Forest refuge areas and carbon emissions from tropical deforestation in the 21st century

Tropical forests are disappearing at an alarming rate due to human activities. Here, we provide spatial models of deforestation in 92 countries covering all the tropical moist forests in the world. Our models confirm the effectiveness of protected areas in displacing deforestation and the negative impact of roads and landscape fragmentation on forest conservation in the tropics. Using our models, we derive high-resolution pantropical maps of the deforestation risk and future forest cover for the 21st century under a “business-as-usual” scenario based on the deforestation rates observed in the 2010s. Although under this scenario, large areas of tropical moist forest should remain in the heart of the Amazon, in the Congo Basin, and in New Guinea in 2100, 48% (39–56%) of all forest cover is expected to disappear during the course of the 21st century, and many countries will have lost all their forests by 2100. The remaining forests will be highly fragmented and located in remote places. As future deforestation will concern forests with higher aboveground carbon stocks, annual carbon emissions associated with tropical deforestation are expected to increase by +0.161 Pg/yr (+35%) between the 2010s and the 2090s. 1

Forecasting forest cover change is paramount as it allows us 28 to foresee the consequences of deforestation (in terms of carbon 29 emissions, biodiversity loss, or water supply) under various tech-30 nological, political, and socio-economic scenarios, and to inform 31 decision makers accordingly (Clark et al. 2001). The "business-32 as-usual" scenario is of particular interest as it makes it possible 33 to predict the likely future in the absence of change in deforesta-34 tion rates, and if necessary, to alert decision-makers to an essential 35 change of course to avoid any potential environmental disaster. 36 While models and scenarios of carbon dioxide emission and climate  (Kremen et al. 2008, Mittermeier et al. 49 2011) vary considerably in space at fine scale. Non-spatial scenar-50 ios of forest cover change (FAO 2020) cannot be used to forecast 51 associated carbon emissions and change in biodiversity accurately, 52 or for systematic conservation planning at the local scale. Spatial 53 forecasts of forest cover change are based on spatial statistical mod-54 els, which enable estimation of a probability of change in space as a 55 function of a set of spatial predictors (Rosa et al. 2014). In addition 56 to forecasts, statistical models can be used to identify the main 57 drivers of deforestation and to quantify their relative effects. For 58 example, models can be used to assess the impact of roads on the 59 risk of deforestation (Laurance et al. 2014) and the effectiveness of

Significance Statement
Under a "business-as-usual" scenario of deforestation (i.e. projecting the 2010-2020 deforestation rates at the country level in the future), three quarters of the tropical moist forests that existed in 2000 will have disappeared around years 2120, 2160, and 2220 in Southeast Asia, Africa, and Latin America, respectively. By 2100, 41 tropical countries, plus 14 states in Brazil and one region in India, will lose all their tropical forests. Remaining forests will be highly fragmented and concentrated in remote areas (far from roads and towns), preferentially in protected areas, and at high elevations. Future deforestation will concern forests with higher carbon stocks. In the absence of change in the deforestation rates, annual carbon emissions associated with tropical deforestation will increase from 0.467 Pg/yr in the 2010s to 0.628 Pg/yr in the 2090s (+35%), making tropical forests a major carbon source in the 21 st century. Maps show the predicted change in tropical moist forest cover in the three continents (America, Africa, and Asia) for the period 2020-2100 under a business-as-usual scenario of deforestation. The horizontal black line represents the Equator. The boundaries of the study area are represented by dark grey lines. For the deforestation projections, we assumed no diffusion of the deforestation between countries. Forest areas in red are predicted to be deforested in the period 2020-2100, while forest areas in green are likely to still exist in 2100. Several countries on the three continents are expected to lose all their tropical moist forests by 2100 (including Nicaragua and Mexico in Central America, Madagascar and Ghana in Africa, and Laos and Vietnam in Asia). We predict progressive fragmentation of the remaining forest in the future, with an increasing number of isolated forest patches of smaller size (e.g., Pará state in Brazil, the Democratic Republic of the Congo, and Indonesia). These maps make it possible to identify both future hotspots of deforestation and forest refuge areas (e.g., concentrated in the heart of the Amazon, West Central Africa, and Papua New Guinea). An interactive map is available at https://forestatrisk.cirad. fr/maps.html. al. 2006, Swann et al. 2015, Aguiar et al. 2016. In this paper, 67 we model and forecast deforestation at the pantropical scale using 68 high-resolution spatial data. This was made possible by the recent 69 availability of pantropical spatial datasets of forest cover change and OpenStreetMap). We combine these extensive datasets in a 74 spatial statistical model to test the effectiveness of protected areas at 75 reducing deforestation and to assess the impact of roads on the risk 76 of deforestation at the pantropical scale. Assuming a business-as-77 usual scenario, we derive high-resolution maps of deforestation risk 78 and future forest cover over the 21 st century in the humid tropics. 79 We also estimate the carbon emissions associated with projected 80 deforestation and conduct an uncertainty analysis.

