Numerous examples show how predicted climate change will affect land suitability and crop yields; and how increased frequency and intensity of extreme weather events can increase fluctuations in crop yields (Schmidhuber & Tubiello, 2007; Lobell et al., 2008; Challinor et al., 2010; Hawkins et al., 2013). Minimizing the negative effects that these fluctuations have on food security, requires an understanding of the effects of climate change on the resilience of crop growth and yield (Fraser et al., 2013) and, where relevant, of crop-pollinator systems. Eighty-seven of the leading global food crops, accounting for about 35% of the global production, benefit from animal pollination, in particular from insects (Klein et al., 2007); yet, pollination is rarely accounted for in projections of the impacts of climate change on crop yields.
The importance of insect pollinators has been valued at €153 billion to agricultural production (Gallai et al., 2009) worldwide. The documented declines in insect pollinators (Biesmeijer et al., 2006; Potts et al., 2010; Carvalheiro et al., 2013) may therefore threaten food security, which in turn may lead to an increased demand for agricultural land (Aizen et al., 2009). Climate change, together with land-use intensification and the spread of alien species and diseases, is one of the main anthropogenic pressures on insect pollinators (Vanbergen & The Insect Pollinators Initiative, 2013). Several authors have investigated how climate change can affect plant-pollinator interactions (e.g. Devoto et al., 2007; Memmott et al., 2007). Recent work by Kuhlmann et al. (2012), for instance, has highlighted how climate-induced range shifts of dominant bee species are likely to affect specialized plant-pollinators mutualisms, with negative consequences for the reproduction of these plants. Other studies have looked at specific crop-pollinator systems, examining the cost of replacing pollination service (Allsopp et al., 2008), the predicted decline in suitable bee habitat due to climate change (Giannini et al., 2012), and the potential environmental suitability of nearby regions allowing the persistence of crop-pollinators mutualisms (Giannini et al., 2010, 2011).
Understanding how climate change will affect crop-pollinator systems helps highlight areas potentially vulnerable to pollinator shortage, or predict potential geographical mismatches between crops and their pollinators. Similarly, such information can be used to identify areas suitable for the persistence of both the crop and its wild pollinators, or direct local interventions to boost pollination service, thereby strengthening food security.
In this study, we examined the impact of projected climate change on the distribution and likelihood of occurrence of commercial orchards and their pollinators in Great Britain (GB). Commercial orchards (orchards hereafter) occupy about 19 000 ha in UK, which equates to nearly 13% of the total area dedicated to the production of fruit and vegetables. Of the area used for fruit production, orchards occupy 65.5%, with the remainder used for the production of soft fruits, such as strawberries. The major products of orchards in the UK are apples, pears, plums and cherries, with dessert apples alone accounting for nearly 28% of the planted area and having a net value of £70 million annually (Department for Environment, Food & Rural Affairs, 2013). Bees and hoverflies are the predominant pollen vectors for these plants (Klein et al., 2007). To ensure marketable fruits, therefore, it is important that the activity of pollinators overlaps spatially and temporally with the flowering period of the fruit trees.
Here, we used the projected climate for 2050 from the SRES A1B Emissions Scenario (Nakićenović et al., 2000) to estimate the nation-wide environmental suitability for orchard species and their pollinators. Recent trajectories of greenhouse gas emissions are higher than those considered under the SRES A1B (Raupach et al., 2007), so the projections used here may be interpreted as conservative estimates of climate change. We used present day distributional data to characterize the climatic conditions most favourable to the orchard species and to their pollinators. We excluded land-cover/land-use information, since its distribution cannot readily be predicted with sufficient certainty beyond the current time. We then projected the potential distribution of orchards and pollinators, on the basis of the climate projected for 2050, and derived a relative measure of pollination service available to orchards for present and future conditions. This approach allowed us to highlight changes in environmental suitability which may threaten the persistence of the orchard-pollinator system, and identify geographical mismatches between orchards and pollinators potentially affecting pollination service provision. We used this information to suggest appropriate intervention measures which could be used to mitigate against future loss of pollination services to orchards.
