Species locality records

We utilized species data provided by the “EMYSystem Global Turtle Database” [25]. These data consist of locality records from Iverson [26,27], collected in a web site by the “Terra Cognita” laboratory (Geosciences Department of Oregon State University in Corvallis, Oregon). Iverson’s works [26,27] are recognized as the state of the art and a milestone in the study of turtle distributions and are still used as references in fundamental papers in this field [24,28,29], being considered the most authoritative, comprehensive, and precise source of turtle distribution data, especially in Africa where frequent updating are not available. The EMYSystem Global Turtle Database contains distribution records for every species of freshwater and terrestrial chelonian that has been collected by a museum, private individual, or referenced in a publication. In the database, locations for sea turtles and for non-marine species described later than 1992 are not included. Locality records are mapped at a spatial resolution of 0.01 geographic degrees (about 1.1 km).

In total, we considered 16 species of Testudinidae: Centrochelys sulcata, Chersina angulata, Homopus areolatus, Homopus boulengeri, Homopus femoralis, Homopus signatus, Homopus solus, Kinixys belliana, Kinixys erosa, Kinixys homeana, Kinixys natalensis, Malacochersus tornieri, Psammobates geometricus, Psammobates oculifer, Psammobates tentorius, Stigmochelys pardalis. These species are only a very small fraction of the overall African biodiversity. Nevertheless, according to the Turtle Taxonomy Working Group of the IUCN/SSC Tortoise and Freshwater Turtle Specialist Group [24], this set of species represents almost the entire fauna of tortoises in Sub-Saharan Africa, only two species being excluded (Kinixys lobatsiana, Kinixys spekii). However, other studies consider additional taxa which were not recognized by Iverson et al. [25] and Rhodin et al. [24]. For instance, some authors considered Kinixys nogueyi as a valid West African species [30,31] rather than a subspecies of Kinixys belliana as in Rhodin et al. [24]. The number of usable records was extremely variable between species (mean = 81.50, SD = 96.62): there was only one record for H. solus and, for the other species, the range was from 11 records (for K. natalensis) to 316 records (for K. belliana).

Environmental predictors

For both continental and regional analyses, we fitted species distribution models (SDMs) utilizing four groups of environmental predictors with high resolution (i.e. < 30 arc seconds): (i) 19 climatic variables from the WorldClim databank [32], (ii) two land morphology descriptors from the UNEP Environmental Data Explorer, (iii) three variables relating to water bodies from the FAO GeoNetwork, and (iv) one land cover variable from the JRC databank (see Table S1 for data sources and description). All the variables were re-sampled in ArcGIS 9.3 at the resolutions of one geographic degree for the entire continent and of five arc minutes for the regional analysis, by calculating the most common class for land cover and the average value for the other parameters. We calculated the variance inflation factor (VIF) to eliminate correlated variables [33]. VIF measures the degree of inflation of the unexplained variance due to the inter-correlation among independent variables [34,35]. We removed variables until all VIFs were below 5 [33], consequently we trained the models utilizing a subset of 11 and 10 predictors for continental and regional analyses respectively (Table S1). This process reduced the comparability of distribution models between scales but, at the same time, allowed to optimize the models by avoiding the introduction of biases due to correlation between predictors.

Modeling procedure

Because the choice of algorithm influences the performance of SDMs, we adopted an ensemble forecasting approach [36]. We used five different techniques for modeling habitat suitability in the R-based (version 2.8.1 [37]) package BIOMOD [38], at two spatial resolutions for the continental and the regional analyses (one degree and five arcmin respectively). We fitted Generalized Boosting Model (GBM [39]), Generalized Linear Models (GLM [40]), Multiple Adaptive Regression Splines (MARS [41]), Flexible Discriminant Analysis (FDA [42]), and random forest for classification and regression (RF [43]) on presence and pseudo-absence data for all the species occurring in the considered area. Generalized Boosting Model are models that combine two different techniques – regression trees and boosting – to optimize predictive performance [39], Generalized Linear Models are regression models that allow non-linear distributions for the response variable [40]. Multiple Adaptive Regression Splines are models that combine linear regression, mathematical construction of splines, and binary recursive partitioning to produce a local model [41]. Flexible Discriminant Analysis are classification models based on the well-known linear discriminant analysis [42]. Random Forest for classification and regression are models that create a suite of models using a classification and regression tree [43]. We selected 500 (in the continental analysis) and 200 (in the regional analysis) random cells across the study area as pseudo-absences in order to have a good representation of the background conditions. With a fixed number of pseudo-absence points, the prevalence changed across species and was proportional to the number of species presences. We validated each model by calculating the Area Under Curve (AUC) through a 10-fold cross-validation procedure [44]. For each species, we produced a single consensus model by calculating the weighted average of any single model with AUC > 0.7 [36] using AUC values as model weights.

