Positive effects of fragmentation per se on the most iconic metapopulation

While habitat loss is a major threat to species, the effects of habitat fragmentation independent of habitat loss (fragmentation per se) are debated. Metapopulation studies often assert negative fragmentation effects, but they do not measure fragmentation per se. We evaluate the effects of fragmentation per se (patch density) across 20 years of patch occupancy patterns of the Åland Islands Glanville fritillary butterfly, Finland, a famous model system in metapopulation studies. Fragmentation per se had mainly positive effects on patch occupancy, the proportion of years occupied per patch, and patch colonization, and negative effects on patch extinction. These results suggest that fragmentation per se does not threaten persistence of the Åland Islands Glanville fritillary butterfly. Our results support the growing body of research challenging the paradigm that habitat fragmentation per se is mostly negative for species, highlighting the value of small patches for species conservation.


INTRODUCTION
Habitat loss is widely considered the most important factor leading to species loss (Díaz et al., 2019;Fahrig, 2003;Watling et al., 2020).However, the effects of habitat fragmentation independent of the effects of habitat loss ("fragmentation per se") are still debated (Fahrig et al., 2019;Fletcher et al., 2018).Metapopulation theory (Hanski, 1994;Levins, 1969) has had a deep impact on researchers' understanding of the effects of fragmentation on species.Metapopulation theory defines habitat fragmentation as "the size and isolation of habitat patches" (Hanski, 1989).This implies that This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.© 2024 The Authors.Conservation Letters published by Wiley Periodicals LLC.
fragmentation is inherently linked to habitat loss at a landscape scale, because a landscape containing less habitat is often characterized by smaller, more isolated patches.The metapopulation concept of fragmentation is essentially opposite to the metapopulation concept of patch connectivity, which measures the likelihood of immigration to a given focal patch from other patches in the metapopulation.The larger and the closer those other patches are to the focal patch, the higher the connectivity, the colonization rate, and the more likely the occupancy of the focal patch.Thus, in the metapopulation construct, habitat fragmentation increases, patch connectivity decreases, and patch occupancy decreases with declining patch size, increasing isolation and, indirectly, declining habitat amount in the landscape surrounding the focal patch (Hanski, 1999).For this reason, metapopulation studies generally infer that increasing fragmentation has strong negative effects on species occurrence (Bascompte & Solé, 1996;Hanski, 1998).
Many researchers in landscape ecology (De Camargo et al., 2018;DiLeo & Wagner, 2016;Fahrig, 2003;Martin, 2018) argue that it is important to determine the effects of fragmentation independent of the effects of habitat loss, that is, fragmentation per se (Fahrig, 2003).Under this perspective, fragmentation operates at a landscape scale.While habitat loss and fragmentation can occur together, they lead to different changes in habitat pattern, namely, a decrease in habitat amount and an increase in the number of patches in the landscape, respectively.Thus, habitat fragmentation effects are measured as the effects of patch density for a given amount of habitat or while controlling for total habitat amount in the landscape (Fahrig, 2017).High fragmentation per se, therefore, implies a landscape containing many small patches.If such a landscape has equivalent biodiversity value to a landscape containing few large patches but the same total habitat amount, then we can infer that conservation policies should target large total amounts of habitat, irrespective of the patch sizes.Studies that assess fragmentation per se most often find weak or positive effects on species responses (Fahrig, 2003(Fahrig, , 2017)).Such positive effects could be related to positive edge effects (Moore et al., 2011) or to a decrease in mean nearest distance among patches in the landscape with increasing fragmentation, which could increase between-patch dispersal and resource availability (Fahrig, 2017).
The Åland Islands (Finland) Glanville fritillary butterfly (ÅGFB) is a famous model system used in more than 100 papers for testing and expanding metapopulation theory over the past three decades (Hanski, 1994(Hanski, , 1998;;Ojanen et al., 2013).Surprisingly, there has not been an investigation of the independent effects of landscape-scale habitat amount and fragmentation on the ÅGFB.Such an investigation is important because of the central role of this iconic metapopulation in our understanding of fragmentation effects.Here, we capitalize on the dataset published by Schulz et al. (2020) to evaluate the independent effects of fragmentation (fragmentation per se) on the ÅGFB over 20 years using a landscape ecology perspective.We expect that fragmentation per se will increase patch occupancy, the proportion of occupied years per patch, and colonization rate, and fragmentation per se will reduce extinction rate.We also determine whether the effects of fragmentation per se are consistent across years, and whether they are sensitive to the spatial extent over which the response is measured.