Materials and methods
We used the species distribution model (SDM) Maxent, (Phillips et al., 2006; Aguirre-Gutiérrez et al., 2013; Polce et al., 2013) to predict the distributions of orchards and pollinators in relation to climatic conditions. Pollinator and orchard data were collected using different methodologies, thus requiring different modelling approaches; we describe these datasets in the following subsections ‘Pollinator data’ and ‘Orchard data’ respectively. Details of the climate data used to characterize the environmental space of orchards and pollinators are provided in the subsection ‘Climate data’. We modelled pollinator species individually, using available records to predict current potential distribution and then future projections. We modelled future distribution of orchard species as an entire category since information on orchard composition was not available from the current agricultural survey data. The model settings for pollinator species and orchards are described separately, under ‘Distribution models’. We then describe how we used the outputs from the SDMs to identify:
- The climatic predictors that contributed most to the final models (subsection ‘Contribution of predictors’).
- How the climatic predictors for the pollinators and orchards are projected to change in 2050 (subsection ‘Similarities between current and predicted climate’).
- An index of relative pollinator availability to orchards, which we used as a proxy for potential pollination service (subsection ‘Pollinator availability’).
We used presence-only sightings of wild bees and hoverflies recorded in GB within the period 2000–2010 (‘Bees, Wasps and Ants Recording Society’, BWARS, http://www.bwars.com/; ‘Hoverfly Recording Scheme’, HRS, http://www.hoverfly.org.uk/). On the basis of literature (Free, 1993; Marini et al., 2012) and knowledge gained by our team members during the past years of pollinator-related field projects (e.g. http://www.reading.ac.uk/caer/Project_IPI_Crops/project_ipi_crops_index.html; http://www.step-project.net/; http://www.alarmproject.net/alarm/; all accessed February 2014), we selected 22 species of wild bees and 8 species of hoverflies, known to be visitors of fruit trees and therefore potential pollinators of orchard crop flowers in GB. The spatial resolution of the records varied from 100 m2 to 4 km2. We aggregated all sightings for each species on the 25 km2 grid (5 by 5 km cells) and removed any duplicate records, so that for each species, there was at the most one entry per grid cell. The number of available records per bee species ranged from 26 to 2096 (mean ± SD = 650 ± 580; median = 471); records per hoverfly species ranged from 150 to 1981 (mean ± SD = 1032 ± 616; median = 1033). Pollinator species and numbers of records are listed in the Data S1 (Table S1 in Data S1).
The current distribution of orchards was derived from the 2010 Defra June Agricultural Survey (http://www.defra.gov.uk/statistics/foodfarm/landuselivestock/junesurvey/junesurveyresults/). Orchards included areas of at least 1 ha, planted with top fruit such as apples, cherries, pears, plums and nuts (walnuts and hazelnuts mainly); their distribution was originally mapped on a grid of 2 × 2 km cells (‘tetrads’) and included information on the extent of the orchards within each tetrad. We superimposed a 5 × 5 km grid onto the crop tetrads, proportionally allocating each tetrad's orchard extent to the overlapping 5 km grid cell(s). The final extent of the orchards within each cell was the sum of the proportional extent from all tetrads intersecting the cell. Of the 9726 grid cells used to represent GB, around 14% contained orchards (1354), with a total mapped extent greater than 12 200 ha. The difference between the Defra figures for orchards and the actual mapped hectares are due to insufficient spatial information for some of the orchard fields to be mapped.