We converted to predicted presence/absence the continuous values of habitat suitability (HS) predicted by the consensus models according to the minimal predicted area criterion [45,46], i.e. HS threshold selected to achieve a sensitivity (i.e. the true positive fraction) of 0.9. We chose this threshold, by producing a set of thresholds from different criteria (i.e. maximum percentage of presence and absence correctly predicted, maximum kappa, maximum TSS) and selecting the criterion that produced the highest threshold, in order to reduce the rate of commission error in SDMs. Grid cells where a species was recorded were treated as presences regardless of the model predictions. In order to exclude the areas well beyond the known range of a species, we removed all areas that do not currently contain locality records and are isolated from other areas that do contain records by a barrier of unsuitable habitat wider than the mean distance between closest pairs of locality records for the species. This procedure produced one continuous area or few separated areas per species that can be considered as the best approximation of the true species distribution at the given resolution. In the case of Homopus solus, which has only one record in our dataset, we considered the unique occupied cell as the entire distribution at both the resolutions.

Measures of conservation priority

We performed all analyses of conservation priority at both continental and regional scale. We used species distributions to estimate three indicators of conservation priority for each grid cell in the study areas: (i) species richness, (ii) conservation value, and (iii) complementarity with existing reserves. Species richness was derived by simply counting the number of species estimated to occur in each cell. In order to evaluate conservation value and complementarity of the grid cells we used an approach based on the principle of irreplaceability [47]. We used the C-Plan Systematic Conservation Planning System, Version 4 to predict irreplaceability [48]. Irreplaceability of each cell was estimated as the number of possible combinations of cells that include the focal cell and meet a predefined set of specific conservation targets, but which would not meet the targets if the focal cell was removed, divided by the total number of possible combinations that meet the targets (see 3 for details). Values close to 1 indicate sites difficult to replace, often containing species endemic to those sites, while values close to 0 indicate easily replaceable sites, containing only widely distributed species. Following the suggestion by Pressey et al. [47], in order to estimate the ‘landscape’ of conservation value, we calculated the irreplaceability on the entire study areas, considering all the cells as non-protected. In order to estimate the potential contribution of each available cell to the improvement of the established network of protected areas, we calculated the level of complementarity of the cells as the irreplaceability value calculated for unprotected cells by taking into account the existing reserves [47].

Dealing with large scale datasets, a problem arises when reserve boundaries, which are mapped as polygons, should be matched with species distribution maps in regular grids [16]. In these cases, it is necessary to define a threshold percentage of intersection between cells of distribution maps and polygons of reserves boundaries for determining whether a grid cell should be considered protected or not. This choice is crucial, because it potentially generates an over- or under-estimation of the protection level. For solving this issue, we tested different thresholds, using the 2010 WDPA annual release of reserve boundaries [49]. Protected areas with only a point location were mapped as circles with appropriate surface. At both the scales of analysis, we chose the thresholds that selected a number of cells with a total surface equal to the total surface of reserves in the study areas [50]. On the basis of this approach, we considered protected any cell with a proportion of park coverage larger than 36.95%.

In the gap analysis, we determined whether each species met a conservation target set in terms of percentage of its distribution intersecting the current reserve system. We defined the conservation target for each species as the minimum number of cells that should be protected to consider the species sufficiently represented. We established the conservation targets for species on the basis of range sizes, following Rodrigues et al. [17]. We set the conservation target to 100% of the cells protected for the species with the smallest extent of occurrence. On the other hand, we set to 5% the conservation target of the most widespread species. Targets for all species with intermediate range sizes were calculated by interpolating the extreme range size targets using a linear regression on the log-transformed number of occupied cells. We considered those species not represented in any protected area as total gap species and species that met only a portion of their conservation target as partial gap species [17].

Cross-levels comparison

In order to explore the potential cross-scale interaction, we compared the outcomes of continental and regional prioritizations. More specifically, we evaluated if (i) the relative degree of species protection and (ii) the land priority values were congruent at the two resolutions. To do this, we adopted a null models approach, contrasting the observed correlation coefficient r with those simulated by 3 × 10⁴ random Monte Carlo permutations in EcoSim 7.0 [51]. This number of permutations ensures that algorithm biases are avoided [52]. The index (r) was calculated for the original data as well as for the simulated matrices and results were compared, calculating the probability (P) of the null hypothesis that the observed index (r_obs) was drawn at random from the distribution of the simulated indices (r_exp) [53]. This means that the observed correlation in the data does not reflect real patterns, but represents chance variation or sampling effects. Non-random correlations were assumed when P_Fobs≥Fexp ≤ 0.05 [54]. In addition, we defined as priority sites the envelope of the cells with priority score higher than the threshold that selects a number of cells as close as possible to the best 1%, and we measured if the priority sites identified at high resolution were nested across the priority sites identified at low resolution.

Continental analysis

At the continental scale, we obtained high validation scores for all the species (mean AUC ± SD = 0.894 ± 0.041). One half of the species appeared localized to relatively small areas (< 100 cells) and most of them are endemic to Southern Africa (Table 1). On the contrary, we evidenced only three very widespread species (> 500 cells). The protection provided to tortoises by the African network of protected areas was highly variable (Table 1). Similarly, the percentage of target met by each species changed. Three species met their respective conservation targets but, at the same time, two species were not represented at all in the reserve system (Table 1).

