1 Introduction

Scots pine (Pinus sylvestris L.) forests at the southern edge of their distribution are currently facing decline and succession, resulting from a combination of climate warming, land use changes, and increased abiotic and biotic disturbances (Gimmi et al. 2010; Vacchiano et al. 2012).

From a physiological standpoint, drought has been identified as the primary driver of pine decline, as it affects foliage production, carbon allocation (Galiano et al. 2010), cambial activity (Eilmann et al. 2011; Oberhuber et al. 2011), hydraulic capacity (Sterck et al. 2008), and the likelihood of xylem cavitation (Martı́nez-Vilalta and Piñol 2002). Additionally, drought can predispose weakened trees to inciting mortality agents, such as mistletoe, bark beetles, or root-rot fungi (Dobbertin et al. 2007; Gonthier et al. 2010; Rigling et al. 2010; Sangüesa-Barreda et al. 2013), and reduce seedling recruitment and survival (Matias and Jump 2014).

On top of this, at the landscape level, Scots pine forests in southern Europe have recently experienced a decrease in management intensity, shifting from open-canopy, even-aged stands maintained by broadleaves coppicing, wood pasture, litter raking, and pitch collection (Gimmi et al. 2007) to denser forests following depopulation of mountain areas and abandonment of traditional land use practices. Under such scenario, succession by mid-tolerant species such as downy oak (Quercus pubescens Willd.) is favored over Scots pine regeneration (Urbieta et al. 2011; Rigling et al. 2013). Both mechanisms, land use changes and climate extremes, are at work at the same time (Gimmi et al. 2010), determining feedbacks and interactions difficult to disentangle and providing a challenge for forecasting future vegetation patterns.

Recession of Scots pine forests in southern European landscapes would affect the provision of important ecosystem services, such as protection from hydrogeological hazards, plant and animal diversity, timber, and recreation. A shift from Scots pine to oak can also be problematic because of the loss of useful life traits, as the ability to rapidly colonize open or disturbed ground (Vacchiano et al. 2013). Predictions of future vegetation changes and knowledge of the suitability of pine vs. oak to expected environmental conditions will help managers in developing adaptation strategies to sustain the fulfillment of the desired forest functions (Chmura et al. 2011).

The aims of this work were (1) to detect drivers of current pine and oak occurrence in a mountain region of the southwestern Alps, by fitting species distribution models (SDM) on climate, soil, land use, stand structure, and disturbance-related predictors; (2) to apply the models using future (2080) scenarios, in order to assess if and where vegetation shifts are likely to occur under climate and management changes; and (3) to compare the effects of the Europe-wide drought events of 2003 and 2006 (Thabeet et al. 2010) on Scots pine vitality against SDM predictions in 2080, in order to assess the potential role of extreme drought response as an early warning of future vegetation changes.

2 Methods

2.1 Study species

Scots pine is the most widespread coniferous species in Europe and the most widespread pine in the world (Mirov 1967). Scots pine is a species of continental climates, able to grow in areas with annual precipitation ranging from 200 to 1800 mm (Burns and Honkala 1990). The upper/northern and lower/southern limits of the species correspond with isotherms −1 °C (mean temperature of the coldest month) and +33 °C (mean temperature of the warmest month), respectively (Dahl 1998), even if pine can tolerate more extreme temperatures without tissue damage, especially at the cold end (−90 °C: Sakai and Okada 1971).

Scots pine is a light-demanding, early seral species that can establish both in acid and limestone soils (Richardson 1998; Debain et al. 2003). Its ecology is largely characterized by stress tolerance. On the one hand, this allows it to occupy a range of habitats that are unfavorable to other tree species, through tolerating various combinations of climatic and edaphic stress (Richardson 1998). On the other hand, this implies that Scots pine is excluded from more favorable sites through competition. In recent decades, it was favored by past fires (Gobet et al. 2003), by heavy forest cuts, and by the recent increase of fallow lands (Farrell et al. 2000; Kräuchi et al. 2000; Caplat et al. 2006; Picon-Cochard et al. 2006). In the absence of disturbances, it will eventually be overgrown or replaced by broadleaves or mixed broad-leaved coniferous forest. However, in the drier, central alpine sectors (<700 mm year−1 rainfall), Scots pine often forms stable communities due to limited competiveness of other conifer tree species (Ozenda 1985).