We used total annual precipitation and monthly minimum and maximum temperature to derive a set of environmental descriptors commonly used in species distribution models (e.g. Hijmans & Graham, 2006; Warren et al., 2010; Warren et al., 2013; Wolmarans et al., 2010): growing degree days greater than 5 °C (GDD5, used only for crop), calculated following Sork et al. (2010); and 19 bioclimatic variables (Hijmans et al., 2005, 2011). This choice reflected the need to satisfy two main criteria: the same predictors needed to be available for both the present and the future projections; they needed to be relevant for the modelled group. The three input variables were obtained from UKCP09 (http://www.metoffice.gov.uk/climatechange/science/monitoring/ukcp09/). Baseline data for pollinator distribution models were made of the 25 km2 gridded monthly averages for the decade 1990–1999, while for the orchard distribution model, we used gridded data for the 30 year period 1977–2006 (the most recent available complete 30 year period). We used a longer time series for orchard crops, to reflect the longer life cycle of fruit trees compared to insect pollinators. Future projections of monthly averages were derived from the UKCP09 projections (Murphy et al., 2010) for the SRES A1B storyline (‘Medium’ Emissions Scenario, as referred in the UKCP09 report). We used the 30 year period from 2040 to 2069; we will refer to the baseline data as the ‘Present’ and to the future projections as the ‘M2050’. These data are located on a rotated-pole grid with a spatial resolution of approximately 25 by 25 km. We rescaled them to the 5 × 5 km British National Grid, to match the orientation and resolution of the baseline data. Additional information on this dataset and details of the rescaling procedure are provided in the Data S1 (Data S1, Material and Methods, ‘Climate data for future projections’).
All the variables were computed within R (R Development Core Team, 2011). To minimize colinearity between predictors (Guisan & Thuiller, 2005), subsets were created for pollinators and orchards. For pollinators, due to lack of a general set of commonly used variables, we used Jolliffe's Principal Component Analysis with the rejection method ‘B2’ and λ0 = 0.70 (Jolliffe, 1972, 1973); we reduced the original set to six predictors (Table 1). For orchards, we based the choice on literature (Thuiller, 2004; Termansen et al., 2006; Sork et al., 2010; Franklin et al., 2013; Warren et al., 2013), and we selected five predictors (Table 1). The Pearson's correlation between the selected crop and pollinator variables is reported in the Data S1 (Data S1, Material and Methods, ‘Correlation between selected climatic predictors’, Tables S2 to S5).
Table 1. Variables used for orchards and pollinators distribution modelling. The table shows the final set of predictors used to model the distribution of orchards (ODM) and pollinators (PDM). The selection of predictors was based on several criteria, including their use in published literature and minimizing multicollinearity
|Bio04||Temperature Seasonality (SD × 100)||TSeasSD||ODM|
|Bio06||Min Temperature of Coldest Month||mTCM||ODM|
|Bio07||Temperature annual range||TAR||PDM|
|Bio08||Mean Temperature of Wettest Quarter||MTWQ||ODM|
|Bio09||Mean temperature of driest quarter||MTDQ||PDM; ODM|
|Bio11||Mean temperature of coldest quarter||MTCQ||PDM|
|Bio15||Precipitation seasonality (Coefficient of variation)||RainSeasCV||PDM|
|Bio18||Precipitation of Warmest Quarter||RainWQ||ODM|
|Bio19||Precipitation of coldest quarter||RainCQ||PDM|
Pollinator distribution models
Detailed settings for the Maxent pollinators’ distribution models (PDM) follow Polce et al. (2013) and are summarized in the Data S1 (Data S1, Material and Methods, ‘Pollinator distribution model’). Model training and testing was performed through 10-fold cross-validation, and ‘10th percentile of training presence’ was used as a threshold to convert probability of occurrence into binary predictions (‘presence-absence maps’). We chose this threshold since it retains as suitable environmental conditions, those characterizing 90% of the training locations, thus excluding records that were found at the extreme of the species’ suitable environment. We assumed unlimited dispersal capability for each species, but we restricted the predicted presence to areas where all 10 model runs had predicted ‘presence’; we assigned average probability of occurrence (p) to these areas, and ‘absence’ outside them.
We assessed the models using the Area Under the Receiver Operating Characteristic Curve (AUC), which, despite known assumptions and limitations (Termansen et al., 2006; Austin, 2007), is commonly used as a threshold-independent measure of model performance within SDMs. With presence-only data such as the pollinators’ sightings, the maximum achievable AUC is <1 (Wiley et al., 2003) so standard thresholds for evaluating goodness of fit do not apply. Instead, we followed Raes & ter Steege (2007) and we compared the average AUC value of each species PDM (AUCPDM) with the average AUC value of a set of null models (AUCNM) where species records were replaced by randomly chosen locations. We expected AUCPDM > AUCNM.
