Which Landscape Size Best Predicts the Influence of Forest Cover On

Home , Applied ecology

Journal of Applied Ecology 2016, 53, 440–448 doi: 10.1111/1365-2664.12590 Which landscape size best predicts the inﬂuence of forest cover on restoration success? A global meta-analysis on the scale of effect

Renato Crouzeilles1,2* and Michael Curran3

1Laboratorio de Vertebrados, Departamento de Ecologia, Universidade Federal do Rio de Janeiro, Rio de Janeiro, Rio de Janeiro CEP 21941-590, Brazil; 2International Institute for Sustainability, Rio de Janeiro, Rio de Janeiro CEP 22460-320, Brazil; and 3Swiss Federal Institute of Technology (ETH) Zurich,€ Chair of Ecological Systems Design (ESD), Institute of Environmental Engineering (IfU), Zurich€ 8093, Switzerland

Summary

1. Landscape context is a strong predictor of species persistence, abundance and distribution, yet its influence on the success of ecological restoration remains unclear. Thus, a primary question arises: which landscape size best predicts the effects of forest cover on restoration success? 2. To answer this question, we conducted a global meta-analysis for biodiversity (mammals, birds, invertebrates, herpetofauna and plants) and measures of vegetation structure (cover, density, height, biomass and litter). Response ratios were calculated for comparisons between reference (e.g. old-growth forest) and disturbed sites (degraded or restored). Using an information-theoretic approach, mean response ratio (restoration success) and response ratio variance (restoration predictability) within each study landscape were regressed against the percentage of overall (summed forest cover) and contiguous (summed pixels of ≥60% forest cover) forest within eight different buffer sizes of radius 5–200 km (at 1-km resolution). 3. We included 247 studies encompassing 196 study landscapes and 4360 quantitative comparisons. The best buffer (landscape) size varied for the following: (i) overall and contiguous forest cover, (ii) biodiversity and vegetation structure and (iii) mean response ratio and response ratio variance. Only plant biodiversity was influenced by overall forest cover (buffer size of 5, 10 and 200 km radii), while plants (10 and 200 km radii), mammals (5, 10 and 50– 200 km radii), invertebrates (5 and 10 km radii), cover (5 km radii), height (5 km radii) and litter (100 km radii) were influenced by contiguous forest cover. Overall, mean response ratio and response ratio variance were positively and negatively nonlinearly related with both overall and contiguous forest cover, respectively. 4. We reveal for the first time a clear pattern of increasing restoration success and decreasing uncertainty as contiguous forest cover increases. We also indicate preliminary recommended buffer sizes for investigating landscape restoration effects on biodiversity and vegetation structure. However, the coarse grain and variability in the data mean the optimal landscape size may not have been detected; thus, further research is needed. 5. Synthesis and applications. When setting targets for ecological restoration, policymakers and restoration practitioners should account for the following: (i) the landscape context, particularly the amount of contiguous habitat up to 10 km around a disturbed site, and (ii) the uncertainty in restoration success, as it increases when contiguous forest cover falls below about 50%. Key-words: amphibians, birds, buffer size, fragmentation, habitat loss, invertebrates, mammals, reptiles, species composition, vegetation structure

*Corresponding author. E-mail: [email protected]

Here, we assess whether, and at which landscape size, Introduction the percentage of both overall (all forest remnants ≥9 ha) Landscape context is a strong predictor of species persis- and contiguous (1-km pixels with ≥60% forest cover only) tence, abundance and distribution (e.g. Ewers & Didham forest cover is relevant to predict restoration success of 2006; Villard & Metzger 2014), but yet to understand how biodiversity and vegetation structure. We hypothesize that it affects ecological restoration remains unclear (Leite et al. the scale of effect of forest cover on restoration success is 2013). Most restoration assessments have been carried out as follows: (i) stronger for contiguous than overall forest at a local scale (<1 ha), potentially missing the effects of cover as contiguous forests are more likely to be less dis- landscape context (>1000 ha) on the success of ecological turbed by humans and consequently harbour populations restoration, that is return to a condition used as reference of dispersers and (ii) more variable and stronger at larger (hereafter restoration success) (Bowen et al. 2007). In scales for biodiversity than vegetation structure as disper- 2004, the Society of Ecological Restoration produced a sal ability may permit species to explore larger portions of document entitled ‘Primer on Ecological Restoration’, rec- a landscape. To do so, we conducted a global meta-analy- ognizing landscape context as an important variable to be sis of restoration success for biodiversity and vegetation considered in restoration assessments (Society of Ecologi- structure in forest ecosystems. We calculated an effect size cal Restoration International Science and Policy Working for each comparison between reference (less disturbed for- Group 2004), yet it is rarely assessed in practice. ests in an old-growth state) and disturbed sites (restored The surrounding forest cover has been documented as a or degraded) within the same assessment. The restored key predictor of where ecosystem recovery take holds sites represent selectively logged forests, or actively or (Leite et al. 2013). Forest areas can act both as a source passively restored forests in an initial or secondary stage of seeds for the colonization of disturbed sites and as crit- of succession, and the degraded sites represent different ical habitat for seed dispersers (Chazdon 2003; Helmer types of human land use like plantation or agriculture – et al. 2008). Contrarily, many forests around the world the starting point of the restoration process (e.g. Rey are theorized to be sink habitats, that is areas where mor- Benayas et al. 2009). We also estimated both overall and tality rates are higher than reproduction rates (Pulliam contiguous forest cover within different buffer sizes at 1998; Bowen et al. 2007); thus, one would expect local each study landscape. Our results revealed for the ﬁrst extinctions unless viable populations of dispersers remain time a clear pattern of increasing restoration success and in nearby source habitat (Brook, Sodhi & Ng 2003; Dunn decreasing uncertainty as the amount of contiguous forest 2004). A recent global meta-analysis showed a slower cover increases. development of plant and animal communities in both replanted and naturally regenerating sites that are isolated Materials and methods from source forests, compared to connected sites (Curran,