82
Using the study by Vancutsem et al. (2021) 92 We observed marked differences in the percentage of forest cover 93 loss at the continental and country scales ( Fig. 2  deforested area is lower (1.9 Mha/yr), this percentage would be 99 53% (45-60%). In Latin America, where the annual deforested 100 area is higher (2.5 Mha/yr), but where the remaining tropical moist 101 forest in 2020 is also much larger than in Southeast Asia and Africa 102 (621 Mha), this percentage would be 38% (31-45%). Under a 103 business-as-usual scenario of deforestation, three quarters of the 104 tropical moist forests that existed in 2000 will have disappeared 105 around years 2120, 2160, and 2220 in Southeast Asia, Africa, and 106 Latin America, respectively, with an average uncertainty of ±45 107 years ( Fig. 2 and Table 1).

108
At the country scale, we predict that 41 countries (16 in Latin 109 America, 21 in Africa, and four in Southeast Asia) out of the 92 we 110 studied, plus 14 states in Brazil and one region in India, will lose 111 all their tropical forests by 2100 (Fig. 1). Among these countries 112 or regions, 19 countries (six in America, ten in Africa, and three 113 We provide past and predicted forest cover for the three continents and for the three countries with the highest forest cover in 2010 for each continent (Brazil in America, the DRC in Africa, and Indonesia in Asia). Past forest cover areas (in thousand hectares, Kha) refers to their status on January 1 st 2000, 2010, and 2020 ("fc2000", "fc2010", and "fc2020", respectively). We provide the mean annual deforested area d (Kha/yr) for the last ten-year period from January 1 st 2010 to January 1 st 2020, and the corresponding mean annual deforestation rate p (%/yr). Projected forest cover areas are given for the years 2050 and 2100 ("fc2050" and "fc2100"). Projections are based on the forest cover in 2020 ("fc2020") and the mean annual deforested area (d) assuming a business-as-usual scenario of deforestation. Column "loss21" indicates the projected percentage of forest cover loss during the 21 st century (2100 vs. 2000). We estimate the year ("yr75") at which 75% of the forest cover in 2000 will have disappeared. from year 2020 to 2400 per continent. For the deforestation projections, we assumed no diffusion of the deforestation between countries. As a consequence, when large countries with high annual deforested areas (Brazil for America, DRC for Africa, and Indonesia for Asia) have no more forest (in 2264, 2181, and 2144, respectively, see SI Appendix, Table S16), deforestation at the continent scale is rapidly decreasing. The horizontal black line indicates a loss of 75% of the forest cover in comparison with the year 2000. Under a business-asusual scenario, this should happen in 2117, 2163, and 2220 for Asia, Africa, and America, respectively. The confidence envelopes around the mean are obtained using the lower and upper bounds of the confidence intervals of the mean annual deforested areas for all study areas.
in Asia), three states in Brazil, and one region in India had more Western Ghats and Sri Lanka (Fig. 1).