The combination of species models produced very clear patterns of species richness, conservation value, and complementarity. We observed the highest values of species richness in Southern Africa (especially in central and southern South Africa) (see Figure S1 for place names), where all but one the 3.6% most specious cells (≥ 5 species; max: 9 species) were restricted (Figure 1a). Secondary peaks of tortoise diversity were also present in East Africa (from Eritrea to northern Tanzania) and along the Gulf of Guinea (from Sierra Leone to Ghana and from Nigeria to Gabon) (Figure 1a). Similarly, all but one the best 6.7% cells in terms of conservation value were located in Southern Africa (especially in central and southern South Africa and southern Namibia) (Figure 1b). At the same time, East Africa reduced its relative importance and the Gulf of Guinea maintained its secondary prominence if conservation value was considered (Figure 1b). When we considered the existence of reserves for calculating the complementarity value, Southern Africa consolidated its great importance, containing the best 6.1% cells (Figure 1c). On the contrary, the Gulf of Guinea and, especially, East Africa further reduced their relative importance (Figure 1c).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Continental and local conservation priorities.

Geographic patterns of: total species richness (a,d), conservation value (i.e. irreplaceability calculated considering all the cells as non-protected; b,e), and complementarity (i.e. irreplaceability calculated considering the presence of the existing reserves; c,f) calculated at low resolution for the entire Africa (a,b,c) and at high resolution for Southern Africa only (d,e,f). Darkness of the pixels is proportional to cell values (see text for ranges).

https://doi.org/10.1371/journal.pone.0077093.g001

Regional analysis

Southern Africa emerged from the continental analysis as the most important region for tortoises. High resolution models for this region obtained high validation scores (mean AUC ± SD = 0.879 ± 0.023). Ten of the 12 species that occur in Southern Africa are endemic to this region and we evidenced that 40% of these species were restricted to a small area (< 2000 cells) (Table 1). At the same time, two species were very widespread across the region (> 10000 cells) (Table 1). At the regional scale, the protection provided by the reserve network was similar to the level of the entire Africa but slightly more constant (Table 1). Similarly, we estimated a great variation among the percentages of target met by each species. Only one species was not represented in any protected area and four other species met their respective conservation targets.

When the priority level was quantified at high resolution across Southern Africa, specific areas emerged as core sites for tortoise conservation. The 4.1% most specious cells (> 5 species; max: 8 species) corresponded to the Cape Fold Mountains, the south-western ranges of the Great Escarpment, and the middle and upper basin of the Great Fish River (Figure 1d) (see Figure S1 for place names). Some other areas along the coasts and the eastern border of South Africa, in south-western Namibia, and in northern Botswana appeared as secondary peaks of species richness. In terms of conservation value, the highest levels (best 0.6% cells) were evidenced in the coastal area south-west of the Cape Fold Mountains and in a single cell in the south-western Namibia (Figure 1e). The Cape Fold Mountains themselves and the south-western ranges of the Great Escarpment maintained an outstanding value (best 4.0% cells), which was also obtained by the south-eastern slopes of the Drakensberg mountain range. On the contrary, the basin of the Great Fish River reduced its importance. Measuring the value of complementarity with respect to the existing reserve network, the coastal area south-west of the Cape Fold Mountains and the cell in south-western Namibia further increased their importance with respect to the other zones (Figure 1f), containing the best 0.4% cells.

Cross-levels comparison

The relative degree of species protection was assessed similarly at the two different resolutions. For the species endemic to Southern Africa, the percentages of protected range measured at low and high resolutions were strongly correlated (arc-sine transformed % protected range_{(LR vs HR)}: r_obs = 0.699, P_{obs ≥ exp} = 0.006). Similarly, considering that species with the percentage of protected range ≥ 100% meet their targets and have, consequently, the percentages of target met = 100%, the percentages of target met calculated at high and low resolutions were correlated (arc-sine transformed % target met_{(LR vs HR)}: r_obs = 0.666, P_{obs ≥ exp} = 0.008). The relative priority values of the cells at high resolution were similar to the relative priority values of the corresponding cells at low resolution. All the three indices of priority (species richness, conservation value, and complementarity) calculated at high resolution were correlated to the values calculated for the corresponding cells at low resolution (r_obs = 0.709, r_obs = 0.336, and r_obs = 0.374 respectively, P_{obs ≥ exp} ≤ 0.001 in every case). The highest priorities identified at the regional level were nested across the priorities defined at the continental scale. Almost all (98.7%) the 1.5% most specious cells at high resolution were contained in the 2.3% most specious cells at low resolution. At the same time, 94.3% of the 0.9% most valuable cells and 90.7% of the 1.9% most complementary cells at high resolution were contained in the best 0.9% and 1.9% cells at low resolution in terms of conservation value and complementarity respectively.