Expected Changes to Alpine Pastures in Extent and Composition under Future Climate Conditions

: As the basis of livestock feeding and related performances, pastures evolution and dynamics need to be carefully monitored and assessed, particularly in the Alps where the e ﬀ ects of land abandonment are further ampliﬁed by climate change. As such, increases in temperature associated with changes in precipitation patterns and quantity are leading to modiﬁcations of grassland extent and composition with consequences on the pastoral systems. This study applied a machine learning approach (Random Forest) and GIS techniques to map the suitability of seven pasture macro types most representative of the Italian Alps and simulated the impact of climate change on their dynamics according to two future scenarios (RCP4.5, 8.5), two time-slices (2011–2040, 2041–2070), and three RCMs (Aladin, CMCC, ICTP). Results indicated that (i) the methodology was robust to map the current suitability of pasture macro types (mean accuracy classiﬁcation = 98.7%), so as to predict the expected alterations due to climate change; (ii) future climate will likely reduce current extend of suitable pasture ( − 30% on average) and composition, especially for most niche ecosystems (i.e., pastures dominated by Carex ﬁrma and Festuca gr. Rubra ); (iii) areas suited to hardier but less palatable pastures (i.e., dominated by Nardus stricta and xeric species) will expand over the Alps in the near future. These impacts will likely determine risks for biodiversity loss and decreases of pastoral values for livestock feeding, both pivotal aspects for maintaining the viability and proﬁtability of the Alpine pastoral system as a whole.


Introduction
The Alpine chain is about 1200 km long and 200 km wide, stretching across eight European countries (i.e., Austria, France, Germany, Italy, Switzerland, Liechtenstein, Slovenia, and Monaco). The location of the Alps (frontal system crossing Europe in a west-east direction), beside the wide variation in exposure, elevation, and its arc-like shape determine a unique climate which also depends on local differences and position of reliefs [1]. This in turn determines the presence of a large variety of habitats and species. Within the whole area, grasslands and pastures represent about 25% of the Alpine vegetation, most of them semi-natural after centuries of human activities [2]. Beside production, these ecosystems are widely recognized as providing several socio-economic services for people living both in mountains and urban areas [3][4][5]. These include cultural services such as recreation and tourism;

Methodology
The methodology applied in this study combines the use of pre-existent environmental (topography, soil) and climate spatial layers as drivers of the distribution of the main pasture macro-types dominating the region, without considering the influence of human activity. Specifically, these spatial layers were used as predictor variables of occurrence of different macro-types by using a machine learning approach (Breiman and Cutler's Random Forests for Classification and Regression-RF, [39]). RF was first calibrated and validated for the present period, so as to test its ability to faithfully predict the to-date spatial distribution of both pasturelands and the relevant macro-types. The calibrated RF was then applied to reproduce the dynamics of pasture macro-types distribution over the Italian Alps region according to RCP 4.5 and 8.5, using three regional circulation models (RCMs) for two future time-slices . The calibration and validation strategy of RF model is described as well as the ensemble approach used to synthesize the outputs for future periods obtained from three RCMs. In the following sections, a brief description of the spatial dataset used is reported, along with the relevant spatial scales and harmonization procedures adopted.

