Topography, more than land cover, explains genetic diversity in a Neotropical savanna tree frog

Effective conservation policies rely on information about population genetic structure and the connectivity of remnants of suitable habitats. The interaction between natural and anthropogenic discontinuities across landscapes can uncover the relative contributions of different barriers to gene flow, with direct consequences for decision‐making in conservation. We aimed to quantify the relative roles of land cover and topographic variables on the population genetic differentiation and diversity of a stream‐breeding savanna tree frog (Bokermannohyla ibitiguara) across its range.


| INTRODUC TI ON
Effective conservation actions rely on information about population genetic diversity, structure and connectivity (Allendorf & Luikart, 2007;Frankham, 2005;Moritz, 1994). Population genetic data provide estimates of individual migration across species' ranges and the factors that have shaped the distribution of genetic diversity among populations; these data in turn can be used to mitigate potential negative effects that arise from isolation (Brauer, Unmack, Smith, Bernatchez, & Beheregaray, 2018;Habel et al., 2015). The higher the functional connectivity among populations, that is the degree to which individuals disperse throughout the landscape, the lower the chance that limited gene flow results in isolation and population inbreeding (Dixo, Metzger, Morgante, & Zamudio, 2009;Murphy, Dezanni, Pilliod, & Storfer, 2010;Spear & Storfer, 2008). For organisms that occupy heterogeneous landscapes, the biological integrity of populations relies on (a) the anthropogenically modified habitats in the landscape matrix and whether the remaining suitable habitats promote natural levels of dispersal and (b) the quality of the natural habitat itself (Angermeier & Karr, 1996;Coster, Babbitt, Cooper, & Kovach, 2015;Peterman, Rittenhouse, Earl, & Semlitsch, 2013).
Different historical and contemporary landscape features can facilitate or hinder animal dispersal, resulting in diversification of phylogeographic lineages and genetic differentiation even in pristine or non-fragmented habitats (Dixo et al., 2009;Funk et al., 2005;Titus, Bell, Becker, & Zamudio, 2014). Organisms are not homogeneously distributed throughout their geographic range, but rather live in patches of suitable habitat within a matrix (Levins, 1969;Smith & Green, 2005). Aside from geographic distance (Wright, 1943), other factors that can affect divergence in this heterogeneous matrix are historical factors such as topography, including elevation and degree of topographic relief, and contemporary factors such as land cover, including various types of vegetation and microhabitats that are selected by different species (Coster et al., 2015;Garroway, Bowman, Carr, & Wilson, 2008;Ovaskainen & Hanski, 2004;Peterman et al., 2013). For example, if a species' preferred environment is found at higher elevations or mountain ranges, this may yield more divergent lineages and deeper phylogeographic structure (Guarnizo & Cannatella, 2013;Rodríguez et al., 2015;Zamudio, Bell, & Mason, 2016) than species selecting microhabitats that are more homogeneously distributed. The distribution of divergent lineages is secondarily affected by non-natural contemporary changes to the distribution of habitats over relatively short periods of time (Dixo et al., 2009;Laurence, 2010;Machado et al., 2004). If those changes affect the matrix, but still permit successful dispersal of individuals among selected habitat patches, populations will persist with sustained gene flow, resulting