Does seasonal fine-tuning of climatic variables improve the performance of bioclimatic envelope models for migratory birds?
ABSTRACT
We examined the influence of ‘seasonal fine-tuning’ of climatic variables on the performance of bioclimatic envelope models of migrating birds. Using climate data and national bird atlas data from a 10 × 10 km uniform grid system in Finland, we tested whether the replacement of one ‘baseline’ set of variables including summer (June–August) temperature and precipitation variables with climate variables tailored (‘fine-tuned’) for each species individually improved the bird-climate models. The fine-tuning was conducted on the basis of time of arrival and early breeding of the species. Two generalized additive models (GAMs) were constructed for each of the 63 bird species studied, employing (1) the baseline climate variables and (2) the fine-tuned climate variables. Model performance was measured as explanatory power (deviance change) and predictive power (area under the curve; AUC) statistics derived from cross-validation. Fine-tuned climate variables provided, in many cases, statistically significantly improved model performance compared to using the same baseline set of variables for all the species. Model improvements mainly concerned bird species arriving and starting their breeding in May–June. We conclude that the use of the fine-tuned climate variables tailored for each species individually on the basis of their arrival and critical breeding periods can provide important benefits for bioclimatic modelling.
INTRODUCTION
Recent studies suggest that the ongoing climate change has already caused changes in bird species phenology, migration and distribution patterns (Thomas & Lennon, 1999; Both & Visser, 2001; Ahola et al., 2004; Brommer, 2004; Crick, 2004). Simulations of the impacts of climate change on species distributions are often based on bioclimatic envelope models (e.g. Berry et al., 2002; Thuiller, 2003). These models relate present species distributions to selected aspects of present climate and fit the derived models into different climate change scenarios to predict the potential changes of species’ geographical distributions (Huntley, 1995; Pearson & Dawson, 2003; Thuiller, 2003). The approach has been applied to birds by, e.g. Berry et al. (2001), Peterson et al. (2002), Peterson (2003) and Araújo et al. (2005a,b).
A crucial factor in developing species distribution models, including bioclimatic envelope models, is the use of the most appropriate predictor variables possible. As argued by Austin (2002), modellers should, wherever possible, use direct and more proximal (e.g. rainfall) rather than indirect or distal predictors (e.g. altitude, latitude). However, the demands for careful selection of variables are especially high in bioclimatic models. This is because these models and their projections can have an important role in conservation planning and political debates. Thus, there are high demands on their accuracy and plausibility (Araújo et al., 2005a). Recent research has highlighted the fact that bioclimatic models can be vulnerable to a number of uncertainties (Pearson & Dawson, 2003; Hampe, 2004; Araújo et al., 2005b; Luoto et al., 2005), e.g. critical issues in model building and the selection of climatic variables for a given study (Kadmon et al., 2003; Thuiller et al., 2004; Thuiller, 2004; Araújo et al., 2005a; Beaumont et al., 2005). In the case of migratory birds, one potential source of uncertainty is the delineation of the climatic variables used in modelling bird species responses. Hitherto, very little attention has been paid to the selection of as accurate climatic variables as possible in the migratory birds–climate impacts models. Such attention can be particularly important at high latitudes (e.g. North Europe), from where several bird species migrate south for the winter (Newton & Dale, 1996). The breeding ranges of these species may be largely determined by the climate conditions prevailing at the time of their arrival, courtship and breeding (Virkkala, 1991; Huntley, 1995; Lennon et al., 2000).