127
Using a spatial statistical modelling approach (see SI Appendix, 128 Materials and Methods, Figs. S1-S9, and Tables S1-S13), we obtain , and Tables S14-S17). Three large 133 "blocks" of relatively intact tropical moist forest will remain in 2100 134 ( Fig. 1). One forest block will be located in Latin America and 135 will include the upper part of the Amazonian basin (including the 136 Peruvian, Ecuadorian, Colombian and Venezuelan Amazonia) and 137 the Guiana Shield (Guyana, Suriname, and French Guiana). The 138 second block will be located in the western part of the Congo basin 139 and will include forests in Gabon, Equatorial Guinea, Cameroon, 140 the Central African Republic, and the Republic of Congo. The third 141 block will be located in Melanesia and will include forests in Papua 142 New Guinea, Solomon Islands, and Vanuatu.

143
Apart from these three large and relatively intact forest blocks, 144 the tropical moist forest remaining in 2100 will be highly fragmented 145 (Fig. 1). In Latin America, highly fragmented forests will be found 146 in the Brazilian states of the Amazonian deforestation arc (Acre, 147 Rondonia, Mato Grosso, Para, Amapa) and in the Roraima state 148 in the northern Amazonia. In Africa, forests in the Democratic 149 Republic of the Congo (DRC) will also be highly fragmented (SI 150 Appendix, Fig. S11) and will be completely separated from the large 151 forest block located in the western part of the Congo basin ( Fig. 1). 152 In Southeast Asia, small patches of heavily fragmented forests will 153 remain in Thailand, Indonesia, and the Philippines (Fig. 1). The 154 remaining forests will be concentrated in remote areas (far from 155 roads and towns), preferentially in protected areas, and at high 156 elevations (Figs. 1, 3 and SI Appendix, Tables S4-S9). For example, 157 the remaining forests of Borneo will be concentrated in the Betung 158 Kerihun and Kayan Mentarang National Parks.

159
As tropical forests shelter a large proportion of terrestrial bio-160 diversity and carbon stocks on land, future tropical deforestation 161 is expected to have strong negative impacts on both biodiversity 162 and climate. The impact of projected deforestation on carbon emis-163 sions is discussed below, but rigorous assessment of the impact of 164 projected deforestation on biodiversity is beyond the scope of this 165 study. Such an impact analysis would require accurate species dis-166 tribution and biodiversity maps including a large number of species 167 representative of the biodiversity in the tropics. Such maps are 168 not available to date (Pimm et al. 2014). Nonetheless, as a rough 169 estimate, if we consider only endemic species (Mittermeier et al. 170 2011) in the six biodiversity hotspots where almost all the tropical 171 forest is predicted to disappear by 2100, and assume that most of 172 these species depend on tropical moist forests, deforestation would 173 lead to the extinction of 29,140 species of plants and 4,576 species 174 of vertebrates (including birds, reptiles, amphibians, freshwater 175 fishes, and mammals) which cannot be found anywhere else on 176 Earth (SI Appendix, Table S18). Upper panels: Maps of the spatial probability of deforestation at 30 m resolution for the three continents. Maps of the spatial probability of deforestation at the level of the study area were aggregated at the pantropical level. The horizontal black line represents the Equator. The boundaries of the study area are represented by dark grey lines. Coloured pixels represent forest pixels for the year 2020. Inside each study area, forest areas in dark red have a higher risk of deforestation than forest areas in green. Lower panels: Detailed maps for three 100 × 100 km regions (black squares in the upper panels) in the Mato Grosso state (Brazil), the Albertine Rift mountains (the Democratic Republic of the Congo), and the West Kalimantan region (Borneo Indonesian part). Deforestation probability is lower inside protected areas (black shaded polygons) and increases when the forest is located at a distance closer to roads (dark grey lines) and forest edge. An interactive map of the spatial probability of deforestation is available at https://forestatrisk.cirad.fr/maps.html.  (Fig. 4).

