1, rather than occupancy at time t; this autocovariate term uses the response of the nearest neighbour sites as a predictor of response at site i. The rationale is that if a species is spreading (or declining) as a wave across the landscape, the probability that site i will become populated (or extinct) will be greater in areas where large amounts of colonisation (or extinction) have also happened. βN1 is the strength of the effect of Ni,t+1 on occupancy. Where Ni,t+1 is favoured as the explanatory variable, as opposed to Ni,t, distribution change is better explained by the phalanx-spread model rather than by the localised model.
Ni,t and Ni,t+1 were also simultaneously modelled as covariates, combining the localised phalanx and phalanx-spread models:
Model IV allows for both the initial and subsequent configurations of populated patches to affect the change in occupancy.
To quantify the extent to which the pattern of occupancy influenced distribution changes under the three models, we used the fitted values from the models to look at the effect of varying number of populated neighbour sites on the likelihood of focal site occupancy. This neighbour effect value (Ne) is the difference between the probabilities of a colonisation for maximum neighbours (Nmax) and for minimum neighbours (Nmin). Values were calculated for each of the three models containing a neighbour term outlined earlier (Models II–IV), where P(N) is the predicted probability of occupancy with N populated neighbour sites:
These values were used to examine the importance of neighbour presence on distribution dynamics. To assess the degree to which these contrast with the values that would be expected in the absence of spatial effects, we ran null models for 500 replicates. Each replicate used a randomly sampled distribution from our data set where extinction and colonisation events were randomised to occur independently of neighbour presence. The null distributions were analysed as above using Models II–IV, and their neighbour effect values were calculated. A t-test was carried out on the binomial data, which was arcsine-transformed to test for differences between the null and observed distributions of neighbour effect values.
To examine the role of climate in distribution dynamics, a number of models were fitted including climate terms as covariates. Model IV was used as the base model for these as it accounts for the spatial configuration of patches in each time period as well as spatial autocorrelation through autologistic regression (Lichstein et al. 2002); it should therefore allow a clear climate signal to be identified. The climate terms were added to this model:
In Model V, R is the mean monthly rainfall at site i at time t + 1 and βR is the strength of the effect of rainfall on distribution dynamics, while in Model VI, T is the mean monthly temperature at site i at time t + 1 and βT is the strength of the effect of temperature. Model VII includes both of these climate terms as covariates.
The climate data from time t + 1 was used, rather than that from time t or change in climate between time periods, as it better reflects the biological response of plant species to climate change. For example, a change of several degrees in temperature may not result in distribution change if it is within a species tolerance limits; however, a small temperature change in another instance may reach a biotic threshold forcing a response to climate through distribution change.
The two climate variables will inevitably be negatively correlated because sites that are warm will also tend to be dry, whereas cool sites are likely to be wet. Overall, the correlation between rainfall and temperature was c. 0.6. Correlations of this order are unlikely to present problems for regression modelling (Freckleton 2011). We also checked variance inflation factors to ensure that models were unlikely to be compromised.
The fits of the model set were compared using Bayesian Information Criterion (BIC), which allows the comparison of models with differing numbers of parameters (Hjort & Claeskens 2003; Link & Barker 2006). BIC is similar to Akaike Information Criterion (AIC; see Burnham & Anderson 2002). It is calculated for a suite of models with the best fitting having the smallest BIC value. Absolute BIC values are unimportant; it is the difference between them that indicates relative support for the models. Differences are calculated relative to the best fitting and smallest BIC (BICmin) for each model in the set. For model i, the BIC difference (Δi) is calculated as:
The wi sum to 1 for all R models in the set and can be interpreted as probabilities. The BIC weight is the probability that model i would be selected as the best fitting model were the data collected again under identical circumstances. Bayesian Information Criterion was used instead of the more frequently used AIC method because the AIC is well known to be more inclusive, often favouring more complex models. The difference between the AIC and BIC can be interpreted in terms of Bayesian priors, with the BIC tending to inherently attach more weight to simpler models (Link & Barker 2006). In our analysis, we in fact find that the BIC favours one of the relatively more complex models of distribution change, so we are confident that this is a real signal from the data, not a consequence of our choice of model selection criterion.
