Ethics Statement

The majority of field sampling was conducted non-invasively from tiger scat, without animal capture and handling. Permits for collection of tiger scat samples were obtained from the National Tiger Conservation Authority and the State Forest Departments. Capture and radio collaring of tigers required the approval of the Ministry of Environment and Forests, Government of India and the Chief Wildlife Warden, Madhya Pradesh State, under the Wildlife (Protection) Act 1972. The permissions define the conditions required for capture of tigers, which include an approved protocol and participation by a Park Official and supervision by a qualified veterinarian in the capture and collaring exercise. Both these permissions were obtained and strictly adhered to. Capture operations were conducted by trained veterinarians and wildlife biologists as per the protocols of the Wildlife Institute of India and the National Tiger Conservation Authority. A tiger tissue sample was obtained from Satpura Tiger Reserve where the tiger died due to natural causes (was killed by another tiger in a territorial strife). This research project was conceived and radio collaring reported in this paper was done prior to the formation of an animal ethics committee at the Wildlife Institute of India.

Study Area

The present study was carried out in the global priority tiger conservation landscape of Central India within the states of Madhya Pradesh, Maharashtra, and Chhattisgarh (Figure 1). A forested area of 76,913 km² (20.1–23.5°N and 76.5–81.5°E) covering the seven tiger reserves of Melghat, Satpura, Pench, Kanha, Tadoba, Achanakmar, and Bandhavgarh along with their buffer zones, corridor habitats and adjoining forested habitats were sampled. Sampled sites covered different types of tiger habitats found in Central India ranging from the tropical moist-deciduous Sal (Shorea robusta) forests in Kanha and Bandhavgarh to tropical dry-deciduous teak (Tectona grandis) dominated forests in Pench, Tadoba and Melghat. The topography varied from about 200 meters above sea level (m a.s.l.) in the low-lying hills to the dadar plateaus and meadows in Kanha (500 m a.s.l.) and the rugged Satpura ranges (highest elevation 1,352 m a.s.l.). The rainfall, primarily restricted to the monsoon season (late June to September end) ranged between 1,000 to 2,200 mm per year. The large mammal fauna found in the region included tigers (Panthera tigris), leopards (P. pardus), sloth bears (Melursus ursinus), dholes (Cuon alpinus), gaurs (Bos gaurus), wild pigs (Sus scrofa), sambar deer (Rusa unicolor), chital deer (Axis axis), barking deer (Muntiacus muntjak), swamp deer (Rucervus duvaucelii) and nilgai antelopes (Boselaphus tragocamelus).

Download:

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

Figure 1. Study area map showing sampled sites, genotype locations and habitat connectivity.

The study area of Central India spanning the states of Madhya Pradesh, Maharashtra and Chhattisgarh, showing tiger habitat (forest cover) coded with tiger occupancy probability, protected areas, human habitation (night lights), major roads and least-cost habitat corridors connecting tiger reserves. Individually genotyped tigers (n = 165) are shown as color coded dots at their sampled locations with their colors matching their genetically assigned population.

https://doi.org/10.1371/journal.pone.0111207.g001

Genetic Sampling and Laboratory Work

Blood from 16 radio-collared tigers, one tissue sample from a dead individual and 587 putative tiger scat (faeces) samples were collected between 2006 and 2011 from the seven Tiger Reserves, and at a few intervening forest corridors in the area (Figure 1 and Table 1). The number of samples obtained was largely proportional to the population size of tigers in that region. However, due to logistical constraints the Tadoba population was under sampled. Scats were stored, either dry with silica or in 75% ethanol, and kept at ambient temperature, prior to laboratory analysis. For each scat, a Global Positioning System (GPS) reading was taken and transferred into a GIS. Scat DNA extractions were performed in a room dedicated to low-copy DNA extraction, using the guanidine isothiocyanate - silica extraction protocol [73]. For every extraction, negative controls composed of reagent only without the scat sample were included to monitor contamination. Extractions from blood and tissue samples were carried out using the DNeasy blood and tissue kit (QIAGEN, Germany).

Download:

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

Table 1. Area-wise estimates showing population extents, occupied habitats, sampling effort and number of tiger individuals identified.

https://doi.org/10.1371/journal.pone.0111207.t001

Tiger scats were identified from among the field collected scats through PCR and BamHI restriction enzyme digestion of the mitochondrial DNA cytochrome b gene (Figure S1; see Methods S1 and Tables S10 and S11 for details on protocols and reference species used for PCR). Unambiguously assigned tiger scats and blood and tissue samples were individually identified using a panel of eleven microsatellite loci, derived from domestic cat [74] and tiger [75], [76]. The loci consisted of three dinucleotide (Fca304, Fca954, 6Hdz700), three trinucleotide (Pati01, Pati09, Pati15) and five tetranucleotide repeat markers (Fca441, F85, F53, F124,Pati18), variously labeled at the 5′ end of each forward primer with 6–FAM, PET, VIC and NED dyes. Gender identification in individual tigers was carried out by amplifying the Y chromosome linked SRY gene and an X chromosome microsatellite locus Fca651. PCR amplifications were carried out in 10 µl reactions with a multiple panel of 3 to 4 loci using the Multiplex PCR kit (QIAGEN) according to the manufacturer's instructions. Amplified products were resolved on the ABI 3130 Genetic Analyzer and GENEMAPPER 3.7 (Applied Biosystems, USA) was used to score allele sizes. To limit genotyping error due to allelic dropout (nonamplification of one allele in a heterozygote), multiple PCR replicates were conducted as in Navidi et al. [77]. Heterozygotes were confirmed with at least two independent replicates and homozygotes with five replicates. The genotype data were checked on MICROCHECKER [78] for identifying and correcting genotyping errors such as those that arose from stuttering patterns, null alleles and small-allele dominance.

Identification of Individuals and Descriptive Statistics

In order to investigate the power of the eleven microsatellite loci to distinguish among closely related individuals in the same population, the conservative sibling probability of identity (PI sib) statistic [79], [80] was computed in GIMLET 1.1 [81]. Unique multilocus genotypes were identified using the Identity analysis option in CERVUS 3.0 [82], [83]. Samples that showed mismatches at up to two loci were re-examined for possible genotyping errors and allelic drop-out, and again amplified thrice in order to confirm the multilocus genotypes before assigning them as unique individuals. Multiple replicates of the same individual from the same locality were discarded and only unique multilocus genotypes were used for all further analyses. We used CERVUS 3.0, GENEPOP 4.1 [84], [85], GENALEX 6.3 [86] and FSTAT 2.9.3 [87] to calculate the following descriptive statistics: (i) number of alleles per locus, (ii) observed and expected heterozygosity, (iii) tests for deviation from Hardy-Weinberg Equilibrium (HWE), (iv) significance values for linkage disequilibrium (LD) among all pairs of loci, and (v) estimates of population pair-wise F_ST [88] and R_ST [89] values. Loci with null alleles were flagged by MICROCHECKER, and we tested for deviation from HWE using both null allele adjusted and unadjusted (observed allele frequencies) genotypes. We also tested for associations between genetic variation and estimated population size through Spearman's rank correlation using the pspearman package [90] in R (http://cran.r-project.org).

Population Genetic Structure

We used two types of individual-based analyses to assess genetic differences among individuals and assignment patterns of tigers to populations. First, a Principal Coordinates analysis (PCoA) based on pair-wise Phi_PT genetic distances [91] was carried out in GENALEX and the scatter of population-wise individual assignments was plotted on the first three PCo axes using NCSS (www.ncss.com) to understand broad spatial patterns of populations structure in the landscape. Next, we used the Bayesian individual clustering approach in STRUCTURE 2.3.3 [48] to detect population structure among sampled localities in the area by assigning sampled individuals into a number of clusters (K) based on the multilocus genotype data alone. The clustering process ensures that for identified population clusters, deviations from Hardy Weinberg and linkage equilibrium are minimized. We analyzed our data in STRUCTURE by using the admixed model and correlated allele frequencies option to carry out thirty independent simulations at each K (K = 1 to 10), with a burn-in length of one million Monte Carlo Markov Chain (MCMC) steps and data collection phase of five million MCMC iterations. These run times were sufficient to ensure convergence of the Markov chains, and all runs were carried out both with (locprior  = 1) and without (locprior  = 0) using prior sampling locality information. The true K or most likely number of population clusters in the dataset was inferred from (i) the ad hoc parameter of log-likelihood change in probability of individual assignments to K clusters (Ln P(K), [48]), and (ii) the second order rate of change in the likelihood of K values (delta K, [92]). Both these values were computed from the STRUCTURE output using the program STRUCTURE HARVESTER v0.6.91 [93]. We also carried out an Analysis of Molecular Variance (AMOVA, [91]) in GENALEX to compare the population clusters identified by the STRUCTURE analysis. Genetic variances were partitioned at two levels, viz. among all the STRUCTURE identified population groups (K clusters) and among subpopulations within each group.

Detection of Migrants

We used three Bayesian approaches implemented in STRUCTURE 2.3.3 [48], GENECLASS 2.0 [49] and BAYESASS 1.3 [68], for identification of migrant and admixed individuals. STRUCTURE was used to calculate the posterior probability of whether individuals are residents of their sampled population or migrants from other areas by incorporating the previously identified population cluster information with a priori designation of the migration rate (MIGPRIOR  = 0.05). All other parameters and run times followed previous population clustering runs (as detailed above in the preceding methods section on analyzing population genetic structure). We detected no biases in a priori assignment of the migration rate as the selection of particular MIGPRIOR values (0.001 to 0.1) did not substantially influence the STRUCTURE output, therefore only results for MIGPRIOR  = 0.05 are presented here.

Next, we used the likelihood-based estimator L_h/L_max in combination with the resampling algorithm of Paetkau et al. [94], implemented in the ‘detect migrants' function in GENECLASS to exclusively identify first generation migrants, i.e. individuals assigned to a different population other than the sampled population. The Paetkau et al. [94] routine was selected on basis of its superior simulation scheme which closely mimics natural processes and results in accurate type I error rates. L_h/L_max, is the ratio of L_h, the likelihood of a given individual being assigned to its sampled population to L_max, the greatest likelihood among all sampled populations [94]. We employed the Bayesian criterion of Rannala and Mountain [95] in combination with the re-sampling algorithm of Paetkau et al. [94] using a simulated set of 10,000 area-specific genotypes, to determine migrant thresholds (type I error α levels of 0.01 and 0.05) of L_h/L_max. We considered a minimum log likelihood L_h/L_max ratio of 2.0, which corresponds to a 100 times probability of being cross-assigned, as the threshold level for determining putative migrant status of an individual [7].