Scots pine populations are negatively affected by drought in all demographic processes, i.e., regeneration (Galiano et al. 2013, Carnicer et al. 2014), growth (Vilà-Cabrera et al. 2011), and mortality (Dobbertin et al. 2005; Bigler et al. 2006). On the other hand, downy oak exhibits better ecophysiological adaptations (Nardini and Pitt 1999; Eilmann et al. 2006, 2009; Zweifel et al. 2009) and higher growth (Weber et al. 2008) under comparable climate conditions. Oaks also have an advantage over Scots pine in the regeneration phase following stand-replacing fire, owing to their resprouting ability—as opposed to limitations in Scots pine regeneration due to short dispersal distance and obligate seeder traits (Moser et al. 2010; Vacchiano et al. 2013). Such differences, and the fact that oaks are characterized by lower shade intolerance, make them a suitable species for secondary succession of declining or outcompeted pine stands.

2.2 Study area

The study area covers the Aosta Valley region in Northwestern Italy (3262 km2) (Fig. 1). Topography is shaped by a main east–west-oriented valley with several north–south protrusions. Mean annual temperature in Aosta (45° 26′ N, 7° 11′ E, 583 m a.s.l.) is 10.9 °C (years 1961–1990; Tetrarca et al. 1999). Climate is warm-summer continental (Dfb) according to the Köppen classification (Peel et al. 2007); July and January monthly means may differ by as much as 22 °C. Mean annual rainfall in Aosta amounts to very low values in comparison with localities in other central Alpine valleys (494 mm, years 1961–1990; Biancotti et al. 1998), with a period of water deficit (Bagnouls and Gaussen 1957) extending from June to September. Winter precipitation usually comes as snow. The study area exhibits both crystalline (granites) and metamorphic bedrocks, but most landscape is covered by quaternary deposits of glacial, gravitative, or colluvial origin. Soils belong to the series of western and central Alpine soil on igneous and metamorphic rocks (Costantini et al. 2004) and are mostly represented by shallow soils (Lithic, Umbric, and Dystric Leptosols), eroded soils (Eutric and Calcaric Regosols), acid soils with organic matter, iron oxides and aluminum accumulation (Dystric Cambisols, Haplic Podzols, Humic Umbrisols), or alluvial soils (Eutric Fluvisols).

Fig. 1
figure 1

Location of the study region and area covered by forests (in green)

Full size image

Scots pine stands in the study area cover 5372 ha (Gasparini and Tabacchi 2011), i.e., 6 % of the total forest area, and thrive on both acidic and basic substrates of well-exposed, bottom to mid-elevation slopes. Stands dominated by Scots pine are mostly young, averaging 920 trees per hectare (TPHA) and a basal area (BA) of 26 m2 ha−1 (Gasparini and Tabacchi 2011). Quadratic mean diameter (QMD) is 21 cm, but trees larger than 35 cm are extremely rare (about 2 %) (Camerano et al. 2007). Stand top height can vary from 10 to 25 m according to site fertility (Vacchiano et al. 2008). Depending on successional stage and climatic factors, species composition may range from 100 % pine (especially on recently disturbed sites or dry, southern slopes) to mixtures with Swiss mountain pine (Pinus montana Mill.), European larch (Larix decidua Mill.), Norway spruce (Picea excelsa Karst.), silver fir (Abies alba Mill.), beech (Fagus sylvatica L.), sessile oak (Quercus petraea (Mattus.) Liebl), European chestnut (Castanea sativa Mill.), and mostly with downy oak, which has similar thermal and moisture needs.

Downy oak stands cover 3468 ha in the study area (Gasparini and Tabacchi 2011), at elevations of 300–1200 m (but up to 1500 m on rocky outcrops and 1800 m for isolated individuals), predominantly on shallow soils and carbonatic substrates. Xerophilous stands on south-facing slopes are sparse and slow growing (1000 TPHA, BA 20 m2 ha−1), with young individuals often developed from former coppices, grazed woodland, or after invasion on abandoned fallow lands (QMD 10–25 cm, mean height 5–10 m) (Gasparini and Tabacchi 2011). Just as in Scots pine, meso-xerophilous stands on north-facing slopes exhibit higher growth (mean height 10–15 m) and mixture degree. Scots pine and downy oak can replace each other in the course of forest dynamics, e.g., by regeneration of pine in sparse and degraded oak woodlands or by succession of closed-canopy, or declining, pine forests to more tolerant oak (Zavala and Zea 2004).

