Introduction

To effectively protect biodiversity, it is important to identify areas where high numbers of rare, localised or endangered species are concentrated (Myers 2003). For plant biodiversity, the majority of these areas are located in montane regions (Muellner-Riehl et al. 2019), likely reflecting how mountains contain steep environmental gradients, often harbour unique geological strata, and have complex topography, which leads to a large diversity of niches within relatively small areas (Antonelli et al. 2018; Rahbek et al. 2019). Additionally, many individual mountain tops act as islands in the landscape because of their distinct climatic conditions and are thereby drivers of evolutionary radiation (Riebesell 1982). However, despite their relative isolation, montane ecosystems may face threats in the form of mining, overgrazing and climate change (Schmeller et al. 2022).

Our knowledge of plant diversity in mountains depends on herbarium collections gathered over many years. However, mountain regions are often poorly sampled as the terrain can be largely inaccessible, resulting in large parts of these areas remaining under-explored (Clark et al. 2006; Li et al. 2011; Sainge et al. 2017; Wursten et al. 2017). As a result, these herbarium collections can describe a regions’ overall species diversity and inform broadly about the occurrence of rare and endemic species, but they only partially reveal distribution patterns in mountain environments, without actually characterising species environmental requirements, which is necessary for effective conservation efforts (Greve et al. 2016).

Conservation planning often considers rarity and endemism as criteria for target identification, but this approach overlooks important aspects of plant life in mountains. Accurate assessments of ecosystems must also include the distribution of common species since the functioning of many ecosystems and the associated ecosystem services are only directly influenced by rare species when they are locally abundant or their ecological role is unique (Dee et al. 2019). A focus on common plant species is thus necessary if we want to understand how plant species and their niches are distributed within mountain regions and how these species affect the ecological functioning on a larger scale. Such information will help in understanding plant distributions, but also in monitoring vegetation changes and predicting future floristic responses to climate change (Silveira et al. 2019). In addition, since species distributions are shaped by complex interactions between individual biological traits (e.g., growth forms, morphological adaptations, phylogenetic niche conservatism) and environmental constraints (Carscadden et al. 2020), comparing drivers of distribution across multiple species with different traits can provide important insights of how these components determine range size and niche breadth.

Ecological niche modelling is a tool used to predict suitable environments for a species on the basis of incomplete occurrence data by extrapolating the information from those limited occurrence data to the entire sampled region (Weber et al. 2017). This approach makes use of different statistical modelling algorithms and can be utilised even when complete data (including absence data for a species) is unavailable. It does, however, require detailed environmental data so that the occurrence of the species can be linked to environmental conditions, allowing projecting the species occurrence throughout space and time (Pearman et al. 2008). In the Drakensberg, previous efforts to apply niche models to understand species distributions (Bentley et al. 2019; Gwate et al. 2023) were mainly based on presence-only models, which can limit model evaluation and calibration (Drake 2014), and coarsely interpolated climatic data only, which may fail in representing the topographical climate variation at the microhabitat level (Franklin et al. 2013).

Climatic conditions are among the most important explanatory factors that determine the distribution of plants in any environment at broader scales (Punyasena et al. 2008). Climate plays a particularly large role in mountain environments because there is a high amount of local variation in rainfall, temperature, evapotranspiration and wind direction and speed (Leuschner 2000; Scherrer and Körner 2011; Körner 2021). However, there is a lack of detail in climatic data in the mountains, as weather stations are sparse, and climatic gradients are very steep and not easily modelled (Barry 1992). Therefore, when extrapolations of climatic data are done for mountainous environments, they are typically coarse estimates and often not very reliable (Schmitt et al. 2013; Shahgedanova et al. 2021).

Topography has a large impact on local climatic conditions in mountains, across both large (e.g. temperature variation across elevation) and smaller scales (temperature variation on adjacent north vs. south-facing slopes; Kattel et al. 2015). For example, temperature generally declines with elevation (Montgomery 2006), but sheltering and shading in deep valleys and the effect of anabatic and katabatic airflows may strongly modify this pattern locally (Sturman 1987). Rainfall can have a complex pattern of local variation too, as it depends on the interaction of topography and prevailing wind directions. The effect of air forced to rise up mountain slopes often creates unstable conditions and orographic rainfall, but how this rainfall is distributed over different slopes depends on fine details in topography and air movement through valleys (Barry 1992; Neiman et al. 2002). Air movement in mountains is extremely complex as it involves daily fluctuations in air pressure as a result of differential heating of slopes and valleys, and this, in turn, has an effect on potential evaporation and temperature. In consequence, while we understand that climatic conditions can vary strongly across short distances within mountains, the characterization of this finer scale variation in temperature, rainfall and other variables is in practice limited by data availability and the development of microclimatic models (although see, e.g. Kearney et al. 2020).

