Landscape dynamics in Mediterranean oak forests under global change: understanding the role of anthropogenic and environmental drivers across forest types
Abstract
The Mediterranean region is projected to be extremely vulnerable to global change, which will affect the distribution of typical forest types such as native oak forests. However, our understanding of Mediterranean oak forest responses to future conditions is still very limited by the lack of knowledge on oak forest dynamics and species-specific responses to multiple drivers. We compared the long-term (1966–2006) forest persistence and land cover change among evergreen (cork oak and holm oak) and deciduous oak forests and evaluated the importance of anthropogenic and environmental drivers on observed changes for Portugal. We used National Forest Inventories to quantify the changes in oak forests and explored the drivers of change using multinomial logistic regression analysis and an information theoretical approach. We found distinct trends among oak forest types, reflecting the differences in oak economic value, protection status and management schemes: cork oak forests were the most persistent (62%), changing mostly to pines and eucalypt; holm oak forests were less persistent (53.2%), changing mostly to agriculture; and deciduous oak forests were the least persistent (45.7%), changing mostly to shrublands. Drivers of change had distinct importance across oak forest types, but drivers from anthropogenic origin (wildfires, population density, and land accessibility) were always among the most important. Climatic extremes were also important predictors of oak forest changes, namely extreme temperatures for evergreen oak forests and deficit of precipitation for deciduous oak forests. Our results indicate that under increasing human pressure and forecasted climate change, evergreen oak forests will continue declining and deciduous oak forests will be replaced by forests dominated by more xeric species. In the long run, multiple disturbances may change competitive dominance from oak forests to pyrophytic shrublands. A better understanding of forest dynamics and the inclusion of anthropogenic drivers on models of vegetation change will improve predicting the future of Mediterranean oak forests.
Introduction
Changes in land use and climate are occurring at increasing rates globally, and there is now ample evidence that these fast changes are affecting forest landscapes worldwide (Nelson et al., 2006; Mery et al., 2010; Azevedo et al., 2014a). Major land use changes, such as intensification of agriculture and forestry, often associated with fast population growth, have caused forest loss and fragmentation especially in tropical regions (DeFries et al., 2010; Lobovikov et al., 2010; Hosonuma et al., 2012). In contrast, abandonment of agriculture associated with rural depopulation has occurred most frequently in Europe, where it has been a source of either forest expansion in northern countries or land degradation and shrub encroachment in southern regions (Rey Benayas et al., 2007; Azevedo et al., 2014b). Climate change has been reported as the direct cause of shifts in forest species range in different geographic locations (Walther et al., 2002; Peñuelas & Boada, 2003; Kelly & Goulden, 2008), and for the increased frequency and intensity of disturbances, such as pathogen outbreaks (Harvell et al., 2002; Edburg et al., 2012), extreme weather events (Allen et al., 2010, 2015; Carnicer et al., 2011), and wildfires (Westerling et al., 2006), with the consequent widespread forest mortality across continents (Peterman & Bachelet, 2012; Anderegg et al., 2013, 2015; Allen et al., 2015).
Although the changes in forest landscapes are primarily driven by land use changes (Rounsevell et al., 2012 Azevedo et al., 2014b), it is becoming increasingly evident that land use practices and climate change often interact synergistically to enhance the impacts in forests across spatial, temporal, and organizational scales (Nelson et al., 2006; Allen et al., 2010). For instance, rural exodus and the abandonment of agricultural land in Mediterranean Europe with the subsequent accumulation of fuel loads led to a change in fire regime after the 1970s, with a significant increase in larger wildfires and forest area burned, likely amplified by increasing drought frequency over the past decades (Marques et al., 2011; Moreira et al., 2011; Pausas & Fernández-Muñoz, 2012).
In Europe, the Mediterranean region is projected to be the most vulnerable to global change in the long term (Schröter et al., 2005; Giorgi & Lionello, 2008; Lindner et al., 2010; Kovats et al., 2014), but potential trajectories for the future distribution of typical forest types such as native oak forests remain largely uncertain (Lindner et al., 2014). Some models predict a decrease in Mediterranean oak forest area due to extreme disturbances such as increased droughts and forest fires (Schröter et al., 2005), while others show range expansion of Mediterranean evergreen oak forests favored by the changes in mean climate variables, such as increasing temperature and decreasing rainfall (Hanewinkel et al., 2013; Ruiz-Labourdette et al., 2013; Carnicer et al., 2014). In addition, regional models reveal important oak species-specific differences in demography and vulnerability to climatic changes and other environmental drivers, with deciduous species becoming more vulnerable than evergreen ones (Purves et al., 2007; Benito-Garzón et al., 2008, 2013; Sánchez de Dios et al., 2009; García-Valdés et al., 2013). Nevertheless, oaks are late successional species and their ability to colonize new suitable habitats might be hampered by slow migration rates (Meier et al., 2012; Delzon et al., 2013; Lindner et al., 2014).
Changes in Mediterranean oak forest landscapes are strongly influenced by human activities (Bugalho et al., 2011; Marañón et al., 2012), which will likely affect the predicted trajectories for oak forests (Schröter et al., 2005). However, the role of human drivers has been clearly underestimated in most models of vegetation change (Sinclair et al., 2010; Rounsevell et al., 2012). Furthermore, few studies have investigated the role of multiple drivers in the past dynamics of Mediterranean oak forests, which limits our understanding of forest responses to future conditions (Heller & Zavaleta, 2009; Li et al., 2011). The exception is the recent research for cork oak (Quercus suber L.), which has shown that the interactions between wildfires, topography, and type of land management were important drivers of local landscape dynamics (Acácio et al., 2009, 2010; Costa et al., 2009, 2014; Godinho et al., 2014). However, no studies have included a comparison of drivers of land cover change across co-occurring oak forest types at a large spatial scale (e.g., for the dynamics of one type of oak forest at local or regional scales: Plieninger, 2006 and Plieninger & Schaar, 2008 for Quercus ilex; Kouba & Alados, 2012 for Quercus faginea).
In this study, we compare the long-term (1966–2006) forest persistence and land cover change among evergreen (cork oak and holm oak) and deciduous oak forests in a Mediterranean-climate country (Portugal), and evaluate the relative importance of anthropogenic and environmental drivers on observed land cover change at the national scale. We ask the following questions: (1) ‘Are there differences among oak forest types in persistence and land cover change over time across Portugal?’ (2) ‘What are the most important drivers affecting land cover change and do they have the same importance across oak forest types?’ We hypothesize that oak forest persistence and change over time across mainland Portugal and drivers of change are specific to each type of oak forest due to a combination of distinct management history, economic value, and vulnerability to disturbances and climate change (e.g., increasing temperature and decreasing rainfall, Miranda et al., 2002; Costa & Soares, 2009; Gallego et al., 2011; Ramos et al., 2011). We expect cork oak to be the most persistent forest due to a combination of management oriented to cork production with high revenues (see, for example, Urbieta et al., 2008) and good ecological resilience under climate change (Benito-Garzón et al., 2013; Ruiz-Benito et al., 2013). We also expect deciduous oak forests to be the least persistent due to timber exploitation and the lack of legal protection, combined with a higher vulnerability to climate change as previously suggested by different studies (Benito-Garzón et al., 2008, 2013; García-Valdés et al., 2013). Overall, we expect anthropogenic drivers to be more important than environmental ones for the persistence or change of all oak forest types considered.
