Species–landscape interactions drive divergent population trajectories in four forest‐dependent Afromontane forest songbird species within a biodiversity hotspot in South Africa

Abstract Species confined to naturally fragmented habitats may exhibit intrinsic population complexity which may challenge interpretations of species response to anthropogenic landscape transformation. In South Africa, where native forests are naturally fragmented, forest‐dependent birds have undergone range declines since 1992, most notably among insectivores. These insectivores appear sensitive to the quality of natural matrix habitats, and it is unknown whether transformation of the landscape matrix has disrupted gene flow in these species. We undertook a landscape genetics study of four forest‐dependent insectivorous songbirds across southeast South Africa. Microsatellite data were used to conduct a priori optimization of landscape resistance surfaces (land cover, rivers and dams, and elevation) using cost‐distances along least‐cost pathway (LCP), and resistance distances (IBR). We detected pronounced declines in effective population sizes over the past two centuries for the endemic forest specialist Cossypha dichroa and Batis capensis, alongside recent gene flow disruption in B. capensis, C. dichroa and Pogonocichla stellata. Landscape resistance modelling showed both native forest and dense thicket configuration facilitates gene flow in P. stellata, B. capensis and C. dichroa. Facultative dispersal of P. stellata through dense thicket likely aided resilience against historic landscape transformation, whereas combined forest‐thicket degradation adversely affected the forest generalist B. capensis. By contrast, Phylloscopus ruficapilla appears least reliant upon landscape features to maintain gene flow and was least impacted by anthropogenic landscape transformation. Collectively, gene flow in all four species is improved at lower elevations, along river valleys, and riparian corridors— where native forest and dense thicket better persist. Consistent outperformance of LCP over IBR land‐cover models for P. stellata, B. capensis and C. dichroa demonstrates the benefits of wildlife corridors for South African forest‐dependent bird conservation, to ameliorate the extinction debts from past and present anthropogenic forest exploitation.