As a result of the climatic and topographic heterogeneity, it is a particular challenge to provide realistic assessments of plant distributions within mountain ranges at the landscape scale. In this study, therefore, we link plant distributions directly to fine-scale topographical information that can be obtained from data layers that are directly available from satellite imagery or large-scale maps, assessing (1) the accuracy with which species distributions can be modelled using widely available environmental data in a region with limited botanical survey data and (2) which environmental variables are most frequently related to species occurrence patterns. By understanding the distributions of common species in mountains, we can enhance our understanding of the environmental drivers affecting the occurrence of plants across mountain environments.

Methods

Study area

Our study region is the northern Maloti-Drakensberg mountains in eastern South Africa, adjacent to Lesotho, at the border of the Free State and KwaZulu-Natal provinces. This region is characterised by a temperate climate with warm summers and cold winters that are dry with regular frost and occasional snowfalls. It also experiences summer rainfall in the form of thunderstorms that arise from convective and orographic cloud formation.

The Drakensberg massif is an erosional mountain range that represents the remaining areas of basalt that were deposited after a period of massive volcanic activity following the breakup of Gondwanaland (McCarthy and Rubidge 2005) This mountain range has very steep slopes towards its eastern face, where it rises c. 2000 m above the coastal province of KwaZulu-Natal. The Free State province is situated on an inland high-altitude plain in the interior of South Africa that was formed during a period of uplift at the end of the Cretaceous. The eastern escarpment of the plateau intersects with the Drakensberg escarpment (Fig. 1), and the area around this intersection, therefore, has a topography that is characterised by plains of three different elevations intersecting: inland KwaZulu-Natal at about 1200 m altitude, the eastern Free State at around 1700 m altitude and the Lesotho Highlands at around 3000 m altitude. Rainfall in this area is primarily orographic, coming from the Indian Ocean in the east. Less commonly, cold fronts also sweep in from the west. There are weather stations in several towns in the foothills of the mountains, but only a single weather station in the high mountains (around Sentinel Peak) that was run for a short period of time (Nel et al. 2010). For most of the area, the weather is complex and unpredictable.

Fig. 1
figure 1

a Survey sites in the northern Maloti-Drakensberg mountains in South Africa, bordering Lesotho, showing dominant vegetation types (Mucina and Rutherford 2006). The photographs below show some of the selected species that were recorded in this study: b Gladiolus crassifolius, c Cyperus sphaerocephalus, d Helichrysum chionosphaerum, e Ajuga ophrydis

Full size image

Field sampling

Thirty common grassland forbs were chosen for this study (Table 1; Fig. 1). These species were selected using the following criteria: (1) they should be found regularly in the area, and their distribution should not be confined to a single mountain or valley, (2) they should have a high detectability and be reliably identified, without any confusion with other forbs in the area, and (3) they should remain in this recognisable state throughout the flowering season, either because they flower for a long period, or because their vegetative form is conspicuous and easily recognised. These species covered a broad range of growth forms, including geophytes, succulent forbs and dwarf shrubs. Of the 30 target species, only 19 species were recorded in more than 10 plots, with subsequent data analysis limited to this subset of species.

Table 1 Species selected for this study
Full size table

Transects were laid out in the mountains around the town of Phuthaditjhaba, either following contour lines or moving upslope in valleys or along ridges between 1800 and 2600 m a.s.l. These transects were accessed by foot during the summer growing seasons (November to March) of 2016, 2017 and 2020, and a 10 × 10 m plot was surveyed every 100 m along each transect, in which the occurrence of each of the thirty target species was recorded. Altitude, rock cover, vegetation cover, vegetation height, slope angle and slope aspect were additionally quantified in the field for each plot. Prior to analysis, slope aspect was decomposed into a measure of northness (i.e. 1 = north-facing, − 1 = south-facing, 0 = east- or west-facing) and eastness (following, e.g., Guisan et al. 1999). This means there are six predictor variables that were assessed in the field. A total of 231 plots were surveyed along eight transects.