Orchard distribution models
After running Maxent models using different feature classes (i.e. including different possible relationships between species data and climate variables from linear to hinge to quadratic) to predict present orchards distributions, we retained the models that used hinge features, which were then used to predict orchards’ future probability of occurrence. For the orchard distribution model (ODM), Maxent was trained on 75% sample points, and the remainder was used for testing. This procedure was repeated 10 times. We used areas where at least seven model runs had predicted presence (based on the ‘10th percentile of training presence’), to indicate suitable conditions for crop growing under the M2050 scenario, and assigned to these areas the average probability of occurrence obtained from the 10 model runs. We used a more relaxed criterion for orchards than pollinators (7 vs. 10 model runs to indicate presence), to account for the fact that orchards are a managed resource and so can overcome some of the barriers which would prevent colonization and establishment of wild organisms such as the pollinators that were modelled.
Contribution of different predictors to distribution models
The contribution of each predictor to the final PDMs and ODM was derived from the drop in AUC observed after permuting the values of each variable with those of the background, with larger drops indicating that the model depended heavily on that variable (Data S1, Material and Methods, ‘Contribution of different predictors’). Average and confidence interval for the observed drops were derived through 10 000 bootstrap replicates.
We used a linear mixed effects model (Pinheiro et al., 2013) to test if the contribution of different predictors differed between pollinator species and/or model runs. We used predictors as fixed factors and model run as a random factor. Model run was nested within species and group (bees or hoverflies) when analysing the results from the PDM. Multiple pairwise comparisons of different predictors were then performed using Tukey's post hoc test for a general linear hypothesis (Hothorn et al., 2013).
Similarities between current and predicted climate
The PDM and ODM were required to predict conditions not sampled in training data. Computing the similarity between conditions at training locations and conditions where predictions are to be made can be done within Maxent, through Multivariate Environmental Similarity Surfaces (MESS) (Elith et al., 2010). MESS measure the similarity of any given point to a set of reference points, for each model predictor. The lowest similarity obtained for that point is used as the point's MESS. Negative values indicate conditions that are outside the range of references values, while positive values indicate greater similarity to the set of reference points, with 100 assigned to a point not novel at all (i.e. having a predicted value within the range of reference points). In addition to mapping the MESS across the region of interest (GB), an accompanying map also showed, at any given location, the variable that drove the MESS. We used these two maps to spatially quantify predicted climatic changes.
We used pollinator availability (PA) as a proxy for pollination service. Pollinator availability resulted from the contribution of each species probability of occurrence predicted by the Maxent model. We assumed that all pollinator species are equally efficient in pollinating orchard flowers. For each grid cell, where the presence of orchards was predicted or observed, PA was defined as:
Where: PAm = pollinator availability to the orchards in grid cell m, resulting from all pollinator species; psm = Maxent probability of occurrence for species s on cell m; S = total number of pollinator species. Eqn (1) is loosely based on Lonsdorf et al.(2009) and Polce et al. (2013), with the main difference being that the weighted term allowing pollinators to reach an orchard located on neighbouring cells is excluded, as the model resolution is coarser than the typical pollinator foraging distance.
We used Eqn (1) to derive the PA where:
- Orchards are currently present: the baseline PA;
- Orchards are predicted to occur based on the M2050 future scenario: to assess whether the most suitable areas for fruit trees are also suitable for their pollinators;
- Orchards are currently present, but climatic conditions are those predicted for M2050: to derive the difference in PA between the present and the M2050 scenario, assuming that fruit trees can continue to persist where currently present, despite changes in climatic conditions.