Hellweg & Beck 2014). Therefore, restoration assessments META-ANALYSIS OF RESTORATION SUCCESS should explicitly quantify the effects of landscape context on restoration success. To do this, a primary question We conducted an extensive analysis of all recorded studies in arises: which landscape size (i.e. scale of effect) is most seven key reviews on either biodiversity recovery or ecological relevant to predict the influence of landscape context on succession of vegetation structure in degraded and/or restored restoration success? forest ecosystems (Dunn 2004; Ruiz-Jaen & Aide 2005; Bowen A common procedure to identify the scale of effect is et al. 2007; Rey Benayas et al. 2009; Gibson et al. 2011; Wortley, Hero & Howes 2013; Curran, Hellweg & Beck 2014). Despite a by establishing hierarchical, multilevel buffers around common focus on degraded and/or restored forest ecosystems, sampling locations to detect species’ responses at different many of the studies included in each review were exclusive (i.e. or even multiple scales (Brennan et al. 2002; Boscolo & selected by one review only), indicating variable search and inclu- Metzger 2009). Species traits such as body size, dispersal sion criteria (Table S1, Supporting Information). We included ability and home range size can help to explain such vari- studies that were: (i) conducted in forest ecosystems, (ii) used ability (Jackson & Fahrig 2012; Stevens et al. 2014). Most multiple sampling locations (replicates for both reference and dis- studies applying multiple buffer sizes to identify the scale turbed sites) to measure biodiversity and/or vegetation structure of effect showed highest predictive power at either the lar- and (iii) included less disturbed forest in an old-growth state as a gest or the smallest scales observed, indicating the stron- comparable reference for the disturbed sites under study. In each gest predictive scale was potentially outside the range study, we characterized results for biodiversity into five taxo- included in the study (Jackson & Fahrig 2015). For exam- nomic groups (mammals, birds, invertebrates, herpetofauna and plants) and results for vegetation structure indicators into five ple, the number of species in a restored site can be poorly measures related to ecological succession (cover, density, height, estimated if forest cover is not measured at the strongest biomass and litter) (Ruiz-Jaen & Aide 2005). predictive scale. To identify the strongest scale of effect, To quantify restoration success, we used response ratios, mea- forest cover should be measured encompassing different sured as the standardized mean effect size of each comparison orders of magnitude and based on biological reasons within the same assessment (Hedges, Gurevitch & Curtis 1999). (Jackson & Fahrig 2015). The response ratio is measured as ln(x disturbed/x reference),

© 2015 The Authors. Journal of Applied Ecology © 2015 British Ecological Society, Journal of Applied Ecology, 53, 440–448 442 R. Crouzeilles & M. Curran where x is the mean value for a quantified variable within all sam- all forest cover) and (ii) forest cover defined as those pixels with a pling locations. It ranges from negative to positive values, with minimum forest cover proportion of 0Á6(contiguous forest cover) negative values implying that either biodiversity/vegetation struc- (Fig. 1). The overall forest cover map was constructed based on ture decrease in disturbed compared with reference sites, while the total subpixel proportional forest cover (0–1) within each 1-km opposite holds for positive values. Response ratio values around pixel, summed up across all pixels in the landscape. Contiguous zero are the desired outcome of restoration projects, that is success forest cover was estimated based on Percolation theory (Stauffer in bringing a quantified variable in a disturbed site back to the ref- 1985), which predicts that above 60% habitat within simulated erence state. For our statistical analysis, we inverted response artificial landscapes forest cover becomes highly or completely ratios of biodiversity and vegetation structure that are a priori connected (i.e. contiguous), independent of the landscape size. We expected to exhibit positive values (i.e. higher in disturbed than in classed contiguous forest as a binary forest/non-forest map, in reference sites). These are, for example, measures of openness, which only forest pixels (≥60% cover) were summed up. introduced species, grasses (e.g. Gibson et al. 2011). As our main The percentage of forest cover for both maps (overall and con- interest was to represent restoration outcomes for multiple species, tiguous) was estimated within each study landscape using eight community-level biodiversity data were gathered for ecological different buffer sizes: 5, 10, 25, 50, 75, 100, 150 and 200 km (as a metrics that assess multiple species (richness, diversity and similar- radius of a circle), that is ranging from 7854 to 12 566 400 ha. A ity) or population patterns in a community (abundance). To facili- recent review of comparative diversity studies in restored habitats tate a deeper understanding of biodiversity responses to (Curran, Hellweg & Beck 2014) found the median distance restoration, we analysed the response ratios of each taxonomic between sites to be 5Á3 km (M. Curran unpublished data). Thus, group separately, pooling data from different ecological metrics. we judged a 5 km radius to be a suitable lower bound for this type of analysis. Hence, we increased the range of buffer sizes up to two orders of magnitude following the recommendations of LANDSCAPE DATA Jackson & Fahrig (2015). All area calculations were conducted We mapped forested areas within each study landscape to esti- within the equal-area World Mollweide projection (EPSG: mate the percentage of forest cover. When geographic coordi- 54009). All geographic analyses were carried out using the Geo- nates were not available, we contacted the corresponding graphic Resource Analysis Support System GIS, vers. 6.4.3 authors. As restoration success was calculated for an average (GRASS Development Team 2013), and Quantum GIS, vers. number of sampling locations (replicates) per treatment, we tried 2.6.1 (QGIS Development Team 2014). to position geographic coordinates in the centre of location clus- ters from each study. Thus, we refer to the ‘study landscape’ as MODEL SELECTION the unit of analysis rather than the sampling location. Forest cover data were based on a recent 1-km resolution consensus To avoid spatial pseudo-replication (multiple samples for the land cover data set (Tuanmu & Jetz 2014), derived from combin- same study landscape), we modelled the mean response ratio (i.e. ing three existing land cover products (GLC 2000, MODIS 2005 restoration success) for each study landscape and biodiversity/ and ESA GlobCover 2008). We chose to use the ‘reduced’ data vegetation structure as a function of the percentage of overall set (Tuanmu & Jetz 2014), excluding the DISCover product from and contiguous forest cover within the eight different buffer sizes. 1995 to avoid the influence of pre-2000 deforestation. The land In addition, we modelled the response ratio variance to investi- cover data express subpixel coverage of 12 land cover types (ever- gate the predictability (i.e. uncertainty) of restoration success green/deciduous needleleaf trees, evergreen broadleaf trees, decid- across the gradient of overall and contiguous forest cover. Thus, uous broadleaf trees, mixed/other trees, shrubs, herbaceous only study landscapes with more than one response ratio for the vegetation, cultivated and managed vegetation, regularly flooded same quantified variable were used. Study landscapes could have vegetation, urban/built-up, snow/ice, barren and open water). To multiple response ratios if they were the focus of multiple studies, represent as robustly as possible the extent of forest vegetation or the same study analysed the following: (i) multiple ecological within the landscape, we included only the first three land cover metrics (abundance, richness, diversity and/or similarity), (ii) classes, excluding the ‘mixed/other trees’ that contains both forest more than one guild (e.g. frugivores and insectivores birds), (iii) and non-forest land cover types. lower taxonomic group divisions than applied in our analysis The finest resolution of the land cover products included in the (e.g. hemiptera and hymenoptera) or (iv) different years or sea- consensus data set (ESA GlobCover 2008) has a spatial resolu- sons separately. tion of 300 m, implying subpixel information can detect forest We used an information-theoretic approach (Burnham & Ander- remnants ≥9 ha. Higher resolution forest cover data sets are son 2002) to identify whether and at which buffer size (scale of available (e.g. 30-m resolution), but we considered the consensus effect) the percentage of both overall or contiguous forest cover data set more robust to temporal variation due to the need for best predicts mean response ratio and response ratio variance. We agreement between products of different time periods. In addi- fitted generalized linear models to compare a set of candidate buf- tion, this consensus data set was tested for accuracy in an ecolog- fer sizes for each taxonomic group and measures of vegetation ical modelling context (Tuanmu & Jetz 2014). The relatively structure. The mean response ratio was modelled assuming a nor- coarse resolution of our biodiversity/vegetation structure data mal distribution where the values were continuous and varying (i.e. a collection of locations representing a landscape sampling between negative and positive infinite (Bolker 2008). The response unit) meant high resolution forest cover data was not necessary, ratio variance was modelled assuming a gamma distribution where and the subpixel information of the consensus data set was suffi- the values were continuous but varying between 0 and positive infi- cient for our purposes. nite (Bolker 2008). We used the identity link function for all mod- We estimated percentage of forest cover within each study land- els. Models with normal residuals were fitted separately for each scape in two ways, considering: (i) all forest remnants ≥9ha(over- data set, totalling 40 separate analyses: five taxonomic groups and

