Spatial predictions of land-use transitions and associated threats to biodiversity: the case of forest regrowth in mountain grasslands

Authors


Loic.Pellissier@unil.ch

Abstract

Question

Can we predict where forest regrowth caused by abandonment of agricultural activities is likely to occur? Can we assess how it may conflict with grassland diversity hotspots?

Location

Western Swiss Alps (400–3210 m a.s.l.).

Methods

We used statistical models to predict the location of land abandonment by farmers that is followed by forest regrowth in semi-natural grasslands of the Western Swiss Alps. Six modelling methods (GAM, GBM, GLM, RF, MDA, MARS) allowing binomial distribution were tested on two successive transitions occurring between three time periods. Models were calibrated using data on land-use change occurring between 1979 and 1992 as response, and environmental, accessibility and socio-economic variables as predictors, and these were validated for their capacity to predict the changes observed from 1992 to 2004. Projected probabilities of land-use change from an ensemble forecast of the six models were combined with a model of plant species richness based on a field inventory, allowing identification of critical grassland areas for the preservation of biodiversity.

Results

Models calibrated over the first land-use transition period predicted the second transition with reasonable accuracy. Forest regrowth occurs where cultivation costs are high and yield potential is low, i.e. on steeper slopes and at higher elevations. Overlaying species richness with land-use change predictions, we identified priority areas for the management and conservation of biodiversity at intermediate elevations.

Conclusions

Combining land-use change and biodiversity projections, we propose applied management measures for targeted/identified locations to limit the loss of biodiversity that could otherwise occur through loss of open habitats. The same approach could be applied to other types of land-use changes occurring in other ecosystems.

Nomenclature
Aeschimann et al.

(2004)

Introduction

Land-use change represents a major threat to biodiversity in most ecosystems across the world (Sala et al. 2000; Chemini & Rizzoli 2003; Araújo et al. 2008), with important consequences for ecosystem functions (Lenzen et al. 2008; Niedrist et al. 2009). Improving our understanding of the drivers of land-use change and being able to predict where such change may occur is essential for land management and conservation strategies. Forecasting where changes are likely to cause important species loss could be critical for optimizing applied measures. Yet, statistical models predicting land-use change have rarely been related to spatial biodiversity data at local scale in order to provide decision makers with spatially explicit projections of threats to biodiversity caused by land-use change.

In the last few decades, land-use change has affected natural and semi-natural ecosystems in many different ways. The intensification of human activities has had devastating effects on biodiversity and ecosystems (Tasser & Tappeiner 2002; Falcucci et al. 2007; Gehrig-Fasel et al. 2007; Niedrist et al. 2009). An example of this trend is the intensification of grassland management, with increasing amounts of fertilizer mainly in accessible areas (Tasser & Tappeiner 2002; Niedrist et al. 2009), while the capacity to use mechanized techniques that increase productivity has resulted in landscape homogenization (e.g. Falcucci et al. 2007). In contrast, land abandonment is occurring in more remote areas, especially in European mountains (Begueria 2006; Gehrig-Fasel et al. 2007; Tasser et al. 2007), in mediterranean (Millington et al. 2007) and nemoral (Ejrnæs et al. 2008) ecosystems, and in various other ecosystems (e.g. wet and dry grasslands, heathlands). Land abandonment is of concern in many regions: e.g. southern and northern Europe (Hatna & Bakker 2011; Kuemmerle et al. 2011), and North (Ramankutty et al. 2010; Thompson et al. 2011) and South America (Díaz et al. 2011). Land abandonment can be explained by the recent evolution of socio-economic constraints, particularly a loss of profitability, and the increasing cost of agricultural management in remote areas with low productivity (MacDonald et al. 2000; Tasser & Tappeiner 2002; Gellrich et al. 2007b). The recolonization of forests in many countries since the mid-19th century is a direct consequence of this trend in land use (MacDonald et al. 2000; Falcucci et al. 2007). This can have positive impacts for landscape restoration and potential recovery of faunal richness (Bowen et al. 2007; Falcucci et al. 2008), but can also represent a serious threat to biodiversity through loss of non-woody ecosystems, such as grasslands (Lenzen et al. 2008).