Third, the non-equilibrium Bayesian assignment test of BAYESASS was used to trace each individual's immigrant ancestry within the last two to three generations. Unlike STRUCTURE and GENECLASS analyses which require loci to follow Hardy-Weinberg and linkage equilibrium, BAYESASS is robust to violations from Hardy-Weinberg equilibrium as it measures contemporary gene flow within the last few generations based only on multilocus allele sharing among individuals [68]. To ensure convergence, a total run length of eight million MCMC iterations was performed, of which the first two million runs were the burn-in phase and the remaining six million runs comprised the data collection phase with the MCMC chain being sampled every 2,000 steps. Individual assignments and immigrant ancestries were calculated at a migration rate prior of 0.05, and keeping all other settings at default. Varying the prior rate (0.01 to 0.15) did not affect the results.

Lastly, to corroborate the results of the above assignment based migrant decisions, a likelihood based parentage analysis was carried out in CERVUS 3.0 [82], [96] to identify likely parent-offspring relationships between putative migrants and an individual in the cross-assigned source population based on log of the odds (LOD) scores. The LOD score that is the natural logarithm of the overall likelihood ratio for each candidate parent is calculated by multiplying together the likelihood ratios at each locus. A positive LOD score means that the candidate parent is likely to be the true parent, whereas a negative value means that the candidate parent is less likely to be the true parent. We estimated LOD scores for strict (95%) and relaxed (80%) confidence limits as 7.0 and 4.9 respectively, which were calculated from a simulated set of 10,000 offspring and 300 candidate parent genotypes, assuming 25% of candidate parents were sampled, 93% of loci were typed with typing error of 0.01 to 0.10.

Designation of migrant status to an individual was contingent upon - (i) significant assignment of first generation migrant status in GENECLASS (P<0.01, Lh/Lmax ≥2.0); (ii) observance of>50% migrant or cross-assignment probability in STRUCTURE and BAYESASS; (iii) high assignment probability to first generation immigrant ancestry state (gen1>50%) in both STRUCTURE and BAYESASS; and (iv) high membership (Q>0.8) to a single non-home cluster in the STRUCTURE analysis without prior population information. Further, in most cases of putative migrants, the successful parentage assignment corroborated the migrant status of the individual tiger. We considered a conservative approach by identifying individuals as putative migrants only if all three programs suggested evidence of immigrant ancestry.

Estimation of Contemporary and Historical Migration Rates

In order to study gene flow across different timescales, we used the programs BAYESASS 1.3 [68] and MIGRATE 3.3.2 [97], [98] to compare migration rates over contemporary and historical timescales, respectively. Although the two programs use different approaches to derive estimates of gene flow, both programs generate parameters from which a comparative estimate of the proportion of genetic migrants in the population per generation (m) can be inferred. In BAYESASS, a Bayesian approach incorporating an MCMC sampling scheme is used to estimate migration rates between pairs of populations over the last few (approximately <5) generations back. With an estimated generation time of four to five years in tigers [99], this period corresponds to a timescale of nearly 20–25 years ago. MIGRATE on the other hand, uses the coalescent to estimate the relative mutation-scaled effective population size, theta, θ_Ne (4N_eµ; where N_e is the effective population size and µ is the mutation rate) and asymmetric mutation-scaled immigration rate M (m/µ). The mutation-scaled immigration rate M, which is the immigration rate m divided by the mutation rate µ, is a measure of the importance of immigration events over mutation in contributing to variation in the population [97]. The relative effective population size, theta, is the number of individuals representing an idealized (Wright-Fisher) population that will result in the same amount of genetic drift as in the actual population. The number of migrants per generation, 4N_em, is the product of theta and M. MIGRATE assumes mutation-migration-drift equilibrium with values of M and theta constant over time and parameter estimates in MIGRATE date back nearly 4Ne generations into the past (approximately thousands of years ago). Hence, these migration rates provide estimates of gene flow that post and pre-date the estimated time (approximately 600 years ago) when humans began to significantly alter the habitats in which these tigers currently live.

For both the BAYESASS and MIGRATE runs, we used the STRUCTURE defined population clusters to estimate pairwise migration rates. BAYESASS runs were performed as described in the preceding methods section on detecting migrants. A total of 8×10⁶ MCMC iterations were carried out, by discarding the first 2×10⁶ steps as burn-in and sampling at every 2,000 iteration intervals of the remaining 6×10⁶ MCMC chain. Runs were carried out with a migration rate prior of 0.05 while other parameters were kept at default settings. The average result from three independent BAYESASS runs is presented. Estimates of historical gene flow and effective population size were carried out in MIGRATE by using the Bayesian approach and the Brownian motion model as an approximation for the step-wise microsatellite mutation model. Following initial trial runs, the Bayesian search criteria for the MCMC sampler was set at 10 replicates of one long chain of 50,000 steps with every 100 steps of the chain being recorded, producing a total of 5×10⁷ visited parameter values. The initial 10⁷ steps of the MCMC run were discarded as burn-in, and the remaining 4×10⁷ runs were sampled. To increase efficiency of the sampler, we used four chain-heating temperatures of 1, 1.5, 3 and 10,000, which allows a more efficient exploration of the genealogy space. We used parameter estimates from the initial run to calculate starting values of theta, for use as new parameters during subsequent runs. Parameter estimates from the final run were similar to the results of the initial runs. All MIGRATE runs were carried out on the Bioportal cluster computing facility at the University of Oslo, Norway (https://www.bioportal.uio.no/; accessed 12 May 2013).

Detection of Genetic Bottleneck

To detect past occurrences of genetic bottleneck in the sampled populations, we evaluated three summary statistics - (i) Wilcoxon's sign rank test and (ii) mode-shift test, implemented in the program BOTTLENECK 1.2.02 [100], and (iii) M ratio test [101] implemented in ARLEQUIN 3.1 [102]. We were not interested in quantifying population expansion/decline or dating the time of and therefore avoided using other considerably lengthy Bayesian procedures [103], [104]. Wilcoxon's test detects bottlenecks based on the probability of heterozygosity excess over that expected at mutation-drift equilibrium in a population. It is most effective at detecting historic bottlenecks occurring over approximately 2–4 N_e generations in the past. The mode-shift test is more suited for detecting bottlenecks within the last few dozen generations [105], [106]. This test is based on the premise that a stable population at mutation-drift equilibrium will have a large proportion of alleles at low frequency and a smaller proportion at intermediate frequencies and few at high frequencies. The resulting allele proportions yield an L-shaped distribution. In bottlenecked populations the mode is shifted because of the rapid loss of alleles present at low frequency. We ran 10,000 simulations using the program BOTTLENECK in 5 populations (excluding KPC, ATR and TATR) under both the stepwise mutation model (SMM) and the two-phase mutation model (TPM) with 95% single step mutations and 5% multi-step mutations and a variance of 12 as recommended by Piry et al. [100]. P-values from the Wilcoxon's test were used to examine bottlenecks at each timescale and were assessed at the alpha 0.05 level. The M ratio (M = k/r+1) was calculated from the mean ratio of the number of alleles (k) and the allele size range (r). Assuming loci follow a generalized stepwise mutation model, the loss of rare alleles would diminish the value of k at a faster rate than r thereby a drop in the M ratio below a threshold of 0.68 would be suggestive of populations that experienced a recent bottleneck [101].

Field Data Collection for Occupancy Analysis

Remotely Sensed Variables

Remotely sensed data that depict landscape characteristics and human impacts were obtained from various sources and extracted at the 10×10 km grid resolution. Forest cover was obtained from the Forest Survey of India [109] that is based on IRS 1D LISS III satellite with 4 multispectral band data at 23.5 m resolution. Normalized Differential Vegetation Index (NDVI) information were derived from 1 km² Advanced Very High Resolution Radiometer (AVHRR) data, acquired from the National Aeronautics and Space Administration's (NASA) Television Infrared Observation Satellite (TIROS) (http://science.nasa.gov/missions/tiros/; accessed 23 Dec 2010). Road and drainage information were obtained from Digital Chart of the World (http://statisk.umb.no/ikf/gis/dcw/; accessed 20 Dec 2010). Protected area shape files were obtained from the wildlife database at the Wildlife Institute of India, National Tiger Conservation Authority and State Forest Departments of India. The Shuttle Radar Topography Mission has produced the most complete, high-resolution digital elevation model of the earth [110]. Within each 1 km² grid, this information was used for computing average elevation and the coefficient of variation of elevation used as a measure of terrain ruggedness. Night light data was obtained from the United States Air Force Defense Meteorological Satellite Program (DMSP) and National Oceanic and Atmospheric Administration's (NOAA) Operational Linescan System (OLS) (http://www.ngdc.noaa.gov/dmsp/sensors/ols.html; accessed 18 Dec 2010). Density of roads (length of paved road per km²), and Euclidean distances to roads, protected areas and night lights were computed in ArcGIS 9.3 (www.esri.com) software.

Occupancy Modeling

Though sampling was done at the level of the forest beat (an administrative unit of about 15 km²) so as to ensure an even distribution of sampling effort across the landscape, analysis was done at the scale of 1,851 grids, each of size 10×10 km. This grid size was chosen since it was larger than the average home range size of a tiger [111], [112] and the size was relevant for subsequent administrative and managerial inputs. Sign surveys of 5 km independent spatial replicates within each grid [113] were modeled to address the issue of imperfect detections of tiger sign using the program PRESENCE v6.3 [114]. Detection of tiger signs was likely to depend on the abundance of tigers within a grid [115]. We first modeled the detection process by (i) keeping detection () constant across surveys, (ii) varying across surveys and (iii) as a function of tiger abundance in that grid, wherein we used average encounter rate of tiger sign as an index that surrogated tiger abundance [31], [115]. The model that best explained the detection process based on Akaike Information Criteria (AIC) was then used in all subsequent models of tiger occupancy [70].