Species differ in sensitivity to habitat degradation following fragmentation (Amos et al., 2012;Devictor et al., 2008;Dondina et al., 2017), and landscape configuration change may independently alter local and long-distance dispersal within a species (Freckleton et al., 2005;Richardson et al., 2016). The impacts of these environmental disturbances may take multiple generations to be detected within populations (Epps & Keyghobadi, 2015;Lowe et al., 2015;Samarasin et al., 2017), further complicating assessments of the ecological effects of distinct historic and contemporary anthropogenic activity. Interpretations of species responses to habitat fragmentation can be particularly challenging in populations confined to naturally fragmented habitats, where population complexity may arise naturally (Epps & Keyghobadi, 2015;Fenderson et al., 2020;Richardson et al., 2016). In fragmented landscapes, more vagile species better retain functional connectivity (Amos et al., 2014;Callens et al., 2011;Canales-Delgadillo et al., 2012;DeCamargo et al., 2018;Kalle et al., 2018), as do species that facultatively disperse through otherwise unsuitable intermediary habitats (Keeley et al., 2017). This latter trait is underappreciated in landscape ecology, yet can be critical for understanding structural connectivity between spatially discrete metapopulations of vulnerable species (Driscoll et al., 2013;Kadmon & Allouche, 2007;Kupfer et al., 2006). As the loss of important matrix elements potentially impedes species dispersal, thereby exacerbating the effects of primary habitat fragmentation, the identification and preservation of these elements may prove necessary for longterm species viability, even in cases where the matrix is infrequently utilized. Testing tolerance to both natural and anthropogenic fragmentation is best achieved by comparative research on multiple species which differ in their level of habitat specialization and mobility.
In South Africa, native forests comprise a highly fragmented biome confined to 0.5% of the country's land area (Mucina & Geldenhuys, 2006). This biome is subdivided into Afromontane forests, which are mostly scattered across low-and mid-elevation slopes of inland mountains, and Indian Ocean coastal belt (IOCB) forests, which are discontinuously present along the eastern coast (von Maltitz et al., 2003;Mucina, 2018). In both sub-biomes, forest fragmentation arose naturally through palaeoclimatic shifts (Eeley et al., 1999;, but has been exacerbated by anthropogenic deforestation of over 80% of IOCB forests, and 15% of Afromontane forests during the past two centuries (Berliner, 2009;Olivier et al., 2013). Commercial logging largely ceased by 1940 (Adie et al., 2013;, yet many forest remnants remain degraded, partly due to widespread illegal harvesting of forest products (Leaver & Cherry, 2020a), as well as the reduced structural connectivity of this biome following clearance of small forest patchworks (Kotze & Lawes, 2007), and conversion of the landscape matrix Freeman et al., 2018;Russell & Ward, 2016).
Anthropogenic pressures placed on South African forests have reportedly caused declines in forest-dependent bird species, especially among insectivores (Cooper et al., 2017;Freeman et al., 2018). This group is sensitive not only to forest loss and degradation but also to conversion of the natural vegetation matrix-a trait less apparent in other South African forest-dependent birds (Freeman et al., 2018;Neuschulz et al., 2013;Olivier & van Aarde, 2017). Clearance of coastal thicket, a vegetation type resembling low, recovering IOCB forest, is shown to impede the interforest connectivity of these bird species, as well as forest-dependent mammals, both at the local and regional scale (Ehlers-Smith et al., 2017a, 2017b, 2019). Despite these community-level observations, it is unknown whether elements of the natural landscape matrix facilitate gene flow in forest-dependent insectivorous birds, and the population genetic stability of these species remains unassessed. Accordingly, we conducted a comparative landscape genetic study of four forest-dependent insectivorous songbird species, focussing on the southeast region of South Africa, where ranges declines of forest-dependent birds between 1992 and 2014 have been most substantial (Cooper et al., 2017). Our study aims were to assess contemporary levels of genetic connectivity between forest metapopulations within each species and to infer whether these species facultatively disperse through the regional landscape matrix.
Additionally, we sought to evaluate the historic stability of the effective population sizes within each species. We undertook this study using microsatellite markers and employed an a priori landscape resistance modelling technique developed by Peterman et al. (2014) and Peterman (2018) to conduct our landscape genetics investigation. We hypothesized that (1) connectivity between regional forest fragments would vary between species, depending on known species vagility; (2) each species would exhibit facultative dispersal through well-wooded habitats (thicket) to facilitate gene flow between naturally fragmented forest; and (3) species with greater forest specialization which had experienced greater contemporary range declines would have more rapidly decreasing effective population sizes.

| Field sampling and laboratory procedures
Collected blood samples were preserved in 500 μl 95% ethanol, and genomic DNA was extracted using a Nucleospin Tissue DNA extraction kit (Macherey-Nagel). For each species, we screened a separate microsatellite library available from past literature (9-29 loci per species; 55 loci total) and obtained for each species a unique set of eight informative loci (see AppendixS1 for microsatellite locus screening and amplification conditions). We randomized within-species samples prior to amplification to minimize false-positive discovery from downstream analyses (Meirmans, 2015). Microsatellite alleles were genotyped on an ABI377xlsequencer (CAF, Stellenbosch), against LIZ 500© internal size marker, and scored in GENEIOUS 7.1.4 (©Biomatters), using three positive control individuals per species for each marker to verify scoring accuracy.  (Kawashima et al., 2009). Forest-level deviations from expectations of Hardy-Weinberg equilibrium (HWE) and linkage disequilibrium (LD) within forests were assessed in GENEPOP4.7 (Rousset, 2008); adjusting significance values using a Benjamini-Hochberg correction (Benjamini & Hochberg, 1995) to control for false discovery rate. Forest-level species genetic diversity was estimated as rarefied allelic richness (AR), and private allelic richness (PrAR) in ADZE1.0 (Szpiech et al., 2008); observed (H o ) and expected (H e ) heterozygosity; and inbreeding coefficient (F IS ), in GENETIX4.05 (Belkhir et al., 2001). As a precautionary measure against low sample sizes and the limited number of microsatellite loci employed in this study, we used POWSIM 4.1 (Ryman & Palm, 2006) Table S2), was selected as larger N e are considered more appropriate; the value of t was selected following recommendations by Ryman and Palm (2006) to test for the particular F ST . Simulations were performed assuming two subpopulations (N = 50 and N = 40) over 1000 replicates, and statistical power was measured as the proportion of significant tests.