in outbred and genetically diverse populations (Galbusera, Githiru, Lens, & Matthysen, 2004). On the other hand, if anthropogenic changes are extreme, gene flow may be compromised, resulting in deleterious effects for isolated populations (Couvet, 2002;Dixo et al., 2009). These complex interactions between natural features, the distribution of preferred habitats and rapid anthropogenic change will likely dictate species persistence or extinction. Thus, landscape genetic research on organisms distributed in heterogeneous landscapes is key to examine the relative roles of historical and contemporary barriers to gene flow, how they are mediated in habitat specialist species, and their consequences for conservation (Chiucchi & Gibbs, 2010).
Landscape genetic studies in temperate ecosystems vastly outnumber those in the Neotropics (reviews in Manel & Holderegger, 2013;Torres-Florez et al., 2018). The Brazilian Cerrado is the second largest South American morphoclimatic domain, with an original area of approximately 2.5 million km 2 of highly heterogeneous landscapes (Silva, Farinas, Felfili, & Klink, 2006). It is the most species-rich savanna in the world and one of 35 worldwide biodiversity hotspots (Mittermeier et al., 2004). Before intense human modification over the past six decades (Myers, Mittermeier, Mittermeier, Fonseca, & Kent, 2000), the Cerrado included seasonal savannas, grasslands and gallery/riparian forests (Felfili, 1995;Meave, Kellman, Macdougall, & Rosles, 1991). The riparian forests occur alongside streams and provide suitable habitat for forest-dependent species (Johnson, Saraiva, & Coelho, 1999;Nali & Prado, 2012). In addition, the Cerrado is topographically diverse, including montane regions dominated by high-elevation plateaus (the chapadas or chapadões) separated by a network of low elevation savannas (Ab'Saber, 1983;Cole, 1986). Understanding habitat connectivity in the Cerrado is imperative as almost half of its land area has been transformed into pasture or croplands (Weinzettel, Vačkář, & Medková, 2018) and only 2.2% of the original Cerrado is currently legally protected (Klink & Machado, 2005). The suitability of current conservation policies can be determined only by assessing the contemporary status of genetic differentiation across animal populations (Bergsten et al., 2012), providing information necessary to assign new protected areas to avoid genetic erosion.
The Neotropical region harbours the largest diversity of frogs, and ca. 1,100 species occur in Brazil (Frost, 2020;Segalla et al., 2019).
Here, we used the tree frog B. ibitiguara to study the roles of different landscape features on amphibian differentiation with a conservation focus. Specifically, we investigated whether genetic diversity could be attributed to contemporary (e.g., land use, forest cover) or historical factors (topography). We tested the hypotheses that (a) individuals will be genetically differentiated among populations, because they occur in forested areas within a matrix dominated by open vegetation, characteristic of the Brazilian savanna; (b) gene flow, when present, will be higher in areas of topographic homogeneity, which facilitate dispersal; and (c) genetic diversity will be higher, and differentiation lower, in populations within the SCNP due to the less anthropogenic disturbance of land cover. Combined, our results will elucidate the processes leading to population differentiation in terrestrial vertebrates, test hypotheses about the interaction between topography and land cover in shaping genetic divergences, and how these factors can contribute to effective conservation measures.