Tiger site occupancy was a priori expected to be positively influenced by (a) prey abundance, and (b) amount and quality of tiger habitat, and negatively influenced by (c) human disturbance [22], [115], [116]. We tested these hypotheses by modeling variables representing these factors as covariates using the logit link function in PRESENCE [70], [114]. We initially generated data on 23 site covariates that represented landscape and habitat features (forest area, core forest area, forest patch size, NDVI, elevation, ruggedness, drainage density, rainfall, distance to protected area), prey availability (chital, sambar, wild pig and gaur encounter rates on line transect walks, and wild ungulate dung density), human disturbance (distance to night lights, distance to roads, humans and livestock encountered on transect walks, human/livestock trails within sampled plots, and livestock dung density) that could potentially explain tiger occupancy. These covariates were examined with exploratory data analyses for their interrelationships and relationship to tiger presence (by scatter plots, box plots, and correlation analysis) and based on this a subset of 16 variables was selected for inclusion as site covariates for occupancy modeling (see Table S1 for univariate coefficients). Since many of the covariates had high correlation coefficients between them, their contribution to explaining tiger occupancy would be redundant. To account for this collinearity and to reduce the dimensionality of the covariate matrix we extracted Principal Components (PCs) from 16 variables [117]. The varimax rotated PCs were further used as independent variables in a logit link function to model tiger occupancy in the program PRESENCE ([70], available for download from http://www.proteus.co.nz/). Model selection was done using AIC, and model fit was assessed by comparing the actual detection histories with simulations generated from 50,000 parametric bootstrap runs of the target model in PRESENCE. The over dispersion parameter Ĉ close to one suggests that the model provides an adequate description of the data, values of Ĉ greater than one suggests more variation in the data than expected by the model, and values less than one suggest less variation in the data. The standard errors of model estimates were adjusted by the square root of Ĉ as recommended by MacKenzie et al. [70]. Models were built using PCs that represented prey abundance, human disturbance and habitat quality; these were evaluated against the null model and the full model by their delta AIC values. A total of six models were evaluated for modeling tiger occupancy and coefficient estimates for all models with delta AIC <4 were averaged based on model weights [70] to arrive at occupancy probability (Ψ) in each grid.

Tiger Population Extents and Occupied Habitats

We used two approaches to estimate population extents and area of occupied habitats; (i) a more conservative approach wherein we considered only those grids that detected tiger sign as being occupied (the naïve estimate) and (ii) model inferred occupancy that corrected for detection bias and covariates in PRESENCE. Herein, landscape scale occupancy was computed by sum of cell occupancy probability values and divided by the total number of cells. Tiger habitat (forested area) in each grid was weighted by the tiger occupancy probability of that grid and summed across all grids to arrive at occupied tiger habitat for the landscape [115]. All adjacent tiger sign detected cells were joined and were considered to be occupied by a single tiger population.

Habitat Corridor Modeling

Grid based tiger occupancy probability (Ψ) obtained from PRESENCE was used as a measure of habitat suitability for tigers [69], [70]. A cost surface for tiger habitat suitability across grids was generated as 1-Ψ. For the non-tiger habitat (human land uses) where tiger occupancy values were not available, we considered them permeable to tigers at higher costs than forested habitats, although human habitations (townships) were considered impermeable to tigers (see Methods S2 for details). These costs were used as a resistance layer for modeling habitat connectivity using LCP [63] and circuit theory [118] analyses. Based on these cost surfaces the resulting models would optimize connectivity by selecting high quality tiger habitat and minimize gaps formed by non-tiger habitat. Least-cost pathways (LCP) were modeled using PATHMATRIX [71], and resistance pathways were modeled using CIRCUITSCAPE [72]. Core areas of tiger reserves were considered as “sources” or areas of high potential from which tiger movement across paths of least resistance were modeled across the landscape. PATHMATRIX models several potential routes in a radiating manner from the “source” to connect to another adjacent “source”. It then selects a single “least cost” pathway as the best alternative. CIRCUITSCAPE models connectivity through habitat swaths considering resistance to movement based on pixel cost and corridor width [118]. It provides one to several potential alternatives for joining sources and helps in identifying bottlenecks within the corridors. Since the Central Indian landscape is highly fragmented and human dominated with clearly defined and limited forested habitat, we could overlay LCP on high resolution Google Earth images and align them to match geographic features within occupancy grids, to delineate realistic corridors. These least cost corridors buffered by 1.5 km (LCC) were considered the minimal essential corridors joining two tiger reserves.

Genetic Structure, Migrants and Corridor Cost

The genetic structure between populations is a consequence of the amount of genetic drift to which each local population has been subjected, due to its local effective size, and/or due to its degree of demographic, geographic and ecological isolation [119]. Since all the Central Indian tigers likely belonged to a large, mostly contiguous population till a few hundred years ago [23], pairwise genetic distances between populations should reflect levels of differentiation and barriers to gene flow, i.e., the cost of movement between these populations. The proportion of migrants between population pairs would reflect gene flow in current or recent times while F_ST values would indicate genetic differences over historical, more long-term time scales. We therefore expected to have more migrants detected between geographically closer populations that had lower movement costs between them. This exploratory analysis would corroborate the short-term mechanisms (migration events) that result in long-term (F_ST) genetic differences between populations due to tiger movement across the landscape.

Pairwise F_ST, R_ST and Phi_PT genetic distance estimates, obtained from AMOVA analysis in GENALEX, were linearized using the formula F_ST/(1 - F_ST), as given by Rousset [120]. In order to determine which spatial model best explained genetic structuring, matrices of linearized pairwise genetic distances were correlated against matrices representing geographic distances (GGD), log-transformed geographic distances (log₁₀ GGD), least-cost pathway distances (LCPD) least-cost corridor distances (LCCD), and resistance distances (RD). In addition, the effectiveness of the modeled corridors and spatial distance matrices was compared using partial correlations that allowed one model to be tested, while controlling for other competing models [55]. The biologically realistic model not only exhibited the highest significant, positive correlation, but also displayed significant positive partial correlations after controlling for each of the competing models. Mantel [121] and partial Mantel [122] tests were carried out with 10,000 randomizations in the program zt [123] to evaluate the significance of the correlations.

As patterns of isolation-by-distance (IBD) are known to bias tests of hierarchical structuring and vice-versa, we used the population clusters as a covariate in a partial Mantel test, to model the partial correlations between pairwise genetic distances and spatial distances, while controlling the effect of population clusters (following Meirmans [124]). Partial correlations between matrices representing pairwise genetic distances and spatial distance matrices were calculated with a third matrix describing whether population comparisons were made between (1) or within (0) the STRUCTURE identified clusters. A non-significant or negative partial correlation of genetic with geographic distance, after controlling for population clusters, would rule out patterns of underlying IBD in the observed genetic structure.

Identified Individuals and Descriptive Statistics

Out of 587 scat samples, 330 scats were successfully PCR amplified using the felid specific mitochondrial DNA cytochrome b (mtDNA cyt b) primers, of which 275 were identified as tiger scat based on BamHI restriction enzyme digestion profile of the PCR products (Table 1 and Figure S1). From the 275 tiger scats, only 250 samples yielded microsatellite genotype data at a minimum of seven loci to be considered for Identity analysis in CERVUS. We identified 169 individuals with 81 recaptures from a total of 267 microsatellite loci genotypes (250 scats, 16 blood and 1 tissue). The total number of individuals identified here constitutes roughly 49% of the estimated total tiger population in the entire sampled area (Table 1). Sex identification yielded 74 males and 85 females, with the sex ratio being nearly symmetric in most localities, except in the lower sampled areas. On average, individual multilocus genotypes were 93.2% complete. We included 97% (164 out of 169) of samples, those that had complete or near complete genotypes with a maximum of two missing loci, for further analyses. Two Tadoba individuals with three missing loci (73% complete) and three individuals from the Kanha Pench corridor with four missing loci (64% complete) were also retained since sample size from these areas was small and the genotypes represented unique tigers. The panel of eleven microsatellite loci that was used for individual discrimination had very low cumulative sibling probability of identity (PI-sib) of 1.5×10⁻⁴ (Table S1), indicating very high power to discriminate individuals. Even in the samples with the least amount of genotype information (four missing loci), the cumulative PI-sib value (1.6×10⁻³) of the samples was sufficiently low to determine unique individuals.

No significant evidence of linkage disequilibrium (LD) among all pair-wise loci combinations was observed, when all sampling locations were pooled (p>0.05 at 1,000 permutations). Except for deviation at three loci, Fca441, Pati09 and Pati18 (p<0.001), all other loci were in Hardy-Weinberg equilibrium (HWE) in the pooled population after adjusting the critical P-value using the Bonferroni correction procedure (Table S1). Loci not in HWE or with null allele frequencies>5% appear to be random with respect to population (Table S2). Pati09 showed significant deviation from HWE across three populations, while Pati01 and Fca441 showed significant deviation from HWE in two populations. 6Hdz700, F53 and Fca304 each deviated from HWE in one population. Such departures from HWE could indicate the possibility of genetic structuring among populations and likely presence of related individuals in the data. The deviation from HWE in few populations could also be explained by the presence of null alleles in the data. While MICROCHECKER tests did not show any evidence of scoring errors due to stuttering or small allele dominance in the dataset, five loci (Pati01, Pati15, Pati18, Fca304 and F53) with high null allele frequencies>9% were detected (Tables S1 and S2). Three loci deviating from HWE were likely due to the presence of null alleles in the data - Fca304 in Bandhavgarh, and Pati01 and F53 in the Melghat-Satpura-Tadoba cluster, as they showed significant deviation from HWE only with observed allele frequencies and not with null allele corrected frequencies (not shown). Although Pati15 and Pati18 also contained null alleles at high frequencies (>9%) in Kanha, they did not deviate from HWE in any population. Since these patterns were random with respect to population (e.g., no departures from HWE in Kanha, and null allele frequency <5% in Pench), it was more likely that these deviations could be due to the presence of population structuring in the data. Hence we retained all eleven loci in subsequent analyses.

Most tigers genotyped in this study showed high heterozygosity (Figure S2). The minimum number of heterozygotes observed across loci for an individual tiger's multilocus genotype was two and three heterozygous loci, for one and two tigers respectively. The most polymorphic individuals (n = 5) were heterozygous at all eleven loci. Individuals that were completely homozygous at all eleven loci, were not observed. Nearly 90% of the tigers (i.e. 152 out of 169 individuals) had heterozygous genotypes at five to nine loci, and 12 tigers were heterozygous at ten to eleven loci. Genetic diversity estimates showed a mean number of alleles per locus to be 9.1±2.2 with heterozygosity He to be 75.4%±3.9 and Ho to be 70.1%±5.9 in the pooled population (Table 2). A summary of population-wise genetic variation, revealed significantly higher allelic diversity and heterozygosity levels in the larger populations of Pench and Kanha compared to all other localities (Table 2 and Table S2). The number of population diagnostic private alleles was high in Pench (n = 8) and Kanha (n = 6), while individual frequency for a private allele was highest in Bandhavgarh compared to all other localities (Table S3). Estimated population size was positively and linearly correlated to genetic diversity statistics (Number of alleles, rho  = 0.857, p = 0.011; polymorphism information content, rho  = 0.738, p = 0.0458; Shannon's index of diversity, r = 0.714, p = 0.058). Partial correlations between estimated population size and the allele diversity indices, after controlling for the effect of sample size, were positive and highly significant (Number of alleles, r = 0.82, p = 0.012; polymorphism information content, r = 0.659, p = 0.054; Shannon's index of diversity, r = 0.714, p = 0.036). Heterozygosity values were not correlated with population size (Ho: rho  = −262, p = 0.428; He: rho  = 0.357, p = 0.389). Correlation of estimated population size with actual sampled size was highly significant (r = 0.964, p = 0.001), meaning that our samples were in proportion to the size of the population.