Materials and methods
Study area: oak biogeography and landscape in Portugal
Mainland Portugal is located on southwestern Europe, on the Iberian Peninsula, between latitudes 37° and 42° N, with an area of approximately 9 million hectares. Climate is predominantly Mediterranean with a strong climatic and altitudinal gradient across the country. In the north, 95% of the area is above 400 m, while in the south, 62% of the area is below 200 m (Jones et al., 2011), with a trend of increasing temperature and decreasing precipitation from north to south and from coast (west) to inland (east). Mean annual temperature values vary between 7 °C in the inner highlands of central Portugal and 18 °C in the southern coastal region. Mean annual precipitation has the highest values (above 3000 mm) in the highlands of the northwest region and the lowest values in the southern coast and in the eastern part of the territory (below or around 500 mm). On average, about 42% of the annual precipitation falls during the three-month winter season (December–February) and the lowest values occur during summer, corresponding to only 6% of the annual precipitation (June–August), as it is characteristic of Mediterranean climates (Miranda et al., 2002).
Today's forest area in mainland Portugal is dominated by native oak species, which cover about 1 million ha (13% of the country's area, ICNF, 2013). Oak forests usually occur as patches in the landscape mosaic, with varying tree densities and species dominance along the country's climatic and altitudinal gradient. Evergreen oaks (cork oak and holm oak) are the dominant oak species (23% and 11% of total forest area, respectively), occurring more frequently across southern Portugal. The evergreen cork oak (Quercus suber L.) is mainly located across the southwest where the oceanic influence is stronger or at higher altitudes; evergreen holm oaks (Quercus rotundifolia Lam.) grow under more extreme temperature conditions and are mainly located across the dry plains of the southeast continental region (Castro, 2009). Both evergreen oaks often coexist in their southern distribution edges and with deciduous oaks in the northeast and cental east.
Deciduous oak species (including marcescent oaks) cover only 2% of the total forest area (ICNF, 2013) and occur mainly as small fragmented patches in northern and central regions. Deciduous oaks include three dominant oak species: (1) the pedunculate oak (Quercus robur L.) mainly located in the northwest region with higher rainfall values; (2) the Pyrenean oak (Quercus pyrenaica Willd.) mostly widespread in acidic and colder environments in the higher altitudes of the northeast and central east; (3) and the Portuguese oak (Quercus faginea Lam.) occurring frequently on less acidic soils and lower altitudes along the central coast, and in higher altitudes in the northeast where it usually coexists as dominated species with Pyrenean oak and cork oak. All three deciduous oak species considered have a wide distribution within the Mediterranean-climate zone in the Mediterranean Basin (Blanco et al., 1997; Quézel & Médail, 2003).
The majority of evergreen oak area is managed as agro-silvopastoral systems consisting of an open overstory of oak trees of varying densities with understory management for crops and pastures. Deciduous oak forests are managed as silvopastoral systems with extensive grazing or denser forested areas exploited for timber.
Inventory data and quantification of land cover changes
We used data from two National Forest Inventories (NFI) to quantify the changes in oak forest cover over nearly a 40-year period (1966–2006) for mainland Portugal. The first inventory includes only data on forest cover and naturally occurring vegetation types (such as shrublands and pasturelands) as vector polygons (with a minimum cartographic unit of 2 ha and a width of 30 m, at 1 : 25 000 scale), classified with aerial photographs from 1965 to 1966 until 1978 with field validation (DGSFA, 1968). This is the only ancient detailed forest cartography covering the entire country. The more recent available inventory was conducted during 2005–2006 (AFN, 2010) and is composed of a 2 × 2 km grid of points that contains land cover data (including forest cover and other land cover types, see Table 1) that were classified using aerial photographs (with a minimum cartographic unit of 0.5 ha and a width of 20 m, at 1 : 5000 scale), with field validation only in forest areas (each point being the center of a surveyed plot). For both inventories, an area was classified as forest if it was at least 10% covered with tree canopy.
| Class | Description |
|---|---|
| Cork oak | Forest areas dominated by cork oak (Quercus suber)a |
| Holm oak | Forest areas dominated by holm oak (Quercus rotundifolia)a |
| Deciduous oaks | Forest areas dominated by pure or mixed stands with deciduous (Quercus robur) and marcescent oaks (Quercus faginea, Quercus pyrenaica)a |
| Other (nonoak) forests | Forest areas dominated by nonoak forest species, including eucalypt (Eucalyptus spp.), maritime pine (Pinus pinaster), stone pine (Pinus pinea), and other broadleaf and conifer speciesa |
| Agriculture | Croplands (including irrigated areas), pasturelands, vineyards, and orchards |
| Shrublands | Areas dominated by shrub species (mostly in the Ericaceae, Cistaceae, and Fabaceae families) or in some cases by natural grasslands |
| Other land covers | Areas with scattered trees, burned areas, urban areas, and water bodies |
- a Forest areas were dominated by a given forest species if that species occupied at least 50% of the total tree canopy area.
We intersected the points located at the 2 × 2 km grid from the last NFI (2005–06) with the forest cover data from the initial NFI (1965–66) over mainland Portugal. Only points classified as oak forests at the start of the study period were selected for land cover change analysis, resulting in 1537 points classified as cork oak forests, 1259 points classified as holm oak forests, and 182 points classified as deciduous oak forests. Data from both NFIs differentiate evergreen oaks in separate classes (cork oak and holm oak), but group the remaining oak species into one single class, hereafter named as deciduous oaks. Although deciduous oaks include three different species, this group in Portugal is dominated by Quercus pyrenaica, and the relative importance of Quercus faginea and Quercus robur is small (Oliveira et al., 2001; Capelo & Catry, 2007a,b; Carvalho, 2007). Land cover classes used are presented and described in Table 1.
For each oak forest type (cork oak forests, holm oak forests, and deciduous oak forests), land cover change in the selected points between 1966 and 2006 was classified as one of five categories: (1) oak forest persistence (no changes in land cover); (2) change to shrublands; (3) change to agriculture; (4) change to other forest types (changes to a forest class dominated by another species, including a different oak species or a nonoak forest species); and (5) change to other land covers (including changes to areas with scattered trees, burned areas, urban areas, and water). Figure 1 shows the distribution of cork oak forests, holm oak forests, and deciduous oak forests in mainland Portugal, and the points used in this study.