| Study site
The Serra da Canastra mountain range is located in the state of Minas Gerais, south-eastern Brazil, and has a seasonal climate with hot and rainy summers (October to March) and dry winters (April to September; Queirolo & Motta-Junior, 2007). Historically, the region was characterized by Cerrado vegetation, with patches of semi-deciduous forest, gallery forests alongside streams and grassland fields at higher elevations (Dietz, 1984). Much of the original vegetation, however, has been modified since the nineteenth century, primarily for conversion into agricultural and pasture lands (May et al., 2009;MMA/IBAMA, 2005;Myers et al., 2000). The SCNP is a strictly protected area es-   Table S1.1 from Appendix S1). We collected 273 tissue samples from adults (liver, leg muscle or toe clip) and tadpoles (tail clip), and preserved them in absolute ethanol. We euthanized individuals with 10% lidocaine applied to the ventral region; specimens were then fixed in formalin 10% and preserved in ethanol 70% (McDiarmid, 1994). Tadpoles of B. ibitiguara do not form schools (Cardoso, 1983;R. C. Nali, pers. obs.), and thus, we randomly collected tadpoles of different body sizes throughout the streams, reducing the probability of collecting tadpoles from a single clutch. Voucher specimens and tissues were depos-

| Laboratory protocols and microsatellite data
All individuals were first genotyped at 21 microsatellite loci previously developed and optimized for the focal species (Nali, Zamudio, & Prado, 2014). We extracted whole genomic DNA from tissue samples with DNeasy extraction columns (Qiagen), following manufacturer's protocols. PCR profiles followed those in . Chains (MCMC) with 10,000 dememorization steps followed by 100 batches of 5,000 iterations each (Guo & Thompson, 1992). We also tested for linkage disequilibrium across individuals and all pairs of loci using GENEPOP.

