The population size of Barau’s Petrel is estimated between 10,000 and 30,000 breeding pairs (Virion et al. 2020). The first breeding colony was discovered in 1995 (Probst 1995; Probst et al. 2000). Several breeding sites have been discovered since then and two of them are monitored annually: Bras des Etangs (west side of Piton des Neiges, 2,400 m a.s.l.) and Grand Bénare (2,600 m a.s.l.). The breeding habitat is characterised by steep cliffs between 2,200 and 2,800 m a.s.l. mainly covered by endemic shrubs, such as Erica reunionensis, Stoebe passerinoides and Sophora denudata (Cadet 1977; Probst et al. 2000; Strasberg et al. 2005). Barau’s Petrels breed seasonally between September and April. They completely leave their colonies after breeding. Feral cats have been reported at all known colonies. Faulquier et al. (2009) showed that feral cats established at Barau’s Petrel colonies prey intensively upon adults, inducing a strong negative impact on populations, as this long-lived species is particularly sensitive to any additive mortality on adults (Russell et al. 2009; Dumont et al. 2010). Rats and mice are also present at Barau’s Petrel colonies, at low density, probably because these altitudinal habitats are suboptimal for these species (authors’ unpubl. data).

The population size of Mascarene Petrel is estimated at 250 breeding pairs (Attie et al. 1997). Two breeding colonies were discovered in 2016 and 2017 (Virion et al. 2020; Juhasz et al., in press) and have been monitored annually since. Burrows are located on steep cliffs from 650 to 1,200 m a.s.l. where the habitat is dominated by shrubs, such as the indigenous Olea lancea and the endemic Monimia rotundifolia (Juhasz et al., in press). The Mascarene Petrel breeds seasonally between August and March. They completely leave their colonies after breeding. Preliminary studies conducted at the breeding colonies suggest that predation by exotic mammals (rats and cats) and habitat loss constitute the main threats for this critically-endangered seabird species (Juhasz et al. in press; authors’ unpubl. data).

We sampled feral cats from July 2015 to December 2016 at six sites. Five of these sites are included in the National Park of Réunion Island and located in native mountain forests (Fig. 1A, B). The site of Cilaos is in the southernmost massive caldera of the Piton des Neiges volcano (called “Cirque de Cilaos”). The site of Les Makes is located 9 km to the southwest of Cilaos. Dimitile and Grand Bassin are located respectively at 6 and 8 km to the southeast of Cilaos. These four sites are located between 1,200 and 1,400 m a.s.l. The fifth site, named Maïdo, is a volcanic plateau sloping to the west and located 9 km to the northwest of Cilaos. Maïdo lies between 1,500 and 2,850 m a.s.l. and is characterised by subalpine shrubland. The sixth study site, Grande Anse, is a coastal peri-urban area located 20 km to the south of all other sites at 0 to 110 m a.s.l. (Fig. 1B). All sampled sites are located at less than 10 km from a breeding colony of Barau’s Petrel or Mascarene Petrel (Fig. 1B) and four of them (Cilaos, Dimitile, Grand Bassin and Maïdo) are located less than 3 km from the nearest petrel colony.

Figure 1. 

Maps illustrating A the locality of Réunion Island B genetic sampling sites, in 2015–2016 and C camera trapping sites in 2016. Each colour of dots corresponds to a different geographical area. Area codes and sample sizes are indicated in parentheses. The grey area corresponds to the National Park and the orange areas correspond to the presence of Barau’s and Mascarene Petrels. Triangles correspond to the localities of camera traps on trails (yellow triangles) and under vegetation cover (black triangles). The black lines show the trails.

Cats were live-trapped with Tomahawk cage traps baited with sardines and brought to the veterinary clinic for sanitary inspection. The veterinarian checked for individual pit-tags and tattoos to identify potential owned cats. The behavioural profile of the cat was then evaluated to estimate the possibility to adopt it. If adoption was impossible because the cat was too wild, the cat was euthanised after the legal guard period of four days. The euthanasia was made by intra-venial injection of pentobarbital. Kidney tissue samples were collected from each euthanised cat and stored in 70% ethanol at - 80 °C until laboratory analysis. The protocol was approved by the CYROI institutional ethical committee, certified by the French Ministry of Higher Education and Research (NoAPAFIS#6916-20151 00213267087 v.6). A total of 158 feral cats were trapped including 87 males, 67 females and 4 indeterminate (Table 1). None of them was identified as an owned or adoptable cat.

