Influence of land use and climate on recent forest expansion: a case study in the Eurosiberian–Mediterranean limit of north-west Spain

Authors


Summary

  1. In Mediterranean mountainous areas, forests have expanded in recent decades because traditional management practices have been abandoned or reduced. However, understanding the ecological mechanisms behind landscape change is a complex undertaking because the influence of land use may be reinforced or constrained by abiotic factors such as climate. In this work, we evaluated their combined effects on recent forest expansion across climatic, topographic and management gradients.
  2. We used orthorectified aerial photographs from the second half of the twentieth century (1956, 1974, 1983, 1990 and 2004) to monitor changes in forest distribution in a set of 20 head-water basins in the Cantabrian Mountains of north-west Spain, at the Eurosiberian–Mediterranean limit. In particular, we evaluated the role of land-use history (comparing natural vs. anthropic basins) and microclimate (comparing shaded vs. sunny aspects) of forest gain/loss rates and spatial distribution shifts. Finally, we applied Species Distribution Modelling techniques (MaxEnt and BIOMOD) in the stated scenarios of land-use history and microclimate, to assess habitat suitability for forest expansion on the basis of topography, soil properties and mesoclimatic variables.
  3. Forest cover increased from 10.72% in 1956 to 27.67% in 2004 in the area. The rate of expansion was significantly higher in natural basins and, particularly, on shaded slopes. In all cases, the mean elevation of new forest patches increased during the study period, which was particularly evident on natural sunny slopes. The performance of the models and the magnitude of the effects varied across land-use histories and microclimatic conditions. Soil properties and temperature and precipitation in late spring and early summer were the main drivers of forest expansion in modelling exercises, although expansion rates and upward altitudinal shifts were primarily controlled by land-use history and the biogeographic origin of the forests.
  4. Synthesis. The combination of monitoring and modelling techniques used in this work contributed to the understanding of forest expansion in cultural systems, indicating that ecological succession is not a homogeneous process, but varies spatially due to human and abiotic constraints since historical times.

Introduction

Mountainous territories in the Northern Hemisphere, with a long history of human occupation, have undergone important changes in traditional management practices in recent decades (Lepart & Debussche 1992; MacDonald et al. 2000). Socio-economic adjustments, such as those linked to the EU Common Agricultural Policy (CAP), have led to a dramatic rural exodus and subsequent abandonment of agricultural land, a cessation of coppicing and a reduction in grazing in natural communities (Debussche, Debussche & Lepart 2001; Benayas et al. 2007). As a consequence, natural vegetation regeneration processes have been reactivated (Peñuelas & Boada 2003; Vicente-Serrano, Lasanta & Romo 2004), causing a widespread forest expansion (MacDonald et al. 2000; Capitanio & Carcaillet 2008). This process of secondary ecological succession exhibits a predictable sequence of change, with different species successively gaining and losing predominance (Suárez-Seoane, Osborne & Baudry 2002; Roder et al. 2008). Shrub vegetation develops a few years after abandonment, including both light-demanding species as well as seedlings of shade-tolerant trees. Later, trees develop to form a closed-canopy and early-stage species disappear (Calvo, Tarrega & de Luis 1999).

In abandoned agricultural lands, regeneration patterns can vary spatially with land-use history and climate, making it difficult to disentangle their individual effects (Randall & Pickett 1994; Gimeno et al. 2012). For example, recurrent fire events or grazing pressure may increase soil erosion, depleting soil nutrients and altering the competitive environment, hence forestalling forest expansion. In turn, warmer temperatures, higher evapotranspiration and drier soils may drive forest expansion to higher elevations and latitudes (Peñuelas et al. 2007a). The evaluation of the spatially explicit interactions between both factors is essential to understand distribution shifts because the way in which land-use history modifies landscape patterns may be influenced (reinforced or constrained) by climate or viceversa. Although there is a substantial body of literature on the impacts of both climate and land-use history on species distribution, or community shifts (Thuiller 2003; Araújo et al. 2005; Verburg et al. 2006; Morin & Lechowicz 2008; Barbet-Massin, Thuiller & Jiguet 2012), the influence of their interactions remains unsolved (Dale et al. 2000; Yates et al. 2010; Dale, Efroymson & Kline 2011). In this sense, modelling approaches can be useful tools for understanding the past to present changes of the species and community assemblages, which may then be related to the ecological processes driving the observed changes (Verburg et al. 2002; Álvarez-Martínez, Suárez-Seoane & De Luis Calabuig 2011). This spatially explicit consideration of landscape dynamics may improve the understanding of non-equilibrium distribution patterns of species at wide geographical scales (Franklin 2010; Peterson et al. 2012).

However, to correctly understand the ecological complexity of heterogeneous territories, modelling techniques need to take advantage of long-term and reliable monitoring data (Thuiller et al. 2008). Otherwise, the magnitude of uncertainties could be so great that it may lead stakeholders to question the overall usefulness of predictions for solving management problems (Pearson et al. 2006; Araújo & New 2007; Hanspach et al. 2010). During recent decades, an increasing number of spatially explicit methodologies have been developed to track past to present landscape patterns and changes. Many of these methods are based on remote sensing (RS), such as semiautomatic image classifications (frequently from Landsat TM and ETM+ sensors data) and photointerpretation of historical aerial photographs (Treitz & Rogan 2004). These techniques have commonly been presented as easy tools for deriving land cover inventories providing regional data at different temporal scales with low collection effort. Nevertheless, medium-resolution satellite image classification represents a simplification of landscape complexity associated with a certain degree of confusion, usually due to misclassification (Lewis, Brown & Tatnall 2000; Álvarez-Martínez et al. 2010). Alternatively, the finer scale of the aerial photos compared with satellite imagery allows for developing reliable land cover maps (e.g. at a very detailed spatial scale in a GIS, even small shrubs can be identified), although it requires the application of time-consuming and non-automatic procedures. These photographs are available since the 1950s, being therefore extremely useful for assessing long-term distribution patterns of vegetation at decadal timescales (Carmel & Kadmon 1998). In any case, whatever the origin of the information, spatially explicit time series of land cover data allow for explaining patterns and processes of landscape change (Serra, Pons & Saurí 2008; Álvarez-Martínez, Suárez-Seoane & De Luis Calabuig 2011), enabling landscape managers to design large-scale conservation strategies (Zavala & Burkey 1997).