Predictor Variables: Topographic, Soil, and Climate Dataset
The complexity in estimating current pasture macro-types and their evolution under future climate conditions over the study area was addressed by integrating pre-existing datasets on pasture vegetation with spatial information on soil, topography, and climate. Elevation, slope, and pH were extracted from the Harmonized World Soil Database-HWSD [40], while monthly climate data (Tmin, Tmax, and cumulated rainfall) were provided by the WorldClim database [41] for 1970-2000 as baseline. All these datasets were available at the same spatial resolution, approximately 1 × 1 km (see [42] for further details).
The outputs of three regional circulation models (RCMs) were empirically downscaled using a delta change approach over the WorldClim gridded baseline climate  to provide informative layers for RF application in RCP4.5 and RCP8.5 future scenarios. Specifically, monthly minimum and maximum temperature (Tmin and Tmax, • C), and cumulated rainfall (Rain, mm), as simulated by CNRM-ALADIN (spatial resolution 0.11 • × 0.11 • ), ICTP-RGCM4 (0.44 • × 0.44 • ), and CMCC-CCLM4 (0.44 • × 0.44 • ) from MedCordex [43], were averaged on a monthly basis for the relevant reference period 1970-2000 and for two time-slices: 2011-2040 and 2041-2070 for RCP4.5 and 8.5 scenarios. The absolute monthly differences between each future time-slice and the reference period computed for Tmin and Tmax was then overlaid over the relevant months of the WorldClim dataset. The same Agronomy 2020, 10, 926 4 of 21 approach was used for rainfall, where the change was expressed as the ratio between each future time-slice and the reference period. The final climatic dataset consisted of twelve combinations of Tmin in January, Tmax in July, and seasonal rainfall calculated for three RCMs (CNRM-ALADIN, ICTP-RGCM4, CMCC-CCLM4) two-time slices (2011-2040 and 2041-2070, two RCPs (RCP4.5 and 8.5).
Both datasets (the unclassified pasturelands and seven macro-types) were converted into two grid datasets, spatially consistent with the predictor variables dataset (1 × 1 km spatial resolution, EPSG: 4326). Specifically, we created one dataset related to the overall distribution of pasturelands (1), and a second one related to the distribution of the seven pasture macro-types (2). The first (1), presence/absence of pasturelands, was determined by aggregating those pixels classified as pastures in the CORINE land cover map with those in which at least one macro-type occurred (presence of pasturelands). Other land uses, potentially suited to the seven pasture macro-types, such as agricultural and forested areas (Class 2 and Class 3 of level 1 of CORINE land cover map, respectively) and wetlands (Class 4 level 1), were also incorporated into the spatial dataset as "absence" response variable. The second dataset (2), macro-type dataset, was created by calculating the coverage of each pasture macro-types within the same 1 × 1 km grid cell. The presence of each macro-type was counted when covering at least 10% of the total area within each 1 × 1 km pixel, assigning the most prevalent macro-type (in terms of coverage) to each pixel.
Finally, both datasets, i.e., pasturelands (1) and pasture macro-type (2) were harmonized with the predictor variables dataset (topography, soil and climate) into a coherent comprehensive dataset (49,405 pixels in total), including: (i) the predictive abiotic independent variables (eight in total); (ii) the presence of pasture macro-types (seven response variables); (iii) the presence/absence of unclassified pasturelands; and (iv) the presence/absence of other land uses. The final dataset comprised 19,748 pixels of actual occurrence of pastoral resources, namely 8866 of unclassified pasturelands, 10,882 of classified pastures (pasture macro-types), and 29,657 of other land use (Table 1). Occurrence probabilities for unclassified pasturelands and seven pastoral macro-types was modeled using the Random Forest (RF) machine learning algorithm [39], as implemented in R environment (package "randomForest"). RF is a classification and regression ensemble of decision trees classifiers, commonly used in studies on species distribution and changes analysis [42,[50][51][52][53], acknowledged as giving the most accurate performances in predicting species distributions using a large set of independent variables [54][55][56] also over unbalanced datasets [39]. RF is able to make internal cross-validated accuracy estimates, as each classifier is generated by a bootstrap sample randomly divided into an internal training (66% of data) and testing subset (33% out-of-bag sample, OOB error), thus giving cross-internal classification error estimates. RF predictions are made giving a majority vote to the ensemble of classification trees and the final model prediction is generated taking the most voted model over all the trees generated within the forest. The response variable of RF is binary, i.e., presence (1) and absence (0), and the model provides, as output, the probability scores from model application. Additionally, RF provides, as final product from the calibration mode, the mean decrease accuracy, which puts in evidence the relative importance of each predicting variable of the model in distinguishing macro-types in the calibration phase, as well as the marginal effect of each variable on model prediction (partial dependence plots).
As RF predictions might be biased by unbalanced datasets (proportion between presence and absence classes), likely resulting in over-prediction of the majority-class, a multi-step strategy for RF calibration was adopted. First, the extent of pasturelands was determined by coupling the geographical extent of all macro-types [31] with the unclassified pastures (Corine land cover map) for training RF. Second, within the area identified by RF as suitable for pasturelands at this stage, the most suited macro-type was determined on a pixel-by-pixel basis. Finally, ten sub-sampling datasets were randomly down-sampled from both datasets (pasturelands and macro-type dataset), so as to have a proportion of the majority class (presence of pastureland and/or macro-type) consistent (i.e., with the same number of occurrences) with the minority-class (other land uses) to be used for RF calibration.
Separately, for each unclassified pastureland and seven pasture macro-types dataset, the ten random sub-samples of RF models were then combined into a final model and thus applied over both datasets against abiotic independent variables and current climate. The mtry parameter of RF was set to 3, as derived by the number of input predictive variables (in this case 8), while the number of trees (ntree parameter) was set to 100 [58] for each dataset. The accuracy of simulations was assessed against a validation dataset, effectively sub-sampled, through the OOB error calculation. The probability of occurrence for the eight classes (seven pasture macro-types and unclassified pasturelands), as resulting from RF simulations, was converted into presence/absence computing the true skill statistic index (TSS, which avoids an overestimation of those categories that are mostly represented, [59]) maximizing the prediction accuracy of the model. The TSS threshold was then applied to all validation subsets to test the classification error of the proposed threshold. Accordingly, the RF models were applied over the entire dataset to predict site-level probability of occurrence of unclassified pasturelands and pasture macro-types over the entire Italian Alpine chain in the current time period and under RCP 4.5 and 8.5 future climate scenarios in the near (2011-2040) and distant future (2041-2070), as projected by CNRM-ALADIN, ICTP-RGCM4, and CMCC-CCLM4 RCM models.
Tmax_jul, Tmin_jan, Prec_djf, Prec_jja, Prec_mam, and Prec_son of the selected RCMs calculated for both scenarios and time slices were tested for significance with respect to the relevant current data. For this test, a delta change of these variables was calculated with respect to baseline on a pixel basis for each macro type, and the null hypothesis that these samples were different from zero was tested. As a result, we obtained a total of 504 test cases corresponding to 3 RCMs × 2 scenarios × 2 time slices × 7 macro-types × 6 variables (Table S1).
The final maps were analyzed as an ensemble of the three RCMs outputs, by ascribing the pasture macro-type resulting more frequent/prevalent (i.e., simulated as present by at least two RCMs within each pixel) among the three climatic models and where at least one RCM simulated the presence of unclassified pasturelands. Results were described for the entire Alpine chain and for two main regions, namely Western and Eastern Alps (as characterized by [31]), when relevant. Tables and Figures related to the Western and Eastern area are reported in the SI.