2.3 Drivers of pine and oak distribution

In order to model the occurrence of Scots pine and downy oak in the study area, we used a diverse set of explanatory variables including vectorial as well as raster information at different spatial resolutions. All variables were resampled at a common spatial resolution of 1 km, i.e., the coarsest resolution among all explanatory variables, and clipped to a land use mask of current forest distribution. Since the model could be calibrated against current vegetation conditions only. Rasterization of vector layers and raster resampling were carried out by aggregating to cell means if raster grain was finer than 1 km and by bilinear (for continuous layers) or nearest neighbor interpolation (for categorical layers) if grain exceeded 1 km. Explanatory variables included the following (Figure S1):

  1. (1)

    Elevation, slope, aspect, southness (i.e., a linearization of aspect: Chang et al. 2004), and topographic position index (TPI: Guisan et al. 1999) computed from a 10-m digital terrain model. A higher TPI is indicative of ridges or hilltops.

  2. (2)

    Climate means (years 1961–1990) at a 1-km resolution, extracted from the WorldClim database (Hijmans et al. 2005). These included mean, minimum, and maximum yearly temperatures (TMIN, TMEAN, TMAX), yearly precipitation (P), precipitation cumulated in the growing season (GSP; April–September), and yearly solar radiation (RAD). Additionally, using mean, minimum, and maximum monthly temperature grids, we computed growing degree days (GDD; base temperature = 5 °C) (Fronzek et al. 2011) and an aridity index (AI) as the difference between monthly precipitation and potential evapotranspiration (PET). PET for month i was computed as after Zimmermann et al. (2007).

    $$ \mathrm{PETi}=0.0023\;{\mathrm{RAD}}_i\left(T{\mathrm{MEAN}}_i+17.8\right)\frac{\left(T{\mathrm{MAX}}_i-T{\mathrm{MIN}}_i\right)}{2}\mathrm{days} $$
  3. (3)

    Soil variables at a 1-km resolution, extracted from the European Soil Bureau (European Soil 1999). We selected variables potentially important for tree establishment and growth, namely available water capacity (AWC) of the topsoil, accumulated soil temperature class (ATC), total organic carbon (OC) of the topsoil, base saturation (BS), erodibility (ERO), depth to rock (DR), dominant surface textural class (TEXT), and volume of stones (VS). All variables were coded as dummy values.

  4. (4)

    Natural disturbances, such as landslides or severe soil erosion (source: Corine Land Cover 1990 raster coverage, resolution 500 m), avalanche tracks, and wildfires >10 ha for the years 1961–1991 (sources: Regione Autonoma Valle d′Aosta, Ufficio Neve e Valanghe, and Regione Autonoma Valle d’Aosta, Corpo Forestale Regionale, Nucleo Antincendo Boschivi).

  5. (5)

    Competition by the pre-existing canopy, assessed by extracting the Normalized Difference Vegetation Index (NDVI) from a Landsat 5 Thematic Mapper image (path 195, row 28) taken on June 30, 1987 (resolution 30 m). The acquisition period was chosen as to be at the peak at the growing season; the image had 10 % cloud cover, but clouds were clustered over high elevation, unforested terrain. The image was first converted to top of the atmosphere radiance using standard equations and calibration parameters obtained from the metadata of the scene (Chander et al. 2009). Then, we computed NDVI using band TM4 (near infrared) and band TM3 (visible red) and used it as a proxy of standing forest biomass (Tucker 1979; Pettorelli et al. 2005). As an additional index of competition by forest vegetation, we used percent tree cover from the recently released Landsat vegetation continuous field (VCF) dataset (Sexton et al. 2013), at a resolution of 30 m, based on a Landsat 5 TM image acquired on July 27, 2001.

  6. (6)

    Land use intensity was assessed by using proxy variables, i.e., total road length and total building surface per 500-m pixel, as extracted from a vector regional map. Moreover, the degree of land abandonment was estimated at the municipality level by the percent variation in resident population in the period 1951–1991 (source: ISTAT).

    In order to limit collinearity of independent variables, predictors exhibiting a Pearson’s correlation coefficient >|0.9| were excluded from further analysis.