Agricultural land abandonment occurs mostly in less intensively used, semi-natural grasslands (Niedrist et al. 2009), which represent important sites for many species (Väre et al. 2003; Niedrist et al. 2009). Because the potential natural vegetation of these grasslands is generally forest, the long tradition of less intensive mowing and grazing has generated local hotspots of heliophilous species (Poschlod & WallisdeVries 2002). However, the extent of these semi-natural grasslands has decreased significantly as traditional shepherding has become less profitable, causing the loss of grassland species through competition with trees for light (Anthelme et al. 2001), resulting in a reduced species richness (Öckinger et al. 2006). A loss of richness spanning many clades has been demonstrated, particularly for vascular plants, insects and birds (Dullinger et al. 2003; Laiolo et al. 2004; Freléchoux et al. 2007; Marini et al. 2008; Sirami et al. 2008).

The ability to predict land-use changes and their consequences on biodiversity can allow improved focus of management and conservation strategies. Land-use change, particularly the change from semi-natural grassland to forest, and the underlying responsible mechanisms, have already been investigated using statistical models (Gellrich et al. 2007b; Millington et al. 2007; Pueyo & Beguerìa 2007; Rutherford et al. 2007; Tasser et al. 2007; Ejrnæs et al. 2008; Améztegui et al. 2010; Baumann et al. 2011; Corbelle-Rico et al. 2012). As a general trend, land abandonment and forest regrowth occur mainly in areas with low accessibility, less favourable local conditions for grassland productivity and for cattle grazing, and in areas where farmers and farms have become less frequent (MacDonald et al. 2000). However, these studies have not assessed at a fine scale whether the changes correspond locally to areas with high biodiversity, where management is most needed for conservation. Also, many such studies were explanatory rather than predictive and often relied on a single statistical method, not considering that an ensemble forecasting approach can lead to more robust projection (Araújo & New 2007; Marmion et al. 2009).

The aim of this study was to illustrate how combining predictions of forest regrowth following agricultural land abandonment and predictions of species richness in grasslands can allow identification of potential conflict areas, where rich grasslands may in future be threatened by forest regrowth. Considering the Western Swiss Alps, and using topo-climatic, accessibility and socio-economic variables, we calibrated a model of forest regrowth following land abandonment using data from a first land-use transition period (1979–1992), and evaluated its predictive power using independent data from a second transition period (1992–2004). In parallel, we modelled plant species richness (following Dubuis et al. 2011) across the same study area using data from an exhaustive vegetation survey. We then combined predictions from the land-use transition and species richness models to identify those grassland areas with high species richness that may potentially be threatened by future land-use change. Our two main questions were: can we predict where forest regrowth caused by abandonment of agricultural activities will occur in the near future; and can we assess conflict areas where forest regrowth following land abandonment would threaten species-rich grasslands?

Methods

Study area and field sampling

The study area, spanning 400 m to 3210 m a.s.l., is located in the Western Swiss Alps (Fig. 1). Vegetation in the study area has long been (and still is) influenced by human land use. Pastures and meadows interspersed within forest patches are common in this region, from valley bottoms up to sub-alpine and lower alpine areas. Land use is particularly intense at lower altitudes, where the open areas are intensively used as fertilized arable land and grassland, which are dominated by Arrhenatherum elatius and Poa spp. At intermediate elevations, agriculture is concentrated on cattle for milk and meat production, with fertilized pastures and meadows containing Cynosurus cristatus, Trisetum flavescens and Heracleum sphondylium. In areas having low economic benefit, e.g. steep slopes difficult to access with tractors for fertilization, some species-rich grasslands dominated by Bromus erectus are maintained, although these have been increasingly abandoned in the last three decades. Around the tree line, fertilization is generally absent and management is restricted to summer grazing in low-productive pastures dominated by Poa alpina, Sesleria caerulea or Nardus stricta.

Figure 1.