Total DNA was extracted from a small piece of kidney tissue using the QIAmp Blood and Tissue kit (Qiagen, Hilden, Germany). Genotyping was conducted for 10 polymorphic microsatellite loci (described in Menotti-Raymond and O’Brien 1995 for Fca43 and Fca96, Menotti-Raymond et al. 1999 for Fca31, Fca69, Fca76, Fca173, Fca275, Fca441 and Fca531 and Menotti-Raymond et al. 2003 for Fca1027) on DNA extracts from 158 individuals. A 3-primer PCR approach, using a M13 tail for the forward primer, was used for microsatellite loci amplification following Schuelke (2000). Four different dyes (6-FAM, PET, VIC, NED) were used for the universal M13 forward primer to enable fragment analysis multiplexing. Simple PCR amplifications were performed using a GeneAmp PCR System 9700 (Applied Biosystems, Waltham, Massachusetts, USA) in 10 μl reaction volume containing 5 μl of MasterMix Applied 2× (Applied Biosystems, Waltham, Massachusetts, USA), 0.3 μl of the forward primer with M13 5’-tail (1 μM), 0.3 μl of the reverse primer (10 μM), 0.3 μl of dyes (10 μM), 2.1 μl of sterile deionised water and 2 μl of genomic DNA (20–40 ng/μl). PCR amplifications were carried out under the following conditions: an initial denaturing step at 95 °C for 5 min, followed by 40 cycles of 95 °C for 30 sec, 56 °C for 30 sec and 72 °C for 30 sec and a final elongation at 72 °C for 20 min. Up to four different simplex PCR plates, each with a different dye, were mixed and PCR product sizes were determined, using a 3730XL DNA analyser (Applied Biosystems, Waltham, Massachusetts, USA) at the Gentyane platform (Clermont-Ferrand, France) and were sized with LIZ(500) standard using GeneMapper (Applied Biosystems, Waltham, Massachusetts, USA).

Evidence of null alleles, large-allele dropout and stutter peaks in all microsatellites was examined using MicroChecker 2.2.3 (Van Oosterhout et al. 2004). Each locus-pair combination was tested for linkage disequilibrium with GenePop 4.7.5 (Rousset 2008). The P-values were corrected using the Benjamini and Yekutieli (2001) method for multiple comparisons (Narum 2006). The mean observed number of alleles per locus (AL) and the number of private alleles per area (AP) were computed using GenAlEx 6.5 (Peakall and Smouse 2012). Allelic richness (AR; El Mousadik and Petit 1996), adjusted for discrepancies in sample size by incorporating a rarefaction method as implemented using FSTAT 2.9.3 (Goudet 2001), was used to make comparisons of the mean number of alleles amongst areas. The means of allelic richness amongst areas were compared using pairwise Wilcoxon’s signed rank tests with Bonferroni correction. Observed heterozygosity (H_O), unbiased expected heterozygosity estimated according to Nei (1978) (H_E) and Wright’s F-statistics (F_IS) according to the method of Weir and Cockerham (1984) were calculated for all and each population using GenAlEx 6.5 (Peakall and Smouse 2012). Deviations from Hardy-Weinberg equilibrium (HWE) were tested for each of the six areas using the package pegas 0.13 (Paradis 2010) using R 3.2.0 (R Core Team 2021), with the exact test based on 10³ Monte Carlo permutations.

Assignment tests, based on multi-locus microsatellite genotypes, were evaluated using two different clustering approaches. First, we used a Bayesian genotype clustering procedure in STRUCTURE 2.3.3 (Pritchard et al. 2000). The admixture model was used with the LOCPRIOR setting, which considers sample location and allows structure to be detected when genetic structure is weak or when the number of loci is small (< 20; Hubisz et al. 2009). The r-index was also used to determine the relevance of the sampling location (LOCPRIOR), with low values of r indicating that sampling locations are informative to the overall model (Falush et al. 2003). Correlated allele frequencies were assumed (Pritchard et al. 2000). For each value (1–10) of the number of independent genetic clusters (K), we ran 10⁶ iterations 10 times (after a burn-in of 10⁵ steps). For choosing the optimal number of clusters, two criteria were used; the log likelihood given K (L(K); Pritchard et al. 2000) and the second-order rate of change of mean log-likelihood (∆K; Evanno et al. 2005). Both criteria were calculated using STRUCTURE HARVESTER online Web server (Earl and vonHoldt 2012). CLUMPAK software (Kopelman et al. 2015) was used to find the optimal individual alignments of replicated cluster analyses and to visualise the results.