2.4 Model runs under future scenarios

Simulation experiments for the future projections of species distribution relied on the same set of explanatory variables. However, values for variables used in future scenarios were chosen as follows:

  1. (1)

    Climate means for the 2080 decade were extracted from 30-arcsec gridded simulations by the ECHAM5/MPI-OM model from the Max-Planck Institute for Meteorology, Germany (Raible et al. 2006), under the high emission scenario Special Report on Emissions Scenarios (SRES) A1B. Under the assumption of a constant solar radiation, we computed GDD, GSP, PET, and AI from the ECHAM-5 grids. For the 2080 scenario, we did not extrapolate the model to pixels exhibiting AI values exceeding the range of current ones (Elith and Leathwick 2009).

  2. (2)

    Fire frequency and size are supposedly responsive to climate change (Moriondo et al. 2006). In order to simulate the influence of fire preceding the 2080 decade, we used wildfire polygons for the years 1981–2000, i.e., a period that included several extreme fire seasons resulting in a +39 and +26 % increase in the frequency and total area burned, respectively, by large fires (>10 ha) relative to 1961–1980.

  3. (3)

    We simulated two alternative land use scenarios: (1) urbanization and land abandonment, i.e., every municipality was assigned a “business as usual” scenario of population change using figures for the period 1951–1991, and (2) maintenance of high land use, i.e., all municipalities were assigned 0 % variation in population respective to 1951, thereby assuming a continued presence of man and its activities at all rural settings.

Soil characteristics are also responsive to climate change (Singh et al. 2011); however, we kept these variables at current conditions for the 2080 simulation, since no quantitative scenarios are available to estimate future changes. Altogether, three scenarios were simulated: current conditions, 2080 climate with unchanged land use, and 2080 climate with intense land use.

2.5 Model building

Presence/absence of pine and oak in the years 1992–1994 served as a response variable, which we extracted from a regional forest inventory based on a 500-m regular grid. At every grid node, the species and diameter at breast height (DBH) of each living tree (DBH >7.5 cm) were measured within a variable–radius circular plot (radius 8–15 m depending on tree density). Plot coordinates were recorded to the nearest meter. Scots pine and downy oak were labeled as present where at least one tree of each species was recorded and absent otherwise.

We assumed that both pine and oak distribution are in equilibrium with the environment (Rohde 2005). For this reason, and because our aim was to model potential niche, no migration constraints were included in the model.

We used an ensemble modeling approach (Araujo and New 2007), by fitting and averaging predictions obtained by a generalized linear model (GLM), artificial neural network (ANN), and multiple adaptive regression spline (MARS) using the same set of responses, predictors, and scenarios. Model specifications were as follows: (a) for GLM, a backward stepwise algorithm was used, based on Akaike Information Criterion (AIC); (b) for ANN, the initial number of cross-validations to find best size and decay parameters was set to five; and (c) for MARS, the cost per degree of freedom charge was set to 2, and the model was pruned in a backward stepwise fashion. All models were fit on a binomial distribution with logit link, without interactions between predictors, and using a maximum of 100 iterations.

For each of the three models, we computed variable importance ratings and response curves. To do so, all variables but one are set constant to their median value, and only the remaining one is allowed to vary across its whole range. In the case of categorical variables (e.g., soil), the most represented class was used. The variations observed and the curve thus obtained show the sensibility of the model to that specific variable.

We carried out k-fold cross-validation of the model by subdividing the data into a 3:1 proportion (k = 4). Model specificity and sensitivity were computed for the selected thresholds; the threshold to convert continuous predictions into binary ones was iteratively chosen to maximize the area under the curve (AUC).

The ensemble prediction was computed from all model realizations with AUC >0.75. The probability of occurrence for the ensemble prediction was the mean of the selected models’ predictions, weighted by the model AUC. Model residuals were scrutinized to detect the absence of trends against predicted values and independent variables; a variogram was fitted to assess the degree of residual spatial autocorrelation. Ensemble models were run for the whole study region to obtain a map of potential species distribution under current and future climate, assuming niche conservationism (Wiens et al. 2010). We classified simulated presence/absence of both species using an occurrence probability threshold of 0.6 and assessed projected area changes and elevational shifts in the distribution of pine and oak under the climate change and climate change + intense land use scenarios. All analyses were carried out using the biomod2 package (Thuiller et al. 2013) for R (R Development Core Team 2013).