| Population genetic diversity and structure
Population genetic substructures were investigated through Bayesian clustering using STRUCTURE (Pritchard et al., 2000). The optimal number of genetic clusters per species (K) were tested for K = 1-12 (the number of forest sites +1). Twenty independent runs of 5 × 10 5 Markov chain Monte Carlo (MCMC) iterations and a burn-in period of 5 × 10 4 were performed per K, using the admixture model, correlated allele frequencies, and with LOCPRIOR (grouped by forest). Results of runs averaged in STRUCTURE HARVESTER (Earl & vonHoldt, 2012), and the optimal number of clusters was F I G U R E 1 The distribution of Afromontane (green), Indian Ocean coastal belt (IOCB) (orange) and intermediate scarp (purple) forests across the Eastern Cape and southern KwaZulu-Natal provinces of South Africa, shown alongside sampled forest locations. Coloured asterisks indicate which species of the four study species were sampled within a forest site. Forests without asterisks are represented by all four study species determined using the Evanno ΔK statistic derived from posterior probability of each value of K (Evanno et al., 2005). STRUCTURE results were visualized in the Pophelper R package (Francis, 2017).
To further investigate population structure, we performed principal component analysis (PCA) in the adegenet R package (Jombart, 2008)

| Demographic history
To infer contemporary gene flow disruption, we compared pairwise F ST , estimated in ARLEQUIN 3.5, to the proportions of shared allele statistic D PS , calculated in MSA 4.0 (Dieringer & Schlötterer, 2003). Lag time to detection of new gene flow barriers is shorter when measuring D PS compared to F ST Robin et al., 2015;Savary et al., 2021), and so larger D PS :F ST ratios may suggest recent reductions in gene flow (Robin et al., 2015). We TA B L E 1 Sample sizes, estimates of genetic diversity, and inbreeding coefficients within each forest for the four focal bird species  (Do et al., 2014). We separately assumed random and monogamous mating (typically observed; Hockey et al., 2005). We observed results at 0.02 and 0.01 critical allele frequencies to better accommodate limited data sets (Do et al., 2014) and used a pairwise jackknife approach to assess confidence intervals; within-species samples were pooled to accommodate overlap/interbreeding among the most recent generations. We further inferred variation in focal species N e over the most recent 100 generations using the VarEff R package (Nikolic & Chevalet, 2014). Default parameter conditions were kept across species, adjusting maximum distance between alleles (DMAX = 18 -P. stellata; 17 -C. dichroa; 22 -B. capensis; 10 -P. ruficapilla), number of past N e changes (JMAX = 3); and generations since the most recent common ancestor (GBAR = 1000; reduced from the default GBAR = 5000 given the low population differentiation observed for each species [Nikolic & Chevalet, 2014]). Runs were performed under both single-step mutation model (SMM), and 10% single-step two-phase mutation (TPM) to accommodate a broader range of mutation dynamics within natural populations. Mutation models assumed a mutation rate of μ = 5 × 10 −4 per generation (Brohede et al., 2002;Coetzer et al., 2020), with an acceptance ratio of 0.25.