Population structure was also explored by integrating spatial coordinates of samples using a Bayesian model executed in a Markov Chain Monte Carlo, as implemented in the R package Geneland 4.0.5 (Guillot et al. 2005, 2008). The geographical information was used to detect spatial delineation of genetic discontinuities, where the number of area units is treated as an unknown parameter. We ran the MCMC ten times independently to verify the consistency of the results. We used K from 1 to 10, 10⁵ iterations with 100 burn-in generations, an uncertainty attached to spatial coordinates fixed to 200 m (i.e. the precision of our sample locations) and the maximum number of nuclei in the Poisson–Voronoi tessellation fixed to 300. The analysis was run with correlated allele frequency models, true spatial and no null allele models. Finally, all runs were examined for consistency.

Genetic differentiation amongst all pairs of areas was assessed by calculating pairwise F_ST values following Meirmans (2006) and pairwise Nei’s G’_ST distances (Nei 1978). Statistical significance was tested by 10⁴ permutations of genotypes amongst areas under Bonferroni’s correction, using GenoDive 3.04 (Meirmans and Van Tienderen 2004).

Pairwise genetic and geographic distances amongst sampling locations were used to test patterns of isolation by distance (IBD) using a Mantel test (Mantel 1967). We used the scaleGen function in adegenet 2.1.3 package (Jombart 2008) to calculate the Euclidean genetic distances amongst samples. Euclidian geographical distances between each pair of samples were calculated. The significance of the correlation coefficient between sample pairs was estimated using a Mantel test with 10,000 randomisations in R. In addition, we repeated IBD analyses using only the subset of natural areas to investigate the effect of geographic distance of Grande Anse peri-urban area from other areas, which might disproportionately contribute to IBD patterns.

To determine possible source populations that could be targeted for control (Rollins et al. 2006), we estimated recent migration rates amongst areas using two methods. First, we used Bayesian assignment tests with BIMr 1.0 (Bayesian Inference of Migration rates, Faubet and Gaggiotti 2008). BIMr infers the proportion of recent immigrants in a population from their genotypes and calculates corresponding asymmetrical migration rates amongst pairs of populations. BIMr determines recent migration even amongst weakly-differentiated populations (i.e. F_ST > 0.01) with unequal sample sizes (Faubet and Gaggiotti 2008). Due to overlapping generations in feral cats, BIMr estimates were interpreted as a relative index of recent gene flow rather than a precise estimate of gene flow in the previous generation. A burn-in period of 20,000 iterations followed by 10⁵ iterations for each run was used. The default values were used for all other parameter settings. Migration rate estimates were obtained by choosing the run with lowest Bayesian deviance, measured by the assignment values (Dassign; Faubet and Gaggiotti 2008). Posterior mean and mode migration rates and 95% high density predictive interval (HDPI) were estimated using the package HDInterval (Kruschke 2011) in R.

Estimates of recent migration rates and approximate 95% confidence interval (CI) were also explored by the Bayesian approach as implemented in BayesAss 3.0.4 (Wilson and Rannala 2003). Five runs were first performed by changing the number of seeds (s = 10, 100, 500, 750 and 1000) to obtain a suitable convergence. The number of iterations was 10⁶, of which 10⁵ were burn-in and the sampling frequency was 100. Mixing parameters were 0.6 for allele frequencies, 0.9 for inbreeding coefficients and 0.5 for migration rates. The final run consisted of the same mixing parameters and 100 numbers of seeds.

