bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

1 The erosion of biodiversity and biomass in the

2 biodiversity hotspot

3 Renato A. F. Lima1,2*, Alexandre A. Oliveira1, Gregory R. Pitta1, André L. de Gasper3,

4 Alexander C. Vibrans4, Jérôme Chave5, Hans ter Steege2,6 & Paulo I. Prado1

5

6 1 Departamento de Ecologia, Instituto de Biociências, Universidade de São Paulo. Rua do

7 Matão, trav. 14, 321, 05508-090, São Paulo, .

8 2 Naturalis Biodiversity Center, Darwinweg 2, 2333 CR Leiden, The Netherlands.

9 3 Departamento de Ciências Naturais, Universidade Regional de Blumenau. Rua Antônio

10 da Veiga, 140, 89030-903, Blumenau, Brazil.

11 4 Departamento de Engenharia Florestal, Universidade Regional de Blumenau. Rua São

12 Paulo, 3250, 89030-000, Blumenau, Brazil.

13 5 Laboratoire Evolution et Diversité Biologique, UMR 5174 CNRS, Université Paul

14 Sabatier, IRD. 118, route de Narbonne, 31062, Toulouse, France.

15 6 Systems Ecology, Vrije Universiteit, De Boelelaan 1087, Amsterdam, 1081 HV,

16 Netherlands

17 *e-mail: [email protected]

1

bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

18 Abstract

19 Tropical forests are being deforested worldwide, and the remaining fragments are

20 suffering from biomass and biodiversity erosion. Quantifying this erosion is challenging

21 because ground data on tropical biodiversity and biomass are often sparse. Here, we use

22 an unprecedented dataset of 1,819 field surveys covering the entire Atlantic Forest

23 biodiversity hotspot. We show that 83−85% of the surveys presented losses in forest

24 biomass and tree species richness, functional traits and conservation value. On average,

25 forest fragments had 25−32% less biomass, 23−31% fewer species, and 33%, 36% and

26 42% fewer individuals of late-successional, large-seeded and endemic species,

27 respectively. Biodiversity and biomass erosion were both lower inside strictly protected

28 conservation units, particularly in large ones. We estimate that biomass erosion across the

29 Atlantic Forest remnants was equivalent to the loss of 55−70 thousand km2 of forests or

30 US$2.3−2.6 billion in carbon credits. These figures have direct implications on

31 mechanisms of climate change mitigation.

32

33 Introduction

34 Tropical forests are major stocks of biodiversity and carbon; and these stocks are

35 declining worldwide. Half of their original cover has already vanished and current

36 deforestation rates are about 1% per year1. Human impacts on tropical forests, however,

37 are not restricted to deforestation. Beyond the reduction in habitat availability and

38 connectivity, deforestation triggers a myriad of modifications that can penetrate up to 1.5

39 km into the remaining fragments2–4. In addition, forest fragments are more accessible,

40 increasing their exposure to fire, selective logging, hunting, and biological invasions. bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

41 These human-induced impacts on forest fragments (i.e. forest degradation) impose a

42 long-lasting burden on forest biodiversity and biomass stocks5–12 that can be as severe as

43 deforestation13,14. Protected areas can mitigate the erosion of biodiversity and biomass15–

44 17, but their effectiveness is contingent on the type of management and level of

45 anthropogenic pressure surrounding the protected areas15–17.

46 Forest degradation can be assessed by high-resolution remote sensing (e.g.

47 LiDAR18), but the coverage of this approach is limited and the impact on biodiversity

48 cannot be measured. This is why large-scale quantifications of the impacts of forest

49 degradation are mostly available for biomass10–12. Therefore, field surveys remain

50 essential to quantify the erosion of both biodiversity and biomass2–4,7,8,10,19. The

51 simultaneous evaluation of forest degradation on tropical biodiversity and biomass at

52 large-scales provides crucial knowledge for the conservation and climate change

53 agenda13,17,20 and to refine regional assessments of biodiversity and ecosystem service

54 (e.g. Intergovernmental Science-Policy Platform on Biodiversity and Ecosystem Services

55 - IPBES).

56 Here, we aim at quantifying the impacts of forest degradation on a major

57 biodiversity hotspot located in eastern South America, the Atlantic Forest

58 (Supplementary Fig. 1). Home to 35% of the South American population, the Atlantic

59 Forest is one of the most fragmented tropical/subtropical forests in the world12,21, which

60 may well represent the present or future of other tropical forests worldwide22. To achieve

61 our goal, we create one of the largest data sets of forest surveys ever assembled for the

62 tropics and subtropics23, both inside and outside protected areas. This data set includes

63 data on forest biomass and tree species richness and/or composition, as well as carefully bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

64 curated metadata associated with each survey, representing a total of 1,819 field surveys,

65 1.45 million trees, 3,124 tree species, and 1,238 ha of sampling coverage (Supplementary

66 Table 1, Supplementary Data 1). The data set covers the entire range of environmental

67 conditions, landscape contexts, and disturbance histories of the Atlantic Forest

68 (Supplementary Fig. 2, Supplementary Table 2). It also contains information on multiple

69 species properties, including plant functional traits (i.e., wood density, maximum height,

70 seed mass), ecological groups (or successional status, e.g. pioneer) and their conservation

71 value (i.e., threat status and endemism level), which enable to assess human-induced

72 impacts on community composition sensu lato.

73 Using this unprecedented data set, we quantified forest degradation impacts on the

74 aboveground biomass stocks, tree species richness, and multiple species properties. More

75 specifically, we assess the extent and magnitude of those impacts by asking: (i) how

76 pervasive negative impacts are across this biodiversity hotspot? (ii) how much these

77 biodiversity and biomass losses represent compared to low-disturbance Atlantic Forests?

78 And (iii) can protected areas mitigate those losses? Next, we explore the implications of

79 our results to the conservation of what's left of this biodiversity hotspot by (iv) projecting

80 forest degradation impacts to the remaining Atlantic Forest area to estimate the total

81 amount of carbon lost. We also (v) explore the costs and benefit of two contrasting

82 restoration scenarios: one that focused only on reducing the within-fragment disturbance

83 level (i.e. ‘fragment restoration’ scenario) and another focused on increasing fragment

84 size and landscape connectivity (i.e. ‘landscape restoration’ scenario).

85 In general lines, we quantified forest degradation impacts on Atlantic Forest

86 biodiversity and biomass as follows. First, the variation in forest biomass, species bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

87 richness and species properties were described using linear mixed-effects regression

88 models. These models accounted for the effects of environmental and human-related

89 variables, as well as sampling and biogeographical effects (see Methods and Figs. 3−7),

90 which explained 53% of the variation in biomass, 71% in species richness and 26−44% in

91 species properties (Supplementary Fig. 8, Supplementary Table 3). Next, we used these

92 models to generate baseline predictions in the absence of major human impacts, i.e.,

93 predictions as if all sites were large, low-disturbance forest patches in landscapes with

94 100% of forest cover. We validated the precision of these predictions using simulations

95 (Supplementary Table 4). Finally, we defined an index of loss due to human-induced

96 impacts, defined as the standardized difference between observed values and baseline

97 predictions (Supplementary Fig. 9). Values of the index close to zero indicate little

98 human impact and the more negative the value, the greater the impact.

99

100 Results and discussion

101 Extent and magnitude of human impacts

102 We found that the majority of the Atlantic Forest surveys (83%) presented losses of

103 species richness and biomass (Fig. 1), i.e. negative indices of loss. These losses were

104 correlated with each other (Fig. 1), meaning that fragments that suffer greater losses of

105 biomass also lose more species. In absolute terms, human-induced impacts corresponded

106 to average declines of 23−32% of the richness and biomass relative to low-disturbance

107 Atlantic Forests (Table 1, Supplementary Table 5). Similar estimates (18−57%) have

108 been reported at smaller spatial scales for Neotropical rainforests8,10,13 and at global

109 scale11,14,17, suggesting that our estimates are representative of the Atlantic Forest. We did bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

110 not explicitly model the distances to forest edge (see Supplementary Methods), where

111 biomass erosion can be even greater2,6–8. But assuming that researchers tend to avoid

112 edges while establishing plots, our estimates are probably conservative.