| Population genetic analyses
We calculated pairwise relatedness among individuals in GENALEX v. 6.5 (Peakall & Smouse, 2012) using the r qg coefficient (Queller & Goodnight, 1989). We permuted population genotypes 9,999 times and derived upper and lower 95% confidence intervals to obtain a range of r qg expected if mating was random across all populations. We bootstrapped confidence intervals 9,999 times to obtain F I G U R E 1 STRUCTURE plot (above) and assignments of four genetic clusters found for the 12 sampling localities of Bokermannohyla ibitiguara based on 17 microsatellites. Spatial connectivity (a) based on topographic complexity calculated from a digital elevation model depicted in panel (b). Land cover map (c) is also shown (legend colours: dark green = natural forest formation; medium green = natural savanna formation; light green = forest plantation; yellow = grasslands and pastures; orange = agriculture; red = non-vegetated area; blue = large water bodies). The outlines represent the Serra da Canastra National Park boundaries, and the star on the inset map is the general locality in south-eastern Brazil. Locality acronyms, complete land cover classifications and the spatial connectivity figure based on land cover can be found in Appendix S1 [Colour figure can be viewed at wileyonlinelibrary.com] estimates of mean relatedness within populations. To confirm that our analyses of local population structure were not biased due to tadpole sampling (i.e., increased chance of sampling siblings from a single clutch; Table S1.1 from Appendix S1), we ran a Pearson correlation analysis between the r qg value and the percentage of tadpoles sampled per stream.
We used FSTAT 1.2 (Goudet, 1995) to calculate overall Fstatistics (Wright, 1969): F IT (inbreeding coefficient of an individual relative to the total population), F IS (inbreeding coefficient of an individual relative to the sampling locality) and F ST (effect of sampling locality compared with the total population). We also calculated in GENALEX and used as a measure of genetic differentiation between sampling localities (Rousset, 1997); statistical significance of this measure was assessed by using 9,999 permutations. Following Narum (2006), we reported statistical significance for 66 comparisons using Bonferroni's correction (p = .00076), and also the B-Y method (p = .01037), which is less conservative and more appropriate in conservation genetics (Benjamini & Yekutieli, 2001).
To analyse genetic discontinuities without considering subpopulations a priori, we ran a Bayesian analysis using STRUCTURE v. 2.3.4 to infer the number of genetic clusters (K) and genetic discontinuities among sampled populations. We assessed K values from 1 to 10 using 25 MCMC runs with 200,000 steps discarded as burn-in followed by 1 million steps for each proposed K, assuming an admixture model, and correlated allele frequencies. We then compiled all results from K runs in a single STRUCTURE HARVESTER analysis (Earl & vonHoldt, 2012) and inferred presence of genetic structure when the plotting of delta-K provided evidence of a clear peak. The cluster data for the selected K were permuted using CLUMPP (Jakobsson & Rosenberg, 2007).