Climate variables used in bird–climate models have varied considerably. Some studies have employed only mean annual precipitation and temperature values (e.g. Peterson, 2003; Seoane et al., 2003). By contrast, Araújo et al., (2005a) used, for example, temperature of the coldest and warmest month and July–September precipitation in addition to annual values. Some other multi-species studies have focused more on the relationships between birds and climate conditions of spring or early summer, i.e. precipitation and temperature of April–June (Forsman & Mönkkönen, 2003; Lemoine & Böhning-Gaese, 2003; Virkkala et al., 2005) or May–July (Lennon et al., 2000; Berry et al., 2001). However, it has rarely been examined whether applying climate variables tailored (‘seasonally fine-tuned’) separately for each of the bird species individually instead of using one common set of climate variables for all the studied species would increase model performance (cf. Huntley, 1995). This paucity of bird–climate modelling studies considering the fine-tuning of variables is surprising because autecological studies have indicated that the climate conditions during the period of courtship and the early weeks of breeding can markedly influence the occupancy patterns and breeding success of birds (Redpath et al., 2002; Rodríguez & Bustamante, 2003; Jovani & Tella, 2004). Consequently, many multi-species modelling studies that have included a large number of migratory bird species (e.g. Brotons et al., 2004; Araújo et al., 2005b) may have been exposed to a possible mismatch of climate variables and critical periods in breeding, and to lowered model accuracy for these species.
In this study, we used generalized additive models (GAM) and distribution data of 63 migratory bird species in Finland at the resolution of 10 × 10 km to examine whether the seasonal fine-tuning of climate variables based on the arrival and early breeding periods of the studied species improves the performance of bird–climate models. Specifically, we addressed the following questions: (1) Does the replacement of baseline summer (June–August) climate variables with variables tailored for each species individually significantly improve the accuracy of the models? (2) Are the possible model improvements related to the time of arrival and breeding of the studied bird species?
METHODS
Study area
Finland covers an area of c. 338,000 km2 in northern Europe between latitudes 59°30′ and 70° N. The climate shows characteristics of both an oceanic and a continental climate, the continentality growing inland and eastwards (Tuhkanen, 1984). The majority of the country has a boreal climate, with a decrease in rainfall and temperature from the south-western hemiboreal zone (mean annual temperature c. 5 °C and mean annual precipitation 600–700 mm) to the subarctic region in northernmost Finland (−2 °C and 400 mm). Biogeographically, Finland is located mainly in the boreal coniferous vegetation zone, and the landscape is largely dominated by forests and mires.
Bird data
The studied 63 land bird species bred and/or foraged in the main terrestrial habitats: 19 species occurred primarily in forests, 17 species in agricultural and bushy habitats, 13 species in mires, 9 species in marshes and coastal wetlands and 5 species in mountain heaths (see Appendix S1 in Supplementary material). Nomenclature of the species follows Dickinson (2003). The species were assigned to three groups according to their average arrival–early breeding period: those arriving and having their early breeding period in (1) March–April (2) April–May, or (3) May–June (see Appendix S1). This classification was based on a literature survey of the key publications of the migratory periods of bird species in Finland (e.g. Hildén et al., 1979; Pöyhönen, 1995).
The information on the distribution of species and the level of survey activity was extracted from the second bird atlas survey in Finland, which was carried out in 1986–1989 and included 3800 squares of 10 × 10 km (Väisänen et al., 1998). Recorders and organizers of the survey graded the survey activity in each square according to six categories: 0 = no observations, 1 = occasional observations, 2 = fair survey, 3 = satisfactory survey of the square, 4 = well surveyed and 5 = thoroughly surveyed square (Väisänen et al., 1998). We used only squares with survey activities of 2–5 in our analysis. Consequently, the data used in the analyses consisted of 2861 squares. Väisänen et al. (1998) listed the breeding status of bird species recorded in each of the grid squares in four classes: 0 = not found, 1 = breeding possible, 2 = breeding probable and 3 = confirmed breeding. For the analysis of this study, we combined classes 1, 2 and 3 as a species-present variable.