Desktop predictor variables

Additional predictor variables were assessed from desktop studies: solar radiation, topographic wetness index, distance from watershed, and the categorical variable soil type.

To determine the solar radiation potentially received at each plot in the study area, the “solar radiation analysis tool” from ArcGIS 10:8.2 (ESRI 2023) was used. The tool maps the potential incident solar radiation over a geographic area for specific time periods. It accounts for atmospheric effects, site latitude and elevation, slope steepness and aspect, daily and seasonal shifts of the sun’s angle, and effects of shadows cast by surrounding topography (ESRI 2023). The tool requires a Digital Elevation Model (DEM) of the study area. Therefore, a 30 m DEM for South Africa and Lesotho was obtained from the Shuttle Radar Topography Mission (SRTM; Drusch et al. 2012) and clipped to the study area. The “solar radiation analysis tool” with the following parameters was selected: output files were created for each month; a sky value of 512 (which is recommended as sufficient for calculations at point locations where calculation time is less of an issue and where you are using monthly averages); a viewshed of 32 (due to the complexity of the landscape in the study area); 32 calculation directions (due to the complex terrain); and finally, ground level was selected for reading height.

Topographic wetness index (TWI) was calculated following the methods of Dadjou (2021). The TWI shows a measure of wetness conditions at the catchment scale. It combines local upslope contributing area and slope using the digital elevation model to represent increased soil moisture where the landscape area contributing runoff is large and slopes are low. This process also uses the SRTM-DEM developed for the solar radiation calculation. Once the TWI layer was calculated the survey points were overlayed and a TWI value for each survey plot could be determined.

Soil data were obtained from the Soil and Terrain database for Southern Africa (SOTERSAF version 1.0) at a scale of 1:2 million (Batjes 2004). The vector layer was imported into ArcMap in order to obtain the data appropriate for each plot. This led to the categorization of soils as one of three types: Lithic leptosols, Eutric leptosols or Paraplinthic Acrisols.

The distances for each plot from the both watersheds (Lesotho border and Free State-KwaZulu-Natal border) were calculated because these are a proxy for orographic rainfall, with the distant plots receiving less rainfall than the proximate plots. The tool for calculating this distance was gDistance() function from the rgeos R package version 0.6-2 (Rundel and Bivand 2013), using a base raster layer of 30 m resolution.

Species distribution models

All continuous environmental variables were examined for collinearity using Pearson correlation tests. Due to strong collinearity between potential incident solar radiation across most months, only solar radiation values for June and December were retained for subsequent analyses. Variance inflation factors (VIF) showed that soil type (the only categorical predictor) was not strongly correlated with any of the other predictors. After removing vegetation cover and distance from the Lesotho border, there were 11 candidate predictor variables retained, all with r < |0.6| and VIF < 2.2 (see Online Appendix; Fig. A1). Vegetation height was log-transformed prior to analyses.

To test if quadratic terms should be considered for each of the continuous predictor variables (i.e. to account for non-linear and/or unimodal species–environment relationships), each predictor was fitted to each species’ occurrence data using a generalised linear model with a binomial distribution. Two models were fitted for each species-predictor combination, a linear model and a quadratic model, and a quadratic term was only considered useful when the quadratic model had a lower AIC score than the associated linear model and the quadratic model improved the deviance explained (% DE) by the linear model for the species by > 5%. Quadratic terms only met these two criteria for > 1 species for altitude, rock cover and distance from watershed. As a result, quadratic terms were only considered further for these three predictor variables.

A best subsets model building approach was then used to identify the best combination of predictor variables (from 11 predictors and 3 quadratic terms) for each species, with transect identity included as a random effect to account for the spatial structure of the sampling design. Species occurrence was modelled using a generalised linear mixed effect model with a binomial distribution, with all resulting models ranked using AIC. Model fit was assessed using the area under the curve of a receiver operating characteristic plot (AUC; Fielding and Bell 1997) and the true skill statistic (TSS; Allouche et al. 2006), based on three repeats of three-fold cross-validation (with random division of the dataset). Relative variable importance was then calculated for each variable included in each species’ top ranked model, using the approach of Niittynen and Luoto (2018). As variable importance cannot be determined for random effects, two analytical approaches were used when assessing model fit and variable importance: transect identity was dropped from the model (likely overestimating the relative importance of the remaining predictors, and underestimating model fit) or was treated as a fixed effect (potentially overestimating its importance and overestimating model fit).