It has been widely documented that shifts in species ranges are correlated with climatic change (Parmesan & Yohe, 2003; Root et al., 2003; Chen et al., 2011). By acting on the distribution and survival of single species, climate change is likely to affect ecosystem functions and, as a result, the provisioning of ecosystem services. In this study, we examined the potential consequences of climate change for pollination services as provided by an array of 22 bee and 8 hoverfly species known to be frequent visitors of orchards in Great Britain. Changes in climatic conditions can affect plant-pollinator interaction networks in several ways (Hegland et al., 2009), for instance by causing phenological mismatches (Burkle et al., 2013) or spatial mismatches (Schweiger et al., 2008). Here, we focused on potential geographical mismatches, and showed that under future climate scenario, suitable conditions for orchards and orchard pollinators may not overlap, threatening pollination service. In particular, we examined the potential distribution of pollinators and orchard species grown in GB, based on the SRES A1B Emissions Scenario climatic projections. We used a relative measure of pollinator availability as an indication of potential pollination service, since quantifying service delivery in absence of pollinator abundance data cannot be done. Using the environmental suitability for wild pollinators, as a relative measure of potential pollination service is a commonly adopted approach (Lonsdorf et al., 2009; Lautenbach et al., 2011; Polce et al., 2013; Zulian et al., 2013), in absence of sufficient information to parameterize the relation between pollinator availability and yield.
Pollinator species differ in their efficacy to pollinate flowers (e.g. Bischoff et al., 2013; Castro et al., 2013; Garratt et al., 2014). In addition, altered phonologies due to climate change may result in temporal mismatches between the availability of the most effective pollinators and the onset of flowering, with potential negative consequences on plant reproduction success (Rafferty & Ives, 2012) and service provision. Although approaches to estimate and compare pollinators’ performances have been discussed (Ne'eman et al., 2010; King et al., 2013), there remain practical and theoretical difficulties to apply the proposed methods over large geographical regions and to many pollinator species. Thus, in deriving pollinator availability, we did not take into account species’ identity and we assumed all species be equally efficient in pollinating orchard flowers.
We chose orchards as they represent a major GB fruit crop, and they include top fruits of global economic importance, such as apples. The distribution of orchards is limited to locations having suitable soils, climate and socio-economic conditions. For examples, apples and other fruits trees are known to be vulnerable to frost occurring during bloom stage: the projected climate warming, therefore, raises concerns that the bloom stage might advance in time and coincide with periods where frost spells can happen, thus threatening the quality and possibly the production of fruits (Eccel et al., 2009). Of the predictors used to model orchard distribution in our study, Temperature Seasonality presented the greatest mismatch between present and future, with projections shifted towards greater variability. We cannot assert at this stage that this will directly threaten fruit production, but greater variability may increase the risk of sharp temperature variations, such as the occurrence of frost spells in periods otherwise characterized by milder temperatures, e.g. during flowering. In addition, fruit trees benefit from bud dormancy, which is triggered by a period of exposure to cold weather: the predicted rise in Minimum Temperature of Coldest Month (mTCM) could interfere with the fulfilment of the chill hours per year, potentially affecting the production of leaves, flowers and subsequently fruit. Inspecting Maxent models built exclusively with this variable provide some support to this hypothesis. While current mTCM is most commonly between 1 and 2 °C, this will increase to 4 °C, with peaks up to 8 °C, in M2050.
Of the predictors used to model pollinator distribution, Mean Temperature of the Driest Quarter was the one with the greatest shift, both in terms of mean and shape; but for the majority of the country none of the predictors moved outside the present climatic range. In interpreting this pattern, we must stress that change only refers to the range of values of the predictors, and not to their geographical location; in other words, a location will be mapped as ‘No change’ if the projected values for all predictors have changed, but they have all remained within the range of values observed for present time. The results from the PDMs projected that locations with greatest pollinator richness would shift north-east, suggesting a similar shift in environmental conditions most suitable to pollinators. For Europe, Ohlemüller et al. (2006) have already shown a prevailing north-east shift in the climatic conditions analogous to the 1931–1960 period. For much of the global land shifting climate has been projected to be greater than 1 km yr−1 over the 21st century (Diffenbaugh & Field, 2013), potentially posing alarming challenges for terrestrial ecosystems. The results of the PDMs assume unlimited dispersal of pollinators and predicted range expansion to occur more frequently than range contraction. Indeed, some species are likely to track such changes. The bumblebee, Bombus hypnorum, arrived in SE England less than 15 years ago and since then has reached Scotland. However, if areas of similar climate are farther than the species’ dispersal distance, colonization and persistence may not be possible (Thuiller, 2004; Ohlemüller et al., 2006), and more species would shrink their range. This risk would be further enhanced by other pressures acting on the pollinators, such as habitat fragmentation and degradation, parasites and alien species (Vanbergen & The Insect Pollinators Initiative, 2013), none of which was considered here. Looking at the species that are already predicted to experience range contraction, some of them, like Osmia and Bombus spp., are known to be efficient pollinators of orchard trees, and of apples in particular (Delaplane & Mayer, 2000). Therefore, geographical mismatches between these species and orchards might have noticeable effects on pollination service provision. There could be expansion of orchard pollinators from the continent into GB, although this element was not included in our study; there could also be additional pollination supply from managed pollinators (e.g. honeybees), although the capacity to utilize honeybees for additional pollination services is primarily independent of climate.