Patterns of Distribution Change
The overall picture of change is one of slow dynamics, with the relative change in distribution size being small: in the majority of species, proportional distribution change is <0.2 (53% of species for extinction, 72% for colonisation), while higher levels of change occur in a smaller number of species. Figure 2 summarises the colonisation and extinction dynamics for all species distributions. Extinction is presented as proportional loss of populated sites from the original distribution (Fig. 2a) and colonisation as proportion of available sites newly populated at the time of second census (Fig. 2b).
When considering distribution dynamics in relation to the total number of populated sites in the study area, it is apparent that species with high occupancy are buffered against high extinction rates. Rates of extinction decrease as occupancy increases (Fig. 2c). Decline in occupancy was observed in 15% of species during the time span of this study, all of which had low initial occupancy. The relationship between colonisation and occupancy (Fig. 2d) was in line with the overall pattern of colonisation rates; a high proportion of distributions undergo low levels of change and have low total occupancy (59%: <0.2 colonisation and <20% occupancy); there was also a high level of variance in this relationship. High occupancy often results in high colonisation rates, as this measure is proportional to the empty sites available. In species with high occupancy, there are fewer unpopulated sites remaining, which results in a higher proportional colonisation rate per site colonised.
Where substantial distribution change did occur, it was largely directional; a high extinction rate prevents a high colonisation rate and vice versa. If extinction matched colonisation there would be a turnover of sites leading to an overall stability in occupancy within a changing distribution, this does not occur (Fig. 2e). Figure 2f shows logit-transformed total occupancy at time (t) against total occupancy at time (t + 1). This highlights apparent change in total occupancy between time periods, with low occupancy species diverging the most from the linear relationship between time periods.
Comparing Models for Range Change
To show how the models relate to changes in occupancy, Fig. 3 shows an example of neighbour effects on distribution change, in this case for Eriophorum angustifolium. Fitted values show change in the probability of either extinction or colonisation events as a function of Ni,t. In this example, an empty site, where N = 0, is almost certain to remain unpopulated (Fig. 3a; probability 0.95), while where N = 8, this probability is greatly reduced (0.18) and there is an increased probability (0.82) of colonisation (Fig. 3b). Sites in the distribution that were initially populated have a reduced probability of extinction, relative to the probability of unpopulated sites remaining empty. This is particularly evident where N is low (Fig. 3a: solid line relative to dashed line). This suggests that at this temporal and spatial scale site, extinction is an uncommon event.
In all models, the effect of neighbours (Νe) for initially empty sites departed significantly from a null distribution of 500 species where neighbour presence was randomised (Fig. 4a–c; Model II: two-sample t(1986.8) = −81.2, P ≤ 0.001; Model III: two-sample t(2119.9) = −115.3, P ≤ 0.001; Model IV: two-sample t(1951.6) = −67.9, P ≤ 0.001). Νe at time (t + 1) shows the strongest effect with a mean of 0.74, followed by Νe for the time periods combined (0.62) and Νe at time (t) (0.59). High positive values of Νe indicate a strong relationship between N and distribution dynamics.
For populated sites (Fig. 4d–f), the mean of Νe was reduced relative to empty sites. Νe (t + 1) again shows the strongest effect (mean = 0.35), but in contrast, Νe for the combined periods (0.25) had a larger mean than Νe (t) (0.19). The null distribution remains highly significantly different (Model II: two-sample t(1396.2) = −19.4, P ≤ 0.001; Model III: two-sample t(1140.5) = −31.5, P ≤ 0.001; Model IV: two-sample t(1007.9) = −18.2, P ≤ 0.001). The reduced effect of N on these sites suggests that populated sites are intrinsically buffered against extinction and that site extinction was unlikely in the time frame of this study.