Dataset description
We used the long-term, publicly available dataset of ÅGFB (Melitaea cinxia) occurrence published by Schultz et al. (2020).Over the past three decades, the Åland Islands have been intensively searched for ÅGFB habitat patches.
The number of patches discovered has grown over time, with a large increase in 1999 when a massive search effort occurred, and since then there have been only a few, small additional patches discovered per year (Ojanen et al., 2013).The dataset provides the occurrence of the species in 4656 patches (1999-2018) (3703 mean patches surveyed/year, maximum = 4357, and minimum = 2935), including the geographic coordinates of each patch, and all patch areas.

Study system
The details about the study region, species, and sampling methods are described in previous studies (Hanski, 1994;Ojanen et al., 2013).The study area consists of one main island (685 km 2 ) and several smaller islands (5-85 km 2 ; Ojanen et al., 2013; Figure 1a).The landscape is highly heterogeneous due to long-term human expansion, comprising mainly agricultural areas, managed mixed forests, rocky areas, urban areas, and roads (Ojanen et al., 2013).
The ÅGFB inhabits dry meadows and similar areas, such as pastures and roadside vegetation where one or both of its larval food plants, ribwort plantain (Plantago lanceolata), and spiked speedwell (Veronica spicata) grow (Hanski, 1994).The habitat of the ÅGFB is naturally and anthropogenically patchy.In the Åland Islands, volcanic rocks were formed over 1700 million years ago, with a mosaic of granite outcrops and areas with sedimentary deposits supporting forest and cultivated areas from land uplift since the last glacial period (Hanski, 2015a).Currently, only 1% of the area are small dry meadows often located in rocky outcrops.Habitat patches are mostly small (mean patch size 0.19 ha; minimum = 0.00063 ha; maximum = 10 ha; and median = 0.06), well-defined areas.Most patches are not occupied by the butterfly in any given year.
The ÅGFB has one generation per year.The larval nests are conspicuous in fall, allowing surveyors to identify and count them in each patch (Ojanen et al., 2013).Most individuals fly <1 km in their lifetime, but around 10% move 2-3 km (Hanski et al., 1994).The mean dispersal distance is 1 km (median 400 m) (Fountain et al., 2018) and the maximum known dispersal distance is 5.5 km (Ovaskainen et al., 2008).
F I G U R E 1 Approaches for defining "landscapes" used in this study.Locations of 4656 habitat patches of ÅGFB (Åland Islands Glanville fritillary butterfly), Finland (a).Habitat patches are defined as discrete areas containing host plants of the butterfly.In patch-landscape multi-scale and patch-landscape fixed-scale approaches (b and c), response variables (patch occupancy, and patch colonization and extinction rates) are measured per patch (the unit of replication for the response is the patch), and landscape predictors (habitat percentage and patch density) are measured in the landscape surrounding each patch, either at multiple distances (multi-scale) or within one pre-defined distance (fixed-scale).Black dots in (b and c) are habitat patches.In landscape-scale grid and metapopulation network approaches (d and e), response variables (fraction of occupied patches in each landscape averaged across all years, average colonization rates, and average extinction rates) are measured across all patches in the landscapes (grid cells or networks, i.e., the unit of replication for the response is the landscape).

Landscape predictors
We quantified landscape scale habitat amount as habitat percentage (100 × sum of the areas of all habitat patches in a landscape/landscape area) and fragmentation per se as patch density (number of patches/landscape area) using the dataset "Patch network" in Schulz et al., 2020, which contains a list of all patches with their respective areas and geographic coordinates.The effects of fragmentation per se were then the effects of patch density while controlling for the effects of habitat percentage (Fahrig, 2003(Fahrig, , 2017)).