The overarching objective of this work was to analyse the ecological mechanisms behind forest expansion over space (in a set of head-water basins located throughout La Sierra de Ancares Natural Park) and time (during the second half of the twentieth century), by following an approach based on monitoring and modelling. The study region is particularly well-suited because human activities over the last centuries have resulted in the destruction and fragmentation of the original forest cover. Oppositely, recent land abandonment is allowing regeneration processes through secondary succession. In this two-sided scenario, we explicitly evaluated the combined effects of land-use history (comparing natural vs. anthropic basins) and microclimate (comparing shaded vs. sunny slope aspects) in regard to gain/loss rates and distribution shifts of forests, particularly across altitudinal gradients. Additionally, we modelled spatially forest expansion on the basis of three families of potential drivers (topography, soil and mesoclimate) for the aforementioned scenarios of land-use history and microclimate, both at the basin and regional scales.

Materials and methods

Study area: defining a set of head-water basins

La Sierra de Ancares is a Spanish Natural Park located in the Cantabrian Mountains (Fig. 1), covering approximately 100 000 ha. Elevation ranges from 600 to 2200 m above sea level and relief is moderate to steep. Climate is Atlantic, with a mean annual precipitation of 1300 mm and a mean temperature of 8 °C, but it shows Mediterranean characteristics at lower elevations and latitudes (Rivas-Martinez & Rivas-Saenz 1996). During the twentieth century, a non-significant trend of warming and rainfall variability was detected from a long time series of data provided by the Spanish Meteorological Institute (Fig. 2). This region constitutes the south-westernmost distribution limit of several species of Eurosiberian trees, including beeches (Fagus sylvatica), oaks (Quercus petraea, Q. robur) and birches (Betula spp). At lower latitudes and elevations, especially in areas with higher insolation, vegetation is typically Mediterranean, with dominance of Quercus pyrenaica. A history of over-exploitation through cultivation and grazing, coppicing for charcoal or wood extraction for building and heating resulted in the contraction and fragmentation of the original forest cover. Conversely, during recent decades, rural depopulation and a decrease in human pressure have resulted in the disappearance of traditional management practices. Although deliberate burning still occurs in some areas, land abandonment is allowing for vegetation recovery in old fields, as described in many Mediterranean mountain areas (Poyatos, Latron & Llorens 2003; Pueyo & Beguería 2007).

Figure 1.

Set of 20 head-water basins selected as a study area in the Natural Park of La Sierra de Ancares (Cantabrian Mountains, NW Spain).

Figure 2.

Climatic trends in La Sierra de Ancares during the twentieth century (data provided by the Spanish Meteorological Institute; own elaboration). (a) Monthly temperatures from 1991 to 2006 (grey bars indicate data gaps), (b) annual rainfall and (c) variation coefficient of precipitation from 1974– 2006.

The boundaries of all head-water basins of La Sierra de Ancares were delineated using hydrological GIS tools over a Digital Elevation Model of 5-m of spatial resolution (5-m DEM; ITACYL 2012), assuming a minimum contributing area of 100 ha. To identify groups of basins with different land-use history, we ran a principal component analysis (PCA) on the basis of two variables obtained from Álvarez-Martínez et al. (2010): (i) dominant land cover (i.e. forests or heathlands) in 2004 and (ii) human management during the previous decade (i.e. recurrence of fire events larger than 1 ha). We then selected a subset of 20 basins (Fig. 1) that maximized the differences between the two variables. These basins were homogeneously distributed across the Natural Park and occupied a total area of ca. 6000 ha. Ten of them were labelled as ‘natural basins’, located in general at higher latitude and elevation, with limited human management over the last few decades and well-conserved mature forests, covering around 50% of the total area. The other 10 were labelled as ‘anthropic basins’ and consisted of intensively human-managed basins, closer to population settlements, affected by recurrent fire events and covered by a matrix of heathlands that included some forest patches. Currently, there is no human activity in any of these 20 basins, with the exception of residual grazing in some valley bottoms and subalpine pastures.

Because microclimate varies markedly between different slope aspects, each basin was subsequently split into two parts on the basis of the incoming solar radiation. Despite the fact that insolation drives many physical and biological processes, such as snow melt patterns, soil moisture, evapotranspiration and light available for plant photosynthesis (Häsler 1982; Valladares et al. 2008; Millington et al. 2009), it has been poorly assessed as a factor contributing to plant responses to environmental change (Grace, Berninger & Nagy 2002; Mouillot, Rambal & Joffre 2002; Pueyo & Beguería 2007). We derived an annual incoming insolation model from the 5-m DEM to be used as a comprehensive indicator of microclimatic conditions (Fu & Rich 2002). Output values vary spatially depending upon latitude, elevation, topography, sun angle and atmospheric effects. The highest values of insolation were assigned to sunny slopes and the lowest to shady slopes, with an average threshold for the 20 basins of 1.295 ± 0.143 MWh m−2 year−1. In summary, according to Quinn and Keough (2002), we defined a two-factor (land-use history and microclimate) analysis with 10 replicates for each case.