CMR data of feral cats were obtained during a single season of camera trap survey at Maïdo (Fig. 1C). This site was selected for: (i) the presence of scats and direct observations of feral cats, (ii) its proximity to a monitored Barau’s Petrel colony (about 5 km) where feral cat predation is known to occur (Faulquier et al. 2009), (iii) the proximity of trails which were supposed to maximise the feral cat detection probability (Meek et al. 2014; McGregor et al. 2015) and (iv) the technical feasibility. We deployed and geo-referenced with GPS (Garmin 64 s; 5 m accuracy) 20 camera-traps (9 Scoutguard-MG882K-12mHD, 10 Bushnell Trophy Cam HD and 1 Reconyx HC600 Hyperfire) from 17 February to 25 April 2016 (68 days). Camera-traps were first placed along trails (17 Feb – 23 March), then under vegetation cover (23 March – 25 April). This study period encompassed the second half of the chick rearing and the beginning of the fledging period of Barau’s Petrels. The mean distance (± sd) between cameras was 2,114.0 ± 1,273.0 m (min = 49 m, max = 5,626 m). Neither bait nor lure was used, to maintain homogeneous detection probabilities. Devices were set using a high-sensitivity trigger to capture three images per event at rapid-fire interval (0.13 s), with no delay between trigger events, to maximise feral cat identification. During the first week, half of the capture stations were equipped with two cameras placed on the opposing side of trails to capture both flanks of passing animals. Each observed feral cat was identified, based on natural marking such as spots, stripes and ocelli on both sides when possible (Bengsen et al. 2012; McGregor et al. 2015; Lavery et al. 2020; Palmas et al. 2020). A sampling occasion lasted one day (24 h, hereafter named “trap-night”; Otis et al. 1978; Wang and Macdonald 2009). For each sampling occasion, individuals were photographed (“captured”), identified (“marked”) and “released”. All feral cats previously identified and re-sighted were considered as a recapture. A capture event was defined as all pictures of unique individuals within a 30-min time period (Di Bitetti et al. 2006). Cameras were checked every 10 days to download data from memory cards and replace batteries. No feral cats were captured for the genetic study in the CMR study area during and 6 months before the behavioural study. However, from September to December 2015, four feral cats were captured at 5.9 to 7.8 km from the nearest camera trap.

The trapping effort (in trap-nights) was calculated by adding for each camera the number of days where each camera was active over the study period. The capture efficiency (in number of capture events/100 trap-nights) was calculated by dividing the number of feral cat capture events for all cameras divided by the total trapping effort and multiplied by 100 (Palmas et al. 2020).

Since we designed a short study period (to avoid emigration, immigration or mortality) and we did not consider kitten pictures in the dataset (to avoid recruitment), we applied SECR models that require a demographically closed population. These models assume no emigration or immigration, no mortality and no recruitment during each trap session (Otis et al. 1978; Efford 2004). The matrix of spatially explicit histories of capture-recaptures was constructed for each feral cat by linking each capture of each individual with the coordinates of the camera and with the occasion. Each camera was associated with a spatial covariate (trail vs. under vegetation) to check if trap location affects the detection probability. Data analyses were performed using the SECR package (Borchers and Efford 2008; Efford et al. 2009) in R 4.0.3. First, we estimated the mean maximum distance moved (MMDM) by the individuals between captures. Then, we implemented SECR models. The trap detector type « count » was chosen for the analysis, allowing more than one detection per animal. A habitat mask was used with a buffer width of 3,000 m around each camera-trap (determined with the SECR package; Efford 2021), but excluding the deep cliffs considered as inaccessible for feral cats. This resulted in a sampling area of 60.55 km². We assumed that home ranges of feral cats were distributed following a homogeneous spatial Poisson process during the trapping period (Efford 2004; Borchers and Efford 2008; Efford et al. 2009). The half-normal detection function was selected as the most appropriate for our models. This detection function is defined by two parameters: the animal detection probability considering that the camera-trap is located at its home range centre (g0) and the movement parameter, i.e. the distance scale of the detection function (σ). Models were used to investigate the effects of camera locations (on trail vs. under vegetation) on g0 and σ. Model performances were compared using the difference in Akaike Information Criterion modified for small sample size (AICc). Each model presenting a ΔAICc < 2 was considered a competing best model. Finally, based on the estimates of the best model, we determined: (1) the site-specific population density D; in number of cats/km² and (2) the home range (HR₉₅) and core area (HR₅₀) in km² of feral cats (see Ringler et al. 2014).