| Landscape genetics framework
Landscape genetics frameworks provide a means to investigate relationships between genetic distances and features landscape, by modelling resistance surfaces of spatially arranged cost values to gene flow (Manel & Holderegger, 2013;Manel et al., 2003;Waits et al., 2015). To investigate the regional landscape influences on the interforest connectivity within each species, we adopted an a priori approach of resistance surface parameterization using RESISTANCEGA 4.1 R package (Peterman, 2018;Peterman et al., 2014). This approach circumvents subjectivity of conflicting expert opinion (Charney, 2012;Zeller et al., 2012) and limited applicability of niche-model derivations towards atypical landscape use (Balkenhol et al., 2015;Keeley et al., 2017;Vasudev et al., 2015;Zhan et al., 2017). The RESISTANCEGA 4.1 R package integrates mixed-effects modelling and stochastic genetic algorithms mimicking natural selection (Scrucca, 2013) specifically to maximize the relationship between pairwise genetic distances of samples and resistance surfaces. Models were fitted using maximum-likelihood population effects (MLPE) parameterization (Clarke et al., 2002) in the LME4 R package (Bates et al., 2014) where fitness was assessed using corrected Akaike information criteria (AIC c ). Models with an AICc difference (ΔAIC c ) <2 were considered equivalent (Burnham & Anderson, 2004). We modelled two ecological distances for each landscape surface: isolation-by-resistance (IBR) considers cumula- We separately considered pairwise F ST and pairwise D PS (Tables S1.1-S1.4) as the dependent variable for mixed-effects modelling, and scaled and centred LCP and IBR surfaces as predictor variables.
F I G U R E 2 Demographic trends within B. capensis, C. dichroa, P. ruficapilla and P. stellata across a region of southeast South Africa: (a) ratios between F ST and D PS genetic distance metrics within each species among sampled forest sites; (b) regional CN e size of each species measured at 1% and 2% critical allele frequencies, and assuming monogomous mating (with 95% confidence intervals); (c) VarEff plots showing variation in CN e of each species over the past 100 generations, assuming a single-step mutation model at a constant mutation rate of μ = 5 × 10 −4 per generation. Species demographic trends are inferred from a combination of eight species-specific loci unique to each species

| Genetic population structure
Global genetic differentiation was significant for C. dichroa

| Demographic history
Pairwise D PS :F ST ratios were highest for B. capensis, P. stellata and C. dichroa, and lowest for P. ruficapilla (Figure 2a). Overall CN e appears lowest in C. dichroa and highest in P. ruficapilla, although both have larger 95% CI compared to P. ruficapilla and P. stellata ( Figure 2b; Table S2). Species CN e assuming monogamous mating (typically observed in each species [Hockey et al., 2005]) were twice as high compared to assuming random mating (Table S2).
Disparities between CN e at 1% and 2% critical allele frequencies were minimal in B. capensis, 18% in P. stellata, 33% in C. dichroa and 150% for P. ruficapilla, reflecting lower rare allele frequencies in the last three species (Do et al., 2014). Fluctuations in N e over the past 100 generations varied across the four songbirds, consistent across single-step (Figure 2c), and two-phase ( Figure S2) mutation models. Historically, B. capensis and C. dichroa had the largest N e , but declined to levels comparable to P. ruficapilla and P. stellata, which both appear more temporally stable, but still in decline (Figure 2c). Assuming a two-year generation time (Bird et al., 2020), or three years for P. stellata (Oatley, 1982b), these events relate to the past three centuries, with most declines beginning <100 years (~20-60 generations) ago.

F I G U R E 4
Relative performance of least-cost pathway and resistance distance models based on landscape surfaces for the four focal bird species, inferred from F ST . Univariate optimizations were conducted independently on four land-cover thematic surfaces, modelling leastcost paths (left). Univariate optimizations were also conducted separately for best-supported land-cover, rivers and dams, and elevation; and multivariate optimizations integrated the three landscape layers into a composite surface. Both univariate and multivariate optimizations employed three replicates of least-cost (middle), and resistance distance (right) modelling regimes. Positive ΔAICc values denote improved model performance over Euclidean distances  (Figure 4), but ranked below the geographic distance model, and native forest and dense thicket configuration model by<4 AICc units ( Figure S3).
According to both pairwise F ST and D PS , detailed land-cover configurations were inadequate to explain genetic distances observed in the four species (Figure 4; Figure S3).