RF Calibration and Classification Accuracy
RF classification accuracy was estimated by comparing simulated presence/absence of macro-types against 10,882 grid points of the validation dataset over the climatic baseline period . The cross-calibration ( Table 2) between observed and simulated pastures macro-types distribution over the Alpine chain for the present period was shown to be robust, with an overall mean accuracy of 98.7%. The lowest performance was observed for Festuca gr. rubra (94.2%), while the highest for Nardus stricta (99.2%). The TSS, calculated for each single macro-type, resulted in any case higher than 0.95 (data not shown). The relative importance of every predicting variable in the classification process was determined via the permutation of each variable in the calibration process, and testing the relevant accuracy of the results (mean decrease accuracy, MDA). The more important a variable is in the accuracy of the prediction, the higher is the relevant MDA ( Figure 1). The results indicated that the most important variables were maximum air temperature of July (Tmax_jul), followed by soil pH. By contrast, slope and spring precipitation (Prec_mam) yielded the lowest importance as drivers of the presence of a specific macro-type.
The partial dependence plot resolved for Tmax_jul (see Figure S1) indicated that each single macro type has a specific climatic niche, where xeric species (XS) are the most demanding with monthly temperatures exceeding 20 • C, followed by Festuca rubra (FR) with a higher probability of detection for temperatures higher than 15 • C. The presence of shrubs (SP) is centered on temperatures~12 • C, whereas that of Nardus stricta (NS) has a peak around 10 • C, similar to Sesleria varia (SV). Carex curvula (CC) has a range included between~5 • and~12.5 • C and the same may be applied to Carex firma (CF). change (Δ) of Tmin_jan compared to the present, with a general trend of the median to increase under future climate conditions ( Figure 2) without any relevant difference between the Eastern ( Figure S2a) and Western ( Figure S3a) areas of the chain, although the Western Alps showed higher increases (+3.6 °C) than the East (+2 °C) with respect to the present. According to the median values, the highest Δ of Tmin_jan was found for 2041-2070 under RCP4.5 (+2.7 °C) and RCP8.5 (+3.1 °C) (Figure 2a). The pattern was similar for Tmax_jul, where the highest Δ of the median values were found for 2041-2070 under RCP4.5 (Δ = +2.1 °C) and RCP8.5 (Δ = +2.5 °C) (Figure 2b).