Characterization of potential drivers of oak cover change
Each of the points was characterized using a set of anthropogenic and environmental variables that could potentially influence the oak cover changes, based on the review of existing evidence and on the hypotheses set at the beginning of the work. We selected an initial set of seventeen variables that comprised four anthropogenic variables (wildfire frequency, distance to roads, population density, and trend of population density), one topographic variable (slope), two mean climate variables (annual precipitation and mean annual temperature), and ten variables associated with extreme climate events (two drought indicators based on the Standardized Precipitation Index, four indices of daily climate extremes, and the long-term trend estimate for each of these four indices; see Table 2). Note that wildfires are mainly caused by human activity in Mediterranean Europe (Vélez, 2009).
| Predictor variable (units) | Variable description | Source and description of spatial dataset | |
|---|---|---|---|
| Anthropogenic variables | Wildfire frequency (fires) | Number of times an area burned between 1975 and 2006 | Burned area annual maps for mainland Portugal (vector format) (Oliveira et al., 2012; ICNF, 2014) |
| Distance to roads (m) | Distance from each location of the territory to the nearest asphalt road (main national and regional roads) | Portuguese Itinerary Military Map (vector format at 1: 500 000 scale) (IGEOE, 2005) | |
| Population density (inhabitantsk/km2) | Population density per parish in 2001 | Portuguese Census (INE, 2001) | |
| Trend of population density (%) | % increase or decrease of population density between 1970 and 2001 per municipality | ||
| Topographic variable | Slope (°) | Calculated from a digital elevation model | ASTER Global Digital Elevation Model (GDEM, resolution of ca. 30 m) (METI/NASA, 2011) |
| Mean climate variables | Annual precipitation (mm) | Annual mean total precipitation (averaged for 1966–2003) | Portuguese Institute of Ocean and Atmosphere (PT02 monthly precipitation dataset, 1950–2003; Belo-Pereira et al., 2011) |
| Mean annual temperature (°C) | Mean annual temperature (averaged for 1966–2006) | E-OBS gridded daily dataset for Europe, 1950–2013(http://eca.knmi.nl/download/ensembles/download.php (Haylock et al., 2008) | |
| Extreme climate | Long-term droughts (total number of months with severe 12-month SPI) | Severe 12-month Standardized Precipitation Index (SPI < −1.5): comparison of the precipitation for 12 consecutive months with that recorded in the same 12 consecutive months in all previous years between 1966–2003 | Portuguese Institute of Ocean and Atmosphere (PT02 monthly precipitation dataset, 1950–2003; Belo-Pereira et al., 2011) |
| Dry winters (total number of March months with severe 6-month SPI) | Severe 6-month Standardized Precipitation Index (SPI < −1.5) at the end of March: comparison of the precipitation for October–March period with that recorded in the same period in all previous years between 1966–2003 | ||
| Monthly maximum value of daily maximum temperature (TXx) (°C) | TXkj – daily maximum temperature in month K, year J TXx = max (TXkj) (averaged for 1966–2006) | E-OBS gridded daily dataset for Europe, 1950–2013 (http://eca.knmi.nl/download/ensembles/download.php (Haylock et al., 2008) | |
| Frequency of warm days (Tx90p) (%) | Annual percentage of days (averaged for 1966–2006) with daily maximum temperature >90th percentile for the base period 1950–2012 | ||
| Warm spell duration index (WSDI) (days) | Annual count of days with at least 6 consecutive days (summed for 1966–2006) when daily maximum temperature (TX) > 90th percentile for the base period 1950–2012 | ||
| Consecutive dry days or maximum length of dry spells (CDD) (days) | Annual maximum number of consecutive days with daily precipitation amount (RR) < 1 mm (summed for 1966–2006) | ||
| Long-term trend estimate for TXx | Slope of the regression line with annual values of TXx for the period 1950–2012 | ||
| Long-term trend estimate for Tx90p | Slope of the regression line with annual values of TX90p for the period 1950–2012 | ||
| Long-term trend estimate for WSDI | Slope of the regression line with annual values of WSDI for the period 1950–2012 | ||
| Long-term trend estimate for CDD | Slope of the regression line with annual values of CDD for the period 1950–2012 | ||
We selected the four anthropogenic variables previously described as potential indicators of human management that may have influenced the land cover change across oak forest types. Wildfires are a major disturbance in Mediterranean European forests, where increased wildfire frequency is often associated with land abandonment and the consequent shrub encroachment (Marques et al., 2011; Moreira et al., 2011). Wildfire frequency was estimated using annual burned area maps (vector format) with the minimum cartographic unit of 5 ha for most of the study period (1984–2006) and 35 ha for 1975–1983 (see Table 2). Distance to main roads indicates isolation or accessibility to a particular location, which may determine the changes in management (e.g., land abandonment in isolated areas vs. land intensification in areas with easier access). Distance to roads was calculated in arcgis (ESRI, 2012), producing a raster map with 100-m spatial resolution (see Table 2 for further details). Population density and its trend are an indication of the intensity of land management (e.g., we expect decreasing population density to be associated with land abandonment). We used data on population density per parish (from national census) but calculated the trend of population density per municipality because there were changes in several parishes across the country in the period 1970–2001 (some were aggregated and disappeared, while others changed in area).
Topography (slope) may determine the trajectories of oak forest change because it affects microclimatic conditions (e.g., light, soil moisture) and the consequent oak growth and vulnerability to disturbances; it may also indirectly influence land management choices as steeper slopes may hamper mechanization.
Annual precipitation and mean annual temperature were selected to understand the influence of mean climatic conditions on oak persistence and land cover changes and to compare its importance against extreme climate events. Annual precipitation (averaged for the period 1966–2003) was extracted from a monthly precipitation dataset based on 806 stations with a regular 0.2° resolution (16 × 22 km) over mainland Portugal, which resulted in a gridded dataset with 257 cells. Mean annual temperature (averaged for 1966–2006) was extracted from the E-OBS gridded daily dataset for Europe with a regular 0.25° × 0.25° resolution, which originated a gridded dataset with 169 cells (approximately 22 × 28 km) over mainland Portugal (see Table 2 for further details). Although the E-OBS dataset is only based on 16 stations over mainland Portugal, it is the best available gridded dataset in terms of resolution and time span with daily and monthly temperature data for the national scale.
Disturbances such as extreme climate events have been linked to tree decline and mortality (Allen et al., 2010; Anderegg et al., 2013). We selected two types of extreme climate events that may have influenced the oak cover changes in mainland Portugal: longer-term droughts (with monthly input data) and shorter-term events such as warm spells and dry spells (with daily input data). Drought events were based on the Standardized Precipitation Index (SPI, McKee et al., 1993) and included two drought indicators: (1) the 12-month SPI (hereafter named as long-term droughts), which is related to hydrological droughts and is usually a good indication that dryness is having a significant impact on natural systems; and (2) the 6-month SPI at the end of March (hereafter named as dry winters), which provides a good indication of the shortage of precipitation during the wet season period, which foregoes the main growing season for oaks (April–September) in Mediterranean Europe. Both drought indicators were calculated as the sum of severe drought months for the period 1966–2006 (see Table 2 for further details), and for each of the 257 cells of the precipitation dataset over mainland Portugal, with the Standardized Precipitation Index Calculator from the National Drought Mitigation Center (NDMC, http://drought.unl.edu/MonitoringTools/DownloadableSPIProgram.aspx).
We selected four indices of daily climate extremes: (1) monthly maximum value of daily maximum temperature (TXx, hereafter named as maximum temperature), which shows the magnitude (°C) of absolute maximum temperatures for each month; (2) frequency of warm days (Tx90p), which shows the annual frequency (%) of days with high temperature anomalies; (3) warm spell duration index (WSDI), which shows the annual duration (days) of warm spells; and (4) consecutive dry days (CDD), also known as maximum length of dry spells (days; see Table 2 for further details; Klein Tank et al., 2009; Zhang et al., 2011). These indices are included in the list of core indices recommended by the CCI/CLIVAR Expert Team for Climate Change Detection Monitoring and Indices (ETCCDMI). Indices of daily climate extremes were computed with the r-based software – rclimdex (version 1.0, WMO CCI/CLIVAR/JCOMM – ETCCDI, http://etccdi.pacificclimate.org/software.shtml), using the base period 1950–2012 (62 years) for threshold calculation. Each index is calculated for each year by rclimdex and we averaged the annual index values of frequency of warm days and maximum temperature and summed the annual index values of warm spells duration and maximum length of dry spells for the period 1966–2006. In addition, we used the long-term trend for each of the four indices to determine whether changes (increasing or decreasing trends) in daily climate extreme events were associated with oak cover changes. This trend was estimated from the slope of the linear least square regression line for each index for the period 1950–2012 (calculated by rclimdex) to encompass the longest time span. All variables were computed for each of the 169 cells of the temperature dataset over mainland Portugal.