Climate data
Climate data produced by the Finnish Meteorological Institute, using the same 10 × 10 km grid system, were employed as predictors of the bird distribution data (Venäläinen & Heikinheimo, 2002). The climate data included mean values for the period 1985–1989 for all climatic variables. Because we focused on analysing the impacts of fine-tuning on the performance of bird–climate models, we used a limited number of variables. The baseline set of predictors simulated the commonly used approach of using one set of explanatory variables for all species. It included three variables: mean temperature of the coldest month (MTCO), mean summer temperature (defined as June–August; TEMPSUM), and mean summer precipitation (June–August; PRESUM). For the fine-tuning analysis we calculated the mean values for three pairs of months: mean temperature of March–April (TEMMA), April–May (TEMAM) and May–June (TEMMJ), and mean precipitation of March–April (PREMA), April–May (PREAM) and May–June (PREMJ). MTCO was included to indicate the general harshness of the environmental conditions in each grid square.
Model calibration
Generalized additive models were developed using grasp (Generalized Regression Analysis and Spatial Prediction) (Lehmann et al., 2003) in S-Plus (version 6.1 for Windows, Insightful Corp.). All the GAMs were built using a binomial distribution of error and a logistic link function, and a stepwise selection procedure to select important explanatory variables and the level of complexity of the response shapes of the various species to each variable. A starting model including all continuous predictors smoothed with 4 degrees of freedom was fitted first. Next, the stepwise grasp procedure proceeded in a loop that aimed to eliminate one variable at a time from the full model. At each step, the least significant variable was dropped from the model or converted to a linear form, after which the loop proceeded with the remaining variables (see Lehmann et al., 2003). The final models for different species may thus include different numbers of predictors with either 1 or 4 degrees of freedom for the spline smoother. The variable dropping or conversion to linear form was tested in the first set of GAMs using Akaike's Information Criterion (AIC) (Akaike, 1974).
The modelling procedure included building two GAMs for each species. In the first set of GAMs (‘baseline GAMs’), the three baseline climate variables (MTCO, TEMPSUM, PRESUM) were used as the three explanatory variables. In the second set of GAMs (‘fine-tuned GAMs’), mean summer temperature was replaced by one of the fine-tuned temperature variables (TEMMA, TEMAM, TEMMJ) and mean summer precipitation by one of the precipitation variables (PREMA, PREAM, PREMJ), based on the information on the average arrival-early breeding periods of the studied species (Appendix S1). In all the models, the three climate predictors included were subjected to the grasp stepwise variable selection process in order to develop parsimonious models, as described in the previous paragraph.
Model evaluation
The models were evaluated in two ways. First, the explained deviance in each model (i.e. the ratio of explained deviance vs. the total deviance) was investigated. This was regarded as the ‘explanatory’ power of the model. Second, we examined the cross-validation statistics in each model (the ‘predictive’ power of the model). Cross-validation was made with four spatially random subsets of the entire data set. Each subset was dropped from the model, the model was recalculated and predictions were made for the omitted data points (see Lehmann et al., 2003). Predictive power was then measured using the area under the curve (AUC) of a receiver operating characteristic (ROC) plot (Fielding & Bell, 1997).
Model validation using cross-validation based on spatially random sets of grid squares may be vulnerable to the effects of spatial autocorrelation in the predictor and response variables (e.g. Legendre, 1993; Koenig, 1999). Cross-validation based on random regions (Peterson, 2001) or random blocks of, e.g. 100 × 100 km (cf. Augustin et al., 2001) may circumvent autocorrelation problems better but they may cause other kind of difficulties for validation. For example, use of large random regions may lead to loss of information because not all species necessarily occur in acceptable numbers in both calibration and evaluation data sets (see Peterson, 2001). Random blocks of, e.g. 100 × 100 in in size would work slightly better in this respect. However, as several of the studied species have very narrow distributions in Finland, the model validation using random blocks would also result in losing species information. Validation using both random regions and random blocks may be vulnerable to the risks of comparing different sampling strategies or field survey activities instead of evaluating a model (Lehmann et al., 2003).