| Landscape features and genetic differentiation
To determine the roles of geographic distance, topographic complexity and land cover on genetic differentiation across our 12 sampling localities, we calculated three connectivity indices: Euclidean distance (minimum straight-line distance between each pair of populations) and two surface resistance indices: topographic resistance and land cover resistance. For topography, we used the digital elevation model from the Shuttle Radar Topography Mission (SRTM; pixel resolution = 30 m) to generate a topographic complexity raster using the raster calculator feature in ArcGIS 9.3.1 (ESRI, 2009), where each elevation pixel was assigned the variance of the eight cardinal/intercardinal neighbouring pixels. This metric depicting terrain ruggedness is a proxy for microclimatic turnover and local habitat heterogeneity (Huaxing, 2008;Riley, DeGloria, & Elliot, 1999). For land cover resistance, we used the MapBiomas raster database for the Brazilian Cerrado (MAPBIOMAS, 2020; pixel resolution = 30 m). We chose the map from 2014, as this was the final year of our sampling. We then assigned resistance values for the seven land cover classes according to the current literature on habitat requirements for amphibians. Specifically, we followed the study by Titus et al. (2014), based on forest-dependent salamanders with aquatic breeding, which provided comparable measures with the classes from MapBiomas (see Table S1.3 from Appendix S1 for descriptions and classifications).
We used these two rasters as resistance layers in CIRCUITSCAPE v. 3.3 (McRae & Shah, 2009), in which pixels with high topographic complexity or high land cover resistance imposed higher dispersal costs throughout the landscape. We employed a cell connection scheme connecting each pixel (node) to the eight cardinal/intercardinal neighbouring pixels. Surface resistance weights all possible paths between pairs of our 12 sampling localities and produces a summary connectivity raster. As CIRCUITSCAPE will account for extreme differences among neighbouring pixels, geographic barriers will be automatically incorporated in the analysis. Finally, we used simple Mantel tests (Mantel, 1967)

| Landscape features and within-population genetic diversity
We also analysed the impact of landscape variables on the following We then used a general linear model approach (GLM-standard least squares) running all possible models including explanatory landscape variables (topographic complexity, elevation, per cent of habitat loss and per cent of gallery forest), including one-level interactions, and genetic parameters as response variables in turn. We ranked models based on Akaike information criterion (AICc) and report the most parsimonious significant model for each run.

| Comparisons among populations within versus outside SCNP
Even considering the likely role of geographic distance, we expected

| Microsatellite data
Our initial 21 markers were highly polymorphic (11-58 alleles; average 20.8 ± 2.4). All loci showed evidence of null alleles in MICROCHECKER when considering all individuals together. When divided by population, loci Bi1, Bi1122, Bi1521 and Bi3629 showed evidence of null alleles for more than seven populations, but not always in the same populations. Therefore, we used only the remaining 17 loci for all subsequent analysis. They remained highly polymorphic (11-58 alleles; average 20.8 ± 11.15) and yielded a final database with only 0.88% of missing data. Genotypes were deposited in the Dryad Digital Repository: https://doi.org/10.5061/dryad.c2fqz 615x (Nali, Becker, Zamudio, & Prado, 2020). We had a genotyping error rate of average 3.3%, ranging from 0% to 7.3% across all loci.

| Population genetic analyses
Individuals from every sampling locality showed higher relatedness than expected under panmixia (all p values < .05; Figure 2), indicating that reproduction is limited across the sampled populations. If sampling tadpoles increased the proportion of related kin sampled at each locality, we would expect higher relatedness values for populations with higher number of sampled tadpoles. However, we detected no correlation between relatedness values and per cent of tadpole sampling per stream (Pearson's R 2 = .11; t = 1.12, p = .29).
Thus, our differences are likely due to the genetic structure and isolation of our populations rather than an artefact of sampling tadpoles.
The 12 populations of B. ibitiguara in this study had an average  (Table 1).
In our STRUCTURE analysis, delta-K showed a clear peak for four genetic clusters, with a high average coefficient of membership (percentage of individual assignment to the cluster) of 90.3% (Figure 1).
Most populations within SCNP belonged to a single cluster, while the other populations belonged to the other three clusters; the only exception was CAL, which clustered together with other SCNP populations despite being located outside the park.

| Landscape features and genetic differentiation
Pairwise

| Landscape features and within-population genetic diversity
Landscape variables significantly explained four of the within-population genetic parameters (AR, PAR, H E and ENA), with a marginally non-significant effect in H O ( Table 2). The most parsimonious models consistently included topographic complexity and elevation as explanatory variables for AR, H E, H O and ENA (Tables 2 and 3). Habitat loss and per cent gallery forests were not recovered as explanatory variables in any of the most parsimonious and significant models (le ).
The results of all models are included in the Dryad Digital Repository.