The model selection method using BIC weights shows a clear division in the model set. Models I, II and IV were selected as the best fitting model for relatively few species, 7%, 5% and 6%, respectively (Fig. 5a, b, d). Model III was shown to be the best fitting model for 82% of species (Fig. 5c). This shows that the model of distribution change with greatest support is the phalanx-spread model.
A second model selection process was carried out for the models containing climate terms (Models V–VII). These were compared with each other as well as Model III, the best fitting model of distribution change. The distribution of BIC weights for these models is shown in Fig. 6. The climate terms improve the fit of the model over that of Model III (Fig. 6a). Models V and VI are more heavily weighted than Model VII, which is penalised by the BIC method for its complexity (Fig. 6b–d). Rainfall was clearly favoured as a predictor (BIC weight >0.8) in 18.5% of species. Temperature was favoured in 25.4% of species. Models III and VII were the best fitting model for 3.5% and 1.2% of species, respectively. There is therefore a clear climate signal, favouring either rainfall or temperature, for c. 45% of species. The remaining species, many of which have BIC weights in the intermediate range for Models V and VI, show an effect of climate without clear distinction between the role of rainfall and temperature in distribution change. We suggest that this is a result of smaller distributions restricted primarily by habitat (e.g. coastal species) where the range in climatic variables is too small to identify a clear effect. There is also the potential for colinearity between rainfall and temperature, whereby they would explain overlapping portions of the variance; however, variance inflation factor tests carried out on these models indicate this is not the case.
The models in general showed only weak spatial autocorrelation (see Table S1). There was significant spatial autocorrelation in only c. 32% of species, which, given the relative simplicity of the models, would seem rather low. We did not include a specification of the form of the spatial autocorrelation beyond inclusion of the covariates, and in the majority of cases, this was clearly enough to account for any spatial dependency.
In this study, a long-term perspective has been used to examine how large-scale distributions change at a national scale. We have emphasised the importance of a dynamic perspective in species distribution modelling studies (Pearson & Dawson 2003). We show that in a large group of species, there are broad underlying patterns in the process of changes, despite the enormous variability in rates of colonisation and extinction. Historical spatial patterns of occupancy exhibit strong influences on changes occurring in distributions. This dependency reflects habitat suitability, the nature of dispersal as well as the factors that drive habitat change.
Extinction and Colonisation Dynamics
Population dynamics in plants are known to operate over a long time-scale and are prone to lags in response to environmental changes (Brook et al. 2009). Here, we have identified highly variable but predominantly low rates of extinction and colonisation over the time span of this study. These are systematically linked to occupancy (Fig. 2a–d). Low extinction rate can be attributed to the fact that many plant species are inherently resistant to extinction, through the long life span of individuals in some species, as well as the ability of seed banks and rhizomes to remain dormant during unfavourable conditions (Eriksson 1996). Moreover, the size of individual sampling units in this study is large. Site extinction at this scale is therefore indicative of a high-order change in the suitability of that site rather than temporary local population extinction.
Most plant species have restricted dispersal (Cain, Milligan & Strand 2000), which reduces their mobility and means that they are highly constrained by the availability of suitable habitat. Consequently, colonisation rates are limited for the majority of species, although wide-ranging colonisations are evident in some cases, possibly for recently introduced species with few specific habitat requirements (Preston, Pearman & Dines 2002; Hulme 2009).