| Landscape resistance surfaces
For B. capensis, only the resistance distance model from elevation outranked the geographic distance model (>8 AICc units), although least-cost distance models from both land-cover (native forest and dense thicket configuration), and rivers and dams were comparable to the geographic distance model (Figure 4). For C. dichroa, the null model outranked all landscape models (Figure 4). For P. ruficapilla, the least-cost distance model from elevation, and the resistance distance models from elevation, and rivers and dams were comparable to the geographic model ( Figure 4). Only in P. stellata did univariate least-cost and resistance distance models from each univariate landscape surface consistently outranked the geographic distance model (Figure 4), of which the least-cost distance model from land cover (native forest and dense thicket configuration) ranked the highest (>17 AICc). Across all four species, the least-cost and resistance distance models from the composite landscape surface ranked far lower than the geographic distance model (Figure 4).

| Comparative performance of landscape models
Both B. capensis add P. stellata showed significant IBD according to pairwise F ST (Table 2), whereas C. dichroa and P. stellata showed significant IBD according to pairwise D PS (Table 3). Partial Mantel tests of either LCP or IBR models controlling for IBD suggest that for B. capensis, P. ruficapilla and P. stellata, genetic distances (pairwise F ST ) better correlated with certain landscape elements than geographic distance (Table 2). In B. capensis, genetic distances correlated significantly with the IBD-controlled LCP (LPC|IBD) models for native forest and dense thicket, rivers and dames, and landscape elevation (Table 2). In this species, landscape elevation appears especially pertinent to gene flow, with both LCP|IBD and IBR|IBD models for elevation strongly correlating to genetic distances (Table 2). Causal modelling, however, did not support one ecological distance model of the other (Figure 7). In P. ruficapilla, genetic distances correlated strongly with IBR|IBD models with native forest land-cover, rivers and dams, and especially with landscape elevation (Table 2). In this species, causal modelling showed that the IBR model of landscape elevation remained significant even after controlling for LCP ( Figure 7). In P. stellata, genetic distances correlated with all tested ecological distance models (after controlling for IBD), but especially so for the LCP|IBD model for native forest and dense thicker land cover ( Table 2). The association between genetic distances within P. stellata and these land-cover classes remained significant across both pairwise F ST (Table 2), and pairwise D PS (Table 3). In this species, causal modelling further corroborated the LCP model over the IBR model for these land-cover classes, further showing that IBD may better explain genetic distances than this land-cover IBR model ( Figure 5). Although no landscape model pairwise F ST genetic distances in C. dichroa (Table 2), the LCP model for both native forest, and native forest and dense thicket remained significant after controlling for IBD (Table 3).

| DISCUSS ION
The four forest-dependent insectivorous songbirds in this study ade-

