Higher spring temperatures increase food scarcity and limit the current and future distributions of crossbills

2.1 Geographical and temporal variation in seed fall

We measured Scots pine seed fall at a total of four sites in two different regions in the Iberian System mountain range, Spain. One region was in Teruel Province at the south-eastern limit of the Iberian System, with sites in Valdelinares and Fortanete (40°23′N, 0°36′W). The other was in Soria Province at the north-western limit of the Iberian System, with sites in Duruelo de la Sierra and near Santa Inés mountain pass (42°00′N, 2°50′W). At each site, we systematically selected 15 trees every 30–60 m along 1 km transects and placed one trap to collect falling seeds under each tree. Seed traps consisted of aluminium trays placed on the ground with a catch area of 0.15 m² and covered with 1.3-cm wire mesh to prevent seed removal by birds and mammals. Each year, traps were set up during February, and seeds were collected every 17–20 days until early July. Seed fall was recorded over 3 years (2010–2012).

We used seed fall data from Scots pine at Abernethy Forest, Scotland. Timing of seed fall in this forest has been previously described using part of this dataset (Summers, 2011; Summers & Proctor, 2005). Methods followed are those described in that study. In short, four sites were selected at Abernethy Forest (57°15′N, 3°40′W), three in stands of ancient native pinewood (Ice Wood, Memorial Wood, and Bognacruie) and one area of plantation woodland, grown from seeds of local provenance (Tore Hill). Five seed traps were located equidistantly under trees along 1–2 km transects set through each site. Seed traps were plastic pots with a catch area of 0.13 m². Traps were set up in February, and seeds were collected at the end of each month until September. Seed fall was recorded for 16 seed years (1992–2007).

We used published information for two sites in central Sweden (Hannerz, Almqvist, & Hornfeldt, 2002) to characterize the timing of Scots pine seed fall in Scandinavia. Hannerz et al. (2002) recorded seed fall for 4 years (1993–1996) at Garpenberg (60°16′N, 16°11′E), using three plastic trays with a catch area of 0.14 m² placed on the ground, 50-m apart. Traps were set up in early March, and seeds were collected on eight occasions until late July or early August. Seed fall was recorded for 3 years (1996, 1997, and 1999) at Knivsta (59°43′N, 17°49′E). At this site, seed traps consisted of circular bag nets with a catch area of 0.25 m² placed about 1 m above the ground. Ten traps were set up in late March, with an average distance of 10 m between traps, and seeds were collected on 10–14 occasions until July.

To estimate the timing of seed dispersal, we followed similar calculations to those made by Hannerz et al. (2002). We counted the number of Scots pine seeds per seed trap at each collection date and calculated the number of seeds per square metre. Dates were transformed into Julian days (1 = 1 January), taking into account leap years, in which we added 1 day after February 28. We included seed fall data from March to late July. February was excluded from calculations because seed fall data were not gathered for February in some years in Scotland, and when measured in February, no seeds were trapped in nearly all of the years. We calculated the cumulative proportion of the total seed fall for every collection date at each site and used these values to fit a logistic function for each year and site. Models were fitted using nonlinear least squares estimation in statistica (Statsoft, Inc., Tulsa, OK, USA). Regression coefficients were used to estimate the days when 10%, 50% and 90% of the seed fall occurred (D10, D50, D90, hereafter) (see Appendix S1).

For each year and site, we determined the monthly mean, maximum and minimum temperatures, and monthly precipitation using the European daily high-resolution gridded climate dataset (0.25° grid resolution; Haylock et al., 2008) to relate to the timing of seed fall. To determine the relationship between climate variables and annual variation in timing of seed fall, we used monthly climate variables and the estimated D50 for each year and site in Scotland (i.e., the longest time series dataset). Preliminary correlations indicated that D50 was consistently correlated with climate variables for March and April, and mainly with maximum temperature. We fitted a multiple linear regression model between D50, as the response variable, and March and April maximum temperatures, and March precipitation as the predictor variables. March and April mean and minimum temperatures were not included due to their high correlation with maximum temperatures (r > .75, for all correlations), and April precipitation was excluded because of its high correlation with April maximum temperature (r = −.68). Finally, to evaluate whether the same climate variables were related to annual variation in seed fall for the three European regions, we fitted a linear mixed model between D50 and March and April maximum temperatures, and March precipitation, including region as a random effect (we compared a random intercept model to a random intercept and slope model).