Climate Analysis
Climate data for the future, as an ensemble, reported a low variability in the simulation of delta change (∆) of Tmin_jan compared to the present, with a general trend of the median to increase under future climate conditions ( Figure 2) without any relevant difference between the Eastern ( Figure S2a) and Western ( Figure S3a) areas of the chain, although the Western Alps showed higher increases (+3.6 • C) than the East (+2 • C) with respect to the present. According to the median values, the highest analyzing the entire chain (Figure 2f). This was determined by a contrasting pattern between the two areas, namely a decrease in the Eastern Alps (up to Δ = −90 mm) ( Figure S2f) and increases (up to Δ = +50 mm) in the Western areas ( Figure S3f).
The statistical analysis of the climate across the entire chain evidenced that in more than 90% of test cases, the deltas between present and future climatic variables were statistically different from zero (data not shown).

Altitude Pattern Dynamics of Pasture Macro-Types Suitability
The seven macro-types showed different ranges of variations in their altitudinal suitability in response to the projected climate change (Figure 3  The statistical analysis of the climate across the entire chain evidenced that in more than 90% of test cases, the deltas between present and future climatic variables were statistically different from zero (data not shown).

Altitude Pattern Dynamics of Pasture Macro-Types Suitability
The seven macro-types showed different ranges of variations in their altitudinal suitability in response to the projected climate change ( Figure 3 showed an increasing of altitudes of SP in both time slices and scenarios with respect to the present, more relevant for RCP4.5 in the distant future ( Figure S5). Pastures dominated by xeric species (XS, Figure 3g) exhibited a progressive increase of their altitudinal range from 2011-2040 (1619 m, average of both scenarios) to 2041-2070 (1845 m), more evident across the Eastern than Western Alps ( Figures  S4 and S5, respectively). The absence of information for 2041-2070 under RCP4.5 indicated that the model did not simulate any presence of both macro-types for that period.

Global Dynamics of Alpine Pasturelands Suitability
Under future climatic conditions, the Alpine pastures suitability showed an overall decrease with respect to the present ( Table 3). The modeling ensemble suggested the highest decrease for 2041-2070, with an estimated reduction of 35% and 40% under RCP8.5 and RCP4.5, respectively. In the distant future (2041-2070), this reduction affected more the Eastern area of the chain, losing up to −45% and −40% of pastures under RCP4.5 and 8.5 scenario, than the Western Alps, depicting a loss up to −34% and −30% under RCP4.5 and 8.5, respectively, with respect to the present (Tables S2 and  S3). However, even in the near future (2011-2040), pasture suitability is found to considerably decrease compared to the present, with an estimated reduction in the range 19-23% under both RCPs (Table 3). In the near future (2011-2040), the main losses will affect the Western Alps under the RCP8.5 (−25%) with respect to the Eastern Alps (−16%), while the trend was opposite under RCP4.5

Global Dynamics of Alpine Pasturelands Suitability
Under future climatic conditions, the Alpine pastures suitability showed an overall decrease with respect to the present ( Table 3). The modeling ensemble suggested the highest decrease for 2041-2070, with an estimated reduction of 35% and 40% under RCP8.5 and RCP4.5, respectively. In the distant future (2041-2070), this reduction affected more the Eastern area of the chain, losing up to −45% and −40% of pastures under RCP4.5 and 8.5 scenario, than the Western Alps, depicting a loss up to −34% and −30% under RCP4.5 and 8.5, respectively, with respect to the present (Tables S2 and S3). However, even in the near future (2011-2040), pasture suitability is found to considerably decrease compared to the present, with an estimated reduction in the range 19-23% under both RCPs (Table 3). In the near future (2011-2040), the main losses will affect the Western Alps under the RCP8.5 (−25%) with respect to the Eastern Alps (−16%), while the trend was opposite under RCP4.5 scenario, where suitability of pastures located across the Western Alps decreased by −22% and Eastern Alps by −24% with respect to the present (Tables S2 and S3). Table 3. Current Alpine occurrence (in terms of number of pixels, i.e., 100 ha wide each) of unclassified pasturelands and expected changes (%) of suitability under future climate conditions, as reported by the three regional circulation models (RCMs) and their ensemble, time slices (2011-2070 and 2071-2100), and RCP scenarios (4.5 and 8.5).  Looking at the single RCM employed, the lowest impacts were reported by Aladin, with a decrease of pasture suitability in a range form −6% and −20% for all time-slices and climate scenarios. By contrast, the highest impacts were indicated by CMCC for all time-slices and climate scenarios, with the sole exception of the period 2041-2070 under RCP4.5, where the highest pasture decrease (−55%) was reported by ICTP.