| Population genetic structures
The South African endemic C. dichroa displayed the highest population structuring (Figure 5a), and the large genetic variability unique to Kubusi (Figure 6b) affirms the climatic refugial importance of the eastern Amatole forest complex (Dalton et al., 2015;Kushata et al., 2020;Madisha et al., 2018). The higher population complexity of this forest specialist contrasts with that of the forest generalist B. capensis, which is near endemic to South Africa, and more genetically diverse than C. dichroa (Table 1). Higher historic availability of suitable habitat for B. capensis likely afforded larger populations that were more buffered against Paleoclimatic fluctuations, although both native forest (Ivory et al., 2018;Lawes, Eeley, et al., 2007) and Albany thicket (Potts et al., 2013) biomes were susceptible to contractions during periodic aridity. Consistently higher genetic diversity at Manubi across species (Table 1) corroborates the refugial significance of intermediate (scarp) forests within South Africa (Grass et al., 2015;Lawes, Eeley, et al., 2007;Moir et al., 2020;Moir et al., 2021), which are close to the coast, buffering them from palaeoclimatic extremes. The lower diversity at Oribi Gorge may reflect greater proximity to subtropical IOCB forests (Mucina, 2018;, which re-established only ~8 kya (Huntley et al., 2016) and are generally avoided by these four songbird species. The low genetic diversity observed in P. ruficapilla suggests more recent, or perhaps constrained, colonization of South Africa than does P. stellata, and the unexpectedly low regional complexity within P. ruficapilla contrasts with the strong population insularity observed in East Africa (Callens et al., 2011). This alludes to dispersal behaviour TA B L E 2 Partial Mantel tests comparing Spearman's correlations between landscape features and genetic distances (F ST ) for four forest-dependent songbird species across a region of southeast South Africa. Least-cost pathways (LCP) and isolation-by-resistance (IBR) ecological distances modelled for resistance surface controlled for isolation-by-distance (IBD) (shown in the bottom row). Only landscape resistance surfaces which ranked above or equal to the geographic distance model (Figure 4)

| Population viability and effective population size
Our study findings suggest that historic deforestation and commercial selective logging likely had a large negative impact on the viability of these four species (Figure 2c), more so than contemporary informal forest resource harvesting, despite it being largely unregulated (Leaver & Cherry, 2020a, 2020bLeaver et al., 2019). The contemporary effective population sizes of all four species (Figure 2b; Table   S2) are likely underestimates resulting from pooled generations , although true numbers are likely to remain low. In east Africa, post-fragmentation sensitivity is evident for P. ruficapilla and P. stellata (Callens et al., 2011;Githiru & Lens, 2006; TA B L E 3 Partial Mantel tests (controlling for isolation-by-distance) comparing Spearman's correlations between landscape and genetic distances (D PS ) for focal species populations across a region of southeast South Africa. Only least-cost pathways (LCP) are modelled for each of the two land-cover resistance surfaces which ranked either above or equal to the geographic distance model of Figure S3. Bold indicates significantly supported correlations (p < 0.05) (shown in parentheses)  Sirén et al., 2018), yet the populations of both species in the Eastern Cape appear to have been largely resilient to historic forest exploitation. Afrotropical forest-dependent species have been observed to initially display stable effective population sizes following forest fragmentation (Husemann et al., 2015;Lens et al., 2002), but continued forest degradation eventually undermines population viability (Habel et al., 2014;Korfanta et al., 2012;Lens & Van Dongen, 1999).
In South Africa, this is observable in both P. ruficapilla and P. stellata ( Figure 2c).

| Gene flow among regional forests
Metapopulation dynamics of these four songbird species do not appear to wholly contingent upon observed adult mobility. Observed regional adult vagility in B. capensis and P. stellata (Oschadleus & Ranwashe, 2017) did not preclude long-term (F ST ) isolation-bydistance in these species (Table 2). Local seasonal migration observed elsewhere in C. dichroa (Johnson & Maclean, 1994;Oatley et al., 2017;Oatley, 1966Oatley, , 1969 is contested within southeast South Africa (Craig & Hulley, 2019;Wolmarans, 2015). Stronger population differentiation ( Figure 5) and more recent (D PS ) regional isolationby-distance in this species indicates that regional populations of this species are likely sedentary. (Table 3). The apparent panmixia ( Figure 5) and lack of isolation-by-distance (Table 2) observed within P. ruficapilla strongly indicates underestimated regional dispersal ability of this species.
Intuitively, young P. stellata should seek to minimize exposure and attempt cost-efficient navigation of hospitable matrix vegetation, explaining the high performance of the LCP model of forest and coastal/mesic thicket configuration.