Fig. 1. Example of the eight buffer sizes (ranging from 5 to 200 km radius) in the study landscape of Aerts et al. (2008). Overall forest cover includes a gradient of forest ranging from 0 (white) to 100% (black) and differs in the amount of forest into each 1-km pixel (ﬁne- scaled data). Contiguous forest cover includes each 1-km pixel with ≥60% forest cover (green and hachured green).

five measures of vegetation structure vs. mean response ratio and landscapes and 4360 quantitative comparisons between response ratio variance vs. overall and contiguous forest cover. In reference and disturbed sites. More than 73% of these each analysis, we compared nine models, eight containing the per- comparisons were between reference and restored sites. centage of either overall or contiguous forest cover within each These data were collected in the field from 1981 to 2009. buffer size, plus a null model containing only the intercept and Six of the seven biogeographic realms proposed by Olson error as parameters. We log-transformed the percentage of overall et al. (2001) were included in our analysis, but most stud- and contiguous forest cover because it could be nonlinearly related ies (49Á3%) were conducted in the Neotropics, encompass- with the response variables. We calculated for each model the Á Akaike Information Criterion corrected for small samples (AICc), ing 82 study landscapes. Data on biodiversity (83 2% of samples) were more common than vegetation structure. the Di as AICci À minimum AICc and the Akaike weight (wi), indicating the probability that the model i is the best model within The best buffer size varied for the following: (i) overall the set (Burnham & Anderson 2002). Models with Di < 2 were con- and contiguous forest cover, (ii) taxonomic groups and sidered equally plausible to the top-ranked model, that is the scale measures of vegetation structure and (iii) mean response of effect can range across these scales (Jackson & Fahrig 2015). We ratio and response ratio variance (Tables 1 and 2). Only 2 2 used the generalized R (for normal distribution) and adjusted R plant biodiversity was influenced by overall forest cover, (for gamma distribution) as a coefficient of determination to repre- while plants, mammals, invertebrates, cover, height and sent the goodness-of-fit of the model (Nagelkerke 1991). All analy- litter were influenced by contiguous forest cover (Tables 1 ses were carried out in R 2.12 (R Development Core Team 2011). and 2). Multilevel buffers affected the same taxonomic group, but not the same measure of vegetation structure. Results OVERALL FOREST COVER From 269 selected studies (Table S2), 247 were suitable studies for further analysis (more than one response ratio In general, for all taxonomic groups and measures of veg- for the same quantified variable), encompassing 196 study etation structure, the null model was considered among

Table 1. Performance of models predicting either mean response ratio or response ratio variance as a function of percentage of overall forest cover derived from eight different buffer sizes of radius 5–200 km. Analyses were carried out separately for each taxonomic group 2 and measure of vegetation structure. Buffer = km radius, Di = AICci À minimum AICc and wi = Akaike weight. R = generalized for mean response ratio and adjusted for response ratio variance. R2 was omitted when the null model was among the most plausible models

(Di < 2). In parentheses, the number of study landscapes = sample size. In bold are the plausible models (Di < 2).

Mean response ratio

(A) Plants (78) (B) Mammals (27) (C) Invertebrates (66) (D) Birds (57) (E) Herpetofauna (18)

2 Buffer Di wi R Buffer Di wi Buffer Di wi Buffer Di wi Buffer Di wi

Null 0Á00 0Á20 Null 0Á00 0Á25 100 0Á00 0Á22 Null 0Á00 0Á24 Null 0Á00 0Á27 50 0Á96 0Á12 75 1Á27 0Á13 150 0Á80 0Á15 75 1Á49 0Á11 25 1Á69 0Á12 25 0Á98 0Á12 100 1Á50 0Á12 75 0Á83 0Á14 50 1Á57 0Á11 50 1Á69 0Á12 75 1Á17 0Á11 50 1Á75 0Á10 200 0Á92 0Á14 25 1Á78 0Á10 52Á02 0Á10 200 1Á34 0Á10 25 2Á06 0Á09 Null 1Á60 0Á10 10 1Á89 0Á09 10 2Á02 0Á10 150 1Á56 0Á09 10 2Á29 0Á08 50 1Á82 0Á09 100 1Á92 0Á09 75 2Á10 0Á09 10 1Á56 0Á09 52Á30 0Á08 25 2Á24 0Á07 51Á98 0Á09 200 2Á56 0Á08 100 1Á61 0Á09 200 2Á44 0Á07 10 2Á89 0Á05 150 2Á16 0Á08 100 2Á76 0Á07 51Á70 0Á08 150 2Á44 0Á07 5 2Á91 0Á05 200 2Á19 0Á08 150 2Á90 0Á06 (F) Cover (35) (G) Height (24) (H) Density (34) (I) Biomass (32) (J) Litter (10)