Rate of distribution change is highly variable and species specific in this study, although modal rates are low. This has implications for efforts to generalise dynamics at this scale and predict future distribution change (Mustin et al. 2009). This variability may in part be a result of anthropogenic land use change in Britain. It is known that land use change has affected habitat quality and resulted in increased fragmentation over the 20th century (Thompson & Jones 1999). This phenomenon, which acts to limit species potential for distribution change, allows only habitat generalists or those species dependent on widespread habitat types to expand across the study area. This suggests that predictions of vegetation response to environmental change are greatly improved by prior knowledge of specific habitat requirements and the distribution of suitable and unsuitable habitat in the study area (Freckleton & Watkinson 2002; Thuiller, Albert & Araujo 2008; Willis et al. 2009).
We have identified several characteristics of large-scale distributions, which can be observed from simple presence/absence data, that are important to understanding and predicting distribution dynamics over time. There is a relationship between total occupancy and extinction/colonisation rate, which describes distribution change over time (Fig 2). All species that showed high extinction rates had low occupancy, a pattern that is maintained for absolute number of sites going extinct. This relationship suggests that species with low occupancy and high extinction rates are specialists experiencing a loss of suitable habitat, which is fragmented, or they are simply incapable of colonising suitable areas. The history of land use change and habitat fragmentation in the British Isles suggests that this situation may occur frequently (Bignal & McCracken 1996; Preston 2000; Walker & Preston 2006). Conversely, species with high occupancy, which appear buffered against extinction, are likely to be habitat generalists able to withstand land use or anthropogenic changes in habitat. The factors that limit a species current occupancy should be considered when attempting to model future vegetation distributions (Willis et al. 2009).
Models for Distribution Change
Improving the accuracy of approaches to plant distribution modelling is currently a major issue in ecology (Pearson & Dawson 2003; Elith et al. 2006; Buckley et al. 2010). Ecological niche models have been widely used for this purpose, but their limitations are now being recognised (Thuiller, Albert & Araujo 2008). A critical problem in many species distribution modelling studies is the use of purely static data ignoring species-specific abiotic and biotic processes (Guisan & Thuiller 2005). The use of historical records allows a dynamic perspective on distribution change. Rates of extinction and colonisation can be observed from changes that have already occurred, as can the factors responsible for this change in individual species or groups of species.
Similarly, attempts to apply general models of regional and large-scale dynamics to account for distribution change in plants have had limited success (Freckleton & Watkinson 2002; Ehrlen & Eriksson 2003). Metapopulation theory (Hanski 1998) rarely applies to plants, in part as a result of the lack of dispersal and possession of long-lived vegetative and dormant stages (Freckleton & Watkinson 2003; Freckleton et al. 2005). Here, we have presented three simple models of distribution change, which attempt to identify the evidence for long-distance versus localised colonisation and so dispersal, which in turn indicates the driving population processes. A key result is that the spatial arrangement of sites in a distribution can be used to predict extent and location of distribution change. The density of nearest neighbour sites (N) is a good predictor of the eventual state (populated or unpopulated) of a site in the distribution for all models. For the majority of species (82%), the phalanx-spread model is favoured. Sites are colonised at the immediate periphery of the established distribution; this process repeats to allow colonisation of large continuous areas. This mechanism of change suggests that distributions at this scale exist on and expand over large continuously occupied areas. Local processes are likely to govern extinction and colonisation patterns and dispersal (Freckleton & Watkinson 2002).
These models demonstrate distribution changes that may occur as a result of very different processes and that are limited by different factors. For example, in the species selected simply to provide example distribution maps for each model (Fig. 1), Solanum rostratum (random colonisation) is an introduced weed species, native to the Americas, which exists in short-lived scattered populations possibly as the result of multiple introductions and which has multiple dispersal vectors (Botanical Society of the British Isles). Luzula spicata (localised phalanx) is an alpine or subalpine species restricted to mountainous areas at the southerly limits of its distribution (Amen 1965). Lactuca serriola (phalanx-spread) is a native species whose northerly limit was widely believed to be a result of climatic limitations (Carter & Prince 1985). It has expanded across the UK in recent decades.