Changes of Pasture Macro-Type Suitability
The seven macro-types showed different dynamics under future climate conditions, ranging from a large decrease for some, to slight increases for others (Table 4 and Figure 4 when analyzing the entire Alpine chain), depending on the region analyzed (Tables S3 and S4)

Changes in Pasture Macro-Types Composition
The expected future pasture composition for 2011-2040 and 2041-2070 under RCP4.5 and RCP8.5 indicated a general reduction in pasture biodiversity in the next decades, with the highest decreases expected in the distant future (2041-2070) ( Table 5). Variations in pasture composition were mainly driven by time-slices rather than the climate scenarios, and while most of pasture macro-types showed a general decrease in their suitability, the most hardy and less palatable species, such as Nardus stricta (NS) and xeric species (XS), showed increases in their occurrence over areas currently suitable for other macro-types (Table 6). Shrub species (SP)  Table 4), each contributing 6%, on average, to the future Alpine pastureland composition (Table 5).

Discussion
Alpine pastures are key ecosystems for people living in both mountain and urban areas, as they provide socio-economic and ecosystem services such as tourism, food products, biodiversity, and greenhouse gas mitigation etc. Given the current impacts of climate change on this ecosystem, this study aimed at evaluating the robustness of RF as a tool for estimating the spatial distribution and expected change of the main representative pasture macro-types in terms of their extent, ecological and pastoral importance across the Alps under future climate. In this context, the use of pre-existent spatial datasets coupled with the RF model proved to be robust at predicting potentially suitable macro-type areas of Alpine pastures under current climate conditions. The analysis has primarily taken into account the impact of climate change projections based on two time-slices (2011-2040 and 2041-2070) and most recent climate scenarios (RCP4.5 and RCP8.5) considering the long-lasting influence of species competition, through fitting pasture response to the environmental predictors without covering land abandonment and land-use policies, which strongly affect pastures [36]. Despite the proposed method proving to be robust, some issues should be reported. The delta change approach used for predicting the future climate in this study does not explicitly consider extreme climate events (likely causing negative impacts on pastoral resources, [60,61]) and we considered data of the warmest and coldest months of the year as a proxy for climatic extreme (i.e., frost events and heat waves) during the year. However, frequency and intensity of these events is expected to increase in the near future [19].
Moreover, even though it is unquestionable that the living conditions of herbaceous species are dependent on soil temperature rather than air temperature [62], the latter may be used as a proxy of for the former. As an example, Choler [63] demonstrated that soil and air temperatures are closely related and this relationship may be used to rescale daily mean air temperature. Indeed, in a previous work [31], we observed that actually air temperature is actually able to discriminate different kinds of macro-types across the Alps and this is likely related to the fact that air and soil temperatures are closely related.
Modeling ensemble indicated a global reduction of pastures suitability for all time-slices and scenarios, particularly for 2041-2070 where pasture lands decreased within a range of −35% to −40% with respect to the present period (Table 3), mostly affecting the Eastern Alps (ranges between −40% and −45%) than the Western (ranges between −30% and −34%; Tables S2 and S3). The magnitude of this decrease was likely due to the expected changes in climate conditions determining the suitability of each specific macro-type (i.e., optimal conditions for development) in conjunction with the available altitudinal space for a shift of the vegetation. The results indicated that projected increase of temperatures raised the optimal climate range of each macro-type to a higher altitude, resulting at the same time in a general spatial reduction of suitable areas. This was also suggested by analyzing the simulated range in which the seven macro-types are currently found (1051-2797; 1746 on average), which in the future is found to be reduced by −356 m on average, thus resulting at altitudes ranging from 1225 to 2635 (1411 on average). This warming effect on pasture has been observed by several studies (e.g., [64,65]). The shift of macro-types to a higher altitude clearly resulted in a generalized decrease of the seven macro-types suitability areas across the whole Alpine chain compared to the present. The reduction of this altitudinal range was mainly driven by less space (number of pixels) for expansion of the macro-types typical of the high altitudes (i.e., Carex curvula, Carex firma and Sesleria varia). The altitudinal range reduction found in this study also reflects the observed dynamics in which this shift upwards of pastures is also boosted by the forest line rise [66], lowered by human activities over the centuries [67]. Since the model does not consider pasture management, RF may have simulated the increase of other land-uses (such as forests), which is consistent with studies reporting that with unmanaged pastures the forests tend to move upwards [68]. According to our predictions, the areas currently occupied by pastures will be likely be replaced by forests, because of the coupled effect of treeline upward movement and expected impossibility of greening process at higher elevation [69].