2.6 Effect of extreme drought events

The response of extant Scots pine forests to drought events in years 2003 and 2006 was assessed by the temporal difference in NDVI (ΔNDVI: year of drought − year before drought). NDVI was computed from two 16-day maximum value composite (MVC) MODIS images (resolution 30 arcsec) taken at the end of the summer (Julian days 226–241). Cloud cover of the MVC was between 1 and 4 % for the four images. Pixels with a quality analysis score of 2 and 3 (i.e., targets covered by snow/ice or cloudy pixel) as well as NDVI lower than 0.2 or null (open water) were filtered out (Vacchiano et al. 2012).

In order to distinguish reflectance anomalies from random or systematic error (Morisette and Khorram 2000), we classified as “decline” all pixels with ΔNDVI < (mean−3 standard deviations), as computed from the full scene (Fung and LeDrew 1988; Vacchiano et al. 2012). Finally, we compared the modeled change in pine occurrence probability (2080–current) of decline vs. non-decline pixels by means of Wilcoxon signed-rank test (Sokal and Rohlf 1995).

3 Results

Scots pines were detected in 460 (27 %) out of 1730 inventory plots, and downy oak in 181 (10 %). After screening for collinearity, 18 predictors were retained for subsequent analyses (Table 1). Since most climate-related variables were correlated to each other and to elevation, we retained only aridity index (AI) as the main climate predictor; Pearson’s correlation between AI and WorldClim variables was always higher than 0.95 (e.g., R = −0.995 vs. mean annual temperature, R = 0.993 vs. annual precipitation, R = −0.962 vs. GDD).

Table 1 Explanatory variables used in this study (minimum, maximum, mean, standard error), computed for currently forested areas only
Full size table

AI was the most important predictor for the current distribution of both pine and oak (Table 2), with higher occurrence probability at low water balance levels (Figure S2). However, MARS captured a reduced probability of occurrence for Scots pine at very low values of the aridity index (i.e., very dry sites). Beyond aridity, variables associated to high probability of Scots pine occurrence were southness, TPI, population change, building density, and past fires—the last two only in the ANN model. Soil erosion, NDVI, and road density (in the ANN model) decreased the probability of pine presence (Figure S2a). Explanatory variables of oak distribution exhibited a similar behavior: southness and TPI, but also slope, soil depth, and soil temperature class were associated to high presence probability, while road and building densities produced a low presence probability (Figure S2b).

Table 2 Variable importance (0–1) for SDM of current Scots pine and downy oak distribution fitted by generalized linear model (GLM), artificial neural network (ANN), and multiple adaptive regression spline (MARS)
Full size table

The ensemble models were successfully cross-validated (AUC = 0.865 for pine and 0.944 for oak) and correctly predicted most observations (sensitivity = 83.4 and 96.9 %, specificity = 72.7 and 80.9 %, respectively) (Fig. 2). Residuals were immune from spatial autocorrelation and trends against any of the predictors.

Fig. 2
figure 2

Occurrence probability (0–1) of a Scots pine and b downy oak under current climate. Ensemble model (mean of GLM, MARS, and ANN). Presence points from the regional forest inventory in black

Full size image

In 2080 (SRES A1B emission scenario, continuing population trend), the mean probability of occurrence of Scots pine declined slightly (0.33 vs. a current 0.36 across the whole study area) (Fig. 3). However, it increased under the intense land use scenario (0.45) (Fig. 4). The area with a probability of occurrence of Scots pine >0.6 decreased from 8700 to 8000 ha under the climate warming scenario and increased to 8800 ha under climate warming + intense land use. The probability of occurrence of Scots pine always declined at lower elevations and increased at higher ones (Fig. 5); mean elevation of simulated presence points shifted from 1328 to 1528 m a.s.l. under climate warming and to 1473 m a.s.l. under climate warming intense land use, i.e., an upward shift of the potential niche of 200 and 145 m, respectively.

Fig. 3
figure 3

Occurrence probability (0–1) of a Scots pine and b downy oak under 2080 climate and current land use scenario. Ensemble model (mean of GLM, MARS, and ANN)

Full size image
Fig. 4
figure 4

Occurrence probability (0–1) of a Scots pine and b downy oak under 2080 climate and intensive land use scenario. Ensemble model (mean of GLM, MARS, and ANN)

Full size image
Fig. 5
figure 5

Change in probability of occurrence (2080–current) of a Scots pine and b downy oak for different elevation classes under 2080 climate (above) and 2080 climate + intensive land use scenario (below)

Full size image

Oak increased its probability of occurrence under all scenarios (6100 ha under current conditions, 10,100 ha under climate change only, and 14,700 ha under climate change + intense land use). Mean elevation of simulated presence points (probability of occurrence >0.6) shifted from 705 to 922 and 933 m a.s.l., respectively, i.e., an upward shift of 215 and 222 m.