197
The predicted increase in annual carbon emissions is explained 198 by the fact that the forests which will be deforested in the future have 199 higher carbon stocks. Several studies have shown that elevation is 200 an important variable in determining forest carbon stocks (Saatchi 201 et al. 2011, Vieilledent et al. 2016, Cuní Sanchez et al. 2021).

202
Forest carbon stocks are expected to be optimal at mid-elevation 203 (Vieilledent et al. 2016) due to higher orographic precipitation at 204 this elevation and because the climatic stress associated with winds 205 and temperature is lower at mid-elevation than at high elevation.

206
Here, we show that low-elevation areas are more deforested than 207 high-elevation areas (SI Appendix, Tables S4, S5). This is explained 208 by the fact that low-elevation areas are more accessible to human 209 populations and by the fact that arable lands are concentrated at 210 low elevation, where the terrain slope is usually lower and the soil 211 is more productive (Geist and Lambin 2002). Consequently, the 212 predicted increase in carbon emissions can be explained by the 213 fact that deforestation will move towards higher elevation areas 214 where forest carbon stocks are higher. Moreover, remote forest areas 215 that have been less disturbed by human activities in the past have 216 accumulated large quantities of carbon (Dargie et al. 2017, Brinck 217 et al. 2017. The progressive deforestation of more intact forests 218 also explains the predicted increase in carbon emissions.

219
The decrease in carbon emissions predicted from the period 220 2070-2080 for Southeast Asia, and from the period 2090-2100 at 221 pantropical scale, can be associated with a decrease in carbon stocks 222 of deforested areas (in association with environmental factors, such 223 as lower carbon stocks at very high elevation) or a decrease in the 224 total deforested area at the continental and global scale, as countries 225 progressively lose all their forest. In Southeast Asia, four countries 226 will lose all their forest between 2070 and 2110 (SI Appendix, Table 227 S16). These countries (which include Laos, Myanmar, and Vietnam) 228 account for a significant proportion (20%) of the annual deforested 229 area in Southeast Asia (407,498 ha/yr out of 2,001,803 ha/yr, see 230 SI Appendix, Tables S14, S15). This largely explains the predicted 231 decrease in carbon emissions in Southeast Asia from 2070 on.

275
Here we show that protected areas significantly reduce the risk 276 of deforestation in 70 study areas out of 119 (59% of the study 277 areas). These 70 study areas accounted for 88% of the tropical 278 moist forest in 2010 (SI Appendix, Table S6). On average, protected 279 areas reduce the risk of deforestation by 40% (Figs. 3, 5 and SI 280 Appendix, Table S5). This result clearly demonstrates the efficiency 281 of protected areas at reducing the spatial risk of deforestation in 282 the tropics. In a recent global study, Wolf et al. (2021) found that 283 protected areas reduced deforestation rates by 41%, close to the 284 40% we find here by focusing on tropical moist forests. Most of 285 the previous studies have assessed the effect of protected areas at 286 reducing deforestation in particular countries or regions (Bruner 287 2001, Andam et al. 2008 or at efficiently protecting a particular 288 group of species (Cazalis et al. 2020). Studies at the global scale 289 (Wolf et al. 2021, Yang et al. 2021 were at 1 km resolution and 290 used spatial matching methods and tree cover loss data (Hansen 291 et al. 2013). Our pantropical approach is based on more accurate 292 forest cover change maps in the humid tropics, in particular in Africa 293 (Vancutsem et al. 2021), and accounts for fine scale deforestation 294 factors acting at a much smaller distance than the distance imposed 295 by a 1 km resolution (see the effect of the distance to forest edge 296 discussed below). Moreover, contrary to most spatial matching 297 methods (Andam et al. 2008, Schleicher et al. 2019, the statistical 298 model we use allows us to account for any potential confounding 299 variables which might skew the estimated effect of protected areas. 300 Like other studies reporting the effect of protected areas on 301 deforestation, our study demonstrates that protected areas are ef-302 fective at displacing deforestation outside protected areas in tropical 303 countries, but not necessarily that protected areas play a role in 304 reducing the deforestation intensity per se. Indeed, the factors that 305 drive the intensity of deforestation at the country scale are more 306 socio-economic or political, such as the level of economic develop-307 ment, which determines people's livelihood and the link between 308 people and deforestation (Geist and Lambin 2002), the size of the 309 population (Barnes 1990), or the environmental policy (Soares-310 Filho et al. 2014). In tropical countries with weak governance 311 (where environmental law enforcement is low) and with a low level 312 of development (where the pressure on forest is high), it is very 313 unlikely that protected areas will remain forested. Under a business-314 as-usual scenario of deforestation, we assume that the deforestation 315 intensity will remain constant over time. When all the forest outside 316 the protected areas is deforested, deforestation is expected to occur 317 inside protected areas (Fig. 1). In this scenario, protected areas are 318 efficient at protecting forest areas of high and unique biodiversity in 319 the medium term, i.e., forests will be concentrated in protected ar-320 eas, where the probability of deforestation is lower. In the long term, 321 under a business-as-usual scenario, forests should completely disap-322 pear from protected areas while deforestation continues (Fig. 1). 323 This phenomenon is already clearly visible in countries or states 324 where deforestation is advanced, such as in Rondonia state (Brazil) 325 in South America (Ribeiro et al. 2005), Ivory Coast (Sangne et 326 al. 2015) or Madagascar (Vieilledent et al. 2020) in Africa, or 327 Cambodia (Davis et al. 2015 in Southeast Asia. In these countries, 328 several forested protected areas have been entirely deforested (e.g., 329 the Haut-Sassandra protected forest in Ivory Coast, or the PK-32 330 Ranobe protected area in Madagascar) or severely deforested (e.g., 331 the Beng Per wildlife sanctuary in Cambodia).