Download:

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

Table 2. Summary statistics of mean genetic variation and bottleneck tests across sampled populations.

https://doi.org/10.1371/journal.pone.0111207.t002

Population Genetic Structure

According to the results of the Principal Coordinates Analysis (PCoA) based on Phi_PT genetic distance estimator, individuals in the area were clustered into roughly four groups with varying degrees of population partitioning (Figure 2A). The three coordinate axes accounted for 60% of the variation in the dataset. Tigers were observed to cluster in four major groups. Tigers from Kanha, Pench and Melghat formed three distinct clusters that partially overlapped each other, while Bandhavgarh tigers formed a discrete cluster with minimal overlap. Tigers from Satpura, Tadoba, and Achanakmar were scattered within the clusters formed by Kanha-Pench-Melghat.

Download:

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

Figure 2. Results of individual clustering analyses.

(A) Three dimensional plot showing partitioning between different populations as obtained from PCo analysis based on Phi_PT co-dominant genetic distance among individuals. (B) Summary barplot of STRUCTURE run at K = 4 showing population assignments for each individual. Four distinct population clusters are observed. Sampled populations are Melghat (M), Satpura (S), Pench (P), Kanha-Pench corridor (KPC), Kanha (K), Achanakmar (A), Tadoba (T) and Bandhavgarh (B).

https://doi.org/10.1371/journal.pone.0111207.g002

Calculation of delta K from the output of the STRUCTURE runs using prior population information (locprior  = 1), produced the largest modal value of the statistic at K = 4, suggesting pronounced population subdivision at K = 4 (Figure S3). On the other hand, the log-likelihood L(K) value reached an inflection point at K = 4 before gradually plateauing at K = 6 to 7 and finally leveling off at K = 8 (Figure S3). The variance in L(K) increased at higher values of K, as reported previously with other studies [92], [125]. The disparity in population structuring patterns between delta K and log-likelihood values occasionally occurs in cases where the F_ST values are significant [126]. Examination of individual Q summary barplots (Figure S3) yielded identical clustering patterns at all runs between assumed K = 2 to 4 (carried out both with and without prior sampling location information), and distinct population saturation, indicative of population subdivision was evident at K = 4 in conformance with the delta K approach. Based on the above, the four cluster solution (K = 4) best describes the levels of genetic subdivision in our sample of the Central Indian tiger population (Figure 2B).

The four population clusters identified by STRUCTURE were similar to the clusters observed using PCoA plots, and in consonance with geographic configuration as well (Figure 1). Melghat, Satpura and Tadoba represented populations with patchy connectivity and formed a unique cluster in the western and southern limits of the landscape, with few individuals being cross-assigned to the Pench and Kanha clusters. To the east of Melghat and Satpura, Pench formed a unique cluster. A few individuals in Pench had cross-assignments with Kanha suggesting gene flow between these two population clusters substantiated by a functional habitat corridor between Pench and Kanha. The next cluster in the eastern part of the landscape was represented by tiger populations in Kanha and Achanakmar. Individuals from the forested corridor between Pench and Kanha were also assigned to this cluster, but most had mixed assignments to both populations indicating that this was not a distinct but rather an admixed population. The last cluster was represented by Bandhavgarh, in the north-eastern part of the landscape, which formed a distinct isolated population where all individuals were assigned to the sampled locality.

Partitioning of genetic variation in the AMOVA test indicated low but significant differentiation (p<0.01) between the STRUCTURE identified clusters (Table S4). The major portion of genetic variance was found within populations (88%) with 7% among the population clusters and 5% among populations within clusters. Exact tests showed significant genetic variance on all three levels (p<0.01). Both F_ST and Phi_PT values showed highly significant structuring (p≤0.001), and had relatively similar trends in magnitude with low sampling variance. In contrast, R_ST estimates showed no variation between groups, and had unreliably high sampling variances and mean square error estimates. The results of pair-wise F_ST and R_ST calculations indicated significant (p<0.05) and varied (F_ST 0.049 to 0.241; R_ST 0.000 to 0.330) genetic structuring between all sampled populations in Central India (Table S5). Within cluster F_ST estimates were mostly lower, ranging from moderately low (0.048 to 0.062) to high (0.079 to 0.102), compared to pairwise estimates between different clusters (0.127 to 0.217).

Migrants

(A) STRUCTURE results.

Eight putative migrants were identified in STRUCTURE (Table 3). Of these, four individuals (D954, D955, D958, D1399) were identified with>80% migrant and cross-assignment probability to a single non-home cluster in the STRUCTURE analysis, carried out without prior population information. The remaining four samples represented individuals (D1843, D1892, D1297, D2058) that had weaker migrant probability (P>0.5 to <0.7) and showed variable Q (0.289 to 0.758), and with the majority of samples having cluster memberships to more than one non-home locality.

Download:

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

Table 3. Summary of migrant assignments based on STRUCTURE, GENECLASS (migrants based on **α_0.01 and *α_0.05 type I error levels) and BAYESASS analyses.

https://doi.org/10.1371/journal.pone.0111207.t003

(D) CERVUS results.

Identification of offspring-candidate parent pairs in CERVUS yielded parentage relationships in thirteen out of seventeen putative migrants (Table 3). No evidence of likely parentage (or sibling) relationships in the offspring population was observed. The cross-assigned population in ten of the thirteen individuals matched the parentage assignment in CERVUS, which serves to further corroborate the results of the migrant assignments. Only three mismatches (D1075, D1393 and D1987) were observed between the parent populations identified by CERVUS and the population assignments depicted by the migrant analysis, but this could be due to low information in the data as opposed to incorrect migrant assignment. Except for negative LOD value in two pairs (D955–D1182, LOD = −0.36; D1400–D1168, LOD = −2.15), LOD scores were positive in the remaining eleven putative parent-offspring pairs. The relationship between a potential migrant sampled in Pench (D958) with a candidate parent from Kanha (D1205) was identified with>80% confidence in assignment (LOD = 5.32). LOD scores in remaining offspring-parent pairs were below the 80% confidence limit (<4.9).

The detection of migrants by the above methods yielded a total of seventeen individuals with putative immigrant ancestry (Table 3 and Figure 3). Identical migrant assignment across all three programs was observed in seven individuals (D954, D955, D958, D1297, D1399, D1843, D1892), while there was equivocal assignment in the remaining ten individuals. Sex identification revealed 12 out of 75 males (16%) and 5 out of 84 females (6%) as individuals with immigrant ancestry in the entire area. Figure 3 shows the posterior distributions of individuals assigned to nonimmigrant (gen0), first (gen1) or second generation immigrant (gen2) ancestry states in GENECLASS, STRUCTURE and BAYESASS. All GENECLASS migrants with L_h/L_max>2.0 were classified as 100% first generation migrants. Two individuals (D955, D958) with>90% gen1 assignment and three individuals (D954, D1399, D1843) with relatively high gen1 assignment probability (0.5 to 0.7) were considered as migrants. Five individuals (D525, D1892, D2058, D1297 and D1987) showed moderate levels of migrant assignment and immigrant ancestry patterns are indicative of admixed status. The assignment status of seven more individuals (D1075, D1381, D1383, D1393, D1400, D1140 and D2154) was equivocal. While STRUCTURE could not assign them as migrants, they were identified as potential second generation migrants or admixed individuals in BAYESASS.

Download:

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

Figure 3. Individual ancestry states of putative migrants.

Posterior distributions of individual assignment to nonimmigrant (gen0), and first (gen1) and second generation immigrant (gen2) ancestry states. Suffixes after indvidual names indicate assignment probabilties as obtained in GENECLASS (G), STRUCTURE (S) and BAYESASS (B).

https://doi.org/10.1371/journal.pone.0111207.g003

Contemporary and Historical Migration Rates

The mean posterior distributions of pairwise immigration rates depicting contemporary gene flow estimates in BAYESASS are shown in Tables 4 and 5. Most populations have low migrant proportions with the exception of migration from Pench to Melghat (m = 0.09) and Kanha to Pench (m = 0.07) where the rates were more than 5% (Table 4). Gene flow between Melghat and Pench was likely asymmetric and there appears to be a source-sink relationship because the expected proportion of migrants into the Pench population from Melghat is much smaller (m = 0.015). Asymmetric migration was also visible between the Pench and Kanha population clusters, as the proportion of migrants from Pench to Kanha was negligible (m = 0.006). The Bandhavgarh population was devoid of migrants as suggested by the lack of gene flow between other populations (m<0.01). As in the locality-wise analysis, similar asymmetry and rates of migration were obtained between the various STRUCTURE defined population clusters (Table 5).

Download:

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

Table 4. Locality-wise contemporary migration rates, m, estimated using BAYESASS, showing means (± standard deviation) of the posterior distributions along with the 95% confidence intervals in parentheses.

https://doi.org/10.1371/journal.pone.0111207.t004

Download:

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

Table 5. Cluster-wise contemporary migration rates, m, showing means (± standard deviation) of the posterior distributions along with the 95% confidence intervals in parentheses.

https://doi.org/10.1371/journal.pone.0111207.t005

Results from the MIGRATE analysis showed low estimates of relative effective population size (theta, θ) and historical mutation scaled immigration rate (M) suggesting very low overall migrant proportions in the area over the long-term (Table 6). Theta estimates were low to moderate and ranged from 0.57 (Bandhavgarh), 0.77 (Kanha-Achanakmar, and Melghat-Satpura-Tadoba) to 1.5 (Pench). Estimates of M ranged from a high of 6.17 (Pench to Bandhavgarh) to a low of 0.9 (Bandhavgarh to Melghat-Satpura-Tadoba), revealing limited to no migration among populations in the landscape (Table 6). Of the twelve pairwise population comparisons, ten pairs showed asymmetric migration patterns, with the higher value of M representing immigration from the population with the larger theta value (Kanha and Pench clusters), to the population with the smaller theta value (Melghat-Satpura-Tadoba and Bandhavgarh clusters). The number of migrants per generation ranged from almost zero (Bandhavgarh to Melghat-Satpura-Tadoba cluster) to nine (Pench to Bandhavgarh). The Pench and Kanha clusters represented the largest source populations for immigrants in the area. Though marginally higher migration from Pench to Kanha was visible compared to migration in the other direction, both population clusters had overall symmetric gene flow.