Study design and response variables
We used four approaches for defining "landscapes": a patch-landscape multi-scale approach, a patch-landscape fixed-scale approach, a grid approach, and a metapopulation network approach (Figure 1).The multi-scale approach and the fixed-scale approach are patch-landscape approaches (McGarigal & Cushman, 2002), where the response in each patch is related to the landscape predictors in that patch's surrounding landscape.Landscape predictors (here, habitat percentage and patch density) are measured within landscapes of specified radii from each patch centroid (McGarigal & Cushman, 2002).For both approaches, response variables were patch occupancy (presence/absence) per patch per year, proportion of years a patch was occupied (number of years a patch was occupied/total years sampled), and colonization and extinction rates per patch (1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018).We measured colonization rate as the number of years the patch changed from non-occupied to occupied/total number of years the patch was unoccupied.We measured extinction rates as the number of years a patch transitioned from occupied to non-occupied/total number of years the patch was occupied.Response variables were taken from the dataset "survey data" in Schulz et al. (2020).
The patch-landscape multi-scale approach allowed us to estimate the "scale of effect," that is, the landscape size that yields the strongest species response to landscape predictors (Jackson & Fahrig, 2012).We empirically estimated the scale of effect for each predictor in each year by creating 16 buffers from 500-to 3500-m radii in 200-m increments from the geographic coordinates of each patch.We chose the largest buffer of 3500 m to match the tail of the lifetime movement range of the ÅGFB.For the fixedscale patch-landscape analysis, we selected landscapes of 2100-m radius following Jackson and Fahrig (2012), which predicts that the scale of effect of a species should be about 0.4 times the maximum dispersal distance, which for the ÅGFB is 5.5 km (Ovaskainen et al., 2008).
The grid and network approaches are landscape-scale study designs, where both the response and the landscape predictor variables are measured across each landscape (Brennan et al., 2002).In the grid approach, we defined landscapes as 5 km × 5 km (25 km 2 ) cells in a grid (Figure 1d).For the network approach, we used the semiindependent networks of habitat patches delineated by Hanski et al. (2017) (Figure 1e).We used the "pooled patch area" from Hanski et al. (2017), that is, sum of all patch areas per network to measure habitat percentage.We calculated patch density in each network by dividing the number of patches by the "area of convex hull of network," which is the size of the landscape containing the network of patches (Hanski et al., 2017).As response variables for the landscape-scale approaches, we used the fraction of occupied patches in each landscape averaged across all years, the average fraction of empty patches that were colonized, and the average fraction of occupied patches that went extinct.Averages were calculated across 1999-2018 for the grid approach (Schulz et al., 2020) and across 1999-2016 for the network approach (Hanski et al., 2017), due to data availability.

Data analyses
In the patch-landscape multi-scale approach, we performed logistic regressions (generalized linear models) to identify the scale of effect.The scale of effect was the landscape scale with the best-fit model among the 16 models (16 scales), based on the lowest Akaike's information criterion (Burnham & Anderson, 2002).We identified one scale of effect per landscape predictor and response variable.The scales of effect were quite similar, ranging from 500-to 1500-m radii (0.8-7 km 2 landscape area; Table S5).We used R 4.1.2for all analyses.
To determine the independent effects of habitat percentage and patch density, we performed multiple regressions using the landscape predictors at their predetermined scales.When the response was patch occupancy (presence/absence) in a patch-landscape study design, we used a binomial error distribution.For the other responses in the patch-landscape designs (proportion of years occupied, colonization, and extinction rates) and for all responses in the grid and network study designs (the proportion of patches occupied averaged across all years and the average colonization and extinction rates across patches in the landscape), we used a quasi-binomial distribution.We applied a natural logarithmic transformation to all predictor variables to avoid skewness and improve model fit.For patch-landscape approaches, we also included the logtransformed area of the sampled patches in the models to control for effects of sampling effort, as sampling effort increased in proportion to patch area (Ojanen et al., 2013).
We assessed goodness of fit of models using the Hosmer-Lemeshow test (Table S6) and sensitivity to outliers.Removal of outliers did not cause substantial changes in the estimates of coefficients.We quantified the spatial autocorrelation of model residuals per year using Moran's I (Dormann et al., 2007) (Tables S7 and S8) and the DHARMa package (Hartig & Hartig, 2017).Moran's I ranges between −1 (perfectly dispersed) and 1 (perfectly clustered), with values close to 0 being randomly dispersed.The maximum values of Moran's I were close to zero (0.06-0.12), suggesting little spatial autocorrelation of residuals.