| Comparisons among populations within versus outside SCNP
We found significant differences among pairwise F ST for populations within SCNP, within versus outside SCNP and outside SCNP  (Table 4).
Our path analysis confirmed that areas of high elevation harbour anuran populations with higher AR (Figure 3a). This coincides with our SCNP populations, which are located in a high-elevation plateau with more homogeneous topography (Figure 1). Lower elevation sites, on the other hand, show high topographic complexity, and this complexity predicted lower AR in our non-protected populations. Elevation was negatively associated with habitat loss, but the amount of natural vegetation cover did not significantly predict AR (Figure 3a). We observed higher ENA in populations at higher elevations, with no direct effect from habitat loss or topographic complexity ( Figure 3b). The effects of landscape variables on H O and H E were non-significant ( Figure S1.3 from Appendix S1).

| D ISCUSS I ON
We characterized the genetic structure among populations of a Brazilian Cerrado tree frog and asked how different landscape variables (both historical and contemporary) might contribute to the observed pattern of genetic distribution in this highly threatened savanna. Our focal species provided insight into direct and indirect effects of habitat loss, distribution of suitable habitats, topography and geographic distance on spatial connectivity and genetic diversity of populations. We showed that frog populations distributed in patches of gallery forests embedded in a matrix of open environment are overall genetically differentiated, but the degree of differentiation and genetic diversity were associated with topography (historical factor), but not with land cover (contemporary factors).
As predicted, populations were indeed less differentiated inside the SCNP, but contrary to our expectations, this pattern was not a direct result of land cover, but because of lower topographic complexity and lower geographic distances within the park, which likely facilitate dispersal.

| Genetic differentiation among populations
Our first hypothesis was that populations would show significant ge- Hardy-Weinberg equilibrium we found in many populations (Austin, Gorman, & Bishop, 2011;Nei, 1977). Amphibians show limited mobility, physiological constraints that confine adults to moist environments, and extreme site fidelity, so dispersal and gene flow are likely reduced in this group (Blaustein et al., 1994;Cushman, 2006;Peterman, Connette, Semlitsch, & Eggert, 2014). The Brazilian Cerrado is predominantly an open environment dominated by a mosaic of grasslands and shrubs, where streams with gallery forests are sparsely distributed (Meave et al., 1991;Silva et al., 2006;Figure 1c). In this landscape, forest-dependent animals such as B. ibitiguara may use those gallery forests for reproduction and refugia, instead of dispersing throughout grasslands (Johnson et al., 1999;Redford & Fonseca, 1986). In fact, phylogeographic structure of frog lineages is more pronounced among forest-inhabiting species when compared to species that live and breed in open habitats (Rodríguez et al., 2015).
Specific reproductive characteristics and timing of reproduction can contribute to genetic differentiation in amphibians (Funk, Cannatella, & Ryan, 2009;Lourenço, Gonçalves, Carvalho, Wang, & Velo-Antón, 2019). In B. ibitiguara, tadpoles and juveniles potentially develop in the same streams and gallery forests that adults use as breeding sites. Thus, genetic differentiation among streams and the high relatedness values within populations may be caused by a higher philopatry in this species when compared to species that require migration from foraging areas to breeding sites, such as pond-breeding amphibians (Coster et al., 2015;Gamble, McGarigal, & Compton, 2007;Semlitsch, 2008). It is possible that tadpoles disperse downstream (Eterovick, Yazbeck, Dergam, & Kalapothakis, 2009;Lawson, 2013), but streams inhabited by B. ibitiguara are typically narrow, shallow and frequently partially obstructed (e.g., by fallen trees), which could hamper tadpole dispersal far enough to promote genetic differentiation among patches at a landscape scale. Although detailed quantification of amphibian larval dispersal is sorely needed (Lourenço, Antunes, Wang, & Velo-Antón, 2018;Wahbe & Bunnell, 2001), evidence suggests that migration/ dispersal events of amphibians are primarily made by adults and/ or juveniles instead of larvae (Cushman, 2006;Semlitsch, 2008). A similar pattern is also observed in stream fishes, which tend to show restricted movement during most of their lives, causing increased genetic differentiation (Comte & Olden, 2018;Rodríguez, 2002).
Another factor is that B. ibitiguara reproduces for over six months during the rainy season (Nali & Prado, 2012). Females spend energy to produce eggs to ensure maximum reproductive output, and likely deposit more than one clutch during a season (Nali & Prado, 2012;Wells, 2007;R. C. Nali, pers. obs.). Males, on the other hand, spend energy calling and defending oviposition territories to obtain females, which are the limiting resource, and engage in male-male vocal and physical duels (Nali & Prado, 2012, 2014. It is unlikely that the territorial males or mature females of this species would undertake major dispersal events while breeding. In the remaining non-reproductive months, the environment is much dryer and desiccation is likely a further deterrent to movement (Smith & Green, 2005;Titon & Gomes, 2015). As a result, genetic differentiation will accumulate with TA B L E 2 Estimates of the most parsimonious model explaining each intra-population genetic index of Bokermannohyla ibitiguara from 12 localities within the state of Minas Gerais, Brazil Note: The most parsimonious models are in Table 3, and all models have been deposited in the Dryad Digital Repository. Beta = estimate from the regression analysis; R 2 = coefficient of explanation; t = size of the difference relative to the variation; p = probability value. * Whole model: F = 17.68; p < .001; R 2 = .80.