2.2 Modelling observed crossbill distributions

To examine whether observed crossbill distributions are related to climate variables that influence the timing of Scots pine seed fall and cover of different conifer taxa, we modelled the presence of each crossbill species in relation to March and April maximum temperatures as well as cover of Scots pine, spruce (Picea spp.) and pines (Pinus spp.) other than Scots pine (hereafter referred to as “other pines”). Information on the current distribution of the three crossbill species came from the Atlas of European Breeding Birds (Hagemeijer & Blair, 1997). Data in this Atlas record the presence and absence of species breeding in Europe within cells of 50 × 50 km (Appendix S1).

We used maps of Scots pine, spruce and other pines developed by Brus et al. (2012) for European countries excluding Russia and Iceland that provide percent cover for each tree species (0%–100%) at 1 × 1 km resolution (Brus et al., 2012;). “Other pines” mostly included species regularly foraged on by crossbills (P. halepensis, P. mugo, P. nigra and P. uncinata), but also some species not used by crossbills (e.g., P. cembra, P. pinea; Cramp & Perrins, 1994). Information to map species distributions also included plantations. Thus, groups of species (spruce and other pines) included, for example, non-native plantations in the British Isles comprised mainly of Sitka spruce Picea sitchensis and lodgepole pine Pinus contorta (Brus et al., 2012), which are commonly utilized by crossbills (Summers & Broome, 2012).

We used monthly mean climate data (~1950–2000) from the WorldClim dataset (Hijmans, Cameron, Parra, Jones, & Jarvis, 2005) at 5′ resolution. We extracted data for monthly maximum, mean and minimum temperatures, and precipitation. For each 50 × 50 km grid with crossbill presence/absence data, we calculated mean values for maximum temperatures and percent cover of the different conifer taxa. We modelled the distributions using boosted regression trees (BRTs), a robust algorithm used in species distribution modelling (Elith, Leathwick, & Hastie, 2008). We used a binomial error structure and followed recommendations by Elith et al. (2008) for parameter settings (Appendix S1).

We divided the dataset of each crossbill species by selecting 70% of the data for model calibration and using the remaining 30% of the data to evaluate the predictive performance of the models. To account for spatial autocorrelation in the data, we split the data by spatial blocking (Roberts et al., 2017), using the “checkerboard1” method in the ENMeval package (Muscarella et al., 2014) in R 3.1.2 (R Core Development Team). We provide the percentage of variance explained and the area under the ROC curve (AUC), estimated on the evaluation datasets (Elith et al., 2008), as measures of accuracy for the selected models.

We used the best-fitted BRT model for each crossbill species to estimate their current potential distribution in Europe (excluding Iceland and Russia; as predicted by the models). To compare the observed distribution of each crossbill species with its potential distribution, we also calculated the area where each species is present according to the European Atlas (i.e., the area of the grids with presence of each species in the Atlas excluding Iceland and Russia). To make both maps more comparable, we overlaid a map of Europe on the Atlas grids to exclude areas lying outside of land (e.g., sea) and rasterized grids to 5′ resolution for area calculation. This calculation provides a rough estimate of the similarity between the distribution of crossbill species and their potential suitable habitat under current climate conditions.

2.3 Modelling future shifts in crossbill distributions

To forecast shifts in future crossbill distributions, we first modelled current potential crossbill distributions across Europe as a baseline for comparison. We used the BRT model for each crossbill species using March and April maximum temperatures and conifer cover, as above, to predict their current potential distribution across Europe (including Iceland and Russia). However, we modelled the potential cover of each conifer group (which may not currently be present due to historical factors or human influence) instead of using forest cover maps so that we could also forecast future distributions. We built models with cover of each conifer group as a separate response variable and three environmental variables as predictors: growing degree-days (GDD), absolute minimum temperature (AMT) and water balance, using monthly mean climate data (~1950–2000) from the WorldClim dataset (Hijmans et al., 2005) at 5′ resolution (Appendix S1). Because cover values were integers including many zeros, we modelled conifer cover using BRTs with a Poisson error structure and similar procedures as explained above (Appendix S1).