All statistical analyses were run in R statistical software v 4.2.2 (R Core Team 2022), with additional functions from the corrplot (Wei et al. 2022), biomod2 (Thuiller et al. 2023), car (Fox and Weisberg 2018), glmmTMB (Brooks et al. 2017) and modEvA (Márcia Barbosa et al. 2013) libraries.

Model projection

The best-performing models were used to project the occurrence of species across the entire study region using the function “predict” from the R package “terra” version 1.7-39 (Hijmans et al. 2022). Since some variables used to model the presence of some species were not available for areas beyond the sampled plots, we restricted this step to species in which the top-performing models included only predictors which could be derived from DEMs: elevation, slope, northness, eastness, solar radiation in June, solar radiation in December, TWI, and distance to the watershed. For elevation, we used the same 30 m resolution SRTM-DEM used in the previous step. The solar radiation data was obtained from the Photovoltaic power potential (PVOUT) raster published by the Global Solar Atlas dataset version 2.0 (ESMAP 2019) under the resolution of 250 m. For the Topographic Wetness Index, we used the Africa Soil Information Service (AfSIS): Topographic Wetness Index (TWI) (Vagen 2010) under 90 m resolution. For topographic eastness and northness, we used the DEM raster and the functions “raster.east” and “raster.north” available on the package “red” version 1.5.0 (Cardoso 2023). All raster resolutions were resampled to 30 m resolution using bilinear interpolation and cropped to the extent of the study site using the R package “terra”. The predicted suitable area for each species was calculated using a threshold based on a fixed sensitivity value of 0.9 (Liu et al. 2005; Bean et al. 2012). The model sensitivity was calculated using the presence and absence values from the sampled sites with the function “evaluate” from the package “dismo” version 1.3-9 (Hijmans et al. 2017). We used the calculated thresholds to reclassify the projections into presence/absence maps using the function “classify” in the R package “terra”. Finally, the extent of the suitable area for each species was calculated using the function “expanse” in the same package.

Results

Occurrence models varied strongly in performance, with mean AUC scores from cross-validation ranging from 0.68 to 0.94 and mean TSS scores from 0.22 to 0.80 (when including Transect as a fixed effect; AUC scores were on average 0.03 units lower and TSS scores 0.07 units lower when Transect was excluded from models; Table 2). All predictor variables (and their associated quadratic terms) were included in at least one species’ best fit model, with altitude (13 species), distance to watershed (13 species) and rock cover (8 species) being included in the most best-fit models (Table 2; Fig. 2). None of the predictors that were included in more than one best-fit model had consistent signs for their coefficients (i.e. the nature of the occurrence-predictor relationship was not consistent across all of the species where that predictor was included in the best-fit model). For example, altitude showed positive (3 species), negative (4 species), humped-shaped (3 species) and valley-shaped (2 species) relationships with the occurrence of the modelled species.

Table 2 Best-fit occurrence model for each species, indicating the model fit (AUC = area under the curve, TSS = true skill statistic), and the variables included in the best model and their coefficients (+ = positive linear relationship; − = negative linear; ∩ = hump-shaped quadratic; U = valley-shaped quadratic)
Full size table
Fig. 2
figure 2

Exemplar response curves illustrating contrasting species responses to a altitude, b distance to the watershed, and c rock cover. See Online Appendix B for all response curves for all species

Full size image

Variable importance was highest for distance to the watershed, soil type, and altitude (Table 3). Northness also had a high average variable importance, but this was driven by northness being the only variable included in the best-fit model for Schistostephium crataegifolium. Repeating the variable importance analysis including transect identity as a fixed effect did not alter these results (with transect being the 5th most important variable; Online Appendix Table A1).