Monitoring changes in forest distribution

A total of 348 aerial photographs (scales ranging from 1:10 000 to 1:30 000) for the years 1956 (American flight), 1974, 1983 and 1990 were provided by the Regional Government of Castilla y León. All photographs were scanned at 600 dpi, orthorectified and projected into a common UTM grid. Digital orthophoto images at a 1:5000 scale were also obtained for 2004 from ITACYL (2012). The whole time series was coregistered across years with a RMSE smaller than 2 m. One hundred land cover maps (20 basins, 5 years) were then created at a detailed spatial scale by on-screen digitalization. Eight major categories were recognized, but only broadleaf forests (forests, hereafter) were retained for further analyses. To ensure that patch boundaries matched across the time series, we first mapped land covers for 2004 at a very detailed spatial scale. Then, we edited these maps, deleting or adding boundary lines, to derive new land cover maps for the previous years of the time series. Therefore, we ensured that boundary lines would be identical when landscape does not change for a particular time span, not accounting for false positives during change detection. To validate the 2004 land cover maps, intensive fieldwork was carried out from 2004 to 2008, amending boundary lines and attributes of land cover patches when necessary. Maps were handled as GIS vector layers for calculations of areas and expansion rates.

Changes in forest distribution were then spatially-explicitly analysed through a post-classification comparison (Lambin 1999). All pairwise combinations of maps were studied using transition matrices, which allowed an assessment of the nature, magnitude and direction of changes (Álvarez-Martínez et al. 2010). For each basin and time span, we quantified: (i) percentage of area covered by forest (FO), calculated as the ratio between the area occupied by forest (AF) and the total area (AT) (eqn 1); (ii) percentage of forest change (FOCH), estimated as the difference between forest cover in the year t (FOt) and the previous year t−1 (FOt−1) (eqn 2); (iii) annual rates of forest change (FOAR), calculated as the ratio between the percentage of forest gained or lost (FOCH) and the number of years of the considered time span (T) (eqn 3).

display math(eqn 1)
display math(eqn 2)
display math(eqn 3)

To estimate eventual upwards altitudinal shifts of forests, we overlapped the new forest patches detected at each time span with the 5-m DEM. As the 20 basins were located at different elevations throughout the Natural Park (Fig. 1), the average elevation of new forest patches was centred at each time span by subtracting the average elevation of each basin.

To detect significant differences (< 0.05) in forest gain/loss rates and altitudinal shifts for the scenarios of land-use history and microclimate, we conducted (for each time span) two-way anovas with Bonferroni correction to counteract the problem of multiple comparisons. The majority of data fulfilled both normality (Kolmogorov–Smirnov and Shaphiro–Wilk tests) and homoscedasticity (Levene's test) criteria.

Driving forces of forest expansion

A comprehensive GIS-data base including factors potentially driving forest expansion (mesoclimate, topography and soil properties; 58 variables) was created according to our knowledge of the study area (Álvarez-Martínez et al. 2010; Álvarez-Martínez, Suárez-Seoane & De Luis Calabuig 2011) and other mountain landscapes (e.g. Peñuelas & Boada 2003; Peñuelas et al. 2007a). We avoided any aprioristic selection/rejection.

  1. Fifty-two annual and monthly mesoclimatic variables representing minimum, maximum and mean temperature and rainfall, annual rainfall variability (coefficient of variation of monthly values) and thermal amplitude (i.e. maximum difference between extreme monthly temperatures) were extracted from a climatic data set at a 200-m resolution for the Iberian Peninsula (Ninyerola, Pons & Roure 2007).
  2. Three topographical variables accounting for elevation (indirectly determining temperature and rainfall), slope (accounting for water and nutrient availability in the soil) and curvature (calculated as the second derivative of the surface, indicating whether a given area is convex or concave, which is also related to solar radiation and soil moisture) were calculated from the 5-m DEM.
  3. Three soil property variables were derived from a soil map created ad hoc for the 20 head-water basins. One hundred soil samples (1 kilo each), five for each watershed, were taken from the top 20 cm of soil and analysed for organic matter content, pH and sand percentage. To exclude short distance variability of the soil, a mixed sample was taken from four points in an area of approximately 100 m2. Samples were then analysed in the laboratory, after being air-dried for a couple of days and sieved using a 2 mm sieve, to separate the mineral fraction from gravel, stones and roots. Topsoil organic matter content (SOM) was assessed by loss of the humidity upon oven-drying (105 °C) and ignition at 550 °C for 3 h (Howard & Howard 1990). pH was determined with a soil-to-solution ratio of 1:2 (Hendershot, Lalande & Duquette 2007). Sand percentage was assessed with the soil density method (Benbi, Brar & Kumar 1996). Soil data were analysed using regression models for establishing relationships between soil properties and physical attributes (elevation, Topographic Wetness Index TWI, land cover, fire recurrence and geological unit). Land cover and fire recurrence were obtained from satellite image classification (Álvarez-Martínez et al. 2010) and geological units from IGME (1971). TWI was calculated from the 5-m DEM to assess soil moisture (eqn 4). Seventy-five per cent of the samples were used as training data and the remaining 25% as validation data.
display math(eqn 4)

Area represents the contributing catchment in m2 and β is the slope angle in degrees (Wilson & Gallant 2000). High TWI values indicate shallow slopes and large contributing areas and, thus, a higher probability of soil water saturation.