There is a plethora of techniques available for examining spatial autocorrelation structure and its impacts (e.g. Legendre, 1993; Lichstein et al., 2002; Diniz-Filho et al., 2003). However, we do not examine spatial autocorrelation in our data with such techniques here, mainly because (1) in some biogeographical study settings similar to our work, autocorrelation in species data have been shown to be largely accounted for by the similarly autocorrelated climate predictors (Diniz-Filho et al., 2003), and (2), our climate variables are rather closely correlated (see Table 1) and thus we presume that potential autocorrelation would have a broadly equal impact on the validation of the baseline and on fine-tuned GAMs, and would not alter our comparative results.
| Summer temperature | Summer precipitation | |
|---|---|---|
| Temperature | ||
| March–April | 0.899*** | — |
| April–May | 0.939*** | — |
| May–June | 0.947*** | — |
| Precipitation | ||
| March–April | — | 0.603*** |
| April–May | — | 0.661*** |
| May–June | — | 0.521*** |
- *** P < 0.001.
However, instead of applying more sophisticated techniques to examine autocorrelation issues, we use a simple ad hoc, but prudent, course of action to take its potential impacts into account in the modelling. One of the core impacts of spatial autocorrelation is that type I errors may be inflated, i.e. coefficients declared to be significantly different from zero when in fact they are not (Legendre, 1993). A simple means to take this into account is to reduce the probability levels in model inference tests (Koenig, 1999; Guisan & Thuiller, 2005). Consequently, we re-ran all our GAMs using the fourfold cross-validation process but accompanied with F-tests using a stringent probability limit (0.001) for the selection of the predictors. The results of the AIC-based and F-test–based GAMs were then compared.
To examine the differences in the explanatory and predictive power of the models, we compared the explained deviance and AUC values from the baseline GAMs vs. fine-tuned GAMs for each species. Because deviance and AUC data were not significantly different from normal distribution (Kolmogorov-Smirnov tests, P > 0.05), we used paired t-test for measuring the significance between the baseline and fine-tuned models. Because we expected a priori that fine-tuned climatic variables would improve the model fitting, we used one-tailed t-tests in all analyses. Paired one-tailed t-tests were conducted first for all the 63 studied species simultaneously, and then for the early and mid (March–April and April–May) arriving bird species (‘early and mid arrivers’; n = 29) vs. late spring (May–June) arriving species (‘late arrivers’; n = 34) separately. The early and mid arrivers were lumped in this analysis because of the low number of species (n = 6) included in the first group.
As an example, we produced projected future distributions for selected study species based on both the derived baseline and the fine-tuned GAMs and using the climate data obtained from the HadCM3 general circulation model (GCM) under the business-as-might-be-usual (BAMBU) scenario A2, compiled by the EC FP6 Integrated Project ALARM (http://www.alarmproject.net). This was carried out to examine whether the observed differences in the predicted current distributions from the baseline and fine-tuned models increase when the models are employed to produce predictions of future distributions.
RESULTS
The summer temperature and the three fine-tuned temperature variables were highly correlated with each other, whereas the correlations between precipitation variables were moderate (Table 1). Concurring with our a priori expectations, the explanatory power of fine-tuned models (model selection using AIC) was on average significantly higher (mean percent of deviance explained = 35.1%) than that of baseline models (mean = 34.5%) (paired one-tailed t-test; d.f. = 62; t = −2.135; P = 0.018). In addition, the predictive power (AUC) of the fine-tuned GAMs (mean = 0.887) was slightly but statistically significantly higher than that of the baseline GAMs (mean = 0.861) (paired one-tailed t-test; t = −1.948; P = 0.028).