113 Human-induced impacts also caused a decline in the abundance of late

114 successional, large-seeded, and endemic species (Supplementary Fig. 9, Supplementary

115 Table 4), with reductions of 25‒42% when compared to low-disturbance Atlantic Forests

116 (Supplementary Table 5 and 6). Shifts in species composition caused by human impacts

117 have been reported for tropical forests, including the Atlantic Forest3,7,19. Here we also

118 found greater shifts in species properties in surveys with greater losses of species richness

119 and biomass (Fig. 2, Fig. 10‒12). This means that the erosion of richness and biomass is

120 being accompanied by a parallel decline of species that can enhance the provision of

121 ecosystem services24,25 and safeguard the conservation value of the Atlantic Forest. In the

122 long run, these losses can reinforce each other24, leading to an even greater erosion of

123 biodiversity and biomass.

124 The highest impact we found was related to the decline in the abundance of

125 endemic species, meaning that endemics are being replaced by species with wider

126 geographical ranges. Because species with wider ranges tend to occur over a wider

127 variety of environmental conditions (i.e., generalists)26, widespread species can benefit at

128 the expense of endemic and more specialist species in degraded fragments. Eventually,

129 this process leads to the decrease of beta-diversity through time and thus to the biotic

130 homogenization19,27–29. Thus, our results support the argument that human-induced

131 impacts are driving the biotic homogenization of Atlantic Forest fragments19,30. Our

132 approach based on CWMs of species properties, does not allow us to distinguish which bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

133 groups of species are causing this homogenization, but we can draw some propositions.

134 Widespread tree species in the Atlantic Forest are often small-seeded pioneers that

135 proliferate at forest edges and small fragments19,31,32. However, we found only a weak

136 correlation between the loses of endemism level and of seed mass and ecological groups

137 (r= 0.14 and 0.17, respectively; Supplementary Fig. 12). This suggests that not all species

138 proliferating in disturbed Atlantic Forests are small-seeded pioneers. In addition, because

139 exotics represented only 0.3% of the trees in our data set, the proliferation of widespread

140 native species is the most probable cause of the Atlantic Forest homogenization.

141 Wood density and maximum tree height, both related to carbon storage potential,

142 presented the smallest changes. The latter even presented a positive mean index of loss,

143 which may be explained by an increase in the abundance of late-successional,

144 understorey species as human-impacts decrease. In the Atlantic Forest, species with those

145 characteristics (common within Celastraceae, Erythroxylaceae, Myrtaceae, and Rutaceae)

146 often have wood densities above 0.7 g cm-3, explaining the negative correlation found

147 between maximum height and wood density losses (Fig. 2, Supplementary Fig. 12).

148 Although taller species often have larger seed mass and higher wood density across

149 vascular plants33,34, within trees there is still a wide spectrum of trait variation related to

150 different regeneration strategies. Tree species that demand high irradiance for their

151 development tend to have smaller seed sizes than species able to regenerate under mature

152 canopies35, which was confirmed here by the relatively high correlation (r= 0.51) between

153 the indices of loss of seed mass and ecological groups (Supplementary Fig. 12).

154 Moreover, some canopy and emergent trees are long-lived pioneers, which have

155 relatively light wood and small seeds35,36 (e.g., Albizia, Ceiba, Ficus, Gallesia, Jacaratia, bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

156 Parkia, Phytolacca, Piptadenia, Tachigali). Altogether, these results suggest that declines

157 in forest carbon stocks can be more easily explained by changes in forest structure than in

158 its trait composition37.

159

160 Influence of protected areas

161 Human-induced losses were lower inside than outside protected areas (Fig. 3,

162 Supplementary Fig. 13) and they decreased as the size of the protected area increased

163 (Supplementary Fig. 14). Thus, large protected areas are important to reduce human-

164 induced degradation, besides preventing deforestation itself. However, even inside

165 protected areas, we detected pervasive losses of species richness, species properties and

166 forest biomass (Fig. 3), revealing the practical limits of conservation policies focused

167 solely on the establishment of protected areas16. In addition, human-induced impacts in

168 areas where human settlements and use of resources are allowed but regulated (i.e., the

169 Brazilian "Environmental Protection Areas") were as high as, or higher than, in other

170 areas of private land (Fig. 3). Thus, not all types of protected areas are equally effective at

171 conserving biomass and biodiversity15. This pervasiveness of human impacts, irrespective

172 of land protection category, implies that strengthening the regulation within and

173 surrounding existing protected areas is as important as the creation of new protected

174 areas16. Although the land conservation category explained a small portion of the

175 variation in all indices of loss, this factor explained more variation than other metrics of

176 human pressure (Supplementary Fig. 13). This means that the magnitude of biodiversity

177 and biomass loss is difficult to predict in space based on human presence indices (see

178 Supplementary Discussion). Moreover, we found differences in average loss across the bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

179 Atlantic Forest biogeographical regions for most indices of loss (Supplementary Fig. 15),

180 reinforcing the idea that strategies to overcome biodiversity and carbon loss in the

181 Atlantic Forest should be planned regionally30 (see discussion below).

182

183 Total carbon losses in the Atlantic Forest

184 The erosion of biodiversity and biomass was pervasive across the entire Atlantic Forest

185 hotspot. We projected the biomass loss across the remaining Atlantic Forest area and

186 obtained losses of 451−525 Tg of carbon ‒ equivalent to the deforestation of 55−70

187 thousand km2 (Table 2, Supplementary Fig. 16, Supplementary Table 7). This is about

188 1.4 times the carbon loss due to deforestation of the Atlantic Forest between 1985 and

189 2017, and about a quarter of its remaining forest area (26%). Assuming a value of US$5

190 per Mg C in international markets, this human-induced carbon loss translates into

191 US$2.3−2.6 billion in carbon credits alone (see Supplementary Methods). Although

192 carbon-related impacts can be priced, biodiversity loss is far more difficult to value. The

193 loss of a species locally does not necessarily translate into regional extirpation. Moreover,

194 impacts on tree diversity may be protracted over decades5 and the recovery of tree

195 diversity and species composition occurs at a much slower rate than forest biomass9.

196 Biodiversity losses can be valued indirectly from their negative impacts on ecosystem

197 services20,24,25, but the underpinning of these relationships are currently not understood

198 well-enough to provide accurate economical assessments, particularly regarding shifts in

199 species properties within ecological communities.

200

201 Implications for the Atlantic Forest restoration bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

202 To explore strategies to mitigate biodiversity and carbon losses in Atlantic Forest

203 remnants, we used our models to compare the outcomes of two contrasting restoration

204 strategies: ‘fragment restoration’ that aimed at halving the within-fragment disturbance

205 levels of Atlantic Forest remnants; and ‘landscape restoration’ that aimed at restoring

206 20% of landscape forest cover around them. The first scenario aimed at simulating

207 restoration activities such as the reduction of forest-edge effects, control of invasive

208 species and enrichment plantings, while the second aimed at increasing fragment size and

209 thus landscape connectivity, e.g., forest corridors (see Supplementary Methods for

210 details). We found that proportional gains were greater for the ‘fragment restoration’

211 scenario for all variables, with the exception of wood density, maximum height and

212 endemism levels (Fig. 4). This means that the most effective strategy to restore the

213 remaining Atlantic Forest fragments is to reverse forest degradation inside them. This

214 statement was particularly true for forest biomass, seed mass, ecological groups and

215 extinction level (Fig. 4). Despite being an expected result, it reinforces the role of forest

216 disturbances as an important driver of biodiversity and biomass losses13,14.