Association between each category of oak cover change and predictor variables was made using postgis and qgis 2.8 (Obe et al., 2011; QGIS Development Team, 2014) by intersecting the layer containing the points with data on oak cover changes with the layers with data on predictor variables (separately for each oak forest type). Descriptive statistics for each oak dataset is presented in Table S1.
Data analysis
We quantified the proportion of the five categories of land cover change for each oak forest type as the percentage of sampling points that persisted or changed into a different land cover, and used G-tests of independence to test for the differences between categories of change across oak forest types. G-tests were performed using an r for Windows (R Core Team, 2015) script developed by Peter Hurd, available at http://www.psych.ualberta.ca/~phurd/cruft/g.test.r. Change to other land covers was not included in further analysis because it represented a very small number of occurrences for all oak forest types (see Fig. 2).
For each oak forest type, spatial patterns of persistence and land cover change were characterized using ordinary kriging (ArcGis, Geostatistical wizard, default options except 10–25 neighbors with eight sectors; ESRI, 2012) to produce interpolated maps of prevalence, based on binary information (1/0) separately for each category of persistence/change. We also explored spatial autocorrelation patterns for persistence and the different categories of land cover changes (to shrublands, agriculture, and other forest types), using spline correlograms based on Moran's I, with 95% confidence intervals (Bjørnstad & Falck, 2001; Rhodes et al., 2009). With this method, a plot of the correlogram function against distance is produced, and 95% confidence (or null) envelopes are superimposed. Additionally, local indicators of spatial association (LISA; Anselin, 1995) were used to explore the spatial patterns in the average autocorrelation, using the 20 nearest neighbors and 100 resamples to generate significance values. LISA enables the detection of local spatial clusters of presences or absences for a given transition category, also known as hotspots, as sets of contiguous locations for which the LISA is significant (Anselin, 1995). r package ‘ncf’ (Bjørnstad, 2015) was used to estimate the spatial autocorrelation. The results of these analyses are shown as Supporting Information.
Before proceeding to analyze the role of the different variables in observed oak cover changes, we used the variance inflation factor (VIF) as a measure of the degree of multicollinearity between predictor variables for each oak dataset. Variables that presented a VIF > 3 (Zuur et al., 2010) were eliminated from the set of variables for the subsequent analysis, which resulted in a final set of 13 variables for cork oak, 13 variables for holm oak, and 12 variables for deciduous oaks. r package ‘usdm’ (Naimi, 2015) was used to calculate the VIF. Correlation matrices and VIF values of predictor variables for each oak dataset are shown in Supporting Information (Tables S2–S4).
For each oak forest type, land cover changes were modeled using a multinomial logistic regression analysis with the response variable specified as four categories: oak forest persistence (no changes) as the reference category, plus three main categories of oak land cover change (to shrublands, to agriculture, or to other forest types). r package ’nnet’ (Venables & Ripley, 2002) was used to apply multinomial logistic regression analysis. To assess the goodness of fit of the multinomial logistic model, we used the test statistic Cg, with g = 10 (Fagerland et al., 2008), which is an adaptation of the Hosmer & Lemeshow (1980) test to multinomial logistic model. This test was performed with the r package DAMisc (Armstrong, 2015).
We used an information theoretical approach with Akaike's information criterion (ITA) to select the best model or set of models explaining the land cover changes. The ITA looks for simplicity and parsimony of several working hypotheses and is based on measuring the strength of evidence of each model, for a set of candidate models. The Akaike's information criterion (AIC) is a measure of information loss of each candidate model and is used to rank models, with the best approximating model having the lowest AIC value (Burnham & Anderson, 2002). We applied AICc for the deciduous oaks dataset (the AIC for small sample sizes, recommended when n/K < 40, n being the sample size and K the number of fitted parameters; sample size of deciduous oaks dataset is 182).
We calculated AIC values for every possible combination of variables and intercept. AIC differences (∆AICi) between the AIC value of the best model and the AIC value for each of the other models were used to assess the relative support for the different alternative models (Burnham & Anderson, 2002). ∆AICi values were also used to calculate the Akaike weight (wi) for each model, with the sum of Akaike weights of all models in the candidate set being 1. Akaike weights were used to estimate the relative importance of predictor variables under consideration by summing the Akaike weights for each model in which that variable appears, and considering all models in the candidate set. A Friedman rank test was used to test whether the rank of the importance of predictor variables was similar among the three oak forest types. The Wilcoxon–Mann–Whitney test (Siegel & Castellan, 1988) was used to test whether the rank of the importance of anthropogenic variables (wildfire frequency, distance to roads, population density, and its trend) was different from the rank of environmental variables (the remaining variables) for each individual oak forest type. Both tests were performed using r for Windows (R Core Team, 2015).
As no single model was clearly better than the set of alternatives, parameters were estimated with model averaging, which uses the average of parameter estimates from each candidate model, weighted by its Akaike weight, across a subset of the best models (Burnham & Anderson, 2002). We used the 95% confidence set of the cumulative Akaike weight to select the candidate models for model averaging (Burnham & Anderson, 2002; Symonds & Moussalli, 2011). This represents a subset of models for which we have 95% confidence that it contains the best approximating model to the true model. We applied full-model averaging combined with natural averaging because the best AIC model was not strongly weighted (Burnham & Anderson, 2002). Full-model averaging takes into account all models of the selected subset for parameter estimation, and models not containing the parameters of interest have the corresponding coefficient set to zero (i.e., contribute zero to the calculation of the average; Symonds & Moussalli, 2011). Parameters estimates from both types of model averaging (full-model averaging and natural averaging) converge for dominant variables, but for variables with a weak relationship to the response variable, the coefficients resulting from full-model averaging are smaller than those from natural averaging (Lukacs et al., 2010). We used confidence intervals from natural averaging to assess the magnitude of the effect of each parameter; if the confidence interval excluded 0, we assumed that there is an effect.
A problem in estimating Akaike weights for individual variables is that poor predictors are not expected to have selection probabilities close to zero (Burnham & Anderson, 2002). To overcome this, we followed the approach suggested by Whittingham et al. (2005) and added a randomly generated predictor uncorrelated with the response variable to the existing dataset. The selection probability for this randomly derived predictor (mean and 95% confidence interval) was obtained by performing 100 simulations and estimating summed Akaike weights for all models containing this null predictor. We used the upper bound of the confidence interval for this null predictor as a cutoff value to assume that all predictor variables with lower selection probability had low importance. We provide the overall predictive accuracy for the averaged model, which shows the ratio between correctly classified observations and the total number of observations, calculated from a contingency table of observed and predicted frequencies (highest probability of occurrence). r package ‘MuMIn’ (Bartoń, 2015) was used for model selection and model averaging.