Solely based on climatic projections, the most suitable environmental conditions for orchards shifted north-west, although probability of occurrence for these areas never reached the maxima obtained for the present. Since our projections were only based on climate, however, they must be read with caution: much of the areas identified as suitable for orchards in M2050 occur in uplands that may not be suitable for fruit tree cultivation owing to soil type and topography. In addition, the pollinator availability predicted for these areas was for the vast majority ≤0.2, in contrast with present predictions which showed larger areas with 0.2 ≤ PA ≤ 0.5. These results suggest that, over the next 50 years, the most suitable areas for orchards may not be characterized by pollinator availability as high as now. Furthermore, while the present distribution of orchards largely overlapped areas with the highest pollinator richness, future predictions showed a geographical mismatch between areas most suitable to orchards and areas richest in pollinator species. Pollinator diversity has been observed to increase fruit set in several crop systems (Klein et al., 2003; Hoehn et al., 2008; Garibaldi et al., 2013) and buffer negative effects of extreme weather events such as strong winds (Brittain et al., 2013). Adequate pollination could still be possible by a few species of wild bees with high numbers of individuals, but such a community would be more vulnerable to stressors and stochastic variation. Landscape management to increase pollinator diversity and abundance in these areas of future orchard production could be implemented to improve the stability of pollination services, such as preservation of seminatural landscapes or increasing pollinator habitat and forage resources (Ricketts et al., 2008; Scheper et al., 2013), or additional inputs from managed pollinators might become necessary to achieve optimal yields.
In contrast, the areas currently occupied by orchards are predicted to become even more suitable to pollinators in M2050. Under this scenario, however, due to unfavourable conditions the predicted probability of the occurrence of orchards will decrease. New top fruit varieties could be developed with future climatic conditions in mind, particularly breeding for resistance to those factors identified in this study as key to driving the shift in orchard distribution, namely Temperature Seasonality and Minimum Temperature of the Coldest Month.
In this study, we have used species distribution models and climate projections to derive the environmental suitability for the orchard-pollinator system in Great Britain, under different scenarios. Due to the characteristics of the pollinator data, we used a relative measure of pollinator availability which cannot (yet) be translated into units of service delivery (Maes et al., 2012). Our approach, however, detected a geographical mismatch in climatic suitability for orchards and pollinators, which may potentially lead to low pollination service provision, unless production is moved towards more (climatically) suitable north-westerly areas. However, we found that wild pollinator availability may be preserved and possibly enhanced in areas already used for orchards. The implications of trading off between wild pollinator availability and lower climatic suitability need further research. In particular, methods of boosting wild pollinators through improving landscape resources (Scheper et al., 2013), supplementing wild pollination service with managed pollinators, or choosing fruit tree varieties that are adapted to changed climatic conditions may provide a combination of adaptation options to support top fruit production in GB over the next 50 years. The methods underlying our study can be applied to other regions and crop systems, and expanded to include different climatic scenarios. Some of the most urgent challenges that need to be addressed, are the inclusion of other factors limiting future crop cultivation (e.g. soil type), and the translation of the relative measure of pollinator availability into units of service delivery.