To avoid statistical problems due to multicollinearity (i.e. variance inflation and parameter bias; MacNally 2000; Freckleton 2011), we checked Pearson bivariate correlations among the pool of 58 predictors. The nine best explanatory variables (i.e. with more than a 10% contribution in exploratory univariate Maxent models, the heuristic estimate of the relative contributions of predictors; Phillips, Anderson & Schapire 2006) were retained from each correlated pairwise (r > 0.7; Randin et al. 2006) for further analyses.

Distribution modelling

We modelled forest expansion for all possible scenarios of land-use history and microclimate. To ensure a sufficient representation of the environmental heterogeneity of the study area, we calibrated the models using a random sample of ca. 5000 points that covered the set of 20 basins under study (6000 ha). Sample size was defined after several tests on the effects of background sample size on model structure, avoiding high spatial autocorrelation (Diniz-Filho, Bini & Hawkins 2003). One thousand of points corresponded to new forest patches and were labelled as positive cases (i.e. those present in 2004 but not in 1956). The remaining 4000 points were negative cases (i.e. locations not covered by forests from 1956 to 2004). The number of positive and negative cases varied among the scenarios, but a prevalence of ca. 1/5 (i.e. the ratio of the number of presences to the total number of data points used in model training) was always guaranteed (Jiménez-Valverde, Lobo & Hortal 2008). Jimenez-Valverde & Lobo (2006) argued that even if a prevalence of 0.5 has been extensively recommended in SDM, unbalanced data can provide equally good models if we choose the right predictors and avoid low-quality data affected by false absences, low sample size or unrepresentativeness of the environmental and spatial gradient. Additionally, despite the potentially higher omission rates in models developed for samples with low prevalence (Cramer 1999; Hosmer & Lemeshow 2000), this low-prevalence data set may produce a more conservative and reliable estimate of forest expansion (Álvarez-Martínez, Suárez-Seoane & De Luis Calabuig 2011), acknowledging that overestimation is frequent in land cover change assessments of heterogeneous landscapes (Bradley & Mustard 2005; Álvarez-Martínez et al. 2010). Finally, to state the purpose of the model (i.e. to predict forest expansion, but not persistence; Pontius & Pacheco (2004)), the areas covered by forests in 1956 were eliminated from training data.

We then applied two different SDM techniques for modelling forest expansion under each scenario of land-use history and microclimate: MaxEnt (Phillips, Anderson & Schapire 2006; Phillips & Dudík 2008) and BIOMOD (Thuiller 2003; Thuiller et al. 2009). SDM are correlative approaches commonly used to test or generate hypotheses about mechanisms controlling spatial patterns of species distributions at different scales (Franklin & Miller 2009; Peterson et al. 2012). One of their main achievements is to generate continuous maps of habitat suitability on the basis of the relationship between species occurrence data and environmental predictors (Araújo & New 2007). Models are usually calibrated with data on species occurrence (presence/absence) and environmental limiting factors, coarse-grained layers collected at present or averaged from past time spans that do not change over time or change very slowly at geographical scales (as climate, topography and soil properties; Elith et al. 2006; Elith, Kearney & Phillips 2010).

The maximum entropy method (MaxEnt) is one of the best-performing algorithms for modelling species distribution (Elith et al. 2011), despite certain limitations (Haegeman & Loreau 2008). We ran this algorithm on the full training data set to provide the best estimate of forest distribution for visual interpretation (full models). We evaluated model performance and variable contribution through a fivefold cross-validation on the training dataset (Verbyla & Litvaitis 1989). Both full and cross-validated models were evaluated by means of the area under the receiver operating characteristic (ROC) curve (AUC; Pontius & Schneider 2001). Continuous outputs were converted into Boolean maps of suitable/unsuitable areas for forest expansion using the ‘equate entropy of thresholded and original distributions’ value (Phillips, Anderson & Schapire 2006; Morán-Ordóñez et al. 2011). Spatial maps were obtained for the 20 basins and further extrapolated to the whole Natural Park, to allow a comprehensive evaluation of forest expansion patterns.

BIOMOD is a mixed method combining a range of statistical techniques for examining the species-environment relationships: Generalized Linear Models (GLM), Generalised Additive Models (GAM), Classification Tree Analysis (CTA), Artificial Neural Networks (ANN), Surface Range Envelope (SRE), Generalized Boosting Model (GBM), Breiman and Cutler's random forest for classification and regression (RF), Mixture Discriminant Analysis (MDA) and Multiple Adaptive Regression Splines (MARS). Models were run under default settings and parameters (Thuiller et al. 2009). Using a permutation procedure, we assessed the relative importance of each predictor across models, which is a difficult task given that each model relies on different algorithms, techniques and assumptions (Triviño et al. 2011). The multimodelling approach of BIOMOD provides a complementary sensitivity analysis to MaxEnt about the trend and magnitude of the most relevant variables driving forest expansion (Fang, Gertner & Anderson 2007; Álvarez-Martínez, Suárez-Seoane & De Luis Calabuig 2011; Alonso, Magaña & Álvarez-Martínez 2012).

GIS analyses were done in ArcGIS 10.1 and Orthobase Erdas imagine 8.5. We used IBM spss 19 and r 2.11.1 (R Development Core Team 2010) for statistics.

Results

Monitoring changes in forest distribution: the effects of land-use history and microclimate

Secondary succession has been the dominant process during the second half of the twentieth century in the study area. On average, forests covered 10.72 ± 12.88% of the basins in 1956, reaching 27.67 ± 24.10% in 2004 (Table 1 and Fig. 3). However, the percentage of forest differed noticeably between natural and anthropic basins. In natural basins, forests occupied 20.31 ± 11.97% in 1956, rising 47.53 ± 17.38% in 2004, while in anthropic basins, the increase was from 1.12 ± 1.57% to 7.80 ± 6.93%. In turn, shady slopes showed higher forest cover than sunny slopes, even when considering natural and anthropic basins independently (Table 2). Forest expansion rates were always above the general trend in natural basins, while they were constantly below in anthropic basins, with significant differences between them for all time spans (Table 3 and Fig. 4a). Additionally, expansion rates increased progressively during the four time spans investigated, with an R2 of 0.87 and 0.58 for natural and anthropic basins, respectively. The difference between basin types also increased through the study period with an R2 of 0.40 (Table 1 and Fig. 4a). In turn, differences between slopes become significant only during the most recent decades. The interaction between climate and land use was never statistically significant.