217 The ‘landscape restoration’ scenario, however, can recover much more carbon at

218 the landscape-scale, since it includes the restoration of non-forest lands. Predicted carbon

219 gains and restoration yield (i.e., carbon gains divided by total costs of restoration) for the

220 ‘fragment restoration’ scenario were about 5‒9% and 19‒33% of the gains and yield for

221 the ‘landscape restoration’ scenario (Supplementary Table 6), with restoration outcomes

222 depending on the fragmentation and disturbance levels of the different Atlantic Forest

223 regions (see Supplementary Results and Discussion). The ‘landscape restoration’ strategy

224 can thus sequester carbon more efficiently, but it had little impact in reducing forest bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

225 degradation within the remaining forest fragments (Fig. 4). Thus, restoration will require

226 regionally-planned strategies that combine landscape with fragment restoration to

227 efficiently attenuate biodiversity and carbon losses in human-modified tropical forests30.

228 The two restoration scenarios predicted small improvements for hardwood, tall

229 and/or endemic tree species within fragments (Fig. 4). Greater improvements would

230 probably require the restoration of fragments to their full pre-disturbance conditions and

231 targeting landscape restoration to forest covers much larger than 20%. A 50% of

232 landscape forest cover can decrease the uncertainty in restoration success38, but

233 establishing such a high target for restoration of rural Atlantic Forest landscapes would

234 probably not be economically feasible. In this context, selecting the appropriate species to

235 maximize restoration outcomes becomes a more efficient strategy39. To help the species

236 selection for restoration we provide a list of 242 common trees species (Supplementary

237 Data 2) that have a high potential to increase the carbon storage and the conservation

238 value of disturbed Atlantic Forest fragments (see Supplementary Methods for details).

239 The re-introduction of these species could mitigate effects of local extirpation in the short

240 term, while the connectivity of the landscape can be restored and allow the frugivore

241 fauna to return in the long run. Seedlings from hardwood, tall or endemic species may be

242 less available from nurseries and thus increase restoration costs39,40. So, local seed

243 production cooperatives and nurseries able to propagate these species will be essential for

244 the support of the Atlantic Forest restoration39, particularly regarding the re-introduction

245 of large-seeded Atlantic Forest endemics.

246

247 Implications for biodiversity and carbon conservation bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

248 The imprint of the human-induced degradation on the Atlantic Forest may be an indicator

249 of the future condition of other tropical forests. Thus, the fate of tropical forests depends

250 not only on avoiding the deforestation of intact forests or on promoting the reforestation

251 of degraded lands9,12. It also depends on mitigating forest degradation in the remaining

252 forest fragments10,13,17,25. The impacts of forest degradation are hard to quantify at

253 regional scales and have therefore received less priority in the climate change and

254 conservation agendas. Despite recent changes in the Brazilian government's

255 environmental priorities, the Brazilian pledges to mitigate greenhouse-gas emissions

256 (e.g., Atlantic Forest pact, Paris Agreement or Bonn challenge) do not include actions to

257 reverse forest degradation or to re-introduce species with certain properties (e.g. species

258 with a high conservation value or carbon storage potential). In human-modified

259 landscapes, the mitigation of forest degradation can be a more cost-effective strategy than

260 reforestation depending on the outcomes desired. Moreover, it conflicts less with other

261 land uses, such as agriculture41 and thus has a good potential for engaging stakeholders.

262 Therefore, besides safeguarding part of the biodiversity and catalysing the

263 restoration of nearby areas42, restoring disturbed fragments is an opportunity to enhance

264 tropical biodiversity and biomass. In regions such as the Atlantic Forest, where most of

265 the forest remnants are on private land21, this opportunity has direct ramifications

266 implications to the compensation mechanisms for climate change mitigation. In Brazil,

267 funds to reduce carbon emissions from deforestation and forest degradation (REDD+) are

268 mainly concentrated on Amazonia and focused on avoiding deforestation (e.g., the

269 currently inactive Amazon Fund). Currently, only the State of has a fund

270 for the Atlantic Forest (http://www.fmarj.org). Our results indicate that the potential for a bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

271 biome-wide compensation fund is in the order of billions of dollars. The creation of

272 national policies to mitigate forest degradation, although highly dependent on the vision

273 and will of politicians, could be the key to attracting funds to the Atlantic Forest and

274 therefore be decisive for the future of this biodiversity hotspot43.

275

276 Methods

277 Forest survey and species data

278 We obtained 1,819 tree community surveys of natural Atlantic Forests available from the

279 Neotropical Tree Community database (TreeCo)23. We considered all types of forest

280 formations in Brazil, Paraguay and Argentina (Supplementary Fig. 1, Supplementary

281 Data 1), with the exception of dry deciduous and early secondary forests. Moreover, we

282 considered only surveys including trees with diameter at breast height (dbh) ≥3, ≥5 and

283 ≥10 cm and using plots or the point-centred quarter method. Surveys represented a total

284 effort of 1.45 million trees and 1,238 hectares (Supplementary Table 1) and they covered

285 a wide spectrum of environmental conditions and patch/landscape metrics

286 (Supplementary Table 2).

287 For each survey we extracted the tree density (trees ha-1), basal area (m2 ha-1),

288 species richness, sampling method (plots and point-centred quarter method), arrangement

289 of sample units (contiguous and systematic/random), sampled area (ha), dbh inclusion

290 criteria (cm) and geographical coordinates. We verified the precision of the geographical

291 coordinates provided to ensure that they corresponded to the forest fragment studied,

292 otherwise the survey was discarded. Whenever needed, plot coordinates were corrected, bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

293 based on maps or the site description provided in the study, including internet searches of

294 any valuable information on the fragment, farm or park location.

295 We compiled data on species abundances and basal area for surveys with a

296 sampling effort ≥0.1 ha and with species data presented in an extractable format (see

297 Supplementary Methods). For all surveys providing information on voucher specimens

298 associated with morpho-species, we checked for identification updates performed by

299 taxonomists using the speciesLink network (http://splink.cria.org.br). Although only

300 about one-third of the surveys provided vouchers, this effort improved the taxonomic

301 resolution of around 10% of our records. We checked species names for typographical

302 errors, synonyms, and orthographical variants, following the Brazilian Flora 202044.

303 Names marked as confer were assigned to the species suggested for confirmation, while

304 those marked as affinis were considered at the genus level. We compiled 98,030 species

305 records, totalling 1,171,935 trees measured and 3,124 valid species names.

306 We compiled information on the following species properties: wood density (g

307 cm-3)45, maximum adult height (m)46, seed mass (grams)47–49, ecological group (or

308 successional group, i.e., pioneer, early-secondary, late-secondary and climax), IUCN

309 threat category50 and endemism level. Ecological groups were obtained from information

310 provided in the original surveys and from the specialized literature. Endemism levels

311 were defined based on species occurrences in different continents, countries and Brazilian

312 states44,46. Species with Pantropical, Neotropical and South American distributions were

313 classified as ‘not endemic’, while species restricted to one or two adjacent regions (e.g.,

314 São Paulo and Rio de Janeiro states) were classified as ‘local endemic’. Species restricted

315 to South, South-eastern or North-eastern Brazil were classified as ‘regional endemic’. bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

316 Over 300 different sources of information were consulted to complete these species

317 properties (see Supplementary Notes). Records retrieved for numerical properties were

318 averaged for each species. Maximum adult height was considered here as the 90%

319 quantile of the distribution of maximum height records. For wood density and seed mass,

320 we used genus-level instead of species-level averages if necessary51,52. We also

321 completed missing information on ecological groups for typical pioneer Neotropical

322 genera (e.g., Cecropia, Trema, Vernonanthura).

323 We computed community weighted means (CWM) for the species properties in

324 order to summarize the community composition sensu lato of each survey. For

325 continuous species properties, the CWM was simply the average of each property

326 weighted by the total number of individuals in the community. For wood density and

327 maximum height, we used total basal area instead of the number of individuals of each

328 species to calculate the CWM. For wood density, CWM was obtained after removing

329 palms, palmoids, cacti, and tree ferns. For maximum height, we removed shrubs prior to

330 the calculation of CWM. For seed mass, we removed tree ferns prior to the calculation of

331 CWM. To compute the CWMs for the categorical properties, we treated them as ordinal

332 categorical data: Pioneer < Early secondary < Late secondary < Climax for ecological

333 groups; Not threatened/Not evaluated/Least concern < Data deficient/Near threatened <

334 Vulnerable < Endangered < Critically endangered for extinction level; and

335 Exotic/Naturalized < Not endemic < Northern/Eastern/Southern South-America <

336 Regional endemic < Local endemic for the endemism level. We assigned scores to these

337 ordered categories (see Supplementary Methods) and used them to compute the CWM. bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

338 These scores were chosen so that the higher the CWM, the better the community is

339 regarding species properties.