Null 0Á00 0Á26 Null 0Á00 0Á23 Null 0Á00 0Á15 Null 0Á00 0Á26 Null 0Á00 0Á43 200 1Á46 0Á13 10 1Á43 0Á11 100 0Á10 0Á14 150 1Á80 0Á11 25 3Á50 0Á08 150 1Á66 0Á12 25 1Á46 0Á11 150 0Á14 0Á14 200 1Á80 0Á11 10 3Á50 0Á08 100 2Á13 0Á09 50 1Á49 0Á11 200 0Á41 0Á12 100 1Á86 0Á10 53Á51 0Á07 25 2Á37 0Á08 51Á60 0Á11 75 0Á78 0Á10 5 1Á97 0Á10 150 3Á58 0Á07 50 2Á38 0Á08 75 1Á63 0Á10 50 0Á90 0Á09 10 2Á20 0Á09 200 3Á60 0Á07 10 2Á39 0Á08 100 2Á00 0Á09 25 0Á98 0Á09 75 2Á25 0Á08 100 3Á72 0Á07 75 2Á40 0Á08 150 2Á43 0Á07 51Á08 0Á09 50 2Á37 0Á08 50 3Á73 0Á07 52Á40 0Á08 200 2Á59 0Á06 10 1Á13 0Á08 25 2Á38 0Á08 75 3Á82 0Á06

Response ratio variance

(M) Invertebrates (O) Herpetofauna (K) Plants (78) (L) Mammals (27) (66) (N) Birds (57) (18)

50Á00 0Á30 0Á08 Null 0Á00 0Á23 Null 0Á00 0Á26 10 0Á00 0Á15 Null 0Á00 0Á30 200 0Á60 0Á22 0Á07 150 1Á09 0Á13 5 1Á64 0Á11 5 0Á14 0Á14 75 2Á02 0Á11 10 1Á61 0Á13 0Á05 200 1Á27 0Á12 10 1Á89 0Á10 100 0Á22 0Á13 100 2Á12 0Á10 Null 2Á26 0Á10 – 100 1Á57 0Á10 200 2Á07 0Á09 150 0Á37 0Á12 52Á28 0Á10 150 2Á57 0Á08 0Á03 51Á62 0Á10 100 2Á09 0Á09 25 0Á65 0Á11 50 2Á44 0Á09 25 3Á33 0Á06 0Á02 10 1Á65 0Á10 25 2Á10 0Á09 50 0Á82 0Á10 150 2Á68 0Á08 50 3Á97 0Á04 0Á01 25 2Á08 0Á08 150 2Á13 0Á09 75 0Á85 0Á10 10 2Á76 0Á08 100 4Á24 0Á04 0Á00 50 2Á43 0Á07 75 2Á17 0Á09 200 1Á08 0Á09 25 2Á85 0Á07 75 4Á35 0Á03 0Á00 75 2Á49 0Á07 50 2Á17 0Á09 Null 1Á29 0Á08 200 2Á87 0Á07 (P) Cover (34) (Q) Height (24) (R) Density (34) (S) Biomass (31) (W) Litter (10)

Null 0Á00 0Á26 Null 0Á00 0Á17 Null 0Á00 0Á29 Null 0Á00 0Á22 Null 0Á00 0Á49 51Á23 0Á14 25 0Á60 0Á13 150 2Á27 0Á09 200 0Á14 0Á20 200 3Á62 0Á08 10 1Á82 0Á11 75 0Á60 0Á13 200 2Á34 0Á09 150 1Á27 0Á12 150 3Á79 0Á07 25 2Á15 0Á09 50 0Á61 0Á13 50 2Á38 0Á09 50 1Á74 0Á09 50 4Á14 0Á06 50 2Á36 0Á08 10 0Á98 0Á11 75 2Á39 0Á09 75 1Á82 0Á09 10 4Á15 0Á06 200 2Á40 0Á08 100 1Á33 0Á09 25 2Á41 0Á09 25 1Á87 0Á09 25 4Á22 0Á06 150 2Á40 0Á08 51Á40 0Á09 52Á41 0Á09 10 2Á34 0Á07 100 4Á25 0Á06 100 2Á40 0Á08 150 1Á43 0Á08 100 2Á41 0Á09 100 2Á42 0Á07 75 4Á26 0Á06 75 2Á41 0Á08 200 1Á61 0Á08 10 2Á41 0Á09 5 2Á45 0Á06 5 4Á26 0Á06

the most plausible models (Di < 2) to explain the mean (wi = 0Á3), with buffer sizes of 200 and 10 km radius con- response ratio (wi always ≥0Á10) (Table 1A–J) and sidered equally plausible (Di = 0Á6 and 1Á61, wi = 0Á22 and response ratio variance (wi always ≥0Á08) (Table 1L–W). 0Á13, respectively) (Table 1K). The response ratio variance Only for plants, the top-ranked model to explain response was negatively nonlinearly related with overall forest ratio variance was the buffer size of 5 km radius cover for these three plausible buffer sizes.

Table 2. Performance of models predicting either mean response ratio or response ratio variance as a function of percentage of contiguous forest cover (1-km pixels with >60% forest cover only). Analyses were carried out separately for each taxonomic group and measure 2 of vegetation structure. Buffer = km radius, Di = AICci À minimum AICc and wi = Akaike weight. R = generalized for mean response 2 ratio and adjusted for response ratio variance. R was omitted when the null model was among the most plausible models (Di < 2). In parentheses, the number of study landscapes = sample size. In bold are the plausible models (Di < 2).