TA B L E 3
Generalized linear regression models of four landscape variables versus seven genetic measures for 12 populations of Bokermannohyla ibitiguara in south-eastern Brazil  One of the main drivers of genetic differentiation for every species is geographic distance, and our results corroborated distance as a relevant factor. However, other landscape features also act as potential barriers, disrupting isolation by distance or resulting in a non-stationary pattern of IBD (Duforet-Frebourg & Blum, 2014;Marschalek & Berres, 2014;Murphy et al., 2010;Wright, Bishop, Matthee, & Heyden, 2015). Given that steeper terrain is potentially costly for dispersal, our second hypothesis was that gene flow would be facilitated by topographic homogeneity. Indeed, topographic complexity predicted limited gene flow more than any other variable. Conservation genetic studies have focused much more on habitat quality and distribution, rather than topography (e.g., Dixo et al., 2009;Miller, Bianchi, Mullins, & Haig, 2013;Telles et al., 2007;Titus et al., 2014). However, studies of different taxa have shown that topographic complexity is an important restrictor of gene flow among populations (e.g., Guarnizo et al., 2016;Pérez-Espona et al., 2008). And in fact, this historical factor has had a larger effect on genetic structure than any contemporary factors in B. ibitiguara.

| Landscape features and facilitated gene flow
When analysing contemporary barriers to gene flow, landscape genetic studies must take into account the spatial scale of sampling relative to habitat suitability and disturbance (Epps & Keyghobadi, 2015;Manel, Schwartz, Luikart, & Taberlet, 2003).
Although we did not detect contemporary barriers to gene flow, studies with other amphibians that sampled at scales comparable to ours found genetic signatures of land cover disruption (Homola, Loftin, & Kinnison, 2019;Zancolli, Rödel, Steffan-Dewenter, & Storfer, 2014). Specifically, our sampled populations, which cover the entire species' known range, showed enough variation in the seven land cover classes between population pairs (Figure 1c). Combined with the fact that genetic equilibrium is expected to occur rapidly with molecular markers with high mutation rates (e.g., microsatellites; Epps & Keyghobadi, 2015), this landscape allowed us to ask how land cover changes ( Figure 1c) interact with historical factors such as topography and distance. At first look, some of our results ( Figure 4, Table 4) could indicate a role of the protected status of populations within the SCNP in promoting less genetic differentiation and higher diversity, which supported our third hypothesis.
However, those were related to topography, and not land cover variables. The SCNP is topographically homogeneous at higher elevations (>1,200 m) compared with the localities outside the park (<1,100 m; Figure 1a,b). Moreover, land cover was not directly associated with allelic measures in our path analyses, but topography was ( Figure 3). A similar effect was found in another montane frog, for which differences in elevation explained heterozygosity and allelic richness (Funk et al., 2005). Our results show unequivocallyand for the first time for Neotropical frogs-that topography drives not only connectivity among habitats, but also the maintenance of genetic diversity within-habitat. Thus, topographic differences can play a role in limiting dispersal events even in the presence of pristine vegetation.
While topography was the most important variable in shaping genetic structure of B. ibitiguara, we did find inconsistencies in some populations. First, populations R1 and R2 + R3 separated into two genetic clusters, despite not being located within an area of particularly rough terrain ( Figure 1). Second, population CAL belonged to the same cluster as the populations within SCNP, even though there is no obvious topographic connectivity with populations within the SCNP ( Figure 1). Populations R1, R2 and R3 are found on the outer edge of the range, which normally have lower population sizes than core populations, decreased individual genetic diversity and increased isolation from other populations (Anderson & Danielson, 1997 Peterman et al., 2013). Conversely, CAL is much closer to the core of the distribution, where higher genetic diversity is expected due to higher gene flow relative to peripheral populations and/or retained ancestral polymorphisms (Frankham, 1996;Zancolli et al., 2014). In addition, non-sampled populations may exist in streams between Chapadão da Babilônia and Chapadão da Canastra (SCNP), and wetlands/swamps could serve as stepping stones for dispersal using less complex pathways relative to topography (Coster et al., 2015).
Tropical species are underrepresented in the study of landscape genetics when compared to temperate ones (Manel & Holderegger, 2013), and topography is often not explicitly considered in the few Brazilian frogs studied (e.g., Eterovick et al., 2016;Prado, Haddad, & Zamudio, 2012;Telles et al., 2007

| Conservation implications
Amphibians are one of the most endangered vertebrate groups (Wake & Vredenburg, 2008) (Dixo et al., 2009;Eterovick et al., 2016;Torres-Florez et al., 2018). By adding topography to the scenario, we bring novel information for amphibian conservation, especially for the highly threatened Brazilian Cerrado at both local and broad scales.
Brazilian law mandates that landowners maintain a percentage of undisturbed areas on their properties (Brasil, 2000(Brasil, , 2012 Brazil, is one of the top conservation priority regions in the Cerrado (Werneck, 2011). It is the only true Brazilian cordillera (~1,000 km in extension) with strictly protected units located in areas with topographically complex regions in between. A recent study of Bokermannohyla saxicola in the Espinhaço range showed that human occupancy reduces heterozygosity of populations (Eterovick et al., 2016), but topographic complexity could also be an important underlying mechanism.
Our results have practical implications for decision-making in conservation biology and indicate that safeguarding topographically homogeneous lands may prevent further genetic erosion of the remaining amphibian populations across their range. This is particularly critical for the Cerrado, where agriculture and cattle ranching expansion have accelerated throughout less complex terrains in the last three decades (Hunke, Müller, Schröder, & Zeilhofer, 2015). We recommend that future studies across taxa account for topography, in addition to land cover variables. This will help establish well-informed criteria for the assessment of conservation units that mitigate negative biodiversity impacts in already threatened ecosystems.

ACK N OWLED G EM ENTS
R.C.N. is thankful to colleagues in the Zamudio Lab for help with laboratory work, colleagues that helped with fieldwork, and Rogério C.
de Paula for the help with map resources. We thank two anonymous

PEER R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1111/ddi.13154.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are deposited as • Tissue samples, specimen vouchers and details on collection localities: published as Tables S1.1 and S1.2 from Appendix S1.