Separate analysis of the explanatory and predictive powers of early mid and late arrivers revealed that the differences between the performance of baseline and fine-tuned AIC-based GAMs was caused by the models for May–June arrivers. Both the deviance and AUC values were statistically significantly higher in late arrivers models using fine-tuned climate variables than in the baseline GAMs (Table 2). Moreover, in the models for late arrivers there was a clear general trend towards fine-tuned models providing higher performance (deviance change, 24 vs. 10 species; AUC, 22 vs. 8 species).
| Species group | Baseline climate variables | Fine-tuned climate variables | t | P | Ranks |
|---|---|---|---|---|---|
| (a) Early mid arrivers | |||||
| Deviance | 0.319 ± 0.028 | 0.326 ± 0.026 | −1.117 | 0.137 | 12/17/0 |
| AUC | 0.844 ± 0.014 | 0.852 ± 0.011 | −1.258 | 0.109 | 13/14/2 |
| (b) Late arrivers | |||||
| Deviance | 0.367 ± 0.026 | 0.373 ± 0.026 | −2.422 | 0.011 | 10/24/0 |
| AUC | 0.876 ± 0.010 | 0.880 ± 0.010 | −3.171 | 0.002 | 8/22/4 |
The results of the GAMs based on model selection using the F-test and a probability level of 0.001 (Appendix S2 in Supplementary material) were very similar to those of AIC-based GAMs. Thus, correcting the model inferences for the potential autocorrelation effects simply by reducing the probability levels did not alter our results, and in the remaining part of the paper we will focus on AIC-based modelling results.
The set of species for which the fine-tuning resulted in the highest increase in model performance included birds from different taxonomical and habitat type groups, including turtle dove Streptopelia turtur, arctic warbler Phylloscopus borealis, hawfinch Coccotraustes coccotraustes, ring ouzel Turdus torquata, grey heron Ardea cinerea, moorhen Gallinula chloropus and black-tailed godwit Limosa limosa. The highest (24%) increase in the predictive power was in the case of the black-tailed godwit. As an example, the predicted distributions are presented for the grey heron and the arctic warbler (Fig. 1). The overall pattern of predicted probability of occurrence for the two species did not vary greatly between the baseline and fine-tuned GAMs. However, there were clear regional spatial differences in the model accuracy. The fine-tuned models predicted better, e.g. the northernmost occurrences of the grey heron and certain agglomerations of occurrences of both species along the eastern border of the country.
Recorded and projected distributions of two species: (a) recorded occurrence of the grey heron in grid squares of 10 × 10 km in 1986–1989, and probability of occurrence of the species based on (b) the ‘baseline’ bioclimatic envelope model, and on (c) the ‘fine-tuned’ bioclimatic model; (d) recorded occurrence of the arctic warbler in 1986–1989, and probability of occurrence of the species based on (e) the ‘baseline’ bioclimatic envelope model, and on (f) the ‘fine-tuned’ bioclimatic model. Probability is shown on a three-level scale. D2 = percentage of explained deviance; AUC = the area under the curve of a receiver operating characteristic (ROC) plot.
Figure 2 shows the predicted future distributions for four species based on fitting the derived species-specific baseline and fine-tuned GAMs in the climate data obtained from the HadCM3 GCM under the BAMBU scenario A2. The projected distributions are, in the case of the arctic warbler, the same (both models predict that a suitable climate for the species will not exist in Finland in the future). In the case of the red-breasted flycatcher Ficedula parva the differences are relatively small. However, the projections for the grey heron and the turtle dove show that the slight differences in the predicted current distributions may be more important when the models are fitted to the climate scenarios. The difference in the AUC between the baseline and fine-tuned GAMs was greatest in the case of the grey heron (Appendix S1), which also showed the greatest difference in the projected future distribution between the two models.