Mean response ratio

(E) Herpetofauna (A) Plants (60) (B) Mammals (22) (C) Invertebrates (51) (D) Birds (51) (15)

2 2 2 2 Buffer Di wi R Buffer Di wi R Buffer Di wi R Buffer Di wi Buffer Di wi R

10 0Á00 0Á66 0Á13 5 0Á00 0Á30 0Á19 25 0Á00 0Á20 Null 0Á00 0Á25 Null 0Á00 0Á38 25 2Á82 0Á16 0Á09 10 1Á06 0Á18 0Á15 50 0Á58 0Á15 50 1Á64 0Á11 150 3Á09 0Á08 54Á13 0Á08 0Á07 100 2Á06 0Á11 0Á11 Null 0Á90 0Á13 5 1Á83 0Á10 10 3Á09 0Á08 50 6Á03 0Á03 0Á04 75 2Á42 0Á09 0Á10 51Á19 0Á11 25 1Á85 0Á10 200 3Á14 0Á08 Null 7Á21 0Á02 – 50 2Á71 0Á08 0Á08 75 1Á53 0Á09 75 1Á87 0Á10 100 3Á15 0Á08 75 7Á29 0Á02 0Á02 150 2Á75 0Á08 0Á08 10 1Á66 0Á09 100 1Á99 0Á09 50 3Á17 0Á08 100 7Á89 0Á01 0Á01 Null 3Á01 0Á07 – 100 1Á90 0Á08 150 2Á11 0Á09 25 3Á17 0Á08 150 8Á14 0Á01 0Á00 25 3Á08 0Á06 0Á07 150 2Á00 0Á07 200 2Á18 0Á08 75 3Á18 0Á08 200 8Á33 0Á01 0Á00 200 3Á75 0Á05 0Á04 200 2Á07 0Á07 10 2Á19 0Á08 5 3Á18 0Á08

(F) Cover (26) (G) Height (16) (H) Density (28) (I) Biomass (27) (J) Litter (8)

50Á00 0Á55 0Á15 Null 0Á00 0Á35 Null 0Á00 0Á22 Null 0Á00 0Á30 100 0Á00 0Á44 0Á56 Null 2Á93 0Á13 – 52Á29 0Á11 200 1Á17 0Á12 10 2Á29 0Á09 Null 2Á17 0Á15 – 10 4Á10 0Á07 0Á01 10 2Á85 0Á08 100 1Á22 0Á12 200 2Á37 0Á09 25 2Á75 0Á11 0Á38 75 4Á96 0Á05 À0Á02 25 2Á95 0Á08 51Á22 0Á12 75 2Á43 0Á09 50 3Á27 0Á09 0Á34 150 5Á00 0Á05 À0Á02 200 3Á07 0Á08 150 1Á29 0Á11 100 2Á47 0Á09 10 4Á07 0Á06 0Á27 100 5Á09 0Á04 À0Á03 75 3Á07 0Á08 75 1Á54 0Á10 50 2Á49 0Á09 75 4Á13 0Á06 0Á26 50 5Á25 0Á04 À0Á03 100 3Á07 0Á08 50 2Á11 0Á08 25 2Á51 0Á09 150 4Á40 0Á05 0Á23 200 5Á33 0Á04 À0Á03 150 3Á07 0Á08 10 2Á35 0Á07 5 2Á53 0Á08 200 5Á24 0Á03 0Á15 25 5Á48 0Á04 À0Á04 50 3Á08 0Á08 25 2Á41 0Á07 150 2Á54 0Á08 5 5Á72 0Á03 0Á1

Response ratio variance

(O) Herpetofauna (K) Plants (60) (L) Mammals (22) (M) Invertebrates (51) (N) Birds (51) (15)

200 0 0Á60Á15 100 0 0Á27 0Á34 5 0 0Á33 0Á17 Null 0 0Á21 Null 0 0Á2 150 2Á74 0Á15 0Á10 75 0Á38 0Á22 0Á32 10 1Á22 0Á18 0Á12 10 0Á53 0Á16 150 0Á29 0Á17 100 4Á37 0Á07 0Á06 150 0Á51 0Á21 0Á32 Null 2Á15 0Á11 – 21Á36 0Á11 200 0Á31 0Á17 Null 4Á99 0Á05 – 200 1Á78 0Á11 0Á25 200 2Á43 0Á10 0Á08 25 1Á43 0Á1 100 0Á76 0Á14 75 5Á57 0Á04 0Á04 50 1Á79 0Á11 0Á25 150 3Á01 0Á07 0Á05 51Á48 0Á1751Á55 0Á09 56Á41 0Á02 0Á02 Null 3Á54 0Á05 – 100 3Á44 0Á06 0Á04 150 1Á62 0Á09 50 1Á94 0Á08 10 6Á44 0Á02 0Á02 25 5Á15 0Á02 0Á07 25 3Á46 0Á06 0Á04 100 1Á92 0Á08 10 2Á22 0Á07 50 6Á67 0Á02 0Á01 10 6Á13 0Á01 0Á01 75 3Á69 0Á05 0Á03 75 2Á08 0Á07 25 2Á83 0Á05 25 7Á19 0Á02 0Á00 5 6Á24 0Á01 0Á00 50 3Á72 0Á05 0Á03 50 2Á19 0Á07 5 3Á08 0Á04

(P) Cover (25) (Q) Height (16) (R) Density (28) (S) Biomass (26) (W) Litter (8)

25 0 0Á24 5 0 0Á56 0Á32 25 0 0Á22 Null 0 0Á24 Null 0 0Á22 Null 0Á76 0Á16 Null 2Á52 0Á16 – Null 0Á72 0Á15 5 0Á98 0Á15 5 0Á31 0Á19 200 1Á61 0Á11 10 4Á12 0Á07 0Á1 100 1Á44 0Á11 10 1Á07 0Á14 25 0Á39 0Á18 10 1Á83 0Á10 200 5Á46 0Á04 0Á02 150 1Á54 0Á1 200 1Á68 0Á1100Á39 0Á18 50 1Á99 0Á09 25 5Á55 0Á04 0Á01 50 1Á55 0Á1 150 2Á06 0Á09 50 2Á19 0Á07 100 2Á07 0Á09 50 5Á55 0Á03 0Á01 10 1Á68 0Á09 50 2Á36 0Á07 100 2Á66 0Á06 75 2Á10Á08 75 5Á60Á03 0Á01 200 1Á97 0Á08 25 2Á47 0Á07 75 3Á46 0Á04 150 2Á22 0Á08 100 5Á61 0Á03 0Á01 75 2Á15 0Á07 75 2Á51 0Á07 150 4Á23 0Á03 52Á92 0Á06 150 5Á67 0Á03 0Á00 5 2Á36 0Á07 100 2Á55 0Á07 200 4Á74 0Á02