Study area, Swiss Western Alps, Canton de Vaud. The 912 inventoried plots are indicated as white points, while the background layer shows topographic variation over the study area.

Following an equal random stratified sampling design (Hirzel & Guisan 2002) from 2002 to 2009, we sampled randomly 912 vegetation plots across a variety of non-forested ecosystems in the study area, based on three main stratifying variables: elevation, slope and aspect. Inventories of all vascular plant species were carried out within 4-m2 plots in grasslands (see Pellissier et al. 2010a; Dubuis et al. 2011).

Explanatory and response variables

Lattice maps of land use, with survey pixels at intersections of a 100-m grid covering the study area, were obtained from the Swiss Federal Office of Statistics (www.geostat.admin.ch) for the years 1979, 1992 and 2004. Attribution of a land-use class to each pixel was based on visual interpretation of aerial photographs for the survey years (www.geostat.admin.ch/geostat). Only one class could therefore be attributed to each pixel and no neighbourhood analysis was needed to interpolate values or rasterize thematic vector maps. For each land-use layer, categories were classified in open agricultural land and forested surfaces from the 72 original classes, following Gellrich et al. (2007b). We compared the reclassified land-cover maps between the three time periods to identify changes from open grassland to forest that occurred in the time frame considered, and obtained three transition maps: transition-1 from 1979 to 1992, transition-2 from 1992 to 2004, and transition-3 from 1979 to 2004 (total transition). Each transition map contains two classes: stable grassland (pixels classified as grassland in both maps) and grassland changed to forest. To calibrate the land-use change model (see details in next section), we selected all pixels where a change from semi-natural grassland to forest occurred in addition to 5000 randomly selected pixels that remained as grassland.

Based on the available studies investigating land-use change in the Alps (see Gellrich & Zimmermann 2007; Gellrich et al. 2007b; Rutherford et al. 2008), we selected a set of variables that we a priori considered to be key factors driving land abandonment/forest regrowth. These variables could be grouped into three categories: environmental conditions, accessibility for humans and socio-economic conditions. To account for productivity and for suitability of the habitat for development of forests, we considered three topo-climatic variables: degree-days (DDEG), total precipitation (PREC) and solar radiation (SRAD). SRAD, calculated using Kumar's approach (Kumar et al. 1997), considers both direct and diffuse components of global solar radiation. DDEG was calculated from monthly average temperatures following Zimmermann & Kienast (1999), who also provided the algorithm to calculate PREC. Assuming that areas closer to existing forests are more likely to be colonized by regrowing forests, we also measured the minimum Euclidean distance to the nearest forest stand.

To account for accessibility and management costs, we calculated slope and the minimum Euclidean distance to roads and to the closest building using digital data available from the Swiss Federal Office of Topography. All variables were calculated for the three time frames. Note that although elevation per se is also a good indicator of management costs, it is highly correlated to DDEG, and thus we did not consider it in the model. All these layers were calculated as grid lattices at a resolution of 100 m. Given that land abandonment is also linked to changes in rural activities and to a decrease in farm employment (MacDonald et al. 2000), we also considered the proportion of farms among the total industries and the density of workers in the rural sector for each county inside our study area. These data were obtained from the Swiss Statistical Office.

Modelling land-use changes

We chose a statistical approach to model land-use transition (Rutherford et al. 2008), using the BIOMOD package (Thuiller et al. 2009) developed in the R environment (R Development Core Team 2009; R Foundation for Statistical Computing, Vienna, AT). Out of the models available in BIOMOD, we considered six modelling techniques that show good predictive accuracy (Elith et al. 2006; Pellissier et al. 2011): generalized boosting models (GBMs), generalized additive models (GAMs), generalized linear models (GLMs), multivariate adaptive regression spline (MARS), mixture discriminant analysis (MDA) and Breiman and Cutler's Random Forest for classification and regression (RF). First, we calibrated the models using transition-1. We evaluated the predictive performance of these models both internally, with a repeated split-sample procedure (50 repetitions using a subset of 70% of the data to calibrate the model and the remaining 30% for the evaluation), and externally, comparing the model fitted over transition-1 with independent data from transition-2. We evaluated the models measuring the area under the curve (AUC) of the relative operating characteristic (ROC) plot (Fielding & Bell 1997) and the true skill statistic (TSS; Allouche et al. 2006). For internal evaluation, we calculated the average AUC from the repeated split-sample procedure, obtaining a robust estimate of the predictive performance of the models. In addition, we measured the importance of the variables in each modelling technique based on a sequential randomization procedure implemented in BIOMOD (Thuiller et al. 2009).