Predictions from the BRT models for each crossbill species provided the probability of presence for each grid cell. To select the threshold for converting the continuous logistic probability scale to a binary prediction of potential presence or absence, we calculated model-specific thresholds that maximized the sum of sensitivity and specificity (Jiménez-Valverde & Lobo, 2007). The restricted distribution of the Scottish crossbill resulted in few presences in the Atlas dataset, and this threshold method usually overpredicts the distribution of species with low prevalence. Thus, we estimated the optimal threshold value for this species using the SDMTools package (http://www.rforge.net/SDMTools/) in R. These thresholds were used for each model of current potential crossbill distribution, and in all models using future climate projections.

To forecast the future distributions of each crossbill species, we used the BRT models built for each species using climate change projections for 2070 (IPCC Fifth Assessment Report data) at 5′ resolution, also available from the WorldClim database. We used the IPCC-CMIP5 climate projections from three Global Circulation Models (GCMs) under two representative concentration pathways (RCP 4.5 and 8.5) (Appendix S1). We extracted monthly temperature and precipitation values for the six scenarios and computed the same three climate variables (i.e., GDD, AMT and water balance) as for current conditions. First, the distribution for each conifer species was forecasted by calculating the set of three predictor variables (i.e., GGD, AMT and water balance) under future conditions and predicting with the BRT models fitted for each conifer species. Projections for each conifer species together with March and April maximum temperatures for each scenario were used to forecast potential crossbill distributions (i.e., potential suitable habitat for each species under future conditions), as done for current conditions. We then calculated the predicted area of distribution for each crossbill species under each projection. In addition, we estimated the area for which projections from the three GCMs for each RCP overlapped (i.e., ensemble of predictions), the area where any of two predictions overlapped, and the sum of the areas predicted by at least one projection. All models were built, evaluated and projected using the dismo, raster and gbm packages in R (Hijmans, 2014; Hijmans, Phillips, Leathwick, & Elith, 2013; Ridgeway, 2013).

3.1 Geographical and temporal variation in seed fall

The estimated time when 10%, 50% and 90% of the seeds were shed (i.e., D10, D50 and D90) was 1.5–2 months earlier in the Iberian Peninsula than in Scotland or central Sweden (Figure 2; Appendix S2). The average date for D50 across all sites and years in each region was 25 March in the Iberian Peninsula, 23 May in Scotland and 14 May in central Sweden (Figure 2; Appendix S2).

The linear multiple regression indicated that inter-annual variation in D50 in Scotland was negatively correlated with April maximum temperature (coefficient ± SE: −5.29 ± 0.99, p < .001), whereas March maximum temperature and precipitation did not have a detectable effect (−1.62 ± 1.05, p = .13; 0.07 ± 0.04, p = .11, respectively). The model explained 62% of the variation (F_3,44 = 23.9, p < .001). The mixed model for all three regions including random variation around the intercept provided a better fit than a mixed intercept and slope model according to Akaike's Information Criterion (510.8 vs. 517.9), indicating that slopes did not differ among regions. The mixed intercept model showed that D50 was negatively associated with April maximum temperature (−4.00 ± 0.59, p < .001; Figure 3) and March maximum temperature (−3.18 ± 0.74, p < .001) and positively associated with March precipitation (0.08 ± 0.03, p = .018). Warmer spring temperatures accelerate seed fall and thereby increase the window of food scarcity for crossbills before the next seed crop develops (Figure 1).

3.2 Observed crossbill distributions

The BRT model for parrot crossbills performed well (% deviance explained: 83.2; AUC: 0.99), the one for common crossbills was less accurate, but still good (% deviance explained: 36.8; AUC: 0.88), with the model fit for Scottish crossbills intermediate (% deviance explained: 69.8; AUC: 0.99). The BRT model for common crossbills indicated that spruce cover (relative influence: 42.9%) and April maximum temperature (23.4%) contributed most to explain its observed distribution (Figure 4). April maximum temperature (63.0%) was the most important variable explaining parrot crossbill occurrence, followed by Scots pine cover (25.1%; Figure 4). The BRT model for Scottish crossbill distribution showed greater relative contributions for cover of other pines (49.1%), followed by spruce cover (22.0%) and April maximum temperature (17.3%; Figure 4). All crossbill species were more likely to occur where April maximum temperatures were lower (Table 1) and thus where seeds were more likely to be retained in the cones in late spring and early summer (Figure 3). Each crossbill species was also more likely to occur where the conifer on which it specializes was more abundant (Norway spruce for common crossbills, Scots pine for parrot crossbills; Table 1; Figure 4), or in the case of the Scottish crossbill where there were more introduced conifers (spruce and other pines; Table 1; Figure 4) that it now currently relies upon (Summers & Broome, 2012).