CONTIGUOUS FOREST COVER ratio (Table 2A–J). The only plausible model was the buf- = Á For plants, mammals, cover and litter, at least one buffer fer size of 10 km radius for plants (wi 0 13), 5 km = Á size was considered plausible to explain the mean response radius for cover (wi 0 55) and 100 km radius for litter

(wi = 0Á44), the last having a small sample size (n = 8) ≥0Á15) (Table 2N–P and R–W). The response ratio vari- (Table 2A, F and J, respectively). For mammals, the top- ance was always negatively nonlinearly related with con- ranked model was the buffer size of 5 km radius tiguous forest cover for plants, mammals, invertebrates

(wi = 0Á19), with the buffer size of 10 km radius consid- and height (Fig. 2B, D, F and J, respectively). ered equally plausible (Di = 1Á06, wi = 0Á15) (Table 2B). The null model was considered among the most plausible Discussion models for all other taxonomic groups and measures of Our global meta-analysis of ecological restoration of biodi- vegetation structure (wi always ≥0Á13) (Table 2C– and G– I). The mean response ratio was negatively nonlinearly versity and vegetation structure revealed the landscape size related with contiguous forest cover for plants (Fig. 2A), (scale of effect) at which the percentage of either overall or but positively nonlinearly related for mammals, cover and contiguous forest cover best explains restoration success litter (Fig. 2C, G and K, respectively). (mean response ratio) and its predictability (response ratio For plants, mammals, invertebrates and height, at least variance). The results support our main hypotheses that the one buffer size was considered plausible to explain the ‘best’ landscape size in terms of explaining restoration suc- response ratio variance (Table 2K–W). The only plausible cess and its predictability by forest cover was as follows: (i) model was the buffer size of 200 km radius for plants stronger for contiguous than overall forest cover and (ii) more variable and stronger at larger scales for biodiversity (wi = 0Á15) and 5 km radius for height (wi = 0Á6) (Table 2K and Q, respectively). For mammals, the top- than for vegetation structure. We revealed for the ﬁrst time ranked model was the buffer size of 100 km radius a clear pattern of increasing restoration success and decreasing uncertainty as the amount of contiguous forest (wi = 0Á34), but other buffer sizes were also considered equally plausible (Table 2L). These ranged from 50 to cover increases. From an applied perspective, we emphasize that contiguous forest cover is integral to effective ecologi- 200 km radius (always Di < 1Á79, wi ≥ 0Á25). For invertebrates, the top-ranked model was the buffer size of 5 km cal restoration, with increasingly unpredictable restoration outcomes associated with low contiguous forest cover. radius (wi = 0Á17), with the buffer size of 10 km radius con- We found a stronger effect of contiguous than overall sidered equally plausible (Di = 1Á22, wi = 0Á12) (Table 2M). For birds, herpetofauna, cover, density, biomass and litter, forest cover affecting restoration processes of both biodi- the null model was considered among the most plausible versity and vegetation structure (Tables 1 and 2). This ﬁnding underscores the need to preserve connected forest models to explain the response ratio variance (wi always (a)Mean RR (b) RR variance(g) Mean RR (h) RR variance 3 1 5 4 0 0 2 3 −1 2 –1 1 1 −2 0 0 –2 −3 0 25 50 75 100 0255075 0 255075100 0 25 50 75 100 (c) (d) (i) (j) 2 0 0·75 0 1·5 −1 0·5 1 −0·5 0·5 0·25 −2 0 0 −1 0 25 50 75 100 25 50 750 255075100 0 25 50 75 100 (e) (f) (k) (l) 1·5 1 2 0 1 1 0 −1 0 0·5

−1 −1 −2 0 0 255075100 0 25 50 75 100 25 50 75 25 50 75 Forest cover (%)

Fig. 2. Relationship between either mean response ratio (RR) or response ratio variance and percentage of contiguous forest cover for both biodiversity and vegetation structures that were affected by at least one buffer size (landscape size in km radius). (a, b) Plants (10), (c, d) mammals (5), (e, f) invertebrates (5), (g, h) cover (5), (i, j) height (5) and (k, l) litter (100). Points represent either mean or variance of all response ratios at each study landscape. Blue line = mean value and grey = 95% conﬁdence intervals.