340

341 Site descriptors

342 For each survey we obtained climatic (e.g., mean air temperature and rainfall) and

343 topographic information (i.e., altitude, slope declivity and aspect) from different sources

344 (see Supplementary Methods). We also obtained soil classes from the original surveys,

345 which were checked for their consistency using soil maps at state and national scales.

346 Missing data and inconsistencies between sources were double-checked to assure soil

347 data quality and homogeneity. We used soil classes to infer average soil properties for

348 plant growth by cross-referencing them with a database of physical and chemical soil

349 properties (see Supplementary Methods). We used forest cover maps (30 m resolution)1

350 to extract 4×4 km landscapes centred on the survey coordinates and we used a 70%

351 canopy closure threshold to classify maps into forest or non-forest pixels. Classified maps

352 were used to calculate the proportion of forest cover and core forest cover, the median

353 edge-to-edge distance between fragments and different landscape aggregation indices

354 (see Supplementary Methods for details).

355 Fragment size was obtained from the original publications and was cross-

356 validated using the 2002 and 2012 maps of the Atlantic Forest fragments53. We

357 completed missing values of fragment size if there was consistency of the fragment size

358 obtained from these maps with the one obtained from the classified 4×4 km landscapes.

359 Inconsistencies between these sources were solved in Google Earth Pro (© Google Inc.),

360 including the manual re-calculation of fragment size, which we conducted for ~400 bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

361 surveys. The computation of the (mean) distance between surveys and forest edges was

362 not possible for the majority of surveys for different reasons (see Supplementary Methods

363 for details). Forest-edge effects are inextricably linked to fragment size and shape. Thus,

364 their effects are indirectly accounted for in other landscape and patch metrics.

365 Forest disturbance level was assigned based on the information on the type,

366 intensity and timing of human disturbances (i.e., selective logging, fire, hunting, thinning)

367 provided by the authors of the surveys. We considered three levels of disturbance: high

368 (i.e., highly/chronically disturbed forests, typically disturbed less than 50 years before the

369 survey); medium (lightly/sporadically disturbed forests, and/or disturbed 50–80 years

370 ago); and low (forests left undisturbed for at least 80 years). This classification is

371 qualitative, with substantial variation in forest structure and diversity expected within

372 classes. However, more objective and detailed information on disturbance histories, main

373 disturbance types and their intensities was lacking (see Supplementary Methods). Thus,

374 these coarse classes are the best information available to allow analysis across the

375 Atlantic Forest.

376 We assigned each survey a biogeographical region54. To avoid an excessive

377 subdivision of regions, we reassigned Atlantic Forest surveys mapped as Cerrado,

378 , Campo Rupestre enclaves, Atlantic Coast and Southern Atlantic

379 Mangroves to the closest region (Supplementary Fig. 1). We included the Uruguayan

380 biogeographical region, representing the forests in the transition to the pampas region of

381 South Brazil as part of the Atlantic Forest21,53. The proportion of each region in our

382 sample was similar to their contribution to the remaining Atlantic Forest area with few

383 exceptions (Serra do Mar: 29% in the data set and 21% of remaining area, Araucaria: bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

384 24% and 13%, Alto Paraná: 21% and 31%, Inland: 8% and 11%, Bahia Coastal:

385 6% and 7%, Uruguay: 5% and 3%, Northeast: 5% and 4%, Atlantic Dry: 2% and 10%).

386 Categories of land conservation were obtained from the original study and/or from

387 maps of conservation units, as follows: protected areas (i.e., strict protection and

388 sustainable use conservation units); other public and protected land (e.g., research

389 centres, university campuses, botanical gardens, military and indigenous lands); private

390 land regulated by government laws, a conservation unit known in Brazil as

391 Environmental Protection Areas (or APA for Portuguese: "Área de Proteção Ambiental");

392 and private lands outside conservation units (e.g., farms). We could not assign the land-

393 use category for 5% of surveys. For surveys conducted inside protected areas, we also

394 obtained the size and type of the protected area (e.g., strict protection, sustainable use).

395 We extracted the Human Influence Index (HII)55 based on the coordinates of each

396 survey. The HII ranges from 0 to 64 (minimum and maximum human influence,

397 respectively) and it accounts for human population density, land use (e.g., urban areas,

398 agriculture) and ease of access (i.e., proximity to roads, railways, navigable rivers, and

399 coastlines). The survey distance from main cities (>100,000 inhabitants) was significantly

400 correlated with the HII (Pearson's r= -0.38, p< 0.0001). Because results using this

401 distance instead of the HII were qualitatively the same, we present only the results using

402 the HII.

403

404 Data analysis

405 We conducted separate analyses for each response variable, because not all surveys had

406 data on forest biomass, species richness and CWM simultaneously. We also removed 50 bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

407 surveys from the analyses of species properties, because in these surveys less than 80% of

408 the sample individuals had information on species properties. Thus, analyses were

409 conducted using 1676 surveys for biomass (92% of surveys), 1790 for species richness

410 (98%) and 1213‒1214 (68%) depending on the species property considered

411 (Supplementary Data 1). Response and explanatory variables were transformed if

412 necessary and candidate explanatory variables were pre-selected based on their co-

413 linearity and on their impact on model performance (see Supplementary Methods).

414 We described forest biomass, species richness, and CWMs of species properties

415 using linear mixed-effects regression models. To select which explanatory variables

416 should be included in these models56,57, we first selected the co-variables composing the

417 random structure of the models, which had survey methodology and biogeographical

418 regions as candidate random effects. These two categorical variables divide the

419 observations into groups, in order to account for correlated observations in the data56,57

420 (e.g., species richness is higher and more similar for some biogeographical regions when

421 compared to other regions). Survey methodology refers to the sampling method,

422 arrangement and dbh inclusion criteria, which were combined to create a methodological

423 categorical variable (e.g., contiguous plots dbh ≥5 cm, systematic plots dbh ≥10 cm, etc).

424 Exploratory data analyses suggested an interaction of the effects of sampling effort and

425 the methodological variable on forest biomass and species richness. Indeed, the addition

426 of a random term composed by the log-transformed sampling effort nested within the

427 methodological categories significantly improved the model fit for both variables. To

428 avoid the artificial inflation of model explanation57, we also included log-transformed

429 sampling effort as a fixed effect in the final models of biomass and richness. The same bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

430 was not true for the random structure of the models of species properties, which had

431 fewer observations and were more prone to overfitting issues (i.e., model singularity).

432 We then selected the fixed effects to compose the optimum regression model. We

433 kept to a minimum the interactions between fixed effects to avoid problems with the

434 interpretation of individual coefficients, including only interactions with a clear

435 biological meaning (e.g., temperature × rainfall seasonality). The optimum regression

436 models contained only fixed effects that improved overall model fit (see Supplementary

437 Methods) and they had the following general structure: y ~ Environment + Human +

438 Effort + (Effort|Method) + (1|Region). The term Environment includes the climate,

439 topography and soil variables (and their interactions). Human includes landscape metrics,

440 fragment size, and disturbance level. Effort is the total survey effort (not included in

441 species property models). As described in the previous paragraph, Method is the

442 methodological categories (nine levels) and Region is the Atlantic Forest biogeographical

443 regions (eight levels, Supplementary Fig. 1). Therefore, models had environmental,

444 human, methodological and biogeographical (historical) components. The number of

445 individuals sampled had, as expected, a strong influence on species richness. During

446 explanatory analyses, we also found a positive relationship between biomass and tree

447 density (individuals ha-1). So, we included number of individuals and tree density as fixed

448 effects to model richness and biomass data, respectively.