Projected future distributions of four bird species in Finland based on models calibrated with climate and species data from the 1980s and fitted to the climate variable data obtained from the HadCM3 GCM under the BAMBU scenario A2 for the year 2050: the occurrence of (a) the grey heron (Ardea cinerea) based on the ‘baseline’ GAM and (b) the ‘fine-tuned’ GAM; (c) the turtle dove (Streptopelia turtur) based on the baseline GAM and (d) the fine-tuned GAM; (e) the arctic warbler (Phylloscopus borealis) based on the baseline GAM and (f) the fine-tuned GAM; and (g) the red-breasted flycatcher (Ficedula parva) based on the baseline GAM and (h) the fine-tuned GAM. Black dots represent the 10-km grid squares modelled as suitable for the species (projected presence) and unmarked were areas modelled as not suitable (projected absence). To determine the probability thresholds at which the predicted values for species occupancy are optimally classified as absence or presence values, we used prevalence of the species as the probability level as suggested by Liu et al. (2005).
DISCUSSION
A number of bioclimatic envelope modelling studies have employed multi-species simulations based on one set of climate variables for all the studied species (e.g. Thuiller, 2003; Brotons et al., 2004; Araújo et al., 2005a, b), although the variables included have varied between the studies. For sedentary organisms, especially plants, this approach may work well. Variables such as mean annual temperature and precipitation, minimum temperature of the coldest month, growing degree days and moisture availability can have strong links with the physiology and growth of plants (Huntley et al., 1995). The distributions and densities of resident bird species may also be adequately modelled using a combination of winter and summer climate variables or mean annual variables (Huntley, 1995; Forsman & Mönkkönen, 2003; Seoane et al., 2003).
However, more mobile species such as migratory birds pose challenges for developing accurate bioclimatic models. As Huntley (1995) argued, it is probable that one set of climate variables will not be applicable to all birds in the manner that one set of key variables (e.g. growing degree days, mean temperature of the coldest month and moisture availability) can be applied to a wide range of terrestrial plant species. Our results suggested that the explanatory and predictive power of the species–climate models especially for birds arriving and breeding in May–June in Finland can be enhanced by replacing the summer climate variables with fine-tuned climate variables. Although the absolute increases in the amount of explained deviance and cross-validation AUC values were generally small, they were rather systematic and statistically significant. We consider these differences important for three reasons. First, as the fine-tuned climate variables always provided better or equally good model performance than the baseline summer variables, we see no reason not to use the ecologically more reasonable fine-tuned variables in modelling the species studied here. In other words, the use of fine-tuned predictors should inherently improve the mechanistic basis of the bioclimatic envelope models for migratory birds. Second, particularly in the results for the late arrivers, there was a clear trend of the fine-tuned GAMs outcompeting the baseline GAMs. Third, recent studies have suggested that even slight differences between current predictions from different bioclimatic models can be exacerbated when the models are used for simulating future distributions (Thuiller, 2003; Araújo et al., 2005b). However, it should be noted that in the work of Thuiller (2003) and Araújo et al. (2005b), the differences between the simulated future distributions were caused by the differences between modelling methods. One of the lessons learned from the results of our study is that a priori selection of the climate predictors can also result in intensified differences in the future projections of species distributions and represents another important source of uncertainty in bioclimatic models (cf. Heikkinen et al. in press).
The reason why the increases in the model performances between baseline and fine-tuned models were not greater is probably the result of the considerably high correlations between the mean spring and summer temperatures for the period of 1985–1989 in Finland (Table 1; see also Järvinen, 1989). Summer climate variables can thus act as statistically reasonable, but ecologically imprecise, surrogate predictors of migratory bird species distributions. Moreover, it is probable that the impact of climatically deviating years (e.g. exceptionally warm or cold springs) on the occupancy and density patterns of migratory birds in boreal regions (see Järvinen, 1989; Virkkala, 1991; Ahola et al., 2004) may be ‘lost’ in the mean climate values of longer-term data, even though their influence in the species distributions may be visible in the atlas maps.