332
Impact of roads and distance to forest edge on the deforesta-333 tion risk 334 Here we find that a longer distance to the road significantly reduces 335 the risk of deforestation in 61 study areas out of 119 (51% of the 336 study areas). These 61 study areas accounted for 90% of the tropical 337 moist forest in 2010 (SI Appendix, Table S7). On average, a distance 338 of 10 km from a road reduces the risk of deforestation by 14% 339 (Figs. 3, 5 and SI Appendix, Tables S5, S9). This said, opening a 340 The dots represent the observed mean probability of deforestation in each forest protection class, either protected or unprotected. Bars represent the mean of the predicted probabilities of deforestation obtained from the deforestation model for all observations in each class. Right: The dots represent the local mean probability of deforestation for each bin of 10 percentiles for the distance. Lines represent the mean of the predicted probabilities of deforestation obtained from the deforestation model for all observations in each bin. Note that for distance to forest edge, the first dot accounts for three bins while for distance to road, bins for a distance > 23 km are not shown. For both left and right panels, confidence intervals for predictions were to small to be represented because of the high number of observations per class and bin. road in the forest leads to the creation of two forest edges and 341 computing a distance from a forest pixel to the nearest road implies 342 the existence of a distance to the forest edge. When studying the 343 effect of roads on deforestation, it is thus impossible to neglect the 344 effect of the distance to forest edge on the risk of deforestation.

345
Here, we find that the distance to the forest edge is the most given the current global context, the business-as-usual deforesta-372 tion scenario we examine here appears to be rather conservative. for the increasing demand for agricultural commodities from the 379 tropics, such as palm oil, beef and soybean, which will likely lead 380 to a significant increase in deforestation (Karstensen et al. 2013, 381 Strona et al. 2018). Our projections using high estimates of the 382 annual deforested area for each study area, corresponding to a total 383 deforestation of 8.3 Mha/yr at the pantropical scale, give an indica-384 tion on the consequences of a 30% increase in the annual deforested 385 area in the future. This would lead to a 56% loss of tropical moist 386 forest cover over the 21 st century and to a much faster increase in 387 carbon emissions, up to 0.793 Pg/yr in the 2070s, corresponding to 388 a +70% increase in annual carbon emissions compared to the 2010s 389 (Fig. 4). The percentage of emerge lands covered by tropical moist 390 forests would then drop to 3.7% (554 Mha) in 2100 (SI Appendix, 391 Fig. S14).