449

450 Evaluation of human impacts

451 We used the optimum models to quantify human-related impacts on species richness,

452 species properties and forest biomass. We first obtained the model predictions in a bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

453 scenario without major human-induced impacts, i.e., model predictions for human-related

454 variables reset at values corresponding to large (>300,000 ha), low-disturbance patches in

455 landscapes with 100% of (core) forest cover and maximum patch aggregation. We

456 assumed current conditions for all other co-variables in the models (i.e., climate, soil

457 conditions, sampling method and biogeographical region) and the deforestation caused by

458 indigenous populations to be negligible. Then, we calculated the standardized difference

459 between observed values and predicted values in the human-free scenario:

460 (Observed-Predicted) ∕ Predicted.

461 The difference between observed and predicted values was standardized to make them

462 comparable across our response variables. We used this standardized difference as an

463 index of loss related to human impacts (Supplementary Fig. 9). Negative indices mean

464 that observed values of biomass, species richness or community weighted means (CWM)

465 of species properties are smaller than the predictions in the human-free scenario at the

466 same forest fragment.

467 Because nearly all Atlantic Forest surveys were conducted after the intensification

468 of its deforestation (1940‒1970s)23,58, it is impossible to obtain measurements of pristine

469 forests, particularly at the scale of our study (~ 1.4 million km2). Instead, we assumed

470 Atlantic Forests previous to human impacts to be large, fully-connected and undisturbed.

471 This assumption has limitations but it is a valid yet conservative representation of past

472 forest conditions. We compared the predictions for the human-free scenario against

473 intervals obtained from 5000 samples of the estimated model coefficients (Supplementary

474 Methods). The precision of these predictions was not strongly sensitive to variations in

475 the parameter estimates (Supplementary Table 4). Predictions of carbon stocks for the bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

476 human-free scenario ranged between 43‒162 Mg C ha-1 for individual surveys, but they

477 were typically between 88‒110 Mg C ha-1 (mean of 99 Mg C ha-1), which is consistent

478 with previous studies in the Atlantic Forest58,59. Finally, because we predict shifts in

479 CWM of species properties and not in taxonomic composition itself, our approach did not

480 provide any indication of which species increased or declined in abundance due to human

481 impacts.

482 We inspected the relationship among indices of loss by plotting one against the

483 other (Fig. 1, Supplementary Fig. 10‒S11) and testing the strength of their relationship

484 using linear regression models. We also inspected the correlation among the pairs of

485 indices of loss for the six species properties (Supplementary Fig. 12). We averaged the

486 indices of loss for the six species properties to generate a mean index of loss, using the

487 conditional R2 of the mixed-effects models of each property as weights (Supplementary

488 Table 3). We then explored the relationships between all indices of loss using the two

489 first axes of a principal component analysis, produced using the scaled indices of loss for

490 species richness, properties and biomass (Fig. 2). We also evaluated if the category of

491 land use and of the Human Influence Index (HII)55 have an effect on these indices (Fig. 3,

492 Supplementary Fig. 13). We plotted them over the Atlantic Forest map to reveal possible

493 hotspots of human impacts. Finally, we evaluated the effect of the size (Supplementary

494 Fig. 14) and type of the protected area on the indices of loss (see Supplementary

495 Methods).

496

497 Absolute losses and projections across the Atlantic Forest bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

498 For each biogeographical region, we estimated the mean values of species richness,

499 species properties and forest biomass for low disturbance surveys, taken here as

500 references to the Atlantic Forest species diversity, community composition and biomass

501 storage. Here, we assumed modern, low-disturbance fragments as proxies of past forest

502 conditions. We acknowledge that past forests may have had more biodiversity and

503 biomass than modern-day forests, but these fragments represent the best information

504 currently available for the Atlantic Forest, given that measurements prior to human

505 impacts are unavailable. Thus, we took a conservative approach; if modern-day fragments

506 underestimate past conditions, absolute losses would be even greater than reported here.

507 These reference values were obtained using a different subset of surveys for each

508 forest descriptors and separately for each dbh inclusion criterion (Supplementary Table 5

509 and 6). Next, we calculated the absolute loss for each survey (i.e., Predicted - Observed),

510 which was then averaged for each region. Because our sample is biased towards large

511 forest fragments (Supplementary Fig. 2), these averages were weighted by the probability

512 of having a fragment of the same size in each biogeographical region. These probabilities

513 were obtained from log-normal distributions fitted to the 2016 size distribution of the

514 remaining Atlantic Forest fragments53. We then calculated the proportion that these

515 weighted average losses represent in respect to the region-specific reference values

516 (Table 1, Supplementary Fig. 15).

517 To obtain the forest area that would match the carbon losses caused by post-

518 deforestation human impacts (i.e., equivalent forest loss), the proportion of carbon loss

519 was multiplied by the remaining Atlantic Forest area per biogeographical region53. Due to

520 missing information on human-related variables for all Atlantic Forest fragments (e.g. bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

521 within forest disturbance level), we assumed the same carbon loss for all fragments

522 within each biogeographical region. Next, we obtained the total amount of carbon loss

523 (i.e., the equivalent carbon loss), which was computed based on the equivalent forest loss

524 and the reference values of carbon storage per region. The mean carbon storage, and

525 proportional losses were averaged across the biogeographical regions, using the area of

526 regions as weights. In contrast, values of equivalent forest and carbon loss were summed

527 across regions. Thus, our estimate of carbon loss and its projection across the Atlantic

528 Forest takes into account the bias towards larger fragments in our sample and the regional

529 differences in Atlantic Forest carbon storage potential (Table 2).

530 We compared the equivalent forest loss to Atlantic Forest deforestation between

531 1985 and 201753. In the Brazilian Atlantic Forest, deforestation was 1.93 million ha for

532 the entire interval of 1985 and 201753. For Paraguay, we found estimates for 1989‒2000

533 and 2003‒201360,61, so we estimated a 1.96 million ha forest loss for 1985‒2017 from the

534 reported annual deforestation rates. For Argentina, we had estimates for 1998‒201462 and

535 used the same procedure to obtain an estimate of 0.1 million ha for 1985‒2017. It should

536 be noted that overall deforestation in Brazil was larger than in other countries, but it

537 occurred more intensely before 1985. Therefore, we estimated a total forest loss of 4.18

538 million ha between 1985‒2017. Finally, we calculated how much money the equivalent

539 carbon loss would represent if it had been traded as carbon credits in international

540 markets (Table 2), assuming US$5 per Mg C paid for projects of forestry and land use63.

541

542 Simulating strategies of forest restoration bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

543 We used the regression models fitted to biodiversity and biomass data to simulated two

544 simple scenarios of forest restoration. The ‘fragment restoration’ scenario assumes no

545 increase in landscape forest cover and aims at restoring forest disturbance levels to 50%

546 of the lower level observed for the Atlantic Forest. This is realistic target, because

547 restoring and maintaining disturbance levels below this threshold, although possible,

548 would be too expensive. The ‘landscape restoration’ scenario mimics restoration

549 activities around existing fragments, increasing fragment size and thus landscape integrity

550 (i.e., greater forest cover) and connectivity (i.e., smaller fragment distance). This scenario

551 implicitly assumes that restoration activities should focus on the restoration around

552 existing forest fragments, which would increase the ‘area’ of restoration and thus

553 maximize the success of the restoration project while minimizing its costs64. The

554 ‘landscape restoration’ aimed at a minimum of 20% of landscape forest cover, a target

555 that relates to the legal requirements for the Atlantic Forest fixed by the Brazilian Forest

556 Code65.

557 The two scenarios were compared by generating model predictions for all the

558 remaining fragments in 201653. For each fragment, we compared the ‘current situation’

559 predictions to the predictions under each restoration scenario, by varying fragment

560 disturbance level and patch/landscape metrics, respectively. Current situation predictions

561 used fragment-specific coordinates to extract spatialized information (e.g., climate) or

562 regional averages for information not available for all fragments (e.g., forest disturbance

563 level). For the ‘landscape restoration’ scenario, the increase in forest cover to reach the

564 20% target and the corresponding changes in landscape metrics between the ‘current’ and

565 ‘restored’ landscapes were derived from 100 iterations of simulated landscapes (see bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

566 Supplementary Methods for more details). We compared the proportional gains of each

567 restoration scenario for forest biomass, species richness, and CWMs of species properties

568 (Fig. 4).