RESULTS
Patch density had mainly positive effects on yearly patch occupancy (Figure 2a,b) and on the proportion of years occupied per patch by the ÅGFB in both the multi-scale and the fixed-scale approaches (Figure 3a,b; see also Tables S1 and S2).These effects were less likely to be significant in the fixed-scale approach than in the multi-scale approach.Patch density also increased the average fraction of occupied patches in the grid and the network approaches (Figure 3c,d; Tables S3 and S4) although nonsignificantly.Patch density increased patch colonization rate (Figure 3e-h; Tables S1-S4) and decreased patch extinction rate (Figure 3i-l; Tables S1-S4) in all four approaches.These results were only significant in the multi-scale and the fixed scale patch-landscape approaches (Figure 3).Regardless of statistical significance, the effects of patch density were consistently positive for population persistence, that is, occupancy and colonization increased and extinction decreased with patch density.Correlations among the predictor values are shown in Supporting Information Results S1.

DISCUSSION
Fragmentation per se had mainly positive effects on patch occupancy.For instance, we found that, for the average Effects of fragmentation per se on patch occupancy of the Åland Islands Glanville Fritillary Butterfly per year from 1999 to 2018.The effect of fragmentation per se is the estimated coefficient for the effect of patch density in the landscape surrounding the patch while controlling for the effect of habitat percentage in the landscape.The size of the landscapes was determined using a multi-scale approach (a) and a fixed-scale approach (b).Bars represent 95% confidence intervals; blue estimates are significant and grey are nonsignificant.
patch density coefficient of 0.31 (Figure 2), increasing patch density from the average of 0.08 patches/ha to 1 patch/ha increases the probability of occupancy of the ÅGFB by approximately 15%.Our results are qualitatively consistent with the finding that 67% of tests of fragmentation per se effects on single species abundance or occurrence are positive (fig.9d in Fahrig [2017]).Fragmentation per se also increased patch colonization rate and decreased patch extinction rate.This means that the ÅGFB has a higher probability of colonizing an empty patch and a lower probability of going extinct from an occupied patch in landscapes with many small patches, for a given total amount of habitat.Thus, fragmentation per se does not limit persistence of this famous metapopulation.We note that our finding of positive effects of patch density is consistent with metapopulation studies.For example, metapopulation theory predicts that the probability that all patches in a metapopulation will simultaneously go extinct decreases with the number of patches in the metapopulation (Nachman, 2001).Also, Hanski (2015b) stated "even species adapted to living in highly fragmented landscapes lose their viability when the density of habitat patches becomes very low."In addition, patch density is included in the "metapopulation capacity" metric, "metapopulation capacity [. . .] increases with an increasing number of patches, with increasing patch areas and with increasing connectivity among the patches" (Ovaskainen & Saastamoinen, 2018), and metapopulation capacity is positively related to population persistence.Nevertheless, because the total area of habitat has not been controlled for in metapopulation studies, an increase in patch density generally also implies an increase in habitat amount.In other words, metapopulation studies do not evaluate effects of fragmentation per se.
We speculate that the positive effects of fragmentation per se on the ÅGFB metapopulation are caused by a decrease in mean nearest patch distances with increasing patch density.This implies higher landscape-scale structural connectivity, a fact that is not widely recognized and may seem counterintuitive, likely because of the persistent confounding of habitat fragmentation with habitat loss.Fragmentation per se means more patches for the same total habitat area in the landscape, which reduces the mean nearest distances among patches (Figure 4).The closer the two patches are, the higher the likelihood of successful dispersal between them.This is true for the patches of ÅGFB habitat (Figure 4).Fragmentation per se also means smaller patches, which may have lower habitat quality (Ries et al., 2004).Yet, the larger number of patches can balance and even outweigh the negative effects of small patches (Riva & Fahrig, 2023a).
We emphasize that our goal was to test the effect of fragmentation per se on this iconic metapopulation.We are not suggesting that habitat amount and fragmentation per se are better predictors of patch occupancy of the ÅGFB than the metapopulation predictors whose importance has F I G U R E 3 Effects of fragmentation per se (patch density while maintaining habitat amount and patch area constant at their mean values) on the proportion of years occupied, colonization rate, extinction rate, the average fraction of occupied patches, the average colonization rate, and the average extinction rate of habitat patches by the Åland Islands Glanville Fritillary Butterfly across 20 years (1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018).Results for the multi-scale patch-landscape (a, e, and i), the fixed-scale patch-landscape (b, f, j), the grid (c, g, k), and the network (d, h, l) approaches are shown.Each point in (a), (b), (e), (f), (i), and (j) represents one of 4656 habitat patches, and its surrounding landscape, in the Åland Islands, Finland.Each point in c,g,k represents a grid, and in d,h,l, a network.Note that the number of habitat patches in colonization and extinction plots (e-l) is lower than in occupancy plots (a-d) because we did not include patches that were empty during the whole study period in the analyses (no patches were occupied during the whole period).Blue lines show the predicted estimates, and light blue areas show 95% confidence intervals.Significance codes: ***0, **0.001, and *0.01.been well demonstrated (Hanski, 1998(Hanski, , 1999;;Hanski et al., 2017;Moilanen & Hanski, 2001).Because the ÅGFB dataset is exceptionally rich, with high spatial and temporal resolution, metapopulation models of this dataset can include effects of patch occurrence and/or abundance in previous years, known movement distances, and extinction and colonization rates estimated over multiple decades (Hanski, 1999;Hanski et al., 2017).Such information is unavailable for almost all other species in most locations, where population occurrence or abundance is typically known only for a small subset of habitat patches (if any) in a single year.In such situations, predictive models of the effects of landscape structure are limited to measures such as habitat amount and patch density.Our models of the ÅGBF metapopulation using a landscape ecology perspective provide insights on both metapopulation biology and landscape ecology, bridging the independent traditions of these disciplines.