Our results have some general implications for the bioclimatic envelope modelling of birds. First, if one combination of climate variables is used in modelling studies including migratory bird species, it is intuitively more appropriate to use the mean temperature or precipitation of periods such as May–July (e.g. Lennon et al., 2000; Lemoine & Böhning-Gaese, 2003) than variables extending to August–September (cf. Brotons et al., 2004). The potential dangers of the mismatch in migratory bird–climate models in boreal regions are most apparent in the case of waders, such as whimbrel Numenius phaeopus and greenshank Tringa nebularia studied here, which finish their breeding and the adult birds start their ‘autumn’ migration already in June–July. In the modelling of such species, late summer climate conditions have a smaller role than spring and early summer climate. However, this does not necessarily mean that the climate of all other months than the migratory or residence periods is completely irrelevant for the distribution and occupancy rates of migratory birds. For example, lack of winter rains or severe cold periods in winter may result in limited vegetation growth or food (e.g. insects) availability and thus lower the suitability of habitats for birds in the following summer (cf. Rodríguez & Bustamante, 2003). However, the relative explanatory power of migratory-residence period climate vs. the climate conditions of the rest of the year has rarely been examined in the same multiple regression settings. The few results available, including those of Lennon et al. (2000), have suggested that the single most powerful explanatory variable for individual species distributions is May–July temperature.
Second, we argue that distribution modelling of migratory birds would benefit from using climate variables tailored for each species individually. Recent autecological studies on migratory birds have shown that particularly the climate conditions during the arrival, courtship and early breeding period affect the occupancy patterns and breeding success of birds (Redpath et al., 2002; Rodríguez & Bustamante, 2003; Jovani & Tella, 2004). In a detailed study of the lesser kestrel Falco naumanni (Fleischer, 1758), Rodríguez & Bustamante (2003) showed that the occupancy rates among the colonies were best explained by the temperature and rainfall during the courtship period in April–May, the rainfall in spring being also a key determinant of the nest success rate and the mean number of chicks per nest. Incorporating species-specific knowledge of the critical climate factors from autecological studies into the multi-species bioclimatic modelling exercises would decrease the uncertainty of the models stemming from the possibly loosely delimited climate variables, and yield information on the response of the species to climate in different parts of its geographical range (cf. Redpath et al., 2002).
In our results, the application of fine-tuned variables improved significantly the overall model performance of the studied species and especially the models of May–June arrivers. This may be related to the phenomenon discussed by Kalela (1952) who argued that migratory birds that overwinter in the tropics and arrive in the breeding regions in the late spring can respond to increased spring temperatures by a prolongation of migration and a rapid occupation of new areas. Such species include insectivorous birds such as grasshopper warbler Locustella naevia and reed warbler Acrocephalus dumetorum, which showed a rapid spread of distribution in the warm periods in the 1930s in Finland. Moreover, Väisänen et al. (1998) argued that the expansion of, e.g. the river warbler Locustella fluviatilis in the 1980s and 1990s can be explained by prolonged migration as a result of the warming of late spring. However, Kalela linked the prolongation of migration-improved spring climate relationships also to certain short-distance migrants that arrive earlier in spring to northern Europe, such as lapwing Vanellus vanellus and moorhen G. chloropus. Similarly, in our results, the group of species with the greatest improvements in model performance did not form a uniform group consisting of long-distance tropical migrants.
We conclude that the use of the fine-tuned climate variables tailored for each species individually on the basis of their arrival and critical breeding periods can provide important benefits. In other words, researchers can improve the accuracy and plausibility of the migratory birds–climate models by applying fine-tuned climate variables instead of the general approach used in multi-species modelling studies (i.e. applying one set of predictors for all species). Taking the species differences in arrival and breeding periods into account is particularly essential at high latitudes, e.g. North Europe, where the proportion of the species migrating for the winter can be 50% or more (Newton & Dale, 1996). Overall, it is evident that bioclimatic modelling of migratory birds poses greater challenges for investigators than developing useful models for more sedentary organisms.
ACKNOWLEDGEMENTS
Niko Leikola and Stefan Fronzek helped in aggregating the climate and bird data. The comments by the four referees helped greatly in improving our paper. This research was funded by the EC FP6 Integrated Project ALARM (GOCE-CT-2003-506675).