The evidence presented here for distribution change through each of the models described highlights the importance of prior knowledge of habitat constraints, dispersal ability and specific life-history information, which contribute to rates and mechanisms of change over time and in response to environmental stimulus.
The fact that distribution change occurs predominantly through spatially restricted, local population processes suggests that the development of ecological networks may be an effective conservation strategy for plant species in the UK (Lawton et al. 2010). An ecological network comprises sites which collectively contain the diversity and area of habitat needed to support species and which have ecological connections between them. The UK is largely made up of semi-natural habitats shaped by human land use, and so consequently, much of the UK’s wildlife is restricted to small fragmented areas of high habitat quality. These can rarely be restored to large unbroken areas of natural habitat. However, making connections between them through wildlife corridors and smaller ‘stepping stone’ sites is a much more feasible option, which would improve species ability to track environmental change through short-range colonisation (Hilty, Lidicker & Merenlender 2006).
Climate is undoubtedly an important factor influencing plant distributions (Walther et al. 2002; Thomas et al. 2004; Hulme 2009), and successful prediction of the response of large-scale distributions to climate change is a fundamental goal in planning for its potential effects. Climate envelope models have been widely applied to this end and are an example of the static correlative models discussed earlier. They match distributions to future climate simplistically by treating distributions and vegetation types as inflexible units, which snap to a new position. Many authors have highlighted the potential flaws in this simplistic approach (Woodward & Beerling 1997; Pearson & Dawson 2003; Jeltsch et al. 2008; Mustin et al. 2009; Wiens et al. 2009). Climate envelope modelling takes no account of changes in species composition or interactions under altered climate conditions. Neither do they incorporate the factors limiting a species ability to track climate change.
Here, we find evidence of a climate signal in our data set, which is of clear importance to distribution change in c. 45% of species. Rainfall and temperature appear to have separate or independent effects on distribution change as modelling them as covariates does not improve the model fit (see Results; Fig. 6). The role of climate in distribution change for the remaining species is unclear, as neither rainfall nor temperature is clearly favoured. Our method of model selection through BIC weights resulted in many models with an intermediate probability of being the best fitting. In such cases, it may be that habitat and land use effects are assuming a dominant role in distribution change that inhibits response to climate change.
The problems associated with understanding large-scale distributions and population dynamics in plants are well documented, and they have ongoing importance in ecological and conservation research (Freckleton & Watkinson 2002; Pearson & Dawson 2003; Freckleton et al. 2005; Guisan & Thuiller 2005; Elith et al. 2006). Ecological niche models and ‘climate envelopes’ have come under criticism and have the potential to produce misleading results. The difficulty arises because of the complexity of plant population processes, which are characterised by immobility, restricted dispersal and the time-scale of colonisation and extinction events.
Here, we have highlighted the specific and unpredictable nature of large-scale plant distribution change over time and identified key distribution characteristics and the relationships between them, which can be observed from simple historical records. Models for distribution change show colonisation, and extinction of sites is dependent on the spatial structure of the established distribution suggesting local population growth results in expansion into new habitat and buffers against extinction while isolated sites have a higher probability of extinction. Long-distance dispersal appears to be rare.
Land use history in Britain, which has resulted in habitat fragmentation, is likely to be the primary limiting factor on distribution size and future responses to environmental conditions. There is a clear link between climate and distribution change in our data set, although this cannot always be attributed unequivocally to either rainfall or temperature and is likely complicated by habitat availability. These results emphasise the limitations of simple static and correlative approaches as well as the complexities involved in mechanistic modelling. Use of historical distribution records, when combined with data on land use history, habitat requirements and life history will help to identify which characteristics of individual species are important to the way they have changed over time and so the potential form of future distributions. Future modelling efforts should attempt to include such data where they are available and be aware of the implications where it is not.
S.W.D. is funded by a PhD studentship via a Leverhulme Trust Research Leadership Award to R.P.F. R.P.F. is funded by a Royal Society University Research Fellowship.