Table 1. Annual rates of forest change (ARCHF) for each basin and time span (see Fig. 1). Forest cover (FO) for 1956 and 2004, averaged forest increase rates for the whole period of 1956–2004 and basin area are also shown
Basin numberBasin area (ha)% Forest cover in 1956Annual rates of forest change (% gain/loss per year)% Forest cover in 2004Average rate of forest increase (% gain per year)
1956–19741974–19831983–19901990–20041956–2004
1293.0720.480.771.111.140.6761.590.82
2206.9410.460.32−0.160.491.2635.820.51
3295.3322.410.330.180.670.7945.710.47
4378.7417.900.310.280.950.2536.250.37
5314.233.620.320.501.250.8434.430.62
6301.5625.210.070.680.250.4941.260.32
7691.0517.200.400.570.610.6943.510.53
8747.2814.940.210.350.280.8034.940.40
9147.0449.270.960.931.040.6090.580.83
10221.3721.560.230.720.860.9251.170.59
Natural (mean ± SD)359.66 ± 200.5820.31 ± 11.970.39 ± 0.270.52 ± 0.370.75 ± 0.350.73 ± 0.2747.53 ± 17.380.54 ± 0.17
11277.870.600.270.780.370.5522.640.44
12242.895.150.16−0.160.090.2811.100.12
13118.440.310.010.07−0.070.061.440.02
14182.982.300.12−0.12−0.130.073.520.02
15301.570.480.040.160.030.104.160.07
16522.291.180.030.040.210.6212.260.16
17322.340.160.03−0.020.110.285.280.10
18342.671.040.170.330.160.4013.710.25
19119.250.000.010.000.010.091.490.03
20304.340.040.010.040.060.112.450.05
Anthropic (mean ± SD)273.46 ± 119.161.12 ± 1.570.08 ± 0.090.11 ± 0.270.08 ± 0.140.26 ± 0.217.8 ± 6.930.13 ± 0.13
All (mean ± SD)316.56 ± 166.5510.72 ± 12.880.24 ± 0.250.31 ± 0.380.42 ± 0.430.49 ± 0.3427.67 ± 24.10.34 ± 0.26
Table 2. Forest cover (%) for each slope aspect (i.e. different microclimate) in basins with different land-use history. The average annual rate of forest increase for the whole period of 1956–2004 is also shown
Microclimate (Slope aspect)Land-use history% Forest cover (FO) (mean ± SD)Annual rate of forest increase (% gain per year)
195619741983199020041956–2004
Sunny slopesNatural5.27 ± 4.378.23 ± 6.2310.28 ± 7.9312.04 ± 9.3616.31 ± 10.810.22 ± 0.17
Anthropic0.30 ± 0.430.62 ± 0.931.15 ± 1.641.19 ± 1.721.99 ± 2.480.03 ± 0.05
All2.79 ± 3.954.43 ± 5.835.71 ± 7.286.62 ± 8.69.15 ± 10.60.13 ± 0.15
Shady slopesNatural15.05 ± 9.1719.16 ± 10.2821.77 ± 10.725.29 ± 10.3631.30 ± 9.180.32 ± 0.1
Anthropic0.82 ± 1.252.00 ± 2.162.48 ± 2.523.03 ± 3.035.82 ± 4.660.10 ± 0.09
All7.94 ± 9.6910.58 ± 11.3912.12 ± 12.4614.16 ± 13.6218.56 ± 14.870.21 ± 0.15
Table 3. Results of two-way anova with Bonferroni correction for comparing: (a) forest increase rates and (b) average elevation occupied by new forest patches, in natural-anthropic basins and sunny-shady slopes, as well as their interaction, for each time span
 1956–19741974–19831983–19901990–20041956–2004
F-testP-valueF-testP-valueF-testP-valueF-testP-valueF-testP-value
  1. Significant differences are shown in bold. The averaged elevation of new forest patches for each basin and time span was standardized by subtracting the average elevation of each basin before running the statistical tests.

(a) Forest increase rates
Land-use history (natural vs. anthropic basins)18.64 0.00 12.67 0.00 27.6 0.00 19.4 0.00 35.59 0.00
Microclimate (sunny vs. shady slopes)2.440.130.240.636.47 0.02 5.99 0.02 6.07 0.02
Interaction microclimate * land use0.050.820.360.551.920.180.030.870.30.59
(b) Average elevation of new forest patches
Land-use history (natural vs. anthropic basins)1.040.320.220.640.930.341.640.210.230.63
Microclimate (sunny vs. shady slopes)3.760.061.760.191.320.261.540.221.80.19
Interaction microclimate * land use0.040.841.230.280.10.750.340.560.20.65
Figure 3.

An example of two time series of land cover maps: (a) for a natural basin (number 9, Fig. 1) and (b) an anthropic basin (number 16, Fig. 1).

Figure 4.

(a) Annual rates of forest expansion. (b) Averaged elevation of new forest patches for each time span, land-use history and microclimate (i.e. slope aspect).

For any time span, the new patches of forest were always located at higher elevations in natural basins and sunny slopes (Fig. 4b). In anthropic basins, the effect was much smaller, with no major trends. However, none of these differences were significant (Table 3).