We also explored spatial autocorrelation patterns in the residuals from model averaging to check whether spatial autocorrelation potentially found in the raw data (observed categories of persistence or land cover change) was accounted for by the explanatory variables, and avoid the biased parameter estimation. For this, we used spline correlograms, as previously described, with r package ‘ncf’ (Bjørnstad, 2015). The results of these analyses are shown in Supporting Information.
Results
Persistence and transitions across oak forest types
Differences in persistence and land cover change among the three oak forest types were significant (G-test, G = 162.88, P < 0.001, Fig. 2). Cork oak was the most persistent forest over the 40 years (62% sampling points without the changes in land cover). The main transition was to other forest types (19.2%), namely to forests dominated by eucalypt, holm oak, maritime pine, and stone pine [in decreasing order of magnitude (Figure S1)]. Changes to shrublands and agriculture occurred in similar proportions (8.4% and 7.6%, respectively, Fig. 2). About half (53.2%) of the sampling points classified as holm oak forests remained as such after four decades with the most frequent type of change being to agriculture (19.5%). Changes to other forest types were the second most frequent type of change (17.5%) mostly to cork oak forests (Figure S1), while changes to shrublands occurred in ca. 6% of the points (Fig. 2). Deciduous oaks were the least persistent oak forest with less than half (45.7%) of sampling points remaining as such in 40 years. Deciduous oak forests changed mostly to shrublands (26.6%), and the second most frequent change was to other forest types (16.8%), mainly to forests dominated by maritime pine, evergreen oak forests, and mixed forest types (Figure S1), while changes to agricultural land were comparatively low (9.8%). Changes from all oak forest types to other land covers were the least frequent change (2.9% for cork oak forests, 4.2% for holm oak forests, and 1.1% for deciduous oak forests, Fig. 2).
Overall patterns of persistence and land cover change in cork oak forests showed a significant, although low, spatial autocorrelation up to ca. 40 km distance, and local patterns (LISA) showed a prevalence of cork oak forest persistence in the south, of changes to shrublands in Algarve Province and northeast, of changes to agriculture south of Lisbon, and of changes to other forest types in the central coast (Figures S2–S4). Overall patterns of persistence and land cover change in holm oak forests also showed a significant but low spatial autocorrelation up to ca. 10 km distance, and local patterns showed a prevalence of holm oak forest persistence toward the core area of the species distribution (southeast), of changes to shrublands in Algarve Province and northeast, of changes to agriculture in southeast, and of changes to other forest types in central southwest (Figures S5–S7). Overall patterns of persistence and land cover change in deciduous oak forests did not display a significant spatial autocorrelation and main local patterns included the prevalence of persistence in central east and of changes to shrublands along the northern border (Figures S8–S10).
Drivers of land cover change in oak forests
For all oak forest types, the goodness-of-fit test C10 from Fagerland et al. (2008) indicates that the global multinomial logistic model fits the data well: χ2(24) = 28.343, P = 0.246 for cork oak forests; χ2(24) = 24.783, P = 0.418 for holm oak forests; and χ2(24) = 23.309, P = 0.502 for deciduous oak forests.
Figure 3 shows the selection probability of predictor variables considering all models in the candidate set for each of the three oak forest types (4096 models for cork oak and holm oak forests and 2048 models for deciduous oak forests, resulting from all possible combinations of variables and intercept). Across oak forest types, averaged selection probability shows that the most important predictor variables were wildfire frequency and trend of population density (both with 0.99), followed by distance to roads (0.89) and slope (0.85). Differences in the rank of importance of predictor variables among the three oak forest types were significant (Friedman chi-squared = 18.1, P = 0.034), with climatic variables showing the largest differences across oak forest types. Deciduous oak forests were the only forest type for which the importance of anthropogenic variables was significantly larger than the one of environmental variables (Wilcoxon test, W = 29, P = 0.033; W = 28, P = 0.137; W = 21, P = 0.697, for deciduous oak forests, cork oak forests, and holm oak forests, respectively).
Tables 3-5 show the model-average estimates of coefficients based on the 95% confidence model set (cumulative Akaike weight, acc wi ≤ 0.95) for the likelihood of changes from each oak forest type to shrublands, agriculture, or other forest types, in 1966–2006.
| Variables | Importance (selection prob.) (∑wi) | N. models | Change to shrublands | Change to agriculture | Change to other forest types | |||
|---|---|---|---|---|---|---|---|---|
| β (full-model) | Confidence Interval | β (full-model) | Confidence Interval | β (full-model) | Confidence Interval | |||
| Wildfire frequency | 1 | 87 | 0.545 | 0.227; 0.864 | −0.549 | −0.653; −0.444 | −0.250 | −0.586; 0.086 |
| Slope | 1 | 87 | 0.020 | −0.012; 0.052 | −0.061 | −0.110; −0.011 | 0.054 | 0.027; 0.081 |
| Trend of population density | 1 | 87 | −0.012 | −0.020; −0.005 | 0.006 | 0.001; 0.011 | −0.003 | −0.008; 0.002 |
| Maximum temperature | 1 | 87 | −0.558 | −0.619; −0.496 | −0.137 | −0.209; −0.066 | 0.159 | 0.051; 0.266 |
| Distance to roads | 0.99 | 81 | 0.0002 | 5.43*10 − 5 ; 0.0003 | −0.0001 | −0.0003; 1.18*10−5 | 5.70*10−5 | −2.95*10−5; 1.4*10−4 |
| Annual precipitation | 0.84 | 61 | 0.003 (0.002) | −3.93*10−5; 0.005 | 0.002 | −0.001; 0.005 | 0.003 | 0.001; 0.006 |
| Trend of warm spells duration | 0.80 | 60 | 2.557 (2.049) | 1.877; 3.237 | 0.441 (0.353) | −0.159; 1.041 | −3.019 (−2.419) | −3.907; −2.130 |
| Population density | 0.55 | 46 | 0.002 (0.001) | 0.0003; 0.003 | −0.0003 (−0.0002) | −0.003; 0.002 | 0.001 (0.0004) | −0.001; 0.002 |
| Frequency of warm days | 0.41 | 45 | −0.092 (−0.038) | −0.363; 0.179 | −0.300 (−0.123) | −0.476; −0.124 | −0.841 (−0.345) | −0.977; −0.705 |
| Long-term droughts | 0.17 | 30 | 0.014 (0.002) | −0.012; 0.040 | 0.021 (0.004) | −0.009; 0.051 | −0.001 (−0.0002) | −0.022; 0.020 |
| Mean annual temperature | 0.07 | 23 | −0.093 (−0.006) | −0.300; 0.114 | −0.069 (−0.005) | −0.345; 0.207 | −0.051 (−0.004) | −0.317; 0.214 |
| Trend of maximum length of dry spells | 0.06 | 19 | 0.461 (0.028) | −0.043; 0.964 | 0.207 (0.013) | −0.011; 0.426 | −0.355 (−0.022) | −1.155; 0.444 |
| Dry winters | 0.06 | 17 | −0.053 (−0.003) | −0.287; 0.181 | −0.063 (−0.004) | −0.297; 0.171 | −0.053 (−0.003) | −0.231; 0.124 |
- Variables (see Table 2 for description) are shown by decreasing importance (selection probability), which is calculated by summing all wi scores for all possible models in which the predictor was included. N. models show the number of models that include the predictor variable. Parameter estimates (β) are presented as natural averaging (only models with the parameter of interest are considered) and full-model averaging (all 87 models are considered, number between parentheses), except when estimates are the same for both methods. Estimates shown in boldface have confidence intervals that do not include zero.