No null alleles, large-allele dropout nor stutter peaks were detected for the 10 microsatellite loci. The percentage of missing data was 1.58%. Linkage disequilibrium amongst loci was detected for four of the 45 loci pairs (P < 0.05), but no significant linkage disequilibrium was observed amongst any of the loci after the Benjamini and Yekutieli (2001) correction for multiple tests, suggesting that all loci were independent. The mean allelic richness (AR), based on a minimum sample size of 15 individuals, ranged from 4.7 (Maïdo) to 6.4 (Cilaos) alleles per locus and was relatively similar amongst areas, except for the less variable Maïdo area (Table 1). All areas contained one (Maïdo) to four (Makes, Dimitile) private alleles (Table 1). Observed heterozygosity for Maïdo (H_O = 0.58) was similar to Grand Bassin, but it was lower than all other areas (Ho range 0.65–0.74; Table 1). Deviations from HWE were not significant for all areas (all Ps > 0.05). The raw microsatellite genotypes of the 158 individual feral cats are available in the supporting information (Appendix 1).

Clustering of microsatellite genotypes using STRUCTURE analysis indicated that mean values of the r-index used to determine the relevance of the sampling location in the clustering analysis was low (0.31 ± 0.14), suggesting that sampling locations are informative to the model. Analysis clearly showed that the best-supported model contained three genetic clusters (maximum value of Evanno’s likelihood at K = 3, maximum value of L(K) and minimum standard deviation of L(K) at K = 3, Appendix 2). The first genetic cluster was almost exclusively detected for samples from Maïdo (91% of the genetic pool from Maïdo samples). The second cluster was detected amongst samples from Dimitile (77% of Dimitile’s samples), Grand Bassin (38%) and Cilaos (25%). The third cluster was shared between Makes and Grand Bassin (more than 90% of samples), then Cilaos and Grand Bassin (about 70% of samples) and finally Dimitile (about 18% of samples; Fig. 2, Appendix 2).

Analysis using Geneland yielded a modal number of populations with a higher proportion of three putative genetic groups (K = 3; Appendix 3: Fig. A2F). The run with the highest average posterior density was selected. Sampled feral cats were clustered into five groups. Two inferred groups (part of Cilaos and Grande Anse) had very low posterior probabilities (Appendix 3: Fig. A2D and E, respectively) and the areas of these groups were already represented in groups with strong posterior probabilities (Appendix 3: Fig. A2A–C). The last three putative groups roughly corresponded to the areas defined using the topography of Réunion Island (Appendix 3: Fig. A2A–C); Maïdo, Dimitile, and the other areas, and corresponded to the results obtained using non-spatial analysis with STRUCTURE.

Figure 2. 

Distribution of microsatellite clusters based on Bayesian clustering analysis using STRUCTURE (pies) and Geneland (coloured areas) and map of the migratory pathways suggested by BIMr and BayesAss (black arrows, the thickness is proportional to the amount of gene flow) of the feral cats (N = 158 individuals) in Réunion Island, 2015-2016. Area codes: MAI for Maïdo, CIL for Cilaos, MAK for Makes, GB for Grand Bassin, DIM for Dimitile and GA for Grande Anse.

Pairwise F_ST values ranged between 0.011 and 0.149 with a global F_ST of 0.026 (P < 0.001). The highest values were for the comparison of Maïdo to the other areas. Nei distances showed the same pattern. For both indexes, 8 of the 10 loci showed p-values less than 0.001. Based on the two differentiation index values, three groups were distinguishable: (1) Maïdo, (2) Dimitile and (3) all other areas, suggesting an isolation of Maïdo particularly and Dimitile to a lesser extent, as previously suggested by the clustering analysis (Appendix 4).

Genetic distance amongst individuals showed no significant relationship with geographic distance either at the global scales (Mantel r-test, P = 0.189) or after excluding Grande Anse (Mantel r-test, P = 0.385).

Recent mean migration rates determined using BIMr ranged from nearly zero amongst most pairs of areas to nearly 0.05 between Grand Bassin and Dimitile (Appendix 5). Based on non-overlapping 95% HPDIs, we only recorded significant asymmetric dispersal between Dimitile and Cilaos, with clearly highest migration from Dimitile to Cilaos (Appendix 5). Globally similar results were obtained using BayesAss, suggesting asymmetric dispersal between these two areas. Moreover, all the mean values of recent migration rate were clearly higher compared to those from BIMr and six values had a confidence interval not including zero (Appendix 5), suggesting the migratory pathways presented in Fig. 2.

During the camera-trapping survey, we collected 41,905 pictures including 376 (0.9%) pictures of feral cats. All photographed feral cats were identified and included in the study. Ten individuals were identified. There was no picture capturing more than one feral cat simultaneously (see details in Table 2).