The area of Scots pine pixels classified as decline was 147 ha in year 2003 and 102 ha in year 2006. However, in either year we did not observe a significant difference between decline and non-decline pixels in the modeled probability of occurrence of Scots pine for 2080 (Fig. 6).

Fig. 6
figure 6

Change in probability of occurrence (2080–current) of Scots pine for decline and non-decline pixels in dry years 2003 (left) and 2006 (right), under 2080 climate (above) and 2080 climate + intensive land use scenario (below)

Full size image

4 Discussion

Published models of Scots pine distribution under scenarios of climate change have produced contrasting results (e.g., Casalegno et al. 2011; Meier et al. 2011), probably as a result of different datasets and processes being included or not in the models (e.g., dispersal constraints, biotic competition, choice of climate, and drought-related variables).

Many processes are at work in determining pine decline. Drought is either a direct or a predisposing factor of mortality (Rebetez and Dobbertin 2004; Choat et al. 2013). At low elevations, Scots pine reaches more rapidly decay stages, since trees weakened by drought are easily killed by “inciting” or “contributing” biotic agents (Dobbertin et al. 2005; Bigler et al. 2006; Vacchiano et al. 2012). Also, land use change may eventually result in competitive exclusion of light-demanding Scots pine.

Climate warming and drought are related (i.e., the frequency of drought spells is expected to increase under climate change: Allen et al. 2010); however, extreme drought events may be more important than average climate trends in determining plant population viability (Katz and Brown 1992; Bréda and Badeau 2008), and they can induce shifts in species composition and distribution (Jentsch et al. 2007).

In order to take into account the different factors governing drought sensitivity, we included in our models its meteorological, topographic, and soil-related component. At the resolution and extent analyzed, the probability of occurrence of Scots pine increased under climatic and topographic aridity. This is consistent with the biogeography of the species that forms pure stands in most inner-Alpine valleys such as the study area, preferentially on south-facing slopes and ridge positions (Ozenda 1985). Accordingly, low aridity reduced the probability of presence of Scots pine. In Aosta valley, temperature and precipitation are strongly correlated to elevation (which for this reason was excluded from the analysis); therefore, the AI variable contained also information regarding the upper elevational limits of the habitat suitable for Scots pine.

Another important driver of Scots pine occurrence was biotic competition, as expressed by NDVI of the forest canopy. As expected, the early seral pine cannot establish successfully under thick canopy cover (Vickers 2000). In contrast, it can also establish successfully on non-forested land, such as abandoned pastures and meadows (Poyatos et al. 2003), but this process could not be taken into consideration in future simulations, since our correlative models were calibrated on current vegetation conditions only.

In addition to topo-climatic and competition variables that are routinely assessed in SDM, we also evaluated the effect of soil properties (albeit using a coarse resolution and dummy coding) and natural and anthropogenic disturbances (Matias and Jump 2012). Scots pine did not exhibit any soil preference, consistently with its edaphic plasticity (Médail 2001). However, its occurrence was moderately associated to the absence of steep slopes and severe land erosion, which should be adverse to permanent vegetation cover, and to recurring wildfires. Wildfire polygons were not labeled as surface or crown fires; however, surface fires are more common in the study area, especially at low elevations on south-facing slopes (Vacchiano et al. 2013).

We also evaluated the effect of human land use on species distribution by using proxy variables (Garbarino et al. 2009). Increased population and road density resulted in increased occurrence of Scots pine. Management practices such as timber harvesting, litter collection, and forest grazing may in fact prevent succession to more competitive late-seral species (Weber et al. 2008; Gimmi et al. 2010). The association between pine and population/road density may also be due to recent establishment of Scots pine after agricultural abandonment (Poyatos et al. 2003). Building density was negatively correlated to the probability of occurrence of both Scots pine and downy oak, likely due to the spatial segregation of forests vs. developed or urbanized areas in the main valley.