Second, we calibrated the models using transition-3, i.e. the total transition representing all land-use changes in the last 30 yr (1979–2004), and evaluated them internally (see above). To produce more parsimonious models, we used only the variables that were important in the previous models (distance to settlements, distance to roads, density of workers in the primary sector, proportion of industry in the primary sector were not considered, see Results section). We projected the models for transition-3 over the existing semi-natural grasslands in the study area. We combined the single projections of each model for transition-3 using the ensemble forecasting approach, by calculating an average of the six models weighted by their predictive power (Araújo & New 2007). This procedure has been shown to significantly improve the predictive power of projections (Marmion et al. 2009). We obtained a final map displaying the probability of each pixel classified as agriculture in 2009 to be occupied by forests in the near future.

Biodiversity assessment

We modelled plant species richness over the entire study area using a generalized additive model (GAM) with four degrees of freedom and a Poisson distribution considering the 912 vegetation plots. We used the following explanatory variables: DDEG, SRAD moisture index of the growing period (MIND), slope and curvature calculated from a DEM (Zimmermann & Kienast 1999). MIND was calculated as the difference between monthly precipitation and potential evapotranspiration (Zimmermann & Kienast 1999). These five variables are usually considered as meaningful physiological and ecological constraints explaining species richness patterns (Pausas & Austin 2001).

To evaluate the resulting GAM, we calculated the explained variance and measured model calibration using a repeated split-sample procedure (Dubuis et al. 2011). For each 50 split-sample repetitions, a Spearman rank correlation between observed and predicted species richness was calculated on the evaluation data set. We rescaled the predicted species richness values between 0 and 1, with 1 corresponding to the pixels with the highest richness prediction.

To identify the areas with both high-predicted species richness and a high probability of changing from semi-natural grassland to forest, we multiplied the probabilities of land-use change by species richness predictions. In the resulting map, values close to 0 correspond to areas with low biodiversity and/or a low probability of changing from semi-natural grassland to forest, while values close to 1 correspond to areas with high species richness and high probability of being colonized by forest.

Results

Models fitted using the first transition showed a fair predictive power of forest regrowth when evaluated both internally and externally. For transition-1 (1979–1992), AUC values ranged from 0.77 to 0.78 for the internal evaluation (mean = 0.77), and from 0.73 to 0.77 (mean = 0.75) for the external evaluation (Table 1).

Table 1. Area under the curve calculated for the six models calibrated using transition-1 (1979–1992) and transition-3 (1979–2004) validated using internal evaluation (IE) and external evaluation (EE, transition-2, 1992–2004): generalized additive models (GAM), generalized boosting models (GBM), generalized linear models (GLM), multivariate adaptive regression spline (MARS), mixture discriminant analysis (MDA) and Breiman and Cutler's Random Forest for classification and regression (RF)
 Model of transition-1Model of transition-3
IEEEIE
GAM0.7780.7510.77
GBM0.7730.7570.775
GLM0.7740.7480.768
MARS0.7680.7260.738
MDA0.7670.7240.731
RF0.7650.7470.757
Mean0.770.740.757

The DDEG and distance to forest were the most important variables in explaining land-use change, closely followed by slope and PREC. SRAD and distance to urban settlements had lower importance, while distance to roads and proportion of rural industry and farmer density had almost no effect in all models. This indicates that the widely available topoclimatic variables were sufficient to predict land abandonment followed by forest regrowth (Table 2). According to the ensemble forecast of transition-3, the semi-natural grassland areas most likely to be colonized by trees are primarily gaps situated in large forest patches on steep slopes from intermediate elevations to the tree line (Figs 2, 3). The GAM model of plant species richness explained 44% of the variance and showed a good correlation between observed and predicted values (average Spearman correlation = 0.59). Species richness was predicted to be higher at intermediate elevations (refer to Dubuis et al. (2011) for a more detailed discussion of the drivers of plant species richness in the study area). In general, areas that might suffer high species loss due to recolonization of forest were mainly located in the middle of large forest patches at intermediate elevations (Fig. 3).