In total, we obtained 60 feral cat detections between 17 February and 25 April 2016, corresponding to 10 captures and 50 recaptures for 10 individuals. For cameras on trails, we obtained six captures and 34 recaptures of five individuals at 14 of the 20 cameras. When camera traps were placed under vegetation, eight feral cats were first detected (including four feral cats previously identified on trails) and five of them were recaptured (12 detections) at eight of the 20 cameras (Table 2; Appendix 6). The global trapping effort was 1,082 trap-nights with a capture efficiency of 5.5 detections/100 trap-nights. The mean maximum distance moved (MMDM ± se) by feral cat was 1,926 ± 589 m. We tested the effect of camera types and the linear time trend over occasions on the detection probability g0. No effect was detected.

The model with the greatest support was the null model (Table 3). This model had a maximum detection probability at each camera trap (g0) of 0.06 (95% CI [0.03; 0.09]) and a spatial scale of movement (σ) of 971 m (95% CI [791; 1,193]; Appendix 7). The estimated population density of feral cats was 0.25 feral cats/km² (95%CI [0.12; 0.47]). The mean home range was estimated at 15.0 km² (HR₉₅) with a core area of 2.5 km² (HR₅₀).

The genetic diversity of feral cats on Réunion Island is similar to that observed on cat populations recently introduced on other islands (Kerguelen, Pontier et al. 2005; Hawai’i, Hansen et al. 2007; Christmas & Cocos Island, Spencer et al. 2016). It is also similar to the diversity observed in non-insular contexts, in isolated populations with low dispersal rates (France, Say et al. 2003; Australia, Cowen et al. 2019). It is assumed that most feral cats of Réunion Island are descendants of cats introduced from France (Dreux 1990). Interestingly the genetic diversity of feral cats on Réunion Island is lower than that observed in Europe (Pierpaoli et al. 2003), which may be explained by a founding effect leading to a genetic drift as expected in such isolated contexts (Slatkin and Excoffier 2012; Bélouard et al. 2019).

Microsatellite analysis and Bayesian clustering analysis suggested significant structuring amongst studied populations. Genetic structure was strong compared to populations of Hawai’i (three sampled sites, F_ST < 0.05; Hansen et al. 2007) and the Kerguelen Archipelago (four sampled sites, F_ST ≤ 0.09; Pontier et al. 2005). We found three genetic clusters amongst which, one was observed only at our highest sample site, Maïdo. This suggests very low gene flow between this site and other lower populations. The isolation of Maïdo was also supported by a lower allelic richness compared to other areas, as expected in isolated populations (Frankham 1996; Peter and Slatkin 2015). This pattern is probably due to the very rough topography of the island. Indeed, Maïdo is separated from other sites by a vertical cliff of more than 1 km, which probably represents a geographical obstacle for feral cat dispersal.

The second and third genetic clusters were detected in the five other areas (Dimitile, Cilaos, Grand Bassin, Makes, Grande Anse, Fig. 2). Although differentiation indexes indicated an isolation of Dimitile, feral cats sampled in this area were mostly assigned to a genetic group that was also detected in the four other areas. The low F_ST and Nei distance estimates, coupled with a lack of isolation by distance, suggest that Grande Anse, Makes, Cilaos and Grand Bassin areas were weakly isolated from each other, despite large geographical distances between Grande Anse and the others (minimum of 20 km). This result suggests either or both natural and human-mediated dispersal of feral cats amongst these areas. The human-mediated dispersal hypothesis is reinforced by the lack of genetic isolation through geographical distances, which would be expected if a progressive colonisation process occurred amongst neighbouring sites (Kimura and Weiss 1964; Slatkin 1993).

Comparing home range of feral cats from the literature is challenging because of the large diversity of the methods used, ranging from GPS tracking to SECR modelling (Jones and Coman 1982; Nogales et al. 2004; Bengsen et al. 2012; McGregor et al. 2015). However, our results suggest that feral cats at Maïdo are present at low density (0.25 feral cat/km²) with large home ranges (15 km²). To our knowledge, the only cases where feral cats live in such low densities in the tropics are also in mountainous habitats (Hawai’i, Smucker et al. 2000; Goltz et al. 2008). This suggests that some bottom-up limitation due to low densities of prey are occurring in these extreme habitats, resulting in a low carrying capacity for feral cats (see Liberg et al. 2000; see Bengsen et al. 2016). Furthermore, in the case of feral cats living at seasonal seabird breeding colonies (which is the case of both petrel species), the carrying capacity of their habitats fluctuate in relation to the phenology of the petrels.