Download:

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

Table 6. Means of posterior distributions of mutation scaled immigration rate, M, along with the 95% confidence limits (before comma) and mean number of migrants per generation (after comma) estimated from Bayesian runs in MIGRATE.

https://doi.org/10.1371/journal.pone.0111207.t006

Genetic Bottleneck

Except for Bandhvagarh, which had a low M ratio of 0.66, all other populations showed M values above the critical threshold of 0.68 (Table 2). Wilcoxon's sign rank test of heterozygosity excess were not significant (p>>0.05) regardless of the mutation models used and all localities tested showed normal L-shaped allele distributions in the mode-shift test, indicative of stable non-bottlenecked populations (Table 2). Only, the Pench and Kanha populations showed significant (p<0.05) evidence of heterozygosity deficiency (Table S6), which could suggest recent events of expansion in these two populations.

Tiger Occupancy, Habitat, Prey and Human Disturbance

Ten Principal Components (PCs) had Eigen values greater than one and together explained 86.6% of the variability of the original variables (Table 7). The factor loadings of the components permitted unambiguous ecological interpretation of PCs (Table 7). The detection probability of tiger sign was 21.98% (SE 0.04). Detection of tigers was best explained by index of tiger abundance. The closest competing model for detection differed by a ΔAIC value of 1207 (Table 8). Tiger occupancy was best explained by a model that included PC1 to PC9 and all the covariates conformed to the a priori predictions of their influence on tiger occupancy. The best model differed from the full model by ΔAIC of 0.45 and from the null model by ΔAIC of 1682. Tiger occupancy in a grid was best explained by (a) PC1 that represented availability of large undisturbed good canopy forests, (b) PC2 and PC4 that represented grids that had low human and livestock disturbances, (c) PC3, PC6 and PC9 that represented high ungulate prey especially in the form of chital, sambar, wild pig and gaur, (d) PC5 representing lower elevation, (e) PC7 representing grids within or in the proximity of legally protected areas and (f) PC8 that gave greater loadings for grids with higher rainfall (Tables 7 and 8). The sign of the coefficients of the best model for each PC was the same when models were run using original variables in a univariate model (Table S7). The goodness-of-fit test for observed data against 50,000 model based bootstrap samples failed to show lack of model fit (χ² P value  = 0.11), Ĉ statistic was estimated at 1.4 and standard errors shown in Table 9 are corrected for over-dispersion by a factor of 1.18. Out of the 185,100 km² area of the sampled grids, 76,913 km² was forested or potential tiger habitat. The naïve estimate of grid occupancy was 17.5% while the model-averaged estimate of occupancy was 20.87% giving a 3.4% increment over the naïve estimate. The naïve estimate of tiger occupied habitat was 19,845 km² while the model inferred occupancy of forests was 21,290 km² giving an increment of 1.8% in occupancy estimate. The habitat variables of forest area, forest core area, rainfall, NDVI, and terrain ruggedness had significantly higher values in tiger occupied grids compared to unoccupied cells. All wild tiger prey indices were higher, while livestock abundance indices were lower in tiger occupied grids. Relevant human impact indices had significantly lower values in tiger occupied grids (Table 9 and Figure S4).

Download:

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

Table 7. Principal component (PC) loadings of covariates relevant for modeling tiger occupancy, Eigen values of the components, the percent variation of the original data explained by the PC, their ecological interpretation and a priori effect on tiger occupancy.

https://doi.org/10.1371/journal.pone.0111207.t007

Download:

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

Table 8. Model selection results for estimating tiger occupancy within the Central Indian Landscape incorporating imperfect detections and covariates of landscape characteristics, prey abundance, and human disturbance represented by 10 Principal Components.

https://doi.org/10.1371/journal.pone.0111207.t008

Download:

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

Table 9. Coefficient estimates for the best model selected for estimating tiger occupancy in the Central Indian Landscape.

https://doi.org/10.1371/journal.pone.0111207.t009

Tiger Population Extents and Occupied Habitats

Tiger populations within the landscape were primarily located in and around tiger reserves (Figure 1). The Pench-Kanha-Achanakmar tiger population was located in the largest patch of contiguous forest comprising 16,063 km² with intermittent tiger presence recorded throughout this patch even outside the legal reserve boundaries. The Satpura-Melghat forest patch was 12,720 km², while the forest patches that contained Tadoba and Bandhavgarh Tiger reserves were smaller 2,088 km² and 1,902 km² respectively and connected to larger adjacent patches by fragmented forests (Figures 1 and 4). Melghat Tiger Reserve had the largest critical core area that is legally mandated to be made free of human habitation (Table 1). The total protected area in the landscape under the tiger reserve system was 13,054 km² with 6,395 km² as core area. Tiger population extent was largest for the Tadoba population at 3,519 km², while the smallest area occupied (904 km²) was recorded in Achanakmar (Table 1).

Download:

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

Figure 4. CIRCUITSCAPE model of cumulative current flow used to estimate landscape permeability to tiger movement.

Tiger movement modeled as current flow within the Central Indian Landscape using tiger occupancy probability and drainage systems as conductance layers and human settlements as high resistance barriers in CIRCUITSCAPE. Light colors indicate potential habitat corridors. Note the prominent bottlenecks observed in the Kanha-Pench, Kanha-Tadoba, and Tadoba-Melghat habitat corridors.

https://doi.org/10.1371/journal.pone.0111207.g004

Corridor models – maps and corridor cost values between source pairs

The least-cost corridor plot joining tiger reserves (Figure 1) shows that the longest corridor was between Bandhavgarh and Melghat; and the shortest corridor was between Kanha and Achanakmar (Table S8). The maximum number of barriers in the form of crossings of national highways was five for the corridor connecting Bandhavgarh to Melghat. CIRCUITSCAPE results detected bottlenecks in connectivity in corridors connecting Kanha with Pench, Tadoba with Kanha and Tadoba with Melghat (Figure 4). Alternative habitat connectivity besides the least cost corridor was detected by CIRCUITSCAPE to exist between Bandhavgarh and Melghat along the sparse ridge forests on the southern banks of the Narmada river as well as between Tadoba and Melghat with patchy connectivity observed through Bor wildlife sanctuary. Habitat connectivity between Achanakmar and Bandhavgarh was diffused with no clearly defined flow pathways being observed (Figure 4).

Genetic structure, migrants and corridor costs

Mantel's r correlations between pairwise genetic distances and spatial distance metrics were positive, and showed a similar trend across all three genetic distance estimators (Table S9). The highest correlations were observed for IBR and LCC distances with genetic distance while geographical distances showed the lowest correlations with genetic distance. Linearized R_ST had weak non-significant correlations with spatial distances, compared to F_ST and Phi_PT estimates. A significant linear relationship was observed between population pairwise F_ST/(1−F_ST) genetic distances vs. resistance (r = 0.549, p = 0.019) and least-cost corridor distances (r = 0.533, p = 0.009). In comparison to Euclidean geographic distance correlations, this corresponded to a 29.5% and 25.7% increase in model fit for the circuit theory based IBR and the LCC models respectively. The proportion of migrants was higher between population pairs that had lower corridor costs (Figure 5). Linearized Phi_PT estimates also produced significant positive correlation with both least-cost corridor (r = 0.416, p = 0.035) and resistance distances (r = 0.462, p = 0.023). This meant a 46.2% and 31.6% increase over log-transformed Euclidean distance correlations for the circuit theory and corridor models respectively.

Download:

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

Figure 5. Regression of population pair-wise linearized F_ST values with corridor cost.

The size of each circle is representative of the proportion of migrants shared between each population pair. Depicted corridors are Kanha-Bandhavgarh (KB), Kanha-Achanakmar (KA), Pench-Kanha (PK), Melghat-Kanha (MK), Satpura-Kanha (SK), Satpura-Pench (SP), Melghat-Pench (MP), Melghat-Satpura (MS), Kanha-Tadoba (KT), Pench-Achanakmar (PA), Melghat-Achanakmar (MA), Achanakmar-Bandhavgarh (AB), Pench-Bandhavgarh (PB), Satpura-Bandhavgarh (SB) and Melghat-Bandhavgarh (MB).

https://doi.org/10.1371/journal.pone.0111207.g005

Correlations in partial Mantel tests showed a significant, positive relationship between genetic and resistance distances, after controlling for all competing spatial distance metrics (linearized F_ST: r>0.48, p<0.02; linearized Phi_PT: r>0.39, p<0.05). In contrast, after controlling the effect of resistance distance, significant positive correlation was only observed between linearized F_ST vs. geographic and least-cost corridor distances, while partial correlations between all other spatial and genetic distance metrics were either non-significant or negative. Although significant positive correlation was observed between F_ST and geographic distances in the standard Mantel tests, genetic and geographic distances were uncorrelated in the partial Mantel tests, after controlling the effect of population clusters, thus affirming that isolation-by-distance (IBD) pattern was absent in the data. Only resistance (IBR) and LCC distances retained a significant, positive partial correlation with genetic distances, when controlled for the effect of population clusters (linearized F_ST: r>0.4, p<0.03; linearized Phi_PT: r>0.4, p<0.04).

Genetic diversity and population structure

The present study is illustrative of the general strengths and challenges in using non-invasive genetic samples to assess spatial distribution patterns in an endangered and cryptic species living in fragmented habitats. Although maternally inherited mitochondrial DNA could elucidate patterns of female gene flow, the sequencing of mitochondrial genes was not attempted as the primary objective of the study was to assess the functionality of corridors through the detection of first and second generation migrants. Hence, highly polymorphic, bi-parentally inherited, co-dominant and selectively neutral nuclear loci such as microsatellites were most suited for this purpose. Microsatellite marker amplification success was low from scats, reflecting the often poor quality of samples, typical of tropical environments [127]. Even though standardized methods of sample preservation and DNA extraction were used for the study, finding fresh scats (which may have improved results) was rare despite repeated surveys, due to the low tiger density of many sampled areas, the remoteness of the terrain, and further compounded by the heat and humidity of the region which enhanced degradation of scats. Despite these constraints a relatively sizeable sampling was achieved with the total number of identified individuals (n = 169) representing about 49% of the estimated population [31] in the entire landscape (Table 1). The gender proportions observed in our data are similar to the estimates obtained from camera-trap studies in the area [31]. The relatively large sample size in relation to the estimated population enhances the significance of our results.