| Species-landscape interactions
Landscape resistance modelling suggests that the configuration of both native forest, and dense (coastal/mesic/valley) thicket is important to gene flow in B. capensis, C. dichroa and P. stellata (Figures 4 and 7; Tables 2 and 3). Pogonocichla stellata demonstrates higher gene flow resistance through these thicket habitats compared to B. capensis and C. dichroa (Figure 7), potentially indicating that P. stellata disperses facultatively through select thicket habitats, while B. capensis and C. dichroa are more inclined to inhabit these habitats (especially mesic Albany thicket adjacent to native forests) in southeast South Africa (Johnson, 1997;Oatley, 1997a). Dispersing P. ruficapilla appear not to discriminate landcover beyond forest configuration (Figure 4). This weak land-cover association could be due to a type I error derived from low sample size (Winiarski et al., 2020), but the near panmixia within P. ruficapilla ( Figure 5), and equilibrium between historic and contemporary gene flow (Figure 2b) supports the notion of high dispersal within this species, and tolerance towards anthropogenic landscape transformation.
The significant influence of regional elevation on the population structure of B. capensis, P. ruficapilla and P. stellata (Figure 7;  (Table 2). This landscape genetic association is similarly observed in a forest-associated pipistrelle (Moir et al., 2020).

| Implications for regional Afromontane forest bird conservation
Stronger performance of the LCP models of native forest and dense thicket configuration over respective IBR models for P. stellata (Figure 7e), as well as B. capensis and C. dichroa (Tables 2 and 3) demonstrate the potential utility of conservation corridors in the Eastern Cape and southern KwaZulu-Natal to preserve genetic integrity within regional Afromontane forest birds. Such corridors should promote resilience anthropogenic climate change, as recommended by Colyn et al. (2020). The highest priority forests for conservation are the scarp forests present along the Wild Coast, as well the eastern Amatole Afromontane forests, and Afromontane (eastern mistbelt) forests in southern KwaZulu-Natal. These forests harbour the largest overall and unique genetic diversity (Table 1) and are therefore the most likely to serve as future climatic refugia. Effective creation of conservation corridors could incorporate forest and coastal/ mesic thicket vegetation at lower elevations, particularly where these two land-cover classes coincide with rivers and dams, to ensure the preservation of optimal dispersal pathways beneficial for these four species. The utility of coastal thicket in facilitating movement of forest-dependent taxa is already recognized, and many authorities regard coastal thicket as secondary forest (Ehlers-Smith et al., 2018a, 2019Ehlers-Smith, Ehlers-Smith, Ramesh, et al., 2017;Olivier et al., 2013). Beyond forest and dense thicket configuration, matrix land cover was not shown to impact the gene flow of these four songbird species significantly. Avian connectivity between IOCB forests in KwaZulu-Natal can remain high across anthropogenically transformed areas, but minimally in forest-dependent insectivores (Neuschulz et al., 2013). Matrix transformation, however, can diminish regional South African forest ecological integrity Ehlers-Smith et al., 2019Ehlers-Smith, Si, Ehlers-Smith, Kalle, et al., 2018;Freeman et al., 2018) and undermine the population viability of forest-dependent birds within forests.

| CON CLUS ION
Our results show that reported forest-dependent range contrac- In all four species, landscape resistance modelling suggested that regional gene flow within each of the four species is likely affected by landscape features. Native forest and dense thicket configuration is important to gene flow in P. stellata, B. capensis and C. dichroa, with B. capensis seeming most averse to thicket degradation. Beyond dense thicket, all four species, but particularly P. ruficapilla, do not facultatively disperse through matrix land cover. Finally, we propose that by conserving optimal dispersal routes through the two landcover classes, predominately within low-elevation regions and coinciding with prominent river systems, should effectively ameliorate gene flow disruption and mitigate extinction debts culminating from historic forest exploitation.

CO N FLI C T O F I NTE R E S T
The authors declare no conflicts of interest.

DATA AVA I L A B I L I T Y S TAT E M E N T
Data for this study will be available at the Dryad Digital Repository: to be completed after manuscript is accepted for publication.