3.3 Potential crossbill distributions

The predicted potential distributions for common and parrot crossbills were similar (8%–9% difference) to the observed distributions within the area included in the European Atlas (Table 2; Figure 5). However, the predicted distribution for the Scottish crossbill was around half of the area calculated for the Atlas grids (Table 2; Figure 5), which it is expected due to the over-estimation of the area calculated at the grid resolution of the Atlas (see a detailed survey in Summers & Buckland, 2011). Extrapolating to all of Europe, the potential distribution of common crossbills increased 2.5-fold (Table 2), occupying Iceland and most of Russia (Figure 5); the distribution of parrot crossbills increased by 2.6-fold (Table 2), potentially occurring in parts of Iceland and the northern part of Russia (Figure 5); and the potential distribution of Scottish crossbills increased 2.8-fold (Table 2), potentially occupying Great Britain south of Scotland and scattered areas mainly in mountain ranges in the north of the Iberian Peninsula (Figure 5).

3.4 Future crossbill distributions

Projections for the potential crossbill distributions under future (2070) climate change scenarios indicate a reduction in area for all three species, particularly for parrot crossbills (Table 2; Figure 6). The distribution of common crossbills was projected to expand in northern Russia, but decline overall by an average of 20% (Table 2) because of decreases in southern and central Europe (Figure 6). The future distribution for parrot crossbills contracted and shifted northwards in northern Fennoscandia and Russia (Figure 6) and declined overall by an average of 57% (Table 2). The future distribution for Scottish crossbill decreased on average by 18%, although one scenario predicted an increase of 6% (Table 2). Future projections predicted suitable areas remaining in Scotland, but also more distant areas in Iceland, along the Scandinavian coast and south-central Europe (mainly the Alps and the Pyrenees; Figure 6), areas the relatively sedentary Scottish crossbill is unlikely to colonize.

4.1 Climate factors influencing the phenology of seed fall

Our results are consistent with the important role of (spring) temperature for different phenological events of plants (Cleland, Chuine, Menzel, Mooney, & Schwartz, 2007; Gordo & Sanz, 2010; Sparks, Jeffree, & Jeffree, 2000). We found that higher spring (April) maximum temperatures were related to earlier seed fall in Scotland, in agreement with previous observations (Hannerz et al., 2002; Worthy, Law, & Hulme, 2006) and with the fact that cone scales spread apart, releasing the seeds from their base, in response to dry conditions and high temperatures (Dawson et al., 1997). Surprisingly, early spring precipitation was not correlated with the timing of seed fall in Scotland. However, higher spring (March) precipitation and lower spring (March and April) maximum temperatures were related to delayed seed fall across all three regions. Perhaps precipitation was less relevant to seed dispersal in Scotland because of its more consistently humid maritime climate. Consistently high humidity in Scotland might also account for why maximum spring temperatures contributed relatively little to the model for the distribution of Scottish crossbills (Figure 4).

4.2 Climate and food plants as determinants of crossbill distributions

Early spring (April) maximum temperature, which best explained the timing of Scots pine seed fall, and Scots pine cover contributed most to predicting the distribution of parrot crossbills (Figure 4). These results were consistent with the occurrence of parrot crossbills depending on Scots pine retaining seeds in its cones in late spring and into summer, and with specialization by parrot crossbills on Scots pine (Cramp & Perrins, 1994; Newton, 1972). In contrast, the distribution of the Scottish crossbill, which presumably evolved in isolated Scots pine forests in the British Isles (Knox, 1990; Nethersole-Thompson, 1975; Summers et al., 2010), was more related to other non-native pines and spruce cover than to Scots pine cover. This result is consistent with the now extensive use of introduced pine and spruce by Scottish crossbills (Summers & Broome, 2012). However, it is worth noting that the restricted distribution of Scottish crossbills and the coarse grid resolution of the Atlas data (Summers & Buckland, 2011) resulted in few occurrences for model building. As mentioned above, the minimal importance of spring temperature on the distribution of Scottish crossbills is perhaps related to the relatively humid marine climate in Scotland. Spruce cover was the main determinant, and April maximum temperature the second-most important determinant of the distribution of common crossbills. This crossbill is widely distributed in Europe and feeds mainly on spruce in central and northern Europe (Cramp & Perrins, 1994; Newton, 1972). The importance of April maximum temperature suggests that seed retention in the spring by Norway spruce is related to temperature as in Scots pine.