Most individuals genotyped in this study, possessed heterozygous genotypes at>50% of typed loci, and there was no discernible differences in heterozygosity between migrant and resident individuals (Figure S2). However, unlike resident individuals, we found no migrants with fewer than five heterozygous loci, or completely heterozygous across all eleven loci. Eleven of the seventeen identified migrant individuals were slightly skewed towards a higher heterozygosity distribution, being heterozygous at eight to ten loci, while the rest of the migrants were polymorphic at five to seven loci. This could be due to the contribution by second generation migrant individuals which constituted the bulk of migrant pool. These individuals are more likely to possess a higher number of heterozygous loci than resident tigers, since they are admixed with parental genotypes from different allelic distributions.

The allelic diversity (9.1±2.2) and heterozygosity (Ho = 0.701±0.059, He = 0.754±0.039) observed in our study is typical of genetic diversity prevalent among tigers in the Indian subcontinent [23], [128], [129]. Related studies, using different microsatellite markers than the ones used here, have observed levels of average expected heterozygosity (He) and mean number of alleles (A) to be lower in other subspecies, viz. P. t. altaica (He = 0.26±0.11, k = 2.6±0.84 [130]), P. t. sumatrae (He = 0.493±0.039, A = 3.60±1.48 [129]), P. t. jacksoni (He = 0.571±0.027, A = 3.90±1.18 [129]) and P. t. corbetti (He = 0.670±0.027, A = 6.03±1.81 [129]). Interestingly, allele diversity and heterozygosity in the Indian tiger was comparable to other felid species [129], such as the jaguar (P. onca, He = 0.792±0.0137, A = 8.67±1.72 [131]), Asian leopard (P. pardus, He = 0.790±0.0174, A = 10.71±2.31 [129]), African lion (P. leo, He = 0.610±0.0348, A = 5.0±1.75 [129]), South American puma (Puma concolor, He = 0.774±0.0247, A = 7.0±1.76 [129]), and cheetah (Acinonyx jubatus, He = 0.708±0.130, A = 6.1±1.8 [132]). Although not directly comparable with the present study because of the different markers used, a recent study in the Satpura-Maikal landscape in Central India [64] based on seven microsatellite loci, detected similar levels of heterozygosity (Ho = 0.65±0.09, He = 0.80±0.05), and allelic diversity (A = 7.76±1.96), and very low genetic subdivision (mean F_ST = 0.013±0.006). Another study in the southern parts of the Central Indian landscape, focused primarily in peripheral habitats just outside the scope of this study area [66] also observed high mean heterozygosity (Ho = 0.54, He = 0.81), and allelic diversity (A = 11.71) levels at fourteen microsatellite loci. A range-wide study [23] conducted with five microsatellite loci showed that Indian tigers have higher heterozygosity (Ho = 0.70±0.16) and allelic diversity (A = 12.4±3.6) compared to all other tiger subspecies (Ho = 0.53±0.07, A = 7.2±1.2). The study also reported low and non-significant genetic structuring of the Central Indian tigers with the Northern (F_ST = 0.027, p = 0.063) and Southern Indian populations (F_ST = 0.019, p = 0.054). This could be attributed to a historically large effective population size and inter-population connectivity in the region of Central and Peninsular India, explaining why, despite centuries of immense trophy hunting and continued habitat fragmentation, extant tiger populations in the region currently retain close to 60% of the global genetic variation in the species [23].

By using a combination of classical population differentiation and individual clustering approaches, we were able to detect patterns of population sub-structuring in the region. Results from both the PCoA and STRUCTURE analyses suggest the presence of four genetic clusters in the area. The clustered localities occur in close proximity to each other and the overall pattern of genetic structuring observed in the landscape is concordant with existing habitat connectivity and indicative of the role that habitat fragmentation plays in shaping the distribution of allele frequencies across populations. Significant genetic structuring was also detected in the population pairwise F_ST and R_ST statistics, which may have biased parameter estimations in the STRUCTURE runs carried out without prior population information (locprior  = 0) resulting in higher hierarchical levels of population subdivision at K = 8. In contrast, the runs carried out with population information (locprior  = 1, K = 4) did not detect unnecessary genetic structuring, and ignored the prior sampling location information if the ancestry of individuals was uncorrelated [126]. The cluster solution of K = 4 appeared optimal for the following reasons. In cases where STRUCTURE detects multiple clustering options with similar probabilities, typically the lowest K value which captures much of the biological complexity in the sample is considered to be the most conservative [126]. Additionally, the presence of related individuals in our sample and the model of correlated allele frequencies used for analysis can lead to overestimation of the true K value [126]. Variation across the four STRUCTURE-defined population clusters was weak, but significant, in the AMOVA test, as the major portion of genetic variance was attributed to within population variation. Though the AMOVA results and F_ST statistics indicated significant pairwise structuring across all populations in the area, the pattern shared certain similarities to the population clusters identified by STRUCTURE and the ordination results in PCoA. F_ST values were generally lower for localities within the same cluster compared to pairwise estimates between different clusters. We therefore treated the genetic distance estimators including F_ST and its analogues as relative measures of population differentiation. The estimation of these parameters requires prior identification of populations and unless the population units are clearly known, such a priori designation may not reflect realistic biological patterns as they would only be representing ad hoc division of populations [133]. Importantly, the assumptions of demographic and genetic equilibrium along with long time scales under which F_ST and its analogues are based may not be suited for estimating genetic differences between populations that have undergone fragmentation or demographic fluctuation events only recently. On the other hand, the results of the model based STRUCTURE clustering approach, which partitioned individuals into relatively distinct clusters based on iterative assignments, made much more sense of the biological realities of the area as the structuring observed here is likely an artifact of recent population fragmentation.

Additionally, the differences observed between F_ST and R_ST values can provide valuable insights into the balance between genetic drift and mutation events in the studied populations [134]. In this study, pairwise genetic differences between tiger populations in the landscape showed higher F_ST values compared to R_ST values. The F_ST statistic is based on allelic identity and accounts for gene flow between different populations as the basic premise under which it estimates pairwise genetic differences [134]. In contrast, R_ST relies on allele size and single stepwise mutations are the primary contributors of genetic variation for this statistic [134]. R_ST estimations produced non-significant results across many pairwise comparisons, and had higher sampling variances compared to F_ST. This suggests that populations in the area were not isolated long enough for mutations to have caused the genetic differences between populations and the primary cause of genetic structuring in the area is due to recent genetic drift. Some measure of this drift is evident from the number of private alleles found in each population, especially Bandhavgarh, which has the highest individual allele frequency compared to all other localities (Table S3).

Although the STRUCTURE and PCoA results suggest subdivision at four clusters for the Central Indian tiger population, the significant F_ST structuring observed between all populations could also be indicative of ongoing fine-scale genetic differentiation in the area. Existing patterns of population structuring are a result of past fragmentation effects and fragmentation is not a static process. Gene flow in the area is currently meager and likely to become even lower, due to continued habitat loss and burgeoning anthropogenic activity in the area. Many localities still retain marginal inter-population connectivity, as evidenced by the presence of individuals having immigrant ancestry and further substantiated by camera trapping and radio-telemetry. However, genetic isolation of almost all populations in the foreseeable future is likely if current patterns of habitat fragmentation persist. In the case of Bandhavgarh the extent of fragmentation appears to be so great that the population may have already become genetically isolated for a longer period than the other reserves. Efforts should be made to revitalize the least cost corridor connecting Bandhavgarh with Achanakmar and subsequently to the gene pool of the main Central Indian complex by a combination of restorative ecology and legal instruments (See conservation implications below).

Migration in the Central Indian tiger population

The Central Indian landscape is a mosaic of habitats where tiger population densities [31] varying from high (Bandhavgarh 14 tigers/100 km²) to low (Achanakmar 0.1 tigers/100 km²) are juxtaposed in a matrix encompassing a range of land-use regimes from undisturbed natural forest to high density human settlements (Figure 1). Despite being one of the most fragmented TCL in India [31], some potential for gene flow across populations within the landscape exist because tigers, like most large carnivores, have the ability to disperse across great distances where habitat connectivity is present (e.g., ∼150 km linear distance between two reserves in Nepal [67]). However, in the fragmented human-dominated Terai Arc landscape of Nepal, tigers were not found to disperse across expanses of agricultural land (10 to 20 km wide) though they did traverse through stretches of degraded forest [41]. So far, the only published dispersal estimates from the Central Indian landscape are from recent genetic studies that observed first generation migration and long-range dispersal across protected areas located roughly 200 km [65] and 650 km [66] apart, suggesting that tigers may be more resilient in traversing fragmented habitats than previously reported. Extensive annual camera-trapping since 2006 has recorded tigers dispersing between Pench and Kanha, a geographic distance of more than 150 km (Jhala and Qureshi unpublished data).

Identification of migrants by the methods used in our study requires that the populations being investigated have sufficient genetic separation and low levels of migration. We found in our study that the power to detect migrant individuals varied across localities in the region. There was no power to detect migration between localities in the same cluster (such as Kanha-Achanakmar). In these situations it was difficult to tease apart actual migration events from similar allelic patterns that arose from shared population histories and low genetic separation. On the other hand, migrant identification between localities in different population clusters produced robust assignments because of distinct genetic differences. Migrant detection was highest between localities that had discernible genetic separation and had relatively intact habitat corridors with confirmed tiger presence, such as between Kanha and Pench populations where four migrants and seven likely admixed individuals were detected. Other studies by Joshi et al. [66] and Sharma et al. [65], respectively, detected one and two first generation migrants between Pench and Kanha. The intervening forest patch between Kanha and Pench not only served as a movement corridor but also had some resident tigers as evidenced from camera trap data [135]. A noninvasive genetics study by Sharma et al. [64] detected seventeen individuals from this corridor. At the extreme end in Bandhavgarh, no migrants were detected even though distinct genetic structuring of Bandhavgarh from other population clusters provided sufficient power to distinguish migrants. Between the Melghat and Pench populations where three individuals with migrant ancestry were obtained, there was limited but adequate power to detect migrants as genetic separation was distinct but at a lower level than Bandhavgarh. One individual sampled in Satpura was cross-assigned to Pench. The corridor between Pench and Satpura has fragmented forest connectivity and tigers were reported from this area within the last ten years [42]. Two individuals with admixed ancestry were also detected between Kanha and Tadoba, though one of these had mixed assignments with Pench as well suggesting gene flow in the last few generations. Two individuals with admixed ancestry were also detected between Kanha and Tadoba. Areas of degraded forest and low density tiger occupancy such as Nagzira and Nawegaon Wildlife Sanctuaries located strategically between Tadoba and Kanha, and Bor located between Tadoba, Pench and Melghat may serve to provide stepping stone [136] type corridor connectivity (Figure 1). These areas have resident tigers, and it seems likely that the populations are sustained by sporadic dispersing individuals from larger source populations. Joshi et al. [66] observed that Tadoba, Nagzira and Melghat form a genetic cluster, and based on asymmetric migration patterns in their study it appears that the tiger population in Nagzira acts as a sink population for the high density source population in Tadoba.