Modelling forest expansion: relevant driving forces

MaxEnt models achieved a consistently high AUC for both calibration and evaluation datasets, with values ranging from 70.8% to 91.1% (Table 4). AUC values were always higher in sunny than in shady slopes, being maximal in anthropic basins.

Table 4. AUC values of full and fivefold cross-validated (5CV) MaxEnt models
Land-use historyNaturalAntrhopicAll
Microclimate (slope)BothSunnyShadyBothSunnyShadyBothSunnyShady
Full Models75.4081.5070.8087.9091.1083.3080.4086.2077.10
5 CV75.4481.1671.4083.4485.3175.3680.3184.9776.73

Table 5 shows the relative importance of the nine predictors for modelling forest expansion. The most relevant drivers across management and microclimate scenarios were related to mesoclimate and soil properties. In natural basins, the percentage of sand of the topsoil, maximum temperature in June, annual thermal amplitude and coefficient of variation of annual rainfall were the most contributing variables. In anthropic basins, together with June temperature, rainfall variables become much more important. When comparing the most relevant drivers for different slope aspects, we found only slight differences unless we analysed both basin types independently. In natural basins, May rainfall and annual thermal amplitude were more influential in sunny slopes, whereas the maximum temperature in June and the coefficient of variation of rainfall had more relevance in shady slopes. In turn, annual thermal amplitude was more important in anthropic shady than in sunny slopes. On the contrary, December rainfall and maximum temperature in June were more important in sunny slopes. In regards to the sign of the effects, there were more important differences between basins with different land-use history than between slope aspects. Figure 5 shows how the probability of forest expansion increases with higher values of May rainfall in natural basins, but decreases in anthropic basins. Complementarily, Fig. 6 illustrates, as modelled with MaxEnt, that natural basins had greater suitability for forest expansion than anthropic basins, as shady more than sunny slopes. We point out here the scaling mismatch between the coarser resolution of climatic layers, set at a 200-m pixel, against the 5-m of drivers related to the physical environment. This is inevitable due to data availability, as occurred in similar studies (Wear & Bolstad 1998; Burgi, Hersperger & Schneeberger 2004; Álvarez-Martínez, Suárez-Seoane & De Luis Calabuig 2011).

Table 5. Relative importance of environmental predictors for forest expansion in four different scenarios: within natural and anthropic basins, additionally split into sunny and shady slopes
 Sunny slopesShady slopes
MaxEntBIOMODMaxEntBIOMOD
  1. For MaxEnt models, we show the sign of the effect for each variable within the environmental range of variability of study area: (+) Positive effect, (−) Negative effect, and its relative contribution averaged from fivefold cross validation. For BIOMOD, we show the averaged value ± SD of variable importance obtained from the nine algorithms. For visualization help, dark grey boxes indicate predictor relative importance >20% and light grey boxes 15–20%. Variable codes: Terrain slope (Slope), terrain curvature (Curv), Topographic Wetness Index (TWI), percentage of sand in the topsoil (Sand), rainfall in May (RainMay), rainfall in December (RainDec), coefficient of variation of rainfall (RainCV), maximum temperature in June (TmaxJun), annual thermal amplitude (TAmpl).

Natural basins
Slope−4.04.2 ± 1.8−2.61.9 ± 1.5
Curv−0.71.1 ± 2.6+0.10.5 ± 0.3
TWI+0.61.2 ± 1.9+0.00.4 ± 0.5
Sand−24.921.8 ± 3.6−24.617.4 ± 7.5
RainMay+15.316.8 ± 3.8+10.114.3 ± 7.7
RainDec−1.75.7 ± 3.8−2.09.8 ± 3.6
RainCV−16.56.9 ± 4.7−23.214.4 ± 9.3
TmaxJun+13.122.5 ± 5.9+27.128.8 ± 5.4
TAmpl−23.319.9 ± 6.0−10.412.5 ± 2.6
Anthropic basins
Slope+9.23.7 ± 3.8+11.38.9 ± 9.1
Curv−0.00.2 ± 0.2−2.02.1 ± 3.0
TWI+0.81.1 ± 1.3+6.03.1 ± 2.3
Sand−9.215.9 ± 6.3−9.711.9 ± 3.2
RainMay−22.436.2 ± 14.3−25.324.1 ± 10.8
RainDec+22.512.5 ± 5.6+7.110.3 ± 5.3
RainCV−6.89.9 ± 6.1−9.012.3 ± 5.8
TmaxJun+21.516.5 ± 12.9+13.117.9 ± 7.4
TAmpl−7.73.1 ± 4.1−16.69.4 ± 3.6
Figure 5.

Variable responses (i.e. May rainfall) to forest expansion in MaxEnt models, calibrated with data from both basins and slopes.

Figure 6.

(a) MaxEnt model outputs of habitat suitability for forest expansion calibrated with data from both basins and slopes, compared with current forest cover of the year 2004 derived from satellite imagery (Álvarez-Martínez et al. 2010), in a natural basin (number 5, Fig. 1) (left) and an anthropic basin (number 18, Fig. 1) (right). (b) Sunny and shady slopes of the anthropic basin number 4 (Fig. 1) (left), with its corresponding MaxEnt model of suitable habitat for forests. Maps show a relationship between shady slopes and high habitat suitability.