The analysis of change of macro-type occurrence compared to the present also allowed to indicate the expected changes to the entire Alpine pasture composition, distinctly between two main regions (Eastern and Western Alps), and the suitability change within each macro-type along the chain. Considering the entire Alpine chain, results suggested an expected reduction of pastures diversity driven by a slight decrease or, in some circumstances an increase, of the low-quality macro-types suitability (e.g., mainly those characterized by Nardus stricta and by xeric species), the complete loss of pastures dominated by Festuca gr. rubra and Carex firma, and a decrease in the suitability for all the remaining macro-types (Figure 4 and Tables 5 and 6). While changes in pastoral suitability between the two Alpine regions showed similar patterns for most of the macro types, the projected climatic conditions in the Eastern area determined higher losses of areas suited to Carex firma and Festuca gr. rubra with respect to the Western (Tables S4 and S5). Conversely, the expansion of low-quality pasture macro types (i.e., dominated by Nardus stricta and xeric species) affecting the Western Alps only, was mainly determined by the drier and warmer conditions projected in this area in the future (Tables S4  and S5 and Figures S2 and S3). These results suggested not only that the expected climate change may favor the most hardy and low quality species, but also how the change in future climate conditions, which varies along the chain, over unmanaged pastures may affect the interaction between the current communities. For instance, pastures dominated by Carex firma and Carex curvula showed a considerable reduction compared to the present (Table 5), also resulting as missing or very small communities within the future pasture composition (Table 6). These macro-types, generally found at the highest altitude and over calcareous (Carex firma) or acid (Carex curvula) soils, require a microthermal regime and high summer precipitation [45]. Moreover, predicted temperature increase will anticipate snowmelt date and shift precipitation from snow to rain [70]. This will produce an earlier snow disappearance, less water storage in the soil, and an increasing water demand. As a result of these coupled effects, a higher water stress for vegetation is predicted for the growing season and pastoral macro-types more sensitive to this will probably be those growing at higher altitude [71]. The expected temperature increase in conjunction with spring precipitation decreases would suggest shifts upwards for these macro-types. However, the reduced space availability in conjunction with the difficulty of moving into rocky soils may have strongly reduce their future suitability. By contrast, although pastures dominated by Sesleria varia show similar climatic and ecological requirements as those dominated by Carex firma and Carex curvula, they showed lower decreases (Table 5). This was likely due to the fact that pastures dominated by Sesleria varia cover a wider ecological spectrum and higher altitudinal range than Carex firma and Carex curvula [72], determining reductions in space and soil competition with the other macro-types, thus resulting in more adaptability under the expected future conditions. Concerning the most xeric and low-medium altitude macro-types, areas suited to pastures dominated by Festuca gr. rubra are expected to disappear, while those dominated by Nardus stricta and xeric species showed contrasting results when comparing the two regions. More specifically, while pastures dominated by Nardus stricta showed increases in the Eastern Alps, xeric species are projected to find their optimal conditions in the Western area under future climatic conditions, with relevant expansions in the distant future. The dynamics of the expected suitability change of this macro-type, characterized by general contrasting climatic and ecological requirements with respect to the median values of Carex firma, Carex curvula, and Sesleria varia [31], were clearly driven by intrinsic ecological characteristics of each macro-type as well as by the fact that RF simulations did not consider the management effect on pastures and on the single pastures macro-type. For instance, the simulated increase of Nardus stricta across the Alpine chain in the future, suggested the ability of this community to spread even under changed climatic conditions. This agrees with the ability of this macro-type to develop over poor N-content soils and its hardiness in adapting to a wider spectrum of climatic conditions [10]. This increase might be also favored by the lack of management, since several studies confirmed that Nardus stricta can find the best development conditions under unmanaged or low grazing intensity areas because of its oligotrophic behavior that can favor the spread of this species in areas with reduced return of animal fertility to soil [10,73]. As expected, the simulated increase of xeric species reflected the warmer and drier conditions projected by RCMs for the next decades particularly across the Western Alps, which can widen their ecological spectrum and make them more adaptable and competitive against the other macro-types [74]. By contrast, the losses of pastures suitability dominated by Festuca g. rubra likely involved several factors. First, the expected shift upwards due to the changed climatic conditions may have further decreased the presence of this macro-type already geographically limited on the Alpine chain according to our assessment (Table 1). Second, this macro-type may have strongly suffered the competition with less palatable and hardy species, such as Nardus stricta and xeric species, which may have inhibited the development of pastures dominated by Festuca g. rubra in the Alpine chain. These dynamics can be suggested also by Table 6, where the suitability change of Festuca gr. rubra tends to turn mainly to herbaceous communities dominated by Nardus stricta and xeric species. This agrees with Targetti et al. [44], confirming that pastures dominated by Festuca rubra are inclined to reduce their extents or turn into Nardus stricta macro-types under unmanaged conditions.