Table 2. Importance of the variables (degree-days, DDEG; Slope; solar radiation, SRAD; precipitation, PREC; distance to forest, Forest; distance to settlements, Building; distance to roads, Roads; density of workers in the primary sector, Workers; proportion of industry in the primary sector, Industry) in each model calculated with a permutation procedure (see Methods). Full names for each modelling technique are provided in the caption of Table 1
 DDEGSlopeSRADPRECForestBuildingRoadsWorkersIndustry
GAM0.2430.1480.0110.2080.4550.0010.0020.0080.012
GBM0.0750.1470.0130.0190.5280.0050.0010.0030.002
GLM0.3390.2040.0090.3680.400000.0230.027
MARS0.4590.3750.0590.3330.3300000
MDA0.6260.4060.1190.3870.1580000
RF0.3180.3460.1720.1980.3000.1210.0230.0350.059
Mean0.3430.2710.0640.2520.3620.0210.0040.0120.017
Figure 2.

Relationship between elevation and (a) plant species richness and (b) probability of forest regrowth following abandonment.

Figure 3.

Probabilities of forest regrowth following abandonment multiplied by the values of grassland species richness, as one way of highlighting which sites could be in most urgent need of human intervention to limit erosion of grassland diversity. Predictions are made on a grid of pixels (lattice) separated by a distance of 100 m.

Discussion

Körner (2005) suggested that grassland abandonment can potentially outweigh the impact of climate change on mountain vegetation, and recent field and modelling investigations support this hypothesis (e.g. Albert et al. 2008; Vittoz et al. 2009). In the study area and during the investigated time frame, forest has increased by 7%, mostly in less exploited areas. If this rate persists, an increase of 35% in forests over semi-natural grasslands is expected by 2100. In the same study area, Randin et al. (2010) showed that as a consequence of climate change more that 50% of species will lose 80% of their suitable habitat by 2100. However, while plant species may locally persist in less suitable climate (de Witte & Stöcklin 2010), forest recolonization almost immediately leads to the exclusion of species. Robust methods to assess its effects on biodiversity are still lacking, especially at a fine scale and in complex landscapes, as found in mountain areas. Our results indicate that statistical models provide useful predictive accuracy (sensu Swets 1988) of the transition to forest, which supports their usefulness to predict locations of future land-use transitions and how these may coincide with current surfaces with high species richness. The ability of the models to predict an independent transition occurring one decade later than that used to calibrate them further supports the idea that the factors causing forest regrowth are stable over the time frame considered. When coupled to projections of grassland species richness, the models can highlight areas where both grassland diversity and the likelihood of forest regrowth are high, and therefore where management is most needed to protect these threatened species-rich grasslands.

Our results suggest that forest regrowth occurs where cultivation costs are high and yield potential is low, i.e. on steeper slopes and at higher elevations (Gellrich et al. 2007b; Rudmann-Maurer et al. 2008). Forest regrowth occurred mainly close to existing forests, which can be explained by the generally negative exponential pattern of tree dispersal kernels (Nathan & Muller-Landau 2000), in artificially cleared areas below the natural tree line (Gehrig-Fasel et al. 2007). Also, land abandonment followed by forest regrowth occurred mainly at intermediate to lower values of degree-days and intermediate to higher values of precipitation, where grassland biomass productivity is reduced, but not at very low temperatures that are not suitable for forest growth. Land abandonment has also occurred on steeper slopes less accessible to cattle and mowing. Socio-economic variables overall had low importance in our models. Probably, given the coarse resolution of these variables in our models (county level), they might not be proximal enough to account for land abandonment at the 100-m resolution used. The density of cattle or the size of farmlands would certainly have been better variables, but such information is currently not available in a spatially explicit form. Also, at the scale of the study area as a whole, evolution of the socio-economic context might lead to an increase in land abandonment, but only physical constraints currently contributed to explain where they are expected to occur locally.