This space use strongly contrasts to that observed in a low altitude area of Réunion Island transformed by human activities. A recent study has shown that, at sea level, cat density can reach 27 ± 2 cats/km² (Choeur 2021), with an average home range of 0.12 km² (Choeur et al. 2022). This peri-urban habitat is characterised by extremely abundant food resources for cats including anthropogenic food wastes, supplemental feeding and introduced prey such as mice, rats and lizards.

The combined results of the genetic and behavioural studies of feral cats indicate that, in mountainous habitats of Réunion Island, such as Maïdo, harbouring Barau’s Petrel colonies, cats are likely to be isolated and at low density. This is favourable for long term feral cat control. The genetic isolation implies there might be a low re-colonisation rate from surrounding cat populations (e.g. Lieury et al. 2015; Millon et al. 2019; Palmas et al. 2020).

In terms of feral cat control optimisation, the CMR study produced results that can be used to design the operations. First, in order to increase the number of feral cats exposed to control devices while minimising the number of devices, we propose to use the estimated sigma to optimise the spatial arrangement of the trapping grid (Goltz et al. 2008; Bengsen et al. 2012). In our case, with such an arrangement, each trap should be set every 950 to 1,000 m. This method minimises the number of cages to deploy (which minimises the human effort) while maximising the chance of a cat encountering at least one cage in its home range. Second, we suggest deploying traps near trails as the maximum detection probability has no variation between trail and vegetation cover for this site. This grid design presents the advantage of reducing the workload and the time spent in the field by facilitating the access and maintenance of the traps. Of course, this recommendation is limited to habitats that have trails nearby. This design of device deployment can also be used after an intensive cat control to deploy camera traps for early detection of any re-invasive cats.

In addition to these recommendations, we propose trapping operations be conducted before the breeding season of Barau’s Petrels (i.e. in austral winter, July and August, which correspond to the period when food availability for cats is the lowest, because of the absence of petrels). We also recommend the implementation of an early detection protocol, based on a network of camera traps to detect any re-invasion by cats and to respond with appropriate control actions.

For other sites located at lower altitudes, the absence of genetic isolation indicates strong connectivity between feral cat populations and, thus, a high risk of re-invasion after a feral cat control. Other strategies should be adopted to prevent or limit feral cat impact: 1) permanent feral cat control at colonies and in their vicinity and 2) predator-proof fences around bird colonies (Smith et al. 2020). However, the technical feasibility and financial costs of such operations may limit their implementation on the Island.

Feral cats are known to also prey upon other introduced mammals, such as rats or mice (Faulquier et al. 2009). Thus, in theory, a control of cats may result in the release of these prey, which in turn may impact birds, the so-called “mesopredator release effect” (Courchamp et al. 1999a). However, for such a release to occur, rat or mice populations must be controlled by feral cat predation (top-down control) rather than by their resources (bottom-up control; Courchamp et al. 1999a; Russell et al. 2009; Dumont et al. 2010). In the tropical context, it has been shown that rat and mice populations are controlled mostly by their resources through rain seasonality (Russell et al. 2011), which reduces the risk of a meso-predator release in case of feral cat control (Russell et al. 2011; Ringler et al. 2015). Furthermore, for long-lived animals, like seabirds, the population growth rate is more sensitive to change in adult survival than in breeding success (Le Corre 2008; Russell et al. 2009; Dumont et al. 2010). Feral cats prey upon adults and fledglings, whereas rats prey exclusively on eggs or chicks (Faulquier et al. 2009; authors’ comm. pers.). Thus, even if rat population were released as a consequence of cat control, this would have less impact on the population growth of petrels than the impact of cats. Thus, we recommend to implement cat control at petrel colonies wherever possible.

Another more general recommendation would be to improve the public awareness and sensitisation at the scale of the entire island to stop human-mediated displacement of cats, to stop abandonment of domestic cats or kitten in the wild and to sterilise as many domestic cats as possible (Dias et al. 2017; Russell et al. 2018; Choeur et al. 2022).