Concluding Remarks
The results of this study demonstrated how the projected climatic conditions will determine an overall reduction of the areas currently suited to natural pasturelands of the Italian Alps and variations in the composition of these ecosystems. The main reductions will likely affect high altitude macro types (i.e., pastures dominated by Carex firma, Carex curvula, and Festuca gr. Rubra) over the entire chain, threatening the unique and rare herbaceous biodiversity characterizing the Alps. Contrasting impacts were evidenced between the Eastern and Western Alps on the future distribution of the more hardy and low quality macro types (i.e., pastures dominated by Nardus stricta and xeric species), projecting relevant expansions of areas suited to Nardus stricta and xeric species across the Eastern and Western Alps, respectively, which should be investigated in more detail. All these impacts pose several challenges concerning the related effects on future pasture characteristics and management. First of all, changed climatic conditions may likely lead to variation in biomass peaks and productivity compared to the present, because of advancing of the spring phenology, delay in autumn phenology, and an earlier timing of maximum photosynthesis [75][76][77][78][79]. Furthermore, losses or decreases of suitability of high-quality macro-types, besides leading to a general decrease in pasture diversification and a general homogenization of the landscape, will likely strongly alter the nutritional value of the mountain grazing areas. Moreover, the expected decrease of high-forage quality macro-types such as Festuca. gr. rubra in favor of Nardus stricta, a non-palatable grass, may result in an overall decrease in pasture quality that could be tackled by more appropriate stocking rates or management techniques [80] or by diversifying or mixing livestock species (characterized by different grazing behavior) in order to maximize animal performances and, at the same time, to improve the utilization of these marginal areas [81]. Finally, in the new projected pasture conditions farmers will be forced to provide additional nutrients for livestock feeding in order to guarantee or maintain an adequate production level, though resulting in increase of management costs.
Supplementary Materials: The following are available online at http://www.mdpi.com/2073-4395/10/7/926/s1. Figure S1: Partial dependence plot of the probability of presence of different macro types versus increasing values of Tmax in July. The original data were interpolated using a polynomial function. Figure Table S1: Test for delta change significantly different from 0. The test was performed considering the delta changes samples for each RCMs, variable, and time slice extracted from each macro-type area as delimited in the present period. For each sample, the null hypothesis that these samples were different from zero was tested (* significant for p < 0.05, ** significant for p < 0.01,*** significant for p < 0.001). Table S2: Current Eastern alpine occurrence (in terms of number of pixels, i.e., 100 ha wide each) of unclassified pasturelands and expected changes (%) of suitability under future climate conditions, as reported by the three regional circulation models (RCMs) and their ensemble, time slices (2011-2040 and 2041-2070) and RCP scenarios (4.5 and 8.5). Table S3: Current Western alpine occurrence (in terms of number of pixels, i.e., 100 ha wide each) of unclassified pasturelands and expected changes (%) of suitability under future climate conditions, as reported by the three regional circulation models (RCMs) and their ensemble, time slices (2011-2070 and 2071-2100) and RCP scenarios (4.5 and 8.5). Table S4 Funding: The research leading to these results was supported by the LIFE PASTORALP project (LIFE16 CCA/IT/000060), co-funded by the European Union's LIFE Programme, Climate change adaptation action sub-programme.