569 We also compared the costs and efficiency of the two scenarios in terms of carbon

570 sequestration (Supplementary Table 6). We assumed an average value of US$750 per

571 hectare for within-fragment restoration, which includes activities to reduce edge effects,

572 control of invasive species, and enrichment plantings66,67. We assumed an average of

573 US$1000 per hectare for land opportunity cost (10 years period for cattle-ranching

574 activities)68 and an average of US$1500 per hectare for the restoration costs on degraded

575 lands (from seedling plantation to assisted natural regeneration), which varies between

576 US$350 and US$3000 per hectare for the Atlantic Forest64,68–70. However, we assumed

577 different restoration costs for each biogeographical region (i.e., less disturbed and

578 fragmented landscapes should cost less to be restored ‒ Supplementary Table 6). As

579 before, we monetarized the expected carbon gains in terms of carbon credits in

580 international markets (US$5 per Mg C paid for projects of forestry and land use63). For

581 the ‘landscape restoration’ scenario, we separated carbon gains into those expected inside

582 the remaining fragments and in the restored areas per se (Supplementary Table 6). To

583 calculate the potential of carbon recovery in the restored areas, we used the references of

584 carbon density and the average carbon loss for each biogeographical region

585 (Supplementary Fig. 15), assuming that the drivers of carbon loss in fragments would be

586 similar in restored areas.

587

588 Data availability bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

589 Survey, species abundances and trait data used in this study were extracted from the

590 Neotropical Tree Communities database (TreeCo, version 4.0) and are available upon

591 request at http://labtrop.ib.usp.br/doku.php?id=projetos:treeco:start. The list of surveys

592 extracted from the TreeCo database is provided in Supplementary Data 1, together with

593 the corresponding metadata. The sources of survey and species abundance data are

594 referenced in Supplementary Data 1 and sources of species properties data referenced in

595 the Methods or in the Supplementary Notes. Other relevant data are available from the

596 corresponding author upon reasonable request.

597

598 Code availability

599 Codes used to conduct the analysis is available from the corresponding author upon

600 request.

601

602 References

603 1. Hansen, M. C. et al. High-resolution global maps of 21st-century forest cover

604 change. Science 342, 850–853 (2013).

605 2. Magnago, L. F. S. et al. Do fragment size and edge effects predict carbon stocks in

606 trees and lianas in tropical forests? Funct. Ecol. 31, 542–552 (2017).

607 3. Lopes, A. V., Girão, L. C., Santos, B. A., Peres, C. A. & Tabarelli, M. Long-term

608 erosion of tree reproductive trait diversity in edge-dominated Atlantic forest

609 fragments. Biol. Conserv. 142, 1154–1165 (2009).

610 4. Laurance, W. F. et al. The fate of Amazonian forest fragments: A 32-year

611 investigation. Biol. Conserv. 144, 56–67 (2011). bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

612 5. Uhl, C., Veríssimo, A., Mattos, M. M., Brandino, Z. & Vieira, I. C. G. Social,

613 economic, and ecological consequences of selective logging in an Amazon frontier:

614 the case of Tailândia. For. Ecol. Manage. 46, 243–273 (1991).

615 6. Laurance, W. F. et al. Biomass Collapse in Amazonian Forest Fragments. Science

616 278, 1117–1118 (1997).

617 7. Santos, B. A. et al. Drastic erosion in functional attributes of tree assemblages in

618 Atlantic forest fragments of northeastern Brazil. Biol. Conserv. 141, 249–260 (2008).

619 8. Paula, M. D., Costa, C. P. A. & Tabarelli, M. Carbon storage in a fragmented

620 landscape of Atlantic Forest: the role played by edge-affected habitats and emergent

621 trees. Trop. Conserv. Sci. 4, 349–358 (2011).

622 9. Martin, P. A., Newton, A. C. & Bullock, J. M. Carbon pools recover more quickly

623 than plant biodiversity in tropical secondary forests. Proc. R. Soc. B Biol. Sci. 280,

624 20132236 (2013).

625 10. Berenguer, E. et al. A large-scale field assessment of carbon stocks in human-

626 modified tropical forests. Glob. Chang. Biol. 20, 3713–3726 (2014).

627 11. Chaplin-Kramer, R. et al. Degradation in carbon stocks near tropical forest edges.

628 Nat. Commun. 6, 10158 (2015).

629 12. Haddad, N. M. et al. Habitat fragmentation and its lasting impact on Earth’s

630 ecosystems. Sci. Adv. 1, e1500052 (2015).

631 13. Barlow, J. et al. Anthropogenic disturbance in tropical forests can double

632 biodiversity loss from deforestation. Nature 535, 144–147 (2016).

633 14. Erb, K. et al. Unexpectedly large impact of forest management and grazing on global

634 vegetation biomass. Nature 553, 73–76 (2018). bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

635 15. Scharlemann, J. P. W. et al. Securing tropical forest carbon: the contribution of

636 protected areas to REDD. Oryx 44, 352–357 (2010).

637 16. Laurance, W. F. et al. Averting biodiversity collapse in tropical forest protected

638 areas. Nature 489, 290–294 (2012).

639 17. Newbold, T. et al. Global effects of land use on local terrestrial biodiversity. Nature

640 520, 45–50 (2015).

641 18. Asner, G. P. et al. High-resolution forest carbon stocks and emissions in the

642 Amazon. Proc. Natl. Acad. Sci. 107, 16738–16742 (2010).

643 19. Lôbo, D., Leão, T., Melo, F. P. L., Santos, A. M. M. & Tabarelli, M. Forest

644 fragmentation drives Atlantic forest of northeastern Brazil to biotic homogenization.

645 Divers. Distrib. 17, 287–296 (2011).

646 20. Liang, J. et al. Positive biodiversity-productivity relationship predominant in global

647 forests. Science 354, aaf8957 (2016).

648 21. Ribeiro, M. C., Metzger, J. P., Martensen, A. C., Ponzoni, F. J. & Hirota, M. M. The

649 Brazilian Atlantic Forest: How much is left, and how is the remaining forest

650 distributed? Implications for conservation. Biol. Conserv. 142, 1141–1153 (2009).

651 22. Seymour, F. & Harris, N. L. Reducing tropical deforestation. Science 365, 756–757

652 (2019).

653 23. Lima, R. A. F. et al. How much do we know about the endangered Atlantic Forest?

654 Reviewing nearly 70 years of information on tree community surveys. Biodivers.

655 Conserv. 24, 2135–2148 (2015).

656 24. Bunker, D. E. et al. Species loss and aboveground carbon storage in a tropical forest.

657 Science 310, 1029–1031 (2005). bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

658 25. Bello, C. et al. Defaunation affects carbon storage in tropical forests. Sci. Adv. 1,

659 e1501105 (2015).

660 26. Slatyer, R. A., Hirst, M. & Sexton, J. P. Niche breadth predicts geographical range

661 size: A general ecological pattern. Ecol. Lett. 16, 1104–1114 (2013).

662 27. McKinney, M. L. & Lockwood, J. L. Biotic homogenization: a few winners

663 replacing many losers in the next mass extinction. Trends Ecol. Evol. 14, 450–453

664 (1999).

665 28. Olden, J. D. & Rooney, T. P. On defining and quantifying biotic homogenization.

666 Glob. Ecol. Biogeogr. 15, 113–120 (2006).

667 29. Arroyo-Rodríguez, V. et al. Plant β-diversity in fragmented rain forests: Testing

668 floristic homogenization and differentiation hypotheses. J. Ecol. 101, 1449–1458

669 (2013).

670 30. Joly, C. A., Metzger, J. P. & Tabarelli, M. Experiences from the Brazilian Atlantic

671 Forest: ecological findings and conservation initiatives. New Phytol. 204, 459–473

672 (2014).

673 31. Tabarelli, M., Aguiar, A. V., Girão, L. C., Peres, C. A. & Lopes, A. V. Effects of

674 pioneer tree species hyperabundance on forest fragments in Northeastern Brazil.