Based on our strict criteria of defining migrants, we designated four individuals, three males (D954, D955 and D1399) and one female (D958) that had high cross-assignments (Q>0.8) to a single non-home cluster, as first generation migrants. Further, a fourth male individual (D1853) that had high cross-assignments, albeit with marginally lower Q value (0.527) compared to the above four migrants, was also designated as a first generation migrant. The immigrant ancestry status in the remaining ten putative migrant individuals was classified as admixed based on - (i) intermediate levels of migrant probability in STRUCTURE; (ii) membership to more than one cluster in the individual Q assignments; and (iii) high probability of assignment to second and not to first generation immigrant ancestry state in STRUCTURE and BAYESASS.

Though one of the first generation migrants was a female, most individuals with immigrant ancestry were males (80%: 4 males out of 5 total migrants), confirming that dispersal is male-biased in tigers [41]. The low number of strict first generation migrants suggests that contemporary migration events are minor in proportion to the number of individuals with likely admixed ancestry. The identification of samples with admixed ancestry suggests that most migration events in this landscape have occurred within a few generations and is evidence that migrant tigers are able to reproduce in the new locality. This genetic evidence is further supported by field observations where a sub-adult male tiger photo-captured in Pench tiger reserve in 2006 was observed to be a territorial breeding male in Kanha tiger reserve in 2010 (Jhala and Qureshi unpublished data). The data and analyses show that tigers likely disperse between Kanha-Pench, Pench-Satpura-Melghat, Kanha-Tadoba, Melghat-Tadoba and Tadoba-Pench.

Gene flow rates and past demography

Both the contemporary and historical analyses detected low estimates of migration rate between populations in the landscape. While this result is superficially similar to the findings of Joshi et al. [66], it is in sharp contrast to Sharma et al. [65] where they found both high historical and contemporary gene flow, although historical rates were much higher than contemporary rates. As observed in our study and by Joshi et al. [66], low levels of contemporary gene flow are expected given that the area is highly fragmented and extant populations occupy habitats that are completely surrounded by heavily modified landscapes, altered by agriculture and high density human settlements, thus making it difficult for tigers to disperse between populations. The highest estimate of contemporary migration was seen from Pench to Melghat, but there appears to be an asymmetric source-sink relationship between the two populations with very low migration from Melghat to Pench. Such a pattern is expected given that Pench has high tiger density (4 tigers/100 km²) likely serving as a source population, while the population density in Melghat is lower (2 tigers/100 km²). Asymmetric contemporary rates of gene flow were also apparent between Pench and Kanha, as migration from Kanha (6 tigers/100 km²) to Pench was much higher compared to movement in the opposite direction from Pench to Kanha. Joshi et al. [66] similarly detected high gene flow (>5%) between Kanha and Pench, with Kanha acting as the biggest contemporary source for immigrants. In contrast, Sharma et al. [65] state that only Pench is acting as a contemporary source population and contemporary gene flow from Pench to Kanha and Satpura is very high, and has remained relatively stable since historic pre-disturbance times. A comparative evaluation at historical and contemporary time scales in their study [65] showed that gene flow between the Pench–Satpura and Melghat–Satpura pairs remains similar, whereas there has been a 47–70% reduction in gene flow between Kanha–Satpura, Pench–Melghat and Kanha–Melghat. Our results showed that historical patterns of migration between the two major population clusters of Kanha and Pench were of equal, albeit low proportion in both directions, and that both Kanha and Pench acted as source populations in contemporary times as well. Although Sharma et al. [65] reported higher historic and contemporary gene flow rates compared to our study, both studies similarly observed that Kanha and Pench were important historic source populations and were the main drivers of gene flow in the area.

Contemporary density estimation studies using photographic capture-recapture techniques in the region [31] help explain some of the patterns observed in the gene flow analysis. During 2006, the Kanha population experienced a phase of relative expansion compared to a decline observed later in 2010. The situation was reversed in Pench where the population was relatively low during 2006–2009 compared to a later phase of expansion in 2010 [31], [42]. From the genetic sampling done in Pench during both 2007 and 2010, which pre and post-dated this period of population expansion in the locality, a range of individuals representing likely migrant or admixed ancestry to Kanha was obtained. All first generation migrants from Kanha to Pench were detected during 2007 only, when the population in Kanha was in expansion phase. In 2010, seven individuals showing admixed ancestry from Kanha were obtained in Pench. Dispersal and subsequent breeding by immigrants from Kanha to Pench during periods of population expansion in the former appears to explain the patterns of immigrant ancestry detected in the Pench population.

As parameter estimates of migration rate and theta in MIGRATE average across at least 4Ne past generations, [137] or several thousands of years ago in absolute time, the low long-term gene flow estimates observed in this study could suggest that populations may have been subject to historical fragmentation and genetic drift. Although ancient population fragmentation is perceived to have occurred in response to the forest clearing activity of agro-pastoralist Neolithic people during the mid-Holocene about 5,000–3,000 years ago [138], the effect is difficult to evaluate as early farmers likely did not clear forests at a scale comparable to recent centuries [139]. However, it is conceivable that the shift from hunter-gatherer to agro-pastoralist lifestyle and successful colonization of new areas would have negatively impacted large mammal populations including tigers and prey species through hunting (and reduction of prey base), as seen during human colonization events in other parts of the world [140], [141]. Evidence from an exhaustive range-wide study of tigers has detected a massive decline of about 98% in the number of tigers over the last 200 years in peninsular India [23]. The wanton slaughter of tigers and other wildlife can be glimpsed from historical hunting records of the area, where upwards of 1,000 tigers and 2,000 leopards from this region were killed within just a few decades [25]. The population decline continues apace today, as omnipresent fragmentation and rampant poaching of tigers, other carnivores and prey species have reduced tiger populations such as at Achanakmar to only a few individuals [31], and resulted in their local extirpation in adjacent reserves of Sariska [5] and Panna [29].

Alternatively, the low estimates of historic migration rate raise concerns whether these results are artifacts of sampling or population related. MIGRATE parameter estimations assume coalescent-based models of constant population sizes and mutation-migration-drift equilibrium. Departures from the equilibrium model such as recent and sudden declines in population sizes can negatively bias the posterior parameter distributions of theta and hence migration rate estimations [137]. However, this study and a recent work by Sharma et al. [64] did not find significant evidence of past demographic contraction. In our study, only the population of Bandhavgarh had below par M ratios (<0.68) suggestive of bottleneck, but the evidence was equivocal since no significant heterozygosity excess or a mode-shift in allele frequencies was detected. Though bottlenecks were not observed in this study, the analyses may be undermined in a few populations because of low sample sizes (<20 individuals), as tests such as the mode-shift in allele frequencies is known to be affected by sample size variances [106]. Also, a demographic decline may not necessarily result in a bottleneck as several factors such as duration of decline, pre-bottleneck diversity, and gene flow between populations can affect the probability of detecting a bottleneck [101], [106], [142], [143]. Furthermore, the significant heterozygosity deficit (symptomatic of recent population expansion) observed in the Pench and Kanha populations could mask signatures of population bottleneck as the addition of new individuals could increase the number of rare alleles which can bias allelic and heterozygosity distributions [106]. Demographic expansion in Pench and Kanha is likely, given the recent history of tiger population recovery in the area in the 1970s–1980s [28], [144], or about six to eight generations ago (considering a five year generation time in tigers [99]). This period is well within the window of detection for genetic bottleneck tests of heterozygosity deficiency/excess which assess short term demographic changes only, maximum up to ten generations before present [143]. We could not account for other genetic effects or demographic events which could affect the heterozygosity distributions in these two populations, although our analysis did detect some amount of higher hierarchical STRUCTURE clustering patterns which could be indicative of lineage or extended family structure, since the data definitely contained related individuals. However, even if kin-based segregation did produce a signature of heterozygosity deficit, the specific nature of the effect is difficult to evaluate without in-depth parentage and relatedness analyses, and backed by field investigations, which was beyond the scope of this study. Based on both the theta estimates and absence of bottlenecks, our results suggest that the bigger source populations in the study area have had a relatively stable population history, compared to the smaller populations, as also observed by Sharma et al. [65].

Our study and the recent study by Joshi et al. [66] had higher power to resolve hierarchical genetic structuring in the area compared to Sharma et al. [65] where indistinct structuring was observed. This was most likely due to the higher number of loci used in both studies compared to Sharma et al. [65]. Except for Tadoba where sampling was low due to logistic constraints, the small sample sizes from some other sites were due to small tiger populations. Simulation studies by Paetkau et al. [94] caution regarding the use of MCMC resampling methods implemented in Rannala and Mountain [95], as they tend to over-estimate migrants from a limited data set. In our case, this translates to the possibility that there may be less migration than we report in some of the smaller populations in the study system. Due to low sample sizes, individuals from Achanakmar (n = 4) and the Kanha-Pench corridor (n = 5) were not analyzed separately, as doing so resulted in overestimation of migration rates. Instead these localities were merged with the Kanha cluster, following the results of STRUCTURE assignments. Low sampling was also a problem in the study of Joshi et al. [66] as sites such as Kanha, Pench and Melghat were clearly under-sampled.