Finally, Fig. 7 shows the extrapolation of model predictions across the entire Natural Park. The higher and northernmost areas were more suitable for the expansion of Eurosiberian forests, mainly present in natural basins (Fig. 7a), while low and southernmost areas (more insolated and human-managed) were more suitable for Mediterranean sclerophyllous vegetation, dominant in anthropic basins (Fig. 7b). The visual comparison between the two model outputs with the forest maps developed by image classification in Álvarez-Martínez et al. (2010) (Fig. 7c) highlights a good matching at a regional scale, mainly in Eurosiberian forests. Areas with higher altitudes and slopes, with dispersed rocky areas, were less suitable for forest expansion. Nevertheless, further research is needed to assess the transferability between the two scales.

Figure 7.

Suitability maps of forest expansion in La Sierra de Ancares Natural Park, extrapolated from MaxEnt models calibrated with: (a) natural basins, both slopes, and (b) anthropic basins, both slopes. (c) Forest cover for the year 2004 obtained from image classification in Álvarez-Martínez et al. (2010). Observe that valley bottoms, suitable for forest expansion, remain currently occupied by hedged meadows and croplands.

Discussion

Forest expansion has been coincident in the Cantabrian Mountains with other European landscapes since the beginning of the twentieth century, being mainly linked to the disappearance of traditionally extensive livestock farming and agricultural systems (Jongman 2002; Laiolo et al. 2004; Lasanta et al. 2006). The ongoing process of land abandonment has implied the transformation of large areas of open grasslands into heathlands and woodlands (Vicente-Serrano, Lasanta & Romo 2004; Morán-Ordóñez et al. 2011). These habitats offer new ecosystem services, such as sequestering carbon from the atmosphere, protecting upstream watersheds and soil formation, providing habitat for species and meeting increased demands for landscape beauty (FAO 2012; Morán-Ordóñez et al. 2013). In contrast, other areas remain open because of residual human land use, higher elevation, slope or poor soil conditions, preventing woody species from growing or reducing their spread rates.

In the study area, the more suitable conditions for sprouting and re-growth of tree species in natural basins allowed for a larger and more intensive expansion of forests than in anthropic basins. Therefore, the spread of new forest patches could be primarily explained by historical factors, such as the kind of former woodlands (i.e. primary and secondary) present nearby and their age, as well as the timing of site abandonment. Nevertheless, we highlight here the apparent dampening, but not elimination, of forest expansion by human land-use. Foster, Motzkin & Slater (1998) explained that land abandonment usually starts closer to forested areas on upper slopes and, eventually, may affect productive land in valley bottoms and less pronounced slopes. This would explain the more rapid forestation process of natural basins, which have both a steeper topography and older forests on their slopes and valley bottoms, due to far fewer human disturbances such as burning, agricultural practices and logging activities since historical times. Nevertheless, forest expansion started in all cases from roughly the same altitudinal level (i.e. small patches close to valley bottoms in 1956; see an example in Fig. 3), although new forest patches reached higher elevations on sunny slopes, regardless of whether this was measured in natural or anthropic basins. This issue could be explained by the interplay of climate and land use.

Altitudinal shifts in species distributions have been described for many tree species during this century as a response to changes in climatic conditions (Huntley 1991; Grace, Berninger & Nagy 2002; González et al. 2010). In Spain, Peñuelas and Boada (2003) monitored a 70-m upward shift in beech forests during the last five decades. Kelly and Goulden (2008) found that the average elevation of dominant plant species in California's Santa Rosa Mountains (USA) rose by 65-m between surveys of 1977 and 2006–2007, highlighting a synchronous response that spans across communities rather than focusing on the role of individualistic responses of vegetation to climate change (Breshears et al. 2008). However, other studies have shown certain sluggishness in upward tree line shifts as a response to global warming (Peterson 1998; Peñuelas et al. 2007a). Schwilk and Keeley (2012) stated that species-specific responses to climate change are probably much more complex than simple models predicting vegetation shifts in cooler and/or wetter locations. They demonstrated that, at least for some species, recently reported distribution shifts appear to be an artefact of disturbance history, but not evidence of climate warming. In the Swiss Alps, Colombaroli et al. (2010) reconstructed local fire variability and vegetation dynamics over the last 12 000 years, determining that intensified land use coupled with fire occurrence since the Bronze Age (c. 4000 cal. years bp) had a greater impact on community composition near the tree line than climate change. It is therefore interesting to point out here that the lack of larger differences between slope (and basin) types could be associated with the inherent variability of the system. In this study, our results suggest that distribution shifts caused by climate in human-dominated landscapes are difficult to measure independently because of the masking by land use and related disturbances, such as fire events or agricultural activities. This may disguise pure climatic effects on landscape changes (Fuller et al. 1998; Vicente-Serrano, Lasanta & Romo 2004). Therefore, the upward expansion of forests may not be related to pure climatic effects even in the natural basins, but should be also understood because of human and other abiotic constraints on forest ecotones.

Nevertheless, the distributional patterns of vegetation across the landscape need to be analysed as a first step under the effect of microclimate. The warmer and drier conditions of sunny slopes may hamper forest expansion by reducing seedling survival and sprouting in spring and summer (Pigott & Pigott 1993; Valladares et al. 2008). Therefore, upward expansion of forests on sunny slopes could be understood as a climatic adaptive strategy for compensating thermo-pluviometric deficits in lower areas. Nevertheless, we highlight here the interplay of climate and land use in driving the more intensive upward shifts of new forest patches on sunny slopes. The lower parts of sunny slopes, more accessible and closer to crops of valley bottoms, should have been subject to more fire recurrence and farming activities in the past, causing widespread soil erosion and fertility depletion (Stoorvogel & Smaling 1998; Arnaez et al. 2011). Thus, the combination of more rigorous climatic conditions with poorer soil properties may force trees to shift upward to compensate for both negative effects with more intensity than in shady slopes. According the observed trends, we may expect greater differences in upward shifts between slope aspects during forthcoming decades due to this dynamic interplay of land-use history and abiotic constraints.