© 2015 The Authors. Journal of Applied Ecology © 2015 British Ecological Society, Journal of Applied Ecology, 53, 440–448 Meta-analysis on the scale of effect for restoration 447 habitat as a prerequisite to the restoration of diverse that the scales of effect overlapped (e.g. Tables 1A and 2F ecosystems (e.g. Crouzeilles et al. 2015). Previous work and Q). For litter, we found a large scale of effect (100 km has demonstrated how landscapes with low old-growth radius, Table 2J), but our sample size was small, resulting (source) forest cover exhibit impaired dispersal and colo- in poorly fitted models. Our findings suggest that the opti- nization of plants and seed dispersers (Vellend 2003), and mal scale may not have been detected for both biodiversity higher susceptibility to anthropogenic disturbance (Holl & and vegetation structure (Jackson & Fahrig 2015). The Aide 2011; Markl et al. 2012). This could help to explain scale of effect that we detected is highly dependent on the our results, although it cannot be confirmed as our defini- coarse spatial resolution of the land cover data and on the tion of contiguous forest is not sufficient to differentiate detail of the biodiversity and vegetation structure data. source from sink habitat. In any case, an increase forest Thus, restoration projects with higher resolution data contiguity (i.e. connectivity) is also necessary in landscape should follow the similar methodological procedure to that restoration actions to maximize restoration success for performed here to assess the optimal spatial scale of effect both biodiversity and vegetation structure. for landscape context analyses. We revealed a clear pattern of increasing restoration success and decreasing uncertainty across the gradient of con- CONCLUSIONS AND PRACTICAL IMPLICATIONS tiguous forest cover (Fig. 2). Previous work has shown immigration rates to be high in well-forested landscapes, Restoration initiatives are expanding globally. Overall, we facilitating population persistence/recovery and seed disper- show that ecological restoration is affected by the landscape sal, even in small areas (e.g. Pardini et al. 2010; Leite et al. context, particularly the amount of contiguous habitat up to 2013; Crouzeilles et al. 2014). Contrarily, at lower forest 10 km around a disturbed site. Our results revealed an cover, immigration rates are eroded by increased patch iso- increasing uncertainty in restoration success when contigu- lation and even large populations are subject to local extinc- ous forest cover falls below about 50% (Fig. 2). Policymak- tion (Pardini et al. 2010; Holl & Aide 2011; Leite et al. ers and restoration practitioners should account for this 2013). While almost all of our results are consistent with this when setting targets and accounting for uncertainty. For explanation, for plants we observed the opposite trend: example, halting and reversing deforestation early on will restoration success (but not its predictability) decreased as facilitate recovery and reduce the risk of irreversible biodiver- forest cover increased (Fig. 2A). We attribute this to the sity decline (e.g. Pardini et al. 2010). Ecological compensa- mixing of early and late successional species and non-native tion policies (‘biodiversity offsets’) permit the loss of existing species in highly deforested landscapes (Holl & Aide 2011), habitat in exchange for the restoration of habitat elsewhere. potentially increasing species richness and abundance (two As forest cover declines, the requirements for compensation metrics making up 77Á8% of our biodiversity data set). should be appropriately adjusted upwards to account for Identifying the scale at which landscape context affects lower success and predictability (e.g. Moilanen et al. 2009). biodiversity is a difficult task in empirical studies since a Finally, it is worth highlighting that we have only inves- species may respond to influences at multiple scales (Bowen tigated the influence of one factor (forest cover) on et al. 2007; Boscolo & Metzger 2009; Jackson & Fahrig restoration outcomes. The specific local management con- 2015). We found more than one landscape size explaining text (e.g. restoration technique), characteristics of system either restoration success or its predictability for biodiver- under study (e.g. habitat type, climate), along with habitat sity. Mammals and plants responded strongly to restora- configuration (e.g. patch size and isolation), are likely to tion at smaller scales (5 and/or 10 km radii; Table 2A and play a significant role. Nonetheless, our meta-analysis B), but predictability showed an influence ranging up to a provides globally relevant preliminary recommended land- 200 km radius (Table 2K and L). For larger bodied organ- scape sizes for taxonomic groups and measures of vegeta- isms (e.g. Stevens et al. 2014) and high-range dispersing tion structure, derived from data spanning many realms, plant species (e.g. Kirmer et al. 2008; Kepfer-Rojas et al. habitat types, degradation states, restoration ages and 2014), rare long-dispersal events could reach many kilome- habitat configurations. Our research thus offers a starting tres within time frames relevant for restoration. The over- point for addressing policy issues and mapping out future lapping scales of effect support a role for mammals as the research efforts on landscape restoration. main dispersal agent for zoochoric plants in regenerating areas (e.g. Wright 2003). For instance, it is recognized that Acknowledgements slow recovery of animal communities can hinder the recov- We thank Marcus V. Vieira for suggestions in model selection and Mariana ery of plant communities (Wunderle 1997). S. Ferreira for data collection. Support was provided by CAPES/FAPERJ/ A recent review of landscape restoration showed that the PAPD (R.C.) and the ETH Zurich,€ Research Grant CH1-0308-3 (M.C.). majority of studies demonstrate effects of landscape context on vegetation structure (Leite et al. 2013), but the mecha- Data accessibility nisms by which this occurs was not evaluated. Similarly, we found effects of contiguous forest cover on vegetation cover Mean response ratio, response ratio variance, forest cover data and R and height at the smallest scale (5 km). One obvious mech- scripts are available from Dryad Digital Repository doi: 10.5061/ dryad.v1r34 (Crouzeilles & Curran 2015). anism is an indirect influence of plant biodiversity, given