4.3 Ranges of crossbills under future climates

Climate change scenarios are expected to shift and reduce the distribution of crossbills in Europe, with a dramatic reduction for the species most specialized on Scots pine, the parrot crossbill. These reductions are the result of at least two factors. First, warming conditions during spring are expected to shift the timing of seed fall so as to increase the window of food scarcity for crossbills (Figure 1). Although earlier cone development in summer with increasing temperatures could act to reduce this window of food scarcity, this is an unlikely response given that Scots pine cones start to develop and become profitable to crossbills at similar times from southern to northern Europe (late July-early August; Marquiss & Rae, 2002; Worthy et al., 2006; E. T. Mezquida, pers. obs.). Second, climate scenarios predict a north-eastwards shift in the distributions of the main food plants (pine and spruce) for crossbills, as previously predicted for many European forest trees (Huntley et al., 1995; Thuiller, 2003). The common crossbill's distribution will contract in the central and southern areas as warmer conditions advance seed fall and thereby expand the period of seed scarcity, and as conifers suffer increased mortality due to drier conditions at their warm-edge range margins (Matías & Jump, 2012; Reich et al., 2015). New areas will potentially become suitable in northern Europe allowing northwards expansion. Parrot crossbills will be further restricted to perhaps only northern Fennoscandia and Russia. This range reduction is the result of future range shifts in Scots pine (Huntley, 1995) and the increase in spring temperature causing a longer interval of seed scarcity, especially in southern Fennoscandia.

Our predictions for Scottish crossbills are partly similar to those from previous simulations (Huntley, Green, Collingham, & Willis, 2007). Suitable areas in Scotland are predicted to remain, but additional areas are predicted for Iceland, Scandinavia and south-central Europe. However, it is doubtful that these small and widely spaced locations would be favourable to a relatively sedentary and localized specialist, as Scottish crossbills have been characterized (Knox, 1990; Marquiss & Rae, 2002; Nethersole-Thompson, 1975). Currently, Scottish crossbills often utilize non-native conifers (Summers & Broome, 2012). Although the use of these non-native conifers should aid in their short-term persistence, a much smaller billed crossbill similar to those specialized on these conifers (e.g., Pinus contorta and Picea sitchensis) in their native ranges in North America (Benkman, 1993; Irwin, 2010) will be favoured over the long term (Summers & Broome, 2012). In addition, common crossbills, which also feed on these non-native conifers in Scotland, could co-occur with Scottish crossbills more frequently, potentially increasing both competition and hybridization between this smaller billed crossbill and the Scottish crossbill (Marquiss & Rae, 2002; Summers, Dawson, & Phillips, 2007). We suspect that the outlook for a unique Scots pine specialist in Scotland is even bleaker than that for the parrot crossbill in northern Europe.

These forecasted range shifts for crossbills are consistent with future predictions for different bird species and diverse local scales (Thomas & Lennon, 1999; Tingley, Monahan, Beissinger, & Moritz, 2009) and with avian responses to changes in vegetation during past climates (Holm & Svenning, 2014). Our results suggest that climate change will impact European crossbills in an indirect way through their influence on the distribution of their food plants (Huntley, 1995; Kissling, Field, & Böhning-Gaese, 2008), timing of seed fall and seed availability (Benkman, 2016; Santisteban et al., 2012), and possibly seed production (Matías & Jump, 2012). This is consistent with increasing evidence that climate change has and will have its greatest impact by altering food resources and trophic interactions (Cahill et al., 2013; Ockendon et al., 2014; Pearce-Higgins & Green, 2014).