675 Conserv. Biol. 24, 1654–1663 (2010).

676 32. Thier, O. & Wesenberg, J. Floristic composition and edge-induced homogenization

677 in tree communities in the fragmented Atlantic rainforest of Rio de Janeiro, Brazil.

678 Trop. Conserv. Sci. 9, 852–876 (2016).

679 33. Moles, A. T., Falster, D. S., Leishman, M. R. & Westoby, M. Small-seeded species

680 produce more seeds per square metre of canopy per year, but not per individual per bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

681 lifetime. J. Ecol. 92, 384–396 (2004).

682 34. Díaz, S. et al. The global spectrum of plant form and function. Nature 529, 167–171

683 (2016).

684 35. Clark, D. A. & Clark, D. B. Life history diversity of canopy and emergent trees in a

685 Neotropical rain forest. Ecol. Monogr. 62, 315–344 (1992).

686 36. Hammond, D. S. & Brown, V. K. Seed size of woody plants in relation to

687 disturbance, dispersal, soil type in wet neotropical forests. Ecology 76, 2544–2561

688 (1995).

689 37. Laurance, W. F., Delamônica, P., Laurance, S. G., Vasconcelos, H. L. & Lovejoy, T.

690 E. Rainforest fragmentation kills big trees. Nature 404, 836 (2000).

691 38. Crouzeilles, R. & Curran, M. Which landscape size best predicts the influence of

692 forest cover on restoration success? A global meta-analysis on the scale of effect. J.

693 Appl. Ecol. 53, 440–448 (2016).

694 39. Brancalion, P. H. S. et al. Maximizing biodiversity conservation and carbon stocking

695 in restored tropical forests. Conserv. Lett. e12454 (2018).

696 40. Brancalion, P. H. S., Viani, R. A. G., Aronson, J., Rodrigues, R. R. & Nave, A. G.

697 Improving planting stocks for the Brazilian Atlantic Forest restoration through

698 community-based seed harvesting strategies. Restor. Ecol. 20, 704–711 (2012).

699 41. Verdone, M. & Seidl, A. Time, space, place, and the Bonn Challenge global forest

700 restoration target. Restor. Ecol. 25, 903–911 (2017).

701 42. Turner, I. M. & T. Corlett, R. The conservation value of small, isolated fragments of

702 lowland tropical rain forest. Trends Ecol. Evol. 11, 330–333 (1996).

703 43. Venter, O. et al. Harnessing carbon payments to protect biodiversity. Science 326, bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

704 1368 (2009).

705 44. Ranzato Filardi, F. L. et al. Brazilian Flora 2020: Innovation and collaboration to

706 meet Target 1 of the Global Strategy for Plant Conservation (GSPC). Rodriguésia

707 69, 1513–1527 (2018).

708 45. Chave, J. et al. Towards a worldwide wood economics spectrum. Ecol. Lett. 12, 351–

709 366 (2009).

710 46. Oliveira-Filho, A. T. TreeAtlan 2.0, Flora arbórea da América do Sul cisandina

711 tropical e subtropical: Um banco de dados envolvendo biogeografia, diversidade e

712 conservação. (2010), available at www. cb.ufmg.br/treeatlan.

713 47. Almeida-Neto, M., Campassi, F., Galetti, M., Jordano, P. & Oliveira-Filho, A.

714 Vertebrate dispersal syndromes along the Atlantic forest: broad-scale patterns and

715 macroecological correlates. Glob. Ecol. Biogeogr. 17, 503–513 (2008).

716 48. Bello, C. et al. Atlantic frugivory: a plant–frugivore interaction data set for the

717 Atlantic Forest. Ecology 98, 1729 (2017).

718 49. Kew, Royal Botanic Gardens. Seed Information Database (SID). (2018), available at

719 http://data.kew.org/sid.

720 50. Martinelli, G. & Moraes, M. A. Livro vermelho da flora do Brasil. Instituto de

721 Pesquisas Jardim Botânico do Rio de Janeiro, Rio de Janeiro, 1st ed. (2013).

722 51. Chave, J. et al. Regional and phylogenetic variation of wood density across 2456

723 Neotropical tree species. Ecol. Appl. 16, 2356–2367 (2006).

724 52. ter Steege, H. et al. Continental-scale patterns of canopy tree composition and

725 function across Amazonia. Nature 443, 444–447 (2006).

726 53. Fundação SOS Mata Atlântica & Instituto Nacional de Pesquisas Espaciais (INPE). bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

727 Atlas dos remanescentes florestais da Mata Atlântica: período 2016-2017. Fundação

728 SOS Mata Atlântica (2018).

729 54. Olson, D. M. & Dinerstein, E. The Global 200: Priority ecoregions for global

730 conservation. Ann. Missouri Bot. Gard. 89, 199–224 (2002).

731 55. Sanderson, E. W. et al. The human footprint and the last of the wild. Bioscience 52,

732 891–904 (2002).

733 56. Pinheiro, J. & Bates, D. Linear mixed-effects models: basic concepts and examples.

734 Springer-Verlag, New York. (2000).

735 57. Zuur, A. F., Ieno, E. N., Walker, N., Saveliev, A. A. & Smith, G. M. Mixed effects

736 models and extensions in ecology with R. Springer New York, New York. (2009).

737 58. Leite, C. C., Costa, M. H., Soares-Filho, B. S. & de Barros Viana Hissa, L. Historical

738 land use change and associated carbon emissions in Brazil from 1940 to 1995.

739 Global Biogeochem. Cycles 26, gb2011 (2012).

740 59. Bustamante, M. M. C. et al. Engagement of scientific community and transparency

741 in C accounting: the Brazilian case for anthropogenic greenhouse gas emissions from

742 land use, land-use change and forestry. Environ. Res. Lett. 13, 055005 (2018).

743 60. Huang, C. et al. Rapid loss of Paraguay’s Atlantic forest and the status of protected

744 areas - A Landsat assessment. Remote Sens. Environ. 106, 460–466 (2007).

745 61. Da Ponte, E., Roch, M., Leinenkugel, P., Dech, S. & Kuenzer, C. Paraguay’s

746 Atlantic Forest cover loss – Satellite-based change detection and fragmentation

747 analysis between 2003 and 2013. Appl. Geogr. 79, 37–49 (2017).

748 62. Fundación Vida Silvestre Argentina & WWF. State of the Atlantic Forest: Three

749 countries, 148 million people, one of the richest forests on Earth. Fundación Vida bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

750 Silvestre Argentina, WWF-Brazil and WWF-Paraguay. Puerto Iguazú, Argentina,

751 (2017).

752 63. Hamrick, K. & Gallant, M. Unlocking potential: State of the voluntary carbon

753 markets 2017. Washington, DC. (2017), available at http://www.forest-

754 trends.org/documents/files/doc_5591.pdf.

755 64. Strassburg, B. B. N. et al. Strategic approaches to restoring ecosystems can triple

756 conservation gains and halve costs. Nat. Ecol. Evol. 3, 62–70 (2019).

757 65. Soares-Filho, B. et al. Cracking Brazil’s Forest Code. Science 344, 363–364 (2014).

758 66. César, R. G. et al. Evaluating climber cutting as a strategy to restore degraded

759 tropical forests. Biol. Conserv. 201, 309–313 (2016).

760 67. Mangueira, J. R. S. A., D. Holl, K. & Rodrigues, R. R. Enrichment planting to

761 restore degraded tropical forest fragments in Brazil. Ecosyst. People 15, 3–10 (2019).

762 68. Brancalion, P. H. S., Viani, R. A. G., Strassburg, B. B. N. & Rodrigues, R. R.

763 Finding the money for tropical forest restoration. Unasylva 63, 41–50 (2012).

764 69. Rodrigues, R. R., Lima, R. A. F., Gandolfi, S. & Nave, A. G. On the restoration of

765 high diversity forests: 30 years of experience in the Brazilian Atlantic Forest. Biol.

766 Conserv. 142, 1242–1251 (2009).

767 70. Brancalion, P. H. S. et al. What makes ecosystem restoration expensive? A

768 systematic cost assessment of projects in Brazil. Biol. Conserv. 240, 108274 (2019).