References Markl, J.S., Schleuning, M., Forget, P.M., Jordano, P., Lambert, J.E., Traveset, A., Wright, S.J. & Bohning-Gaese,€ K. (2012) Meta-analysis of Aerts, R., Lerouge, F., November, E., Lens, L., Hermy, M. & Muys, B. the effects of human disturbance on seed dispersal by animals. Conser- (2008) Land rehabilitation and the conservation of birds in a degraded vation Biology, 26, 1072–1081. Afromontane landscape in northern Ethiopia. Biodiversity and Conserva- Moilanen, A., van Teeffelen, A.J.A., Ben-Haim, Y. & Ferrier, S. (2009) tion, 17,53–69. How much compensation is enough? A framework for incorporating Bolker, B.M. (2008) Ecological Models and Data in R. Princeton Univer- uncertainty and time discounting when calculating offset ratios for sity Press, NJ. impacted habitat. Restoration Ecology, 17, 470–478. Boscolo, D. & Metzger, J.P. (2009) Is bird incidence in Atlantic forest Nagelkerke, N.J.D. (1991) A note on a general definition of the coefficient fragments influenced by landscape patterns at multiple scales? Landscape of determination. Biometrika, 78, 691–692. Ecology, 24, 907–918. Olson, D.M., Dinerstein, E., Wikramanayake, E.D., Burgess, N.D., Pow- Bowen, M.E., McAlpine, C.A., House, A.P.N. & Smith, G.C. (2007) ell, G.V.N., Underwood, E.C. et al. (2001) Terrestrial ecoregions of the Regrowth forests on abandoned agricultural land: a review of their habitat world: a new map of life on earth. BioScience, 51, 933. values for recovering forest fauna. Biological Conservation, 140,273–296. Pardini, R., Bueno, A.A., Gardner, T.A., Prado, P.I. & Metzger, J.P. Brennan, J.M., Bender, D.J., Contreras, T.A. & Fahrig, L. (2002) Focal (2010) Beyond the fragmentation threshold hypothesis: regime shifts in patch landscape studies for wildlife management: optimizing sampling biodiversity across fragmented landscapes. PLoS ONE, 5, e13666. effort across scales. Integrating Landscape Ecology into Natural Resource Pulliam, H.R. (1998) Sources, sinks, and population regulation. The Amer- Management, 1st edn (eds J. Liu & W.W. Taylor), pp. 68–91. Cam- ican Naturalist, 132, 652–661. bridge University Press, Cambridge. QGIS Development Team (2014) Quantum GIS. Vers. 2.6.1. Available at: Brook, B.W., Sodhi, N.S. & Ng, P.K.L. (2003) Catastrophic extinctions http://www.qgis.org. follow deforestation in Singapore. Nature, 424, 420–426. R Development Core Team (2011) R: A Language and Environment for Burnham, P.B. & Anderson, D.R. (2002) Model Selection and Multimodel Infer- Statistical Computing. R Foundation for Statistical Computing, Viena, ence: A Practical Information-Theoretic Approach. Springer, New York, NY. Austria. Available at: http://www.R-project.org. Chazdon, R.L. (2003) Tropical forest recovery: legacies of human impact Rey Benayas, J.M.R., Newton, A.C., Diaz, A. & Bullock, J.M. (2009) and natural disturbances. Perspectives in Plant Ecology, Evolution and Enhancement of biodiversity and ecosystem services by ecological Systematics, 6,51–71. restoration: a meta-analysis. Science, 325, 1121–1124. Crouzeilles, R. & Curran, M. (2015) Data from: which landscape size best Ruiz-Jaen, M.C. & Aide, T.M. (2005) Restoration success: how is it being predicts the influence of forest cover on restoration success? – A global measured? Restoration Ecology, 13, 569–577. meta-analysis on the scale of effect. Dryad Digital Repository, http:// Society of Ecological Restoration International Science and Policy Work- dx.doi.org/10.5061/dryad.v1r34. ing Group (2004) The SER International Primer on Ecological Restora- Crouzeilles, R., Prevedello, J.A., Figueiredo, M.S.L., Lorini, M.L. & tion. Society for Ecological Restoration International, Tucson, AZ. Grelle, C.E.V. (2014) The effects of the number, size and isolation of Stauffer, D. (1985) Introduction to Percolation Theory. Taylor and Francis, patches along a gradient of native vegetation cover: how can we incre- London. ment habitat availability? Landscape Ecology, 29, 479–489. Stevens, V.M., Whitmee, S., Le Galliard, J.-F., Clobert, J., Bohning-€ Crouzeilles, R., Beyer, H.L., Mills, M., Grelle, C.E.V. & Possingham, H.P. Gaese, K., Bonte, D. et al. (2014) A comparative analysis of dispersal (2015) Incorporating habitat availability into systematic planning for syndromes in terrestrial and semi-terrestrial animals. Ecology Letters, restoration: a species-specific approach for Atlantic Forest mammals. 17, 1039–1052. Diversity and Distributions, 21, 1027–1037. Tuanmu, M.N. & Jetz, W. (2014) A global 1-km consensus land-cover Curran, M., Hellweg, S. & Beck, J. (2014) Is there any empirical support product for biodiversity and ecosystem modelling. Global Ecology and for biodiversity offset policy? Ecological Applications, 24, 617–632. Biogeography, 23, 1031–1045. Dunn, R.R. (2004) Recovery of faunal communities during tropical forest Vellend, M. (2003) Habitat loss inhibits recovery of plant diversity as for- regeneration. Conservation Biology, 18, 302–309. ests regrow. Ecology, 84, 1158–1164. Ewers, R.M. & Didham, R.K. (2006) Confounding factors in the detection of Villard, M.-A. & Metzger, J.P. (2014) Beyond the fragmentation debate: a species responses to habitat fragmentation. Biological Reviews, 81,117–142. conceptual model to predict when habitat configuration really matters. Gibson, L., Lee, T.M., Koh, L.P., Brook, B.W., Gardner, T.A., Barlow, J. Journal of Applied Ecology, 51, 309–318. et al. (2011) Primary forests are irreplaceable for sustaining tropical bio- Wortley, L., Hero, J.M. & Howes, M. (2013) Evaluating ecological restora- diversity. Nature, 478, 378–381. tion success: a review of the literature. Restoration Ecology, 21, 537–543. GRASS Development Team (2013) Geographic Resources Analysis Support Wright, S.J. (2003) The myriad consequences of hunting for vertebrates System (GRASS) Software. Ver. 6.4.3. Available at: http://grass.osgeo.org. and plants in tropical forests. Perspectives in Plant Ecology, Evolution Hedges, L.V., Gurevitch, J. & Curtis, P.S. (1999) The meta-analysis of and Systematics, 6,73–86. response ratios in experimental ecology. Ecology, 80, 1150–1156. Wunderle, J.M.J. (1997) The role of animal seed dispersal in accelerating Helmer, E.H., Brandeis, T.J., Lugo, A.E. & Kennaway, T. (2008) Factors native forest regeneration on degraded tropical lands. Forest Ecology influencing spatial pattern in tropical forest clearance and stand age: and Management, 99, 223–235. implications for carbon storage and species diversity. Journal of Geo- physical Research, 113, G02S04. Received 11 July 2015; accepted 1 December 2015 Holl, K.D. & Aide, T.M. (2011) When and where to actively restore Handling Editor: Yann Clough ecosystems? Forest Ecology and Management, 261, 1558–1563. Jackson, H.B. & Fahrig, L. (2012) What size is a biologically relevant landscape? Landscape Ecology, 27, 929–941. Supporting Information Jackson, H.B. & Fahrig, L. (2015) Are ecologists conducting research at the – optimal scale? Global Ecology and Biogeography, 24,52 63. Additional Supporting Information may be found in the online version Kepfer-Rojas, S., Schmidt, I.K., Ransijn, J., Riis-Nielsen, T. & Verhegen, of this article. K. (2014) Distance to seed sources and land-use history affect forest development over a long-term heathland to forest succession. Journal of Vegetation Science, 25, 1493–1503. Table S1. The seven key reviews we used as a search criteria are Kirmer, A., Tischew, S., Ozinga, W.A., von Lampe, M., Baasch, A. & van described in detail. Groenendael, J.M. (2008) Importance of regional species pools and func- tional traits in colonization processes: predicting re-colonization after large- scale destruction of ecosystems. Journal of Applied Ecology, 45,1523–1530. Table S2. Reference list for the 269 studies we selected. Leite, M.S., Tambosi, L.R., Romitelli, I. & Metzger, J.P. (2013) Land- scape ecology perspective in restoration projects for biodiversity conservation. Natureza & Conservacßao~ , 11, 108–118.