Implications for conservation
Our results challenge the idea that fragmentation per se has largely negative effects on species persistence.Indeed, the positive response of the ÅGFB to fragmentation per se adds to other work suggesting that small patches often have disproportionately high conservation value relative to their areas (Riva & Fahrig, 2022;Wintle et al., 2019), including very small patches (Arroyo-Rodríguez et al., 2022;Riva & Fahrig, 2023b).We found this same result for a species that is thought to epitomize sensitivity to habitat fragmentation, which indicates that a paradigm shift in conservation is needed.In particular, we see the need for a shift away from policies that place lower limits on the sizes of conserved areas (Riva & Fahrig, 2023b) or that emphasize "defragmenting" landscapes.Past and current lack of protection for small patches of habitat has real consequences.For instance, in Southern Ontario, Canada, most of the wetland area lost between 2002 and 2010 were wetlands smaller than 2 ha, the lower limit for wetland protection in Ontario (Birch et al., 2022).Given the ongoing pace of habitat destruction, and that habitat loss is the main threat to species, efforts to protect habitat should extend to all habitat, irrespective of patch sizes.Increasing fragmentation per se by creating new patches through habitat restoration will also increase habitat amount and dispersal opportunities to enhance persistence of species in anthropogenic landscapes.

A U T H O R C O N T R I B U T I O N S
Carmen Galán-Acedo and Lenore Fahrig developed the study idea.Carmen Galán-Acedo performed the statistical analyses.Torsti Schulz provided some data for the Network approach.Carmen Galán-Acedo wrote the first draft of the manuscript, and all authors contributed substantially to revisions.

A C K N O W L E D G E M E N T S
This work was funded by a grant from the Natural Sciences and Engineering Research Council of Canada (NSERC) to Lenore Fahrig.The authors would like to thank members of the Geomatics and Landscape Ecology Research Laboratory at Carleton University for discussions that helped motivate and improve this work.

C O N F L I C T O F I N T E R E S T S TAT E M E N T
The authors declare no conflicts of interest.

D ATA A N D C O D E AVA I L A B I L I T Y
The data and code supporting the findings of this study are stored and available in the Dryad Digital Repository at https://datadryad.org/stash/share/zJZGmbTWa1mlMU_ WZFvYexN25D2BfNyleuhfG2aWzQE, while the manuscript is currently in peer review and subject to changes.Forthcoming on Dryad at https://doi.org/10.5061/dryad.cvdncjt96.

F
Fragmentation per se reduces mean nearest distance between patches.In (a), three examples of 2100-m radius landscapes in the Åland Islands have the same total habitat area (0.4298 ha) for the Åland Glanville Fritillary Butterfly, but increasing patch densities (23, 44, and 63 patches), smaller patches (size of yellow symbols), and decreasing nearest distances between patches, from top to bottom.In (b), mean-nearest distance decreases with increasing the number of patches in 2100-m radius landscapes.Each point is a landscape surrounding one of the 4656 habitat patches of the Åland Islands Glanville Fritillary Butterfly.