Tiger Occupancy, Habitat, Prey and Human Disturbance

Occupancy has often been used as a straightforward and economical state variable in place of abundance to monitor populations [145]. Occupancy of tigers has been estimated in the Mysore-Wayanad Landscape [115] and the Corbett-Rajaji Landscape [116] and across three major tiger landscapes of India [22]. Correcting for detection bias by replicate surveys within sampling units is especially important when detection probabilities are low, sample sizes and survey lengths are small resulting in naïve occupancy estimates that are substantially negatively biased [115]. The recent availability of powerful analytical tools and software [70], [114] that permit accounting for bias caused by detections being <1 has promoted the use of correcting for such bias post data collection [115], [146] compared to more rigorous designed field methods that minimize such bias in the first place. In this study we use independent spatial replicates, each of 5 km in length, with a minimum effort of one survey for every 5 km² of tiger habitat. Thus a 10×10 km² grid had anywhere between 3 to 35 replicate surveys depending on the amount of tiger habitat present within that grid. The detection probability of tiger sign was high in comparison to other studies [115] due to longer spatially independent surveys as well as higher density of surveys. Therefore, the difference between naïve and bias corrected estimates of occupancy were comparatively small (3.4%). Several assumptions that are difficult to meet underlie the use of occupancy as a state variable to monitor populations [147]. Further, tiger populations are at risk of poaching and they can be severely reduced or even extirpated from prime habitats [5], [29]. Thus, model inferred occupancy from covariates of habitat, prey, and human disturbance can provide misleading inference regarding true occupancy in the case of tigers and other species that can be severely depleted by poaching, and should therefore be interpreted with caution. However, occupancy models serve to provide a good assessment of habitat suitability at a scale relevant for conservation management [70], [146]. It is in this context that we primarily use the occupancy probability in this paper so as to model habitat corridors joining tiger populations.

Detection of tiger sign was better modeled by including encounter rate of sign per km walk. Intensity of sign was found to be a reliable index of tiger abundance [22]. Detection of sign was likely to be higher with greater abundance of tigers and this was reflected in our model. Principal Components that represented distance from protected areas, prey abundance and remote canopied forests had the highest coefficients in the logit-link function that modeled tiger occupancy. All the model coefficients were in concordance with the a priori predictions based on our understanding of tiger ecology. Since PCs are orthogonal to each other [117] coefficients can be interpreted in terms of their sign and magnitude as the models are free from the effects of multicollinearity. The model coefficients of all site variables when used independently to model tiger occupancy as univariate models also conformed to the a priori predicted effects and to the coefficients obtained by the multivariate best model obtained from their PCs by PRESENCE. These results indicate that tigers are a conservation-dependent species, requiring areas having effective legal protection, good forest cover with least human impacts and high prey availability.

Corridor models, genetic structure, migrants and dispersal costs

Although both LCP and circuit theory used tiger occupancy probability as a base layer to parameterize landscape resistance surfaces, LCP defines the optimal minimal route that is required to connect two tiger reserves [53], while the resistance layers obtained from circuit theory analysis provides an understanding of all habitat connectivity currently available between tiger reserves [118]. Circuit theory is especially important to identify bottlenecks in corridors where current is constrained to flow through a narrow channel due to high resistance of the neighboring matrix. Such bottlenecks are highlighted in Figure 4 and exist in almost all corridors connecting tiger reserves. A corridor is as good as its weakest link and if the high resistance habitat matrix surrounding these bottlenecks increase in their extent they could choke the corridor, thus forming a barrier to tiger dispersal.

In our study, the ecological distances that were generated using information from habitat heterogeneity and landscape resistance surfaces (LCCD and RD), correlated significantly with pairwise genetic distances, which validated the effectiveness of the modeled linkages in representing realistic biological scenarios. In particular, IBR models depicted well the ecological costs of movement as they accounted for multiple pathways, irregular patch effects, landscape heterogeneity and wider habitat swaths connecting populations [54]. Although the raw LCP model accounted for habitat heterogeneity while selecting a single pathway as the optimal minimal corridor, it did not correlate significantly with pairwise genetic distances (except F_ST). Instead, the LCCs which were delineated by matching and buffering the pathway results from the raw LCP model with existing ground forest cover maps of the area provided a more realistic relationship of the ecological costs with observed genetic differences. In contrast, Euclidean distances (GGD and log₁₀GGD) did not correlate significantly with genetic distances, due to the null model assumptions of spatial homogeneity where habitats are arrayed in an infinite lattice [124], and hence do not conform to the ground realties of fragmentation in the Central Indian landscape. The apparent IBD pattern observed due to the significant correlation of F_ST (but not Phi_PT or R_ST) with geographic distance is an artifact of metapopulation structure and discrete population clusters present in the area. This observation was affirmed by non-significant partial correlations between genetic vs. geographic distances (but not LCCD and RD) using the STRUCTURE identified population clusters as a control. Apart from fragmentation induced spatial heterogeneity, genetic structure is also a result of population history.

The results of our study imply that population subdivision and genetic structure across most localities in the area was strongly associated with habitat features that offer resistance to dispersal at different intensities such as agricultural land, roads, railway lines, high density human settlements and urban infrastructure and not only by geographic distance between populations. Mantel tests showed significant positive correlation of genetic differentiation with least-cost corridors and landscape resistance surfaces associated in traversing corridors, further confirming this observation. The tiger habitats in the region are patchy with some populations still having connectivity, while being reduced or conspicuously absent in others. The best patches of contiguous forested habitat are present in the corridor between Pench and Kanha, which extends eastward to Achanakmar. Likewise, the Satpura and Melghat populations are also connected through swathes of degraded forests, which are interspersed with agricultural land and medium density human settlements. Connectivity between Pench and Satpura is fragile as parts of the linkage are disrupted by mining activities, and broken up in places by agriculture, habitations, major highways and railway lines. The population in Tadoba is linked with Kanha in a stepping stone connectivity through patchily distributed forests. Even though the intervening matrix between Tadoba, Pench and Melghat is heavily human dominated, the populations are tenuously linked by degraded forest patches and tiger occupied habitats such as Bor Wildlife Sanctuary. Bandhavgarh has linkages with forest habitat and tiger reserves further east [31], but seems isolated from tiger populations in the study area by human settlements and agricultural land. Corridors identified herein need to be given legal status, and mitigated with appropriate green infrastructure [148] for development projects within corridor habitats so as to ensure continued gene flow between populations.

Conservation implications

It is indeed surprising that in spite of being highly fragmented [31], the tiger habitats in Central India still exist as a metapopulation with gene flow occurring between most population clusters in contemporary times. Similar conclusions were also reached in other studies [65], [66], which observed indistinct genetic structuring and low migration amongst populations, although the genetic differences were not substantial to permit unambiguous identification of migrants. These findings suggest that tigers are able to disperse through suboptimal habitat fragments, than was earlier believed [41], [149]. The results of our study underscore the importance of conserving and maintaining corridor connectivity for the continued persistence of tiger populations in the landscape. We found seventeen individuals or about 10% of the sampled tigers with migrant ancestry. As revealed from this study, the sizes of genetic population clusters are clearly beyond the boundaries of protected areas, and have to be managed in a metapopulation framework stressing the need for a landscape conservation policy in place of the current conservation policy focused on protected areas. The functionality of corridors as shown in this study has important implications for the persistence of small populations such as Achanakmar, Satpura and Melghat by the rescue effect of emigrants from the large source populations in the landscape. Apart from poaching, the ecological isolation from other source populations, due to the absence of corridors was one of the prominent causes that led to the local extinction of tigers from the nearby protected areas of Panna and Sariska [5], [29]. Now both of these areas have been repopulated through translocation of tigers from nearby reserves. A range of anthropogenic pressures are fragmenting important habitat corridors such as the ones linking Satpura with Pench, Pench with Kanha, Kanha with Tadoba, and Tadoba with Melghat. Dispersing tigers often have to negotiate passage through suboptimal prey-poor habitats and spend extended periods of time in proximity to human habitations, which greatly reduces survival and successful gene flow. The limited migration occurring in the landscape is probably due to tigers avoiding areas of human activity and high mortality due to commercial poaching or retributive killing in response to predation on livestock and humans. The eventual loss of genetic connectivity between populations in the near future, as observed currently in Bandhavgarh, is likely for most populations in the area if present trends of fragmentation continue unabated. In such situations, last resort alternative strategies such as translocation may be considered, as it could easily enrich genetic diversity by moving tigers across each of the four distinct genetic population clusters, to mimic natural patterns of gene flow. However, maintaining and restoring habitat corridors is the preferred conservation strategy to maintain genetic exchange between tiger populations, since corridors would serve the same function for other biota as well, exemplifying the role of the tiger as a flagship species.

Conservation efforts in such a fragmented and dynamic human-dominated landscape presents enormous challenges, but should be attempted in all areas where substantial habitat is still present. Our study gives renewed hope to tiger conservation efforts within the Central Indian Landscape and similar habitats with small fragmented tiger populations. We show that minimal habitat connectivity permits gene flow between populations, which is essential to maintain metapopulation connectivity. Our findings suggest that tigers could negotiate passage through stepping stone dispersal [136] as observed between Kanha-Tadoba, and Pench-Melghat. The most functional corridor was observed between Kanha and Pench, which has evidence of prey and offers the possibility of resident tigers in some of its larger forest patches [31], [135]. However, infrastructural development in the form of adding lanes to national highways and widening of railway lines, threaten to form permanent barriers even within this corridor unless proper safeguards and mitigation measures [44] are built into these development projects. Tiger range countries like India are heavily investing in infrastructure development and mining to meet the needs of a growing economy [148]. Identification of minimal habitat corridors is vital for conservation efforts of tigers and other wide-ranging fauna like elephants (Elephas maximus), gaurs (Bos gaurus), leopards (Panthera pardus), and dholes (Cuon alpinus). They need to be offered legal protection with smart green infrastructure [148] being the norm of development policies within these corridor habitats. The extinction (nearly two decades ago) of gaur in Bandhavgarh, preceding their recent reintroduction with individuals from Kanha [150], is symptomatic of habitat fragmentation events that have impaired movement with other source populations in the landscape. We demonstrate here an integrated approach to generate reliable information to document metapopulation structure of tigers and the required habitat connectivity [17] to maintain it in Central India. Legal mechanisms to safeguard these minimal corridors could potentially be the eco-sensitive category under the Environment Protection Act (1986) legislation. Currently tigers and other mega-fauna exist as a metapopulation and do exchange genetic material through functional corridors [65], [66], [128], [151]. Such opportunities are likely to be lost rapidly in the wake of new wave of development unless legal sanctity, active restoration and mitigation of development projects become the norm. Conservation policy needs to shift the focus from protected area centered preservation to landscape scale conservation where development policies incorporate a conservation ethic.