| Variables | Importance (selection prob.) (∑wi) | N. models | Change to shrublands | Change to agriculture | Change to other forest types | |||
|---|---|---|---|---|---|---|---|---|
| β (full-model) | Confidence Interval | β (full-model) | Confidence Interval | β (full-model) | Confidence Interval | |||
| Slope | 1 | 100 | 0.121 | 0.074; 0.167 | −0.073 | −0.118; −0.027 | 0.060 | 0.025; 0.095 |
| Maximum temperature | 1 | 100 | −0.960 | −1.114; −0.806 | −0.184 | −0.298; −0.070 | −0.428 | −0.551; −0.305 |
| Frequency of warm days | 1 | 100 | 1.722 | 0.993; 2.450 | 2.436 | 1.926; 2.946 | 0.423 | −0.076; 0.922 |
| Wildfire frequency | 0.99 | 99 | 1.203 (1.202) | 0.815; 1.592 | 0.789 (0.788) | 0.443; 1.135 | 0.789 (0.788) | 0.367; 1.211 |
| Trend of population density | 0.99 | 97 | 0.004 | −0.011; 0.019 | 0.008 | −0.002; 0.018 | 0.018 | 0.009; 0.027 |
| Distance to roads | 0.72 | 58 | 5.01*10−5 (3.61*10−5) | −0.0001; 0.0002 | −9.13*10−5 (−6.58*10−5) | −0.0002; 2.59*10−5 | 0.0001 (7.43*10−5) | −3.18*10−6; 0.0002 |
| Dry winters | 0.35 | 46 | 0.181 (0.064) | −0.137; 0.499 | −0.066 (−0.023) | −0.252; 0.120 | −0.175 (−0.062) | −0.377; 0.026 |
| Trend of warm spells duration | 0.30 | 43 | 1.707 (0.505) | −0.682; 4.096 | −2.203 (−0.652) | −3.761; −0.645 | −2.449 (−0.725) | −4.077; −0.820 |
| Annual precipitation | 0.29 | 38 | 0.001 (0.0002) | −0.002; 0.004 | −0.001 (−0.0004) | −0.004; 0.001 | 0.002 (0.0004) | −0.0004; 0.003 |
| Population density | 0.19 | 35 | −0.004 (−0.001) | −0.032; 0.024 | 0.009 (0.002) | −0.003; 0.021 | −0.003 (−0.001) | −0.018; 0.011 |
| Warm spells duration | 0.11 | 28 | 0.003 (0.0003) | −0.005; 0.011 | −0.003 (−0.0003) | −0.009; 0.003 | 0.002 (0.0002) | −0.004; 0.008 |
| Long-term droughts | 0.06 | 19 | −0.012 (−0.001) | −0.055; 0.032 | −0.010 (−0.001) | −0.036; 0.016 | −0.001 (−5.17*10−5) | −0.028; 0.026 |
| Mean annual temperature | 0.05 | 19 | −0.121 (−0.006) | −0.479; 0.238 | 0.052 (0.003) | −0.310; 0.415 | −0.051 (−0.003) | −0.381; 0.279 |
- Variables (see Table 2 for description) are shown by decreasing importance (selection probability), which is calculated by summing all wi scores for all possible models in which the predictor was included. N. models show the number of models that include the predictor variable. Parameter estimates (β) are presented as natural averaging (only models with the parameter of interest are considered) and full-model averaging (all 100 models are considered, number between parentheses), except when estimates are the same for both methods. Estimates shown in boldface have confidence intervals that do not include zero.
| Variables | Importance (selection prob.) (∑wi) | N. models | Change to shrublands | Change to agriculture | Change to other forest types | |||
|---|---|---|---|---|---|---|---|---|
| β (full-model) | Confidence Interval | β (full-model) | Confidence Interval | β (full-model) | Confidence Interval | |||
| Wildfire frequency | 1 | 205 | 0.529 (0.527) | 0.143; 0.915 | −1.234 (−1.231) | −2.502; 0.034 | −0.149 (−0.148) | −0.656; 0.358 |
| Distance to roads | 1 | 202 | 0.001 | −5.49*10−5; 0.001 | −0.002 | −0.003; −0.001 | 0.0001 | −0.0004; 0.001 |
| Trend of population density | 0.99 | 201 | 0.043 | 0.018; 0.069 | 0.039 | 0.002; 0.077 | −0.001 | −0.029; 0.026 |
| Trend of maximum length of dry spells | 0.96 | 172 | −13.794 (−13.297) | −19.160; −8.429 | 9.199 (8.867) | 3.885; 14.513 | 5.404 (5.210) | 0.387; 10.422 |
| Annual precipitation | 0.72 | 132 | 0.001 | −0.001; 0.003 | 0.004 (0.003) | 0.001; 0.007 | 0.001 | −0.0001; 0.003 |
| Warm spells duration | 0.68 | 108 | −0.017 (−0.012) | −0.027; −0.008 | 0.013 (0.009) | −0.004; 0.029 | 0.004 (0.003) | −0.006; 0.015 |
| Population density | 0.60 | 95 | −0.016 (−0.010) | −0.042; 0.010 | −0.013 (−0.008) | −0.037; 0.011 | 0.009 (0.005) | −0.002; 0.021 |
| Slope | 0.56 | 114 | 0.027 (0.015) | −0.037; 0.091 | −0.170 (−0.096) | −0.326; −0.014 | −0.009 (−0.005) | −0.077; 0.060 |
| Frequency of warm days | 0.27 | 78 | 1.346 (0.362) | 0.901; 1.790 | 0.363 (0.098) | −0.408; 1.135 | 0.443 (0.119) | −0.122; 1.007 |
| Mean annual temperature | 0.20 | 98 | 0.241 (0.048) | −0.144; 0.626 | 0.523 (0.105) | 0.083; 0.964 | 0.422 (0.084) | 0.095; 0.749 |
| Dry winters | 0.17 | 73 | −0.378 (−0.063) | −0.801; 0.045 | 0.072 (0.012) | −0.535; 0.678 | −0.399 (−0.066) | −0.794; −0.004 |
| Trend of warm spells duration | 0.09 | 71 | 8.279 (0.723) | 0.874; 15.683 | 13.831 (1.207) | 1.944; 25.718 | 11.632 (1.015) | 3.115; 20.148 |
- Variables (see Table 2 for description) are shown by decreasing importance (selection probability), which is calculated by summing all wi scores for all possible models in which the predictor was included. N. models show the number of models that include the predictor variable. Parameter estimates (β) are presented as natural averaging (only models with the parameter of interest are considered) and full-model averaging (all 205 models are considered, number between parentheses), except when estimates are the same for both methods. Estimates shown in boldface have confidence intervals that do not include zero.