These factors help explain the response of Scots pine distribution in 2080 under the A1B warming scenario, i.e., a modest reduction of habitat suitability, but a significant increase of its optimum elevation. At low elevations, in fact, aridity could reach the lower limits for the species to persist, as suggested by the MARS response curve (Garzon et al. 2008). This change is partially counteracted in a scenario where land abandonment is avoided: in this case, the probability of occurrence of Scots pine would still decrease at low elevations but, on average, the human factor could be sufficient to prevent the decline of Scots pine throughout its current distribution. This analysis is correlative and does not explore the physiological and successional processes behind such land use/climate change tradeoff. However, it is indicative of the fact that land use changes can be as strong as climate change in determining future species composition and dominance of mountain forests (Dirnböck et al. 2003) and that they deserve a deeper attention in modeling species’ response to future climate conditions.

The distribution of downy oak shared the same topo-climatic features as Scots pine (high aridity/low elevation, southern aspects, low erosion, high soil temperature) but was also associated to lower land use intensity (road density) and higher soil depth. Canopy density (NDVI) and natural disturbances were not influential, since downy oak is more shade-tolerant than pine (Monnier et al. 2013). The response of downy oak to climate warming was different from Scots pine and produced an increased probability of occurrence throughout the study region. Previous research has demonstrated that downy oak is better adapted than Scots pine to both short- and long-term drought, due to its different physiological responses, i.e., stomata closure, resistance to embolism, and seedling vitality (Eilmann et al. 2006; Poyatos et al. 2008; Morán-López et al. 2012).

Population change was not among the most important predictors of current downy oak distribution. However, we detected a moderate association between population increase and higher probability of occurrence of oak. This can be due either to the practice of coppicing oaks for firewood or to the fact that depopulated areas are located in the remotest part of lateral valleys, where elevation and sites are far below optimum for downy oak.

The use of ensemble modeling is justified by the need to reduce model uncertainty due to different modeling approaches (Marmion et al. 2003). Ensemble models in biomod2 are obtained by averaging model predictions and excluding models with low predictive power (AUC <0.75); model predictions are weighted by the AUC of their respective modeling algorithm. In this study, all three model algorithms produced an AUC >0.75. However, differences in importance of explanatory variables and shape of response curves were apparent. MARS are more flexible than GLM as they are fit using piecewise linear splines and are particularly useful when assuming that the shape of species’ responses is not linear (Leathwick et al. 2005). ANN, on the other hand, are not based on specific distribution functions of the response. They are robust to noisy and non-linear responses and allow for categorical predictors (such as soil characteristics in this study). Therefore, they are particularly appropriate in an exploratory context. On the other hand, they are sensitive to multicollinearity and prone to overfitting, and interpretation of causal relationships for individual predictors is not straightforward (Manel et al. 1999). The differences are apparent in species response curves (Figure S2), with MARS and ANN capable of detecting non-linear responses to some explanatory variables that were not picked up by GLM, despite a similar predictive performance. This is reflected by the higher importance of some explanatory variables, such as roads, buildings, TPI, or erosion, under models capable of detecting non-linear species responses (Table 2).

Finally, contrary to our expectations, we did not detect any overlap between drought-induced Scots pine decline in years 2003 and 2006 and change in occurrence probability under a warming scenario. Widespread tree mortality can occur under extreme dry spells, but it is uncertain whether one or two extreme years are sufficient to trigger major shifts in forest composition (e.g., Vicente-Serrano et al. 2013). The effect of extreme years on the realized niche of Scots pine will likely depend on the frequency and severity of droughts, rather than on decadal climate means such as the ones we used in our projections. Other parameters might be important in their extreme yearly or seasonal values, such as high precipitation events promoting a new generation after a mortality episode (Matias and Jump 2012), late frost preventing uphill expansion of sensitive species such as downy oak (Burnand 1976), and natural disturbances such as large, stand-replacing fires (Moser et al. 2010).

What is certain, however, is that downy oak is equipped with better adaptations to drought and is likely to replace Scots pine at lower elevations under a warming scenarios, whereby an increased frequency of droughts is to be expected (Dai 2012). Management actions have the potential to mitigate this shift (Vilà-Cabrera et al. 2013), e.g., thinning to 40–60 % initial basal area to mitigate drought effects on Scots pine on xeric sites (Giuggiola et al. 2013). However, effects of management actions must be more thoroughly explored to evaluate tradeoffs with each species’ resistance and resilience in the face of climate forcing.