Table 3 Variable importance (%) for each fixed effect predictor included in the best model for each species (excluding Transect as a predictor)
Full size table

Habitat suitability and distribution across the mountain region projected by the occurrence models varied considerably across species. Only in the cases where rock cover and vegetation height did not play a major role, could spatial distributions of species be modelled onto a digital elevation model (11 species; see Fig. 3). Some species influenced by similar combinations of variables (e.g., elevation and distance to watershed in Crassula vaginata and Lotononis lotononoides) showed similar response curves, and differed only in terms of the relative contribution of each variable to the overall explanatory power of the model (e.g. the relative importance of distance to watershed was 91% for C. vaginata and 46% for L. lotononoides). The response maps also show that some species, such as Helichrysum aureum and Senecio rhomboideus, were predicted to be approximately equally common across the study region. Species belonging to similar growth forms (Table 1) have yet very different distribution patterns.

Fig. 3
figure 3

Examples of species’ projected occurrence probabilities across the study area. Dark red tones indicate higher predicted probabilities of occurrence, while dark blue tones indicate lower probabilities. All projections are in WGS84 and 30 m resolution

Full size image

Projected suitable habitat across the study area varied strongly between species, ranging between c. 40 and 100% of the area (Fig. 4). Species such as Agapanthus campanulatus, Dianthus basuticus and Senecio macrocephalus are predicted to be relatively range restricted, while Helichrysum aureum, Helichrysum pallidum and Senecio rhomboideous show large suitable areas.

Fig. 4
figure 4

Estimated suitable area for each species that had explanatory models using only topographic variables (i.e. excluding rockiness and vegetation height) across the study region (see Fig. 1)

Full size image

Discussion

For the majority of species, elevation strongly affected species’ occurrence, combined with various other topographic factors, often including rock cover, distance from the watershed, and/or soil type. The impact of altitude on species distribution has been highlighted in many studies (e.g. Wang et al. 2003), as altitude is linked to temperature and often also to orographic rainfall, wind intensity and/or cloud cover (Körner 2007). Interestingly, in this study, some species seem to have a bimodal distribution with regards to altitude. This may indicate a preference for habitats such as relatively flat mountain tops or crests that occur at low altitude on sandstone and again at high altitude on basalt.

In this system distance from the watershed also correlates with the amount of orographic rainfall, as the escarpment edge (i.e. the watershed) is where moist air from the Indian Ocean is forced up due to the steep incline of the topography. Rock cover was an important factor for the occurrence of several species, likely reflecting a gradient of decreasing soil depth, increasing light availability, and greater microclimatic variation. As a result, areas of higher rock cover potentially offer, for example, opportunities for species that are less competitive in a dense grass layer. The influence of soil type is probably predominantly through its effect on nutrient status and hydrological properties of soils. Many studies have emphasized the importance of soil and other edaphic factors in mountain grasslands and, along with topography (including slope, aspect, altitude), are the most important environmental factors driving species occurrence patterns (Pickering and Green 2009; Huo et al. 2015; Buri et al. 2020).

While there tends to be pronounced variability in the environmental variables that are most strongly related to species occurrence patterns, and variation in the nature of the effect of those predictors on species occurrence, some predictors did show strong trends. For example, both eastness and the amount of solar radiation received during summer demonstrated predominantly consistent correlations with the occurrence of the species in which they were included in the top-ranked models (positive and negative correlations respectively). Both of these metrics relate to the solar inputs at site, with west-facing slopes in this system experiencing more direct incident radiation during warmer afternoon times (and with east-facing slopes often having reduced solar inputs due to morning mist). This suggests that in this system many, but not all, species benefit from cooler, less sunny microsites, having a higher probability of occurrence in locations with lower solar inputs. This finding is in strong contrast to observations from most other systems, where warmer, sunnier microsites are typically associated with higher plant richness and more complete vegetation cover (Kulonen et al. 2018). This depends a lot on the choice of species of the study. The ecologically most dominant species, the grasses, are not included in the study and in warm and sunny sites it is likely them that benefit from becoming more dominant. Large forbs and shrubs are typically dominant on scree slopes, drainage lines and rocky areas. There were no growth forms that showed similar responses here but the sample sizes are too small to make definite statements about that.