392
Although some conservation strategies, such as protected areas, 393 can help save some time in the fight against deforestation (being ef-394 ficient at displacing deforestation toward areas of lower biodiversity 395 or carbon stocks), it is extremely urgent to find political and socio-396 economic solutions that are efficient at curbing deforestation in the 397 long term. Several initiatives involving actors from the political 398 and economical world have already been taken to this end, without 399 having so far led to a significant decrease in deforestation rates in 400 the tropics (Vancutsem et al. 2021). Such initiatives include recent 401 national or multinational strategies against imported deforestation 402 (Bager et al. 2021), certification schemes for private companies 403 providing agricultural commodities such as the Roundtable on Sus-404 tainable Palm Oil (Cazzolla Gatti and Velichevskaya 2020), or the 405 REDD+ mechanism (Goetz et al. 2015). The results and products 406 of our study could facilitate the concrete implementation of these 407 initiatives on the ground and help increase their effectiveness. In 408 particular, our deforestation probability map could be used to mon-409 itor areas identified as having a high risk of being deforested. Our 410 projections by country could also be used as reference scenarios of 411 deforestation and associated carbon emissions which are necessary 412 for implementing REDD+ at a wide scale on the basis of a common 413 methodology. Doing so, we hope to contribute to the fight against 414 deforestation and that our map of tropical forest cover projected in 415 2100 will never become a reality.

417
We present below a summary of the materials and methods used in this 418 study. A detailed description can be find in the SI Appendix, Materials and 419 Methods.  Table S2). . CC-BY 4.0 International license made available under a (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 The copyright holder for this preprint this version posted March 25, 2022. ;https://doi.org/10.1101https://doi.org/10. /2022 in the period 2010-2020. Pixels in each category were sampled randomly (SI 448 Appendix, Fig. S7).  Table S3). Eq. S1). To account for the residual spatial variation in the deforestation 460 process, we included additional spatial random effects for the cells of a 10 × 461 10 km spatial grid covering each study-area (SI Appendix, Fig. S8). Spatial the year 2020 for each study-area (SI Appendix, Fig. S10). For each study-494 area, we also estimated the mean annual deforested area (in ha/yr) for 495 the period 2010-2020 from the past forest cover change map (SI Appendix, 496 Tables S14-S15). Using the mean annual deforested area in combination 497 with the spatial probability of deforestation map, we forecasted the forest 498 cover change on the period 2020-2110 with a time step of 10 years, assuming 499 a "business-as-usual" scenario of deforestation (SI Appendix, Fig. S11 and 500 Tables S16-S17). The business-as-usual scenario makes the assumption of 501 an absence of change in both the deforestation intensity and the spatial 502 deforestation probability in the future.

503
Impacts on biodiversity loss and carbon emissions. We estimated the num-504 ber of endemic plant and vertebrate species committed to extinction because 505 of the complete loss of tropical forest by 2100 in 6 biodiversity hotspots 506 (SI Appendix, Table S18). We estimated the carbon emissions associated 507 with past deforestation (2010-2020) and projected deforestation (2030-508 2110) using Avitabile's (2016) pantropical 1 km resolution aboveground 509 dry biomass map (SI Appendix, Fig. S12 and Table S19). We used the IPCC  Table S20). We thus obtained three different predictions of the forest cover 522 change and associated carbon emissions: an average prediction considering 523 the mean annual deforested area, and two additional predictions considering 524 the lower and upper bound estimates of the mean annual deforested area 525 per study area (SI Appendix, Figs. S13-S14, and Data S1, S2).

526
Software. To perform the analyses, we used the forestatrisk Python 527 package (Vieilledent 2021) which has been specifically developed to model 528 and forecast deforestation at high resolution on large spatial scales (SI 529 Appendix, Materials and Methods).