Topography was another interesting factor related to the distribution of forests, mainly in anthropic basins. Steeper areas, frequently excluded from grazing or harvesting activities, maintained a higher suitability for forest expansion. However, in flatter areas, even if residual crops or cattle no longer exist, recurrent fire events for understorey management and historical farming activities may have prevented tree species colonization (Acácio et al. 2010). Consequently, the poorer pattern of soil quality in anthropic basins (i.e. sandier texture or lower organic matter content) may involve that water availability during the plant-growing period, which in turn may control actual evapotranspiration, depends more on precipitation and temperature than in the soil water retention capability (Peñuelas et al. 2007b). Therefore, the most important variables for forest expansion in anthropic basins were climatic. In contrast, in natural basins, sand percentage and climate had an equivalent relevance for forest expansion. In this case, soil pattern may be more heterogeneous, with the best soil patches (i.e. those located on lower shady slopes) associated with greater suitability for forest expansion.

The biogeographic origin of the dominant tree species, or plant diversity, also played an important role in forests altitudinal shifts. Eurosiberian species (more abundant in natural basins) are more sensitive to hot summer droughts than Mediterranean sclerophyllous vegetation (dominant in anthropic basins) (Moreno, Pineda & Rivas-Martinez 1990; Sardans & Peñuelas 2013). Therefore, forest expansion was more related to rainfall (water) availability in the former, while spring temperature had a much higher relevance in the latter. Some authors, such as Barclay and Crawford (1984) and Peñuelas et al. (2007a) have found that high late spring and summer temperatures would favour vegetation shifts as the production of viable seeds at high elevations fails more frequently, except in exceptionally warm years. If global temperatures continue to increase in the future, conditions that are more suitable will appear at higher elevations, even on shady slopes, eventually involving more intensive altitudinal shifts of the forests. Therefore, according to Fig. 7, Eurosiberian forests would tend to colonize primarily higher latitudes and elevations. As explained by Sardans & Peñuelas (2013), the long-term evolutionary adaptation to drought of some species of Mediterranean plants allows them to cope with moderate increases of drought without significant losses of production and survival. However, other species have been shown to be more sensitive, decreasing their growth and increasing their mortality under moderate increases in summer droughts. As a consequence, if climate change follows IPCC (2007) predictions, we would expect continuous vegetation shifts of Eurosiberian forests, as explained above. The lower gaps, where weather conditions will be warmer and drier, would be filled by Mediterranean trees (Peñuelas & Boada 2003; Peñuelas et al. 2007a,b).

However, the simplification of the reality caused by applying static approaches on SDM implies not considering eventual niche expansions, contractions or shifts in both fundamental but also realized dimensions (Stanton et al. 2011). In other words, static SDM are based on the underlying assumption that populations or communities (forests in this particular case) are in equilibrium with contemporary environmental conditions and, therefore, do not account for non-equilibrium situations (Holt & Keitt 2000; Pulliam 2000). To improve the reliability of these approaches, it could be highly relevant to deal with other ecological factors capable of changing at ecological time scales. These factors are usually set at a finer resolution and may control not only species existence at some point, but also migration rates, abundance, reproduction and persistence along the time (Kirkpatrick & Barton 1997; Araújo & Guisan 2006; Holt 2009). Further research is needed in this way to cast light on the relationship between realized and potential distributions of the species at geographic scales, a fundamental question in biogeography.

Conclusions

This study provides some clues for disentangling the combined effect of land use and climate variability on recent forest expansion in mountainous landscapes, paramount to an ecological, conservation and planning perspective. We determined that, even if abiotic constraints were relevant drivers, land-use history primarily controlled forest expansion rates, as well as upward altitudinal shifts. In fact, the high plant diversity that characterizes ecosystems at the boundary between Eurosiberian Mediterranean biogeographic regions is associated with the success of co-existing species with a legacy of land use, climatic and soil resources exploited differentially over space and time (Scarascia-Mugnozza et al. 2000; Sardans & Peñuelas 2013). Therefore, these factors should not be ignored in ecological studies; otherwise, biased conclusions may be reached, misguiding policy decisions (Hanspach et al. 2010). Finally, although our results are site-specific, conclusions could be generalized to other mountainous areas where the effects of global change on terrestrial ecosystems require scientific-based planning and locally tailored sustainable management strategies to maintain their cultural and ecological values (Jongman 2002; Lasanta et al. 2006).

Acknowledgements

J. M. Álvarez-Martínez was awarded by the Regional Government of Junta de Castilla y León (EDU/1490/2003) and the European Social Fund (ESF). Fieldwork was supported by the Regional Government of Junta de Castilla y León (LE039A05 and LE021A08) and the Spanish Government (CGL2006-10998-C02-01). We thank to Fire Ecology and Landscape Ecology research groups of the University of León (Spain) and Soil Geography and Landscape group of the University of Wageningen (The Netherlands), as well as J. Lobo, R. Anderson, J. Hortal, R. García Valdés, I. Pozo, D. Llusía and M. Triviño of National Museum of Natural Sciences (CSIC, Spain) for earlier discussions about modelling. Two anonymous reviewers helped to improve substantially the initial version of the manuscript.

Data accessibility

The following information is available as online data support: (i) Raw climatic data for trend analysis in La Sierra de Ancares. (ii) GIS data about the study area (i.e. natural and anthropic basins, sunny and shady slopes). (iii) Land cover maps from 1956 to 2004. (iv) Environmental variables used for modelling forest expansion. (v) Sampling data base for training the models. Data available from the Dryad Digital Repository (Álvarez-Martínez et al. 2014).

Ancillary