Mountain ecosystems are particularly vulnerable to climatic changes (Thuiller et al. 2005). It is expected that vegetation will generally expand upslope if mountain habitats become warmer and that high-altitude endemic species will therefore be under specific threat (i.e. the ‘escalator to extinction’ (Dirnböck et al. 2011; Pauli et al. 2012; Urban 2018). This is particularly relevant in our study system as elevation was an important predictor in most of the species’ distribution models. Indeed, elevation has always been recognised as a major influence on vegetation patterns in the mountains, and the majority of species are expected to shift their range in response to warming temperatures (Vitasse et al. 2021; Zu et al. 2021). However, given that a diversity of occurrence-elevation patterns were observed (positive and negative, linear and non-linear), a community-wide upslope range shift is unlikely, and instead a more complex reshuffling of grassland composition may occur in response to changing temperatures (as observed by e.g. Le Roux and McGeoch 2008).

The pronounced topographic variation (and, e.g., the associated edaphic heterogeneity) often observed in mountainous areas may, however, also provide some degree of buffering against the impacts of climate change (Scherrer and Körner 2011; Niskanen et al. 2017). For example, aspects of topography, such as solar radiation and rock cover, will not be impacted as a result of global climate change, even though there is a tendency for more rocky places towards higher altitudes. Given that many of the variables identified in this study as being important drivers of forb occurrence patterns will not be directly impacted by changes in rainfall or temperature regimes, these findings suggest variation in susceptibility to change across our study species (in agreement with, e.g., Moeslund et al. 2013). This differential susceptibility to changes in climate could, therefore, cause large differences in the degree to which species ranges shift, potentially leading to a greater mixing of species and an overall homogenisation of vegetation across altitudinal bands (Savage and Vellend 2015; Cramer et al. 2022).

Importance for conservation

Despite study species being selected on the basis of their ‘commonness’ in the region’s mountain grasslands, there was still a wide range in the predicted range of these species within the study region. This variation in species predicted occupancy likely reflects the predictor variables selected in this study (and their resolution), but also differences in species biology. The species that have strong habitat specialisation tend to occur in specific, limited environments. The number of variables involved in determining such habitats cannot be determined at the scale this study operated on. If a species occupies unique or rare environmental conditions, it may respond strongly to the arbitrary choice of threshold, resulting in a smaller predicted area (Boulangeat et al. 2012; Carscadden et al. 2020).

The predicted size of a species occupancy within a specific area does not have to correspond with the actual frequency with which that a species is encountered in the field (Crisfield et al. 2024). The predicted area size indicates the ‘potential’ habitat but because of biotic factors (e.g. presence of competition or the absence of mutualists) and dispersal constraints, the species may not occur in all the places that are indicated as potential habitat. Also, when species have a large potential distribution, neutral models apply and this may lead to a small actual distribution after a series of random dispersal and competition events over several generations, despite the equal tolerance for environmental conditions among all species (Hubbell 2001). This emphasizes why we need to be careful with the interpretation of the predicted range sizes as they do not necessarily reflect the rarity of species on the ground nor give an understanding of the specific threats that a species faces.

When rarity arises due to habitat specialisation and these habitats are at risk, it can help in prioritising which species inhabiting topographically complex mountainous environments require most attention for conservation and known existing populations of such species need to be monitored. Following the procedure outlined in this paper can help in setting such conservation priorities as it becomes evident which species are occurring in a narrow niche, even while they may still be commonly encountered. There are many studies that highlight conservation priorities of mountainous regions (for example Hou et al. 2010) but none of them focus on the landscape-scale of particular mountains and slopes. Knowing the distributions of species within mountain areas is particularly important when we consider the pressures that mountainous regions are subjected to in the face of climatic changes.

The method applied in this study offers an approach to modelling species distributions in montane regions at a landscape level, and can be used to identify species and areas for monitoring that may be most strongly impacted by environmental change. In this approach, the presence of species’ presences and several environmental predictors’ values were directly sampled on-site instead of derived from other geographic databases. This partially overcomes the limitation of the coarse resolution and the uncertainties associated with the low geographic coverage of mountain systems by the global climatic databases (Thornton et al. 2022). Even though some projections include variables that can be considered proxies for climate data derived from weather stations, removing this bias from the model training refines the results compared to models trained using spatially biased datasets.

Species distribution modelling at an intermediate scale can give a good indication of the way in which plant species are distributed within a mountain landscape. This reveals that different plant species, even if they are equally common across the mountain range, may have entirely different habitat requirements and will therefore respond differently to future changes. How plants respond to a complex of environmental predictors in a mountainous landscape is of crucial importance for the conservation of these species.