Changes from cork oak forests to shrublands were more likely in areas with higher wildfire frequency, more distant from roads, with lower maximum temperatures but with an increasing trend in the duration of warm spells for the past 50 years. Such changes also occurred in more populated areas but that have registered population decline in the last 30 years (Table 3). On the contrary, changes from cork oak forests to agricultural land occurred in areas with lower wildfire frequency, flatter slopes and with increasing population density in the last 30 years. Changes to agriculture were also associated with lower maximum temperatures and less frequent warm days. Changes from cork oak forests to other forest types were more likely in areas with steeper slopes, higher precipitation, and higher maximum temperatures, but that registered less frequent warm days and a decreasing trend in the duration of warm spells in the last 50 years. Remaining variables showed low precision and low probability of selection (within or lower than the confidence interval of the null predictor) in the set of candidate models explaining land cover changes in cork oak forests (Table 3). Overall predictive accuracy of the averaged model was 64% for the changes in cork oak forests.
All types of land cover changes in holm oak forests occurred predominantly in regions with lower maximum temperatures and higher wildfire frequency. Changes to shrublands and to other forest types (mainly to cork oak forests, which performs 80% of this type of change) occurred at steeper slopes, while changes to agricultural land took place in flatter areas. A higher frequency of warm days in the observation period was also an important climatic predictor related to the changes to shrublands and agriculture. The replacement of holm oak forests by cork oak forests and other forest types was associated with an increasing population density. Remaining variables showed low precision and low probability of selection (close to or lower than the confidence interval of the null predictor) in the set of candidate models explaining the land cover changes in holm oak forests (Table 4). Overall predictive accuracy of the averaged model was 53% for the changes in holm oak forests.
Deciduous oak forests changed to shrublands in areas with higher wildfire frequency, increasing population density, with a decreasing trend in the maximum length of dry spells for the past 50 years, and with shorter duration of warm spells. Changes from deciduous oak forests to agriculture took place closer to main roads and with increasing population density, at flatter regions with higher precipitation and temperature but that registered an increasing trend in the maximum length of dry spells. Changes from deciduous oak forests to other forest types (dominated by maritime pine, mixed forest types, holm oak, cork oak, and eucalypt) were more likely to occur in areas with an increasing trend in the maximum length of dry spells. Remaining variables showed low precision and low probability of selection (close to or within the confidence interval of the null predictor) in the set of candidate models explaining the land cover changes in deciduous oak forests (Table 5). Overall predictive accuracy of the averaged model was 57% for the changes in deciduous oak forests.
Overall, spatial autocorrelation from model residuals was negligible or not significant for all categories of oak cover change/persistence, for all oak forest types (Figure S2 for cork oak forests, Fig. S5 for holm oak forests, and Fig. S8 for deciduous oak forests).
Discussion
Distinct forest persistence and transitions across oak forest types
Oak forest landscapes in mainland Portugal showed species-specific patterns of persistence and change between 1966 and 2006. Cork oak forests were the most persistent, followed by holm oak forests and deciduous oak forests, in decreasing order of persistence, which matches our initial predictions. Distinct persistence and land cover change across oak forest types during the study period probably reflect the differences in oak economic value, legal protection status, and management schemes. Cork oak forests have been more carefully managed and preserved because cork oak trees have been legally protected since Middle Ages (Natividade, 1950) and tree persistence is needed to produce the highly valued cork (Pinto-Correia & Vos, 2004). Nevertheless, our results also show expansion of fast-growing eucalypt and pine plantations in cork oak areas, probably supported by several afforestation policies (Pinto-Correia & Godinho, 2013; Costa et al., 2014). These changes have taken place mainly in marginal areas where cork oak trees were already under decay (e.g., due to mismanagement or pathogen attacks), in depleted soils or steep slopes, as shown by our results and reported by other studies (Pinto-Correia & Godinho, 2013; Costa et al., 2014).
Holm oak forests were less persistent than cork oak forests. This species has also been legally protected since the past, but has a lower economic value compared with cork oak. The most frequent land cover change observed in holm oak forests was to agricultural land, confirming a strong land management intensification trend. Indeed, livestock production is today one of the main commercial activities in holm oak areas, where the free-ranging Iberian pigs traditionally fed with holm oak acorns have been largely replaced by autochthonous cattle, along with an intensification of crop production (Alagona et al., 2013; Pinto-Correia & Godinho, 2013). Despite the legal protection of holm oak trees, agricultural policies supporting cattle and linked to production have often resulted in too high animal densities and the use of heavy breeds, leading to soil damage and compaction, reduced natural regeneration, and decay of holm oak tree cover (Pinto-Correia & Godinho, 2013). In turn, widespread dieback and mortality of holm oak forests in southern Portugal (Moreira & Martins, 2005) may also have contributed to its replacement by agricultural land and by cork oak forests (the second most frequent change), because holm oak is potentially more affected by dieback than cork oak, due to a higher susceptibility to the pathogen Phytophthora cinnamomi (Robin et al., 2001; Rodríguez-Molina et al., 2002).
With a clear distinct trend, deciduous oak forests were the least persistent and changed mostly to shrublands. Deciduous oak species are not legally protected and can be felled upon landowner's decisions, usually for their high-valued timber (especially from pedunculate oak). Felling of the best trees for timber has likely been eliminating important seed sources and hampering the recovery of deciduous oak forests, already degraded by years of coppicing for firewood and charcoal, followed by shrub colonization (Carvalho, 2005). The main change in deciduous oak forests during the study period, toward shrublands, has been facilitated by wildfires. As a consequence, shrubs have become dominant over oak trees, which likely present a shrub physiognomy and are intermingled with the shrub canopy.
Drivers of land cover change have distinct importance across oak forest types
Explanatory variables associated with land cover changes also had distinct importance across oak forest types. Nevertheless, anthropogenic variables were, on average, more important predictors than environmental variables for the three types of oak forests considered. Wildfire frequency, in particular, was positively associated with the changes from oak forests to shrublands consistently across forest types, in accordance with several studies for the Mediterranean Basin and in particular for Iberian Peninsula (Pausas, 2004; Pausas et al., 2008; Moreira et al., 2011).
The other two most important anthropogenic variables, trend in population density and distance to roads, were distinctively associated with land cover changes across oak forest types. Population decline and difficult accessibility (by roads) were associated with the changes from cork oak forests to shrublands. This pattern of land abandonment in isolated rural areas with the consequent shrub encroachment and increased wildfire activity has been often observed in southern Europe since the 1960s (Pinto-Correia et al., 2004; Rey Benayas et al., 2007; Moreira et al., 2011). Conversely, increased population density was associated with the transition from deciduous oak forests to shrublands, which might be a consequence of timber exploitation and the consequent shrub colonization, as previously hypothesized. Furthermore, increased population density in both cork oak and deciduous oak forests was also associated with land management intensification (change to agriculture). In a distinctive way, increased population density was associated with the replacement of holm oak forests by cork oak forests, which was more pronounced toward west, probably due to a combination of more favorable habitats and forestry practices favoring cork oak.
Topography was an important variable for land cover changes in evergreen oak forests, but with lower importance for deciduous oak forests. One possible explanation is the more severe decline of evergreen oaks in steep slopes caused by the exotic pathogen Phytophthora cinnamomi (Sánchez et al., 2002; Moreira & Martins, 2005), which may have contributed to the replacement of declining oak stands. On the other hand, flatter areas were associated with the changes to agricultural land for all oak forest types. Indeed, intensification efforts for crop production have been mainly localized in flatter and more fertile areas to allow for mechanization and compensate the investment (Pinto-Correia & Mascarenhas, 1999; Alagona et al., 2013).