Millennia of agriculture using fire, mowing and domesticated herds have created large, semi-natural grasslands allowing the settlement of a large number of plant species characteristic of open habitats (e.g. up to 65 vascular plant species in a 4-m2 plot in the study area). The recent societal modification of agricultural practices has led to the intensification of human activities in more productive areas, while less fertile lands have been abandoned (MacDonald et al. 2000; Strijker 2005). As a consequence, extensive semi-natural grasslands tend to constitute the last remnant of occupied area for many species (Fischer & Wipf 2002). Our analyses indicate that these areas, often located at intermediate elevations (700–1500 m a.s.l.), are most sensitive to forest regrowth. In particular, our model showed that land abandonment followed by forest regrowth may cause the highest species loss in grasslands at intermediate elevations, by reducing the heterogeneity of open vs closed areas. To protect these areas, some management measures are required, for instance, through selecting an adequate number of surfaces to be moderately mowed or grazed. Our approach can be considered as a potential tool for optimizing such selection procedures.

One caveat to our approach is the assumption that the socio-economic context will remain identical in future years. This may not necessarily occur, as governments tend to become increasingly sensitive to the problem of mountain farmers and allocate funds to support their activities. Hence, the models applied in this paper remain useful as long as the socio-economic context does not change significantly. Where future societal changes are expected, the approach would need to be combined with scenarios reflecting socio-economic changes to adjust for on-going modifications. A second caveat to our approach is that, to identify areas with both high land-use change probability and high biodiversity, we multiplied predictions by the two models, treating them as independent variables. We used this simplistic approach because it was sufficient to illustrate our approach and because the number of pixels considered was too high to apply more advanced methods. Given that the explanatory variables considered for the two approaches were partly the same, we cannot claim complete independence between the two types of prediction. Our combined predictions should thus not be considered as absolute risk probabilities, but rather as a relative ranking of the pixels in the study area based on a simple combination of forest regrowth and richness probabilities. Future studies could use more advanced approaches, for instance involving Bayesian combinations of the two predictions.

Our approach was successfully applied to one type of land-use change in one geographic area where forest regrowth is the main successional consequence of land abandonment. Considering the increasing importance of forest regrowth following land abandonment in many regions in Europe (e.g. Hatna & Bakker 2011) and in North America (Ramankutty et al. 2010; Thompson et al. 2011), our approach is promising, in that it may help manage ecosystems to reduce biodiversity losses. The approach used in this study should, in principle, also be applicable to any ecosystem around the world facing any type of quantifiable land-use change. For instance, the transition between extensively and intensively managed agricultural land (Reidsma et al. 2006) or deforestation in tropical forests (Dale et al. 1994) or savannas (Velazquez et al. 2003; Jepson 2005) could possibly be similarly modelled using the most relevant predictors of land-use transition. In addition, predictions of species richness, as illustrated here, should likely be complemented with predictions of other community indices, such as phylogenetic and functional diversity or the proportion of species on red lists (e.g. Pellissier et al. 2010b; Pio et al. 2011). Such a combined modelling exercise linking land-use change with biodiversity data should now be tested on other landscape and ecosystem types, to assess more generally if this approach can contribute to optimize restoration or conservation actions, especially in biologically rich and threatened areas.

Acknowledgements

We thank John Dwyer and two anonymous referees for their comments on an earlier version of this manuscript. This research has been supported by the Centre de conservation de la faune et de la nature (Canton de Vaud), the SNF grant Nr. 31003A-125145 (BIOASSEMBLE project), the European Commission (ECOCHANGE project) and the National Centre for Competence in Research (NCCR) ‘Plant Survival in Agro-Natural Landscapes’ in Neuchâtel.

Ancillary