769

770 Acknowledgments

771 We thank Ary T. Oliveira-Filho, Danilo Mori, Markus Gastauer, Melina Melito, Natália

772 Targhetta, Geison Castro, Luiz F.S. Magnago and Carolina Bello for their great help in bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

773 constructing the database. We also thank André M. Amorim, Andréia A. Rezende, João

774 A. Meira-Neto, João L.F. Batista, Márcia C.M. Marques, Maria T.Z. Toniato, Mariana C.

775 Pardgurschi, Mario J. Marques-Azevedo, Ricardo R. Rodrigues, Robson L. Capretz and

776 Victor P. Zwiener for providing published or unpublished data in digital format. Luiz F.S.

777 Magnago provided functional trait information for several species, Mariano C. Cenamo

778 and Mauricio A. Voivodic provided critical comments on the implications of our results,

779 Luciana F. Alves and Simone Viera provided important sources on carbon stock

780 estimates and Pedro, H.S. Brancalion, Ricardo R. Rodrigues, and Ricardo A.G. Viani

781 provided useful advice on restoration costs in the Atlantic Forest. This research was

782 funded by the grant 2013/08722-5, São Paulo Research Foundation (FAPESP). R.A.F.L.

783 was supported by the European Union’s Horizon 2020 research and innovation program

784 under the Marie Skłodowska-Curie grant agreement No 795114. A.C.V. and A.L.G. were

785 supported by the Conselho Nacional de Desenvolvimento Científico e Tecnológico

786 (CNPq), grant 312075/2013-8 and the Fundação de Amparo à Pesquisa e Inovação de

787 (FAPESC), grant 2017TR1922. J.C. acknowledges funding by

788 “Investissement d’Avenir” grants managed by Agence Nationale de la Recherche

789 (CEBA, ref. ANR-10-LABX-25-01 and TULIP, ref. ANR-10-LABX-0041). We

790 gratefully acknowledge the hundreds of researchers, students and technicians that

791 performed the field work, species identification and data analysis that resulted in the

792 surveys included in this study. These surveys were funded by many different agencies

793 (e.g., CAPES, CNPq, FAPESP, FAPEMIG, FAPERJ, FAPESC, etc.), but space prohibits

794 publishing the complete list of grants. Some major plot initiatives used in this study were bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

795 supported by grants 99/08515-0, 99/09635-0 and 04/04820-3 (FAPESP) and grant DE-

796 FC26-01NT411151 (US Department of Energy).

797

798 Author information

799 Contributions

800 This study was conceived and designed by R.A.F.L., with the help of J.C. and P.I.P. Data

801 were compiled and/or validated by R.A.F.L., G.R.P., A.L.G., A.C.V. Data analysis was

802 conducted by R.A.F.L. and P.I.P, with the support of A.A.O. R.A.F.L. drafted the paper

803 and all authors have revised the subsequent drafts. All authors have contributed

804 substantially to the interpretation and discussion of the results.

805 Corresponding author

806 Correspondence to Renato A. Ferreira de Lima

807

808 Competing interests

809 The authors declare no competing interests.

810

811 Supplementary information

812 Supplementary information is available for this paper.

813 bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

814 Tables

815 816 Table 1. The magnitude of the loss of forest biomass, species richness and species 817 properties due to human-induced impacts.

818 The average proportion of losses were obtained from the averages of the absolute loss 819 predicted (predicted - observed values) for each survey normalized by the reference 820 values of low-disturbance Atlantic Forest fragments (Supplementary Table 5). The 821 average of the proportional absolute loss was weighted by the probability of having a 822 surveyed fragment of the same size in the entire pool of Atlantic Forest fragments, a 823 probability obtained from a lognormal distribution fitted to the size distribution of the 824 ~250,000 forest fragments. This procedure was conducted separately for each 825 biogeographical region of the Atlantic Forest and then averaged across regions weighted 826 by the area of each region.

Forest descriptor Dbh cut-off Absolute loss (cm) (%) Forest biomass ≥5.0 25.3 ≥10.0 32.0 Tree species richness ≥5.0 30.9 ≥10.0 22.9 Species properties Wood density ≥5.0 2.2 Max. adult height ≥5.0 -1.9 Seed mass ≥5.0 35.7 Ecological group ≥5.0 32.6 Extinction level ≥5.0 25.2 Endemism level ≥5.0 42.1

827 828 bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

829 Table 2. The average loss of forest carbon and its projection to the remaining 830 Atlantic Forest area.

831 The proportion of carbon loss was obtained per biogeographical region and then 832 multiplied by the remaining forest area of each region to obtain the equivalent forest loss, 833 i.e., the Atlantic Forest area that would match the carbon losses caused by post- 834 deforestation human impacts. The equivalent carbon loss was computed based on the 835 equivalent forest loss and the reference values of carbon storage per biogeographical 836 region. We assumed a value of US$5 per Mg C paid for carbon credits obtained from 837 projects of forestry and land use.

Dbh Carbon Equivalent Equivalent Carbon cutoff loss (%) forest loss carbon loss credits (cm) (km2) (Tg C) (Billion US$) ≥5 25.3 55,082 451.3 2.257 ≥10 32.0 70,279 524.5 2.622 838 bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

839 Figures

840 841 Figure 1. The relationship between losses of species richness and forest biomass due 842 to human-induced impacts. 843 The standardized indices of loss were estimated for each forest survey (points), with 844 biomass on the x-axis and species on the y-axis. By the margin of each axis, the 845 distribution of the index is presented with the mean and the corresponding 95% 846 confidence interval (CI 95%). Dashed lines separate negative indices (losses) from 847 positive ones (gains). The index of loss is dimensionless and is highlighted by different 848 colours ranging from dark red (high losses) to blue (gains). Pearson's correlation 849 coefficient between the two indices was 0.22 (p< 0.0001). bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

850 851 Figure 2. Human-induced impacts on tree species richness, multiple species 852 properties and forest biomass. 853 Two-dimensional representation of the standardized indices of loss for each forest survey 854 (points) plotted on the two first principal component (PC) axes, which explain 64% of the 855 variation in the indices of loss. Arrows and their length represent the vectors of loss, i.e. 856 the direction and strength of each individual index of loss. The more aligned the arrows, 857 the more correlated the indices are. The average of all indices of loss is highlighted by a 858 colour scale for each survey ranging from dark red (high losses) to blue (gains). bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

859 860 Figure 3. The distribution of the indices of loss across the Atlantic Forest and the 861 effects of land conservation category. bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

862 Each point represents one forest survey for which the index of loss was estimated for 863 forest biomass (A and B), tree species richness (C and D) and species properties (E and 864 F). In panels A, C and E, the shaded area is the original extent of the Atlantic Forest. In 865 panels, B, D and F, the box-and-whisker diagrams summarize the distribution of the 866 indices of loss for each land-use category (i.e., bold centre line, median; box limits, upper 867 and lower quartiles; whiskers, 1x interquartile range), organized from left to right in 868 decreasing order of protection. The F-statistics and number of observations of the models 869 are presented, as well as the result of the Tukey test for difference among group means 870 (lowercase letters). Degrees of freedom: 1579 (panel B), 1698 (panel D) and 1150 (panel 871 F). Legend: Protect indicates Strict protection and Sustainable Use conservation units; 872 Other indicates other public and protected lands (e.g., research centres, university 873 campuses, botanical gardens, military and indigenous land); APA indicates private land 874 inside areas with sustainable use of natural resources, locally known as "Environmental 875 Protection Areas"; Private indicates private land. bioRxiv preprint doi: https://doi.org/10.1101/2020.09.08.287938; this version posted September 9, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

876 877 Figure 4. Multiple outcomes of two restoration scenarios for biogeographical region 878 of the Atlantic Forest. 879 In each radar plot, we provide the proportional gain of forest biomass, species richness, 880 wood density, maximum adult height, seed mass, ecological groups, extinction, and 881 endemism levels for the ‘Fragment restoration’ (green line) and ‘Landscape restoration’ 882 scenarios (red line).