Climatic extremes were also among the most important predictors of oak landscape changes, with higher importance than variables expressing average climatic conditions. This result agrees with a wide range of studies suggesting that extreme climate events might have a greater influence on ecosystems than gradual shifts of mean temperature and precipitation (Jentsch & Beierkuhnlein, 2008). In our study, more frequent and increased duration of extreme temperatures combined with higher wildfire frequency was associated with the changes from evergreen oak forests to shrublands, in agreement with previous research for southern Portugal (Acácio et al., 2009). Instead of extreme temperatures, the changes in deciduous oak forests were better explained by extreme deficit of precipitation, probably related to the fact that deciduous oaks are more vulnerable to prolonged periods of water shortage (Baldocchi et al., 2010). Increasing length of dry spells was associated either with the changes from deciduous oak forests to agriculture in flatter areas or with the changes to other forest types better adapted to more xeric conditions (e.g., dominated by maritime pine, evergreen oaks, and eucalypt). In contrast, wetter conditions coupled with higher frequency of wildfires were associated with the changes from deciduous oak forests to shrublands. These results suggest that the most productive forest sites are located at these wetter regions where oaks have been predominantly cut, leading to an arrested succession dominated by shrublands and maintained by more frequent wildfires.
Oak forest vulnerability and resilience under global change
Our study shows that the patterns and drivers of persistence and land cover change in Mediterranean oak forests in mainland Portugal are clearly species specific and that both anthropogenic and environmental variables are important drivers of oak landscape changes. Local-scale management decisions and changes in policies and market values have been the primary drivers of forest changes. As expected, our results also indicate that increased occurrence of extreme climatic events has interacted with human activities to accelerate the changes in cork oak and deciduous oak forests (increased duration of warm and dry spells, respectively). However, contrary to expectations, there was no evidence of a lower impact of observed climate changes on cork oak forests comparatively to forests dominated by the other oak species.
Cork oak forests have been the most persistent due to the high revenues from cork production. Portugal is the largest cork exporter in the world and cork production represents approximately 2% of total Portuguese exports. Despite the global financial crisis of 2007–2008 and the current Portuguese economic crisis, cork exports did not suffer severe losses and have increased since 2010 (APCOR, 2015). Nevertheless, and despite high cork prices, there is an alarming lack of cork oak natural regeneration and tree recruitment in areas with lower tree density, mainly due to livestock overgrazing (Bugalho et al., 2009), which threatens the long-term survival of such stands. If cork loses its market value in the future, cork oak forests might become endangered and additionally threatened by mismanagement (with increasing livestock density), land abandonment, and more frequent wildfires, which may in turn lower tree resistance to increased climatic disturbances and pathogen outbreaks (Branco & Ramos, 2009). Without the value of cork, management of holm oak forests has been highly dependent on agricultural policies and severely affected by the changes in such policies, which led to overuse, holm oak decay, and its replacement by cork oak. Holm oak forests were particularly less persistent in the westernmost range margins of their distribution area (western Portugal). Changes in agricultural policies in the near future and the global economic crisis may lead to a higher pressure for food production (Pinto-Correia & Godinho, 2013), which may accelerate holm oak forest loss. Deciduous oak forests appeared even more vulnerable to global change than holm oak forests. Most of these forests have been already lost or fragmented due to human overexploitation and will be further replaced by shrublands and more xeric forest types (e.g., evergreen oaks and Mediterranean pine forests) with the continuation of human pressure, wildfires, and under climate change such as more prolonged dry spells, as indicated by ours and previous studies (Benito-Garzón et al., 2013; García-Valdés et al., 2013). In this sense, legal protection of native deciduous oak species in mainland Portugal is clearly necessary and would probably help maintaining this forest type.
Forecasted climate change for the Mediterranean Basin projects the occurrence of more frequent extreme events such as wildfires (Moriondo et al., 2006), droughts, and warm spells (Giorgi & Lionello, 2008; Kovats et al., 2014), which will additionally threaten the persistence of Mediterranean oak forests. Increasing wildfire frequency and severity may lower oak resprouting success (Catry et al., 2012, 2013), and recurrent droughts and warm spells may hamper postfire forest regeneration (Moreno, 2009; Catry et al., 2010), limit oak seed production (Pérez-Ramos et al., 2010), and increase oak mortality in the first life stages (Morin et al., 2010). Increased temperature and aridity may further exacerbate drought-induced oak mortality (Lloret et al., 2004; Gea-Izquierdo et al., 2009; Carnicer et al., 2011) and accelerate the spread of the exotic pathogen Phytophthora cinnamomi, considered today as a major cause of evergreen oak dieback across extensive areas of southwestern Iberia (Brasier, 1996; Bergot et al., 2004; Moreira & Martins, 2005; Branco & Ramos, 2009; Ibáñez et al., 2014). In the long run, multiple disturbances (e.g., extreme weather events, pathogen attacks, and wildfires) may change competitive dominance from oak forests to pyrophytic sclerophyllous shrublands (such as Cistus spp.). Such vegetation shifts have also been suggested by previous studies (Acácio et al., 2009; Moreno, 2009; Pinto-Correia et al., 2011; Acácio & Holmgren, 2014) and predicted under future climate scenarios (Ruiz-Labourdette et al., 2013) and may be reinforced by the continuation of land abandonment throughout the 21st century (Rounsevell et al., 2006).
Analysis of more time periods and more explicit abundance or demographic data will improve understanding of processes underlying forest transitions and successional dynamics potentially leading to vegetation shifts. Further research should also test for specific interactions between key drivers of change and for possible causal relations among drivers. Furthermore, models used to predict forest species distribution should account for anthropogenic variables, which are of major importance for the dynamics of highly managed Mediterranean forests (Keenan et al., 2011). A better understanding of patterns and drivers of forest dynamics and the inclusion of human drivers on models of vegetation change will improve predicting the future of Mediterranean oak forests.
Acknowledgements
We acknowledge the Portuguese Sea and Atmosphere Institute (Instituto Português do Mar e da Atmosfera, IPMA, I.P.) for the precipitation data (dataset PT02) used in this study. We acknowledge the E-OBS dataset from the EU-FP6 project ENSEMBLES (http://ensembles-eu.metoffice.com) and the data providers in the ECA&D project (http://www.ecad.eu) Haylock, M.R., N. Hofstra, A.M.G. Klein Tank, E.J. Klok, P.D. Jones, and M. New. 2008: A European daily high-resolution gridded dataset of surface temperature and precipitation. J. Geophys. Res (Atmospheres), 113, D20119, doi: 10.1029/2008JD10201. We acknowledge Xuebin Zhang and Feng Yang at the Climate Research Branch of Meteorological Service of Canada for developing and maintaining rclimdex. We also would like to thank Fireland Project (PTDC/AGR-CFL/104651/2008). We thank J. Carreiras for help with climatic data collection and for comments on a previous version of the manuscript. This research was funded by the Portuguese Foundation for Science and Technology (Fundação para a Ciência e a Tecnologia, FCT), fellowship number SFRH/BPD/80598/2011 to V. Acácio, fellowship number SFRH/BD/69021/2010 to F.S. Dias, and fellowship number SFRH/BPD/93373/2013 to F.X. Catry. F. Moreira was funded by the REN Biodiversity Chair and FCT (IF/01053/2015). The authors thank four anonymous reviewers whose comments strongly improved previous versions of the manuscript.
