Large carbon sink potential of Amazonian Secondary 1 Forests to mitigate climate change 2

17 Secondary forests (SF) have a large climate mitigation potential, given their ability to sequester 18 carbon up to 20 times faster than old-growth forests. Environmental variability and 19 anthropogenic disturbances lead to uncertainties in estimating spatial patterns of SF carbon 20 sequestration rates. Here we quantify the influence of environmental and disturbance drivers on 21 the rate and spatial patterns of regrowth in the Brazilian Amazon, by integrating a 33-year land 22 cover timeseries with a 2017 Aboveground Biomass dataset. Carbon sequestration rates of young 23 Amazonian SF (<20 years old) are at least twice as high in the west (3.0±1.0 MgC ha -1 yr -1 ) than 24 in the east (1.3±0.3 MgC ha -1 yr -1 ). Disturbances reduce SF regrowth rates by 8–50% (0.6 – 1.3 25 MgC ha -1 yr -1 ). We estimate the 2017 SF carbon stock to be 294 TgC, which could be 8% higher 26 by avoiding fires and repeated deforestation. Maintaining the 2017 SF area has the potential to 27 accumulate ~15 TgC yr -1 until 2030, contributing ~5% to Brazil’s 2030 net emissions reduction 28 target. Supporting SF and old-growth forests conservation alongside the expansion of SF in 29 deforested areas is therefore a viable nature-based climate mitigation solution. Our study has quantified the varied and complicated regrowth rates of SF across Amazonia influenced by multiple drivers across Amazonia. Given the uncertain and potentially threatened status of old- growth forest sinks due to ongoing climate change 4 , it is imperative to limit human-induced fire and deforestation disturbance in both old-growth and SF. By preserving the remaining old-growth forest stock and sustainability managing SF we can maintain and increase the carbon sink of this globally important biome and help it to achieve its climate mitigation potential.


Introduction 31
Global forests are expected to contribute a quarter of the pledged mitigation under the 2015 Paris 32 Agreement, by limiting deforestation and by encouraging forest regrowth 1 . The Brazilian Amazon 33 biome (Amazonia) is the largest continuous tropical forest on Earth, occupying 3% of terrestrial land. 34 It stores approximately 10% of the global forest carbon (120 Pg C) 2,3 and between 2000 and 2010 35 sequestered ~150 Tg C yr -1 through natural growth (5% of global land sink), while emitting ~143±56 36 Tg C yr -1 through deforestation (~1.4% of global carbon emissions) 4-6 . As part of their Nationally 37 Determined Contributions (NDC) to the Paris Agreement, Brazil has pledged to restore and reforest 12 38 million hectares of forests by 2030 to contribute to net emission reductions 7 . Part of this reduction can 39 be achieved by the natural regeneration of secondary forest (SF) on abandoned land, which are already 40 regrowing on ~20% of deforested land in Amazonia [8][9][10] . 41 Previous estimates of average net carbon uptake in young (< 20 years old) SF range between 2.95± 0.4 42 and 3.05± 0.5 Mg C ha -1 yr -1 , 11-20 times larger than old-growth primary forests 11,12 . These estimates, 43 which are based on limited field data across the Neotropics, are unable to capture the different spatial 44 patterns and rates of SF carbon sequestration which are influenced by several drivers. This includes 45 environmental drivers such as shortwave radiation, precipitation, soil fertility and forest water deficit, 46 as well as anthropogenic disturbances like fire and deforestation cycles 11,13-16 . The SF carbon stock of 47 regions with very high-water deficit (-1,200 mm yr -1 ) can be up to 85% lower compared to no water 48 deficit (0 mm yr -1 ) regions in the Neotropics 11 . The effects of these drivers are not limited to SF growth, 49 nor are they static over space and time, affecting the magnitude of forest carbon sequestration and 50 stocks 17 . A recent study showed that rising annual mean temperatures and drought reduced tree growth 51 in Amazonian old-growth forests 4 . This effect, coupled with ongoing deforestation suggests that the 52 sink in these forests peaked in the 1990s and is now steadily declining 4 . Considering these changes, it 53 is important to also obtain a wider spatial and temporal understanding of drivers affecting the magnitude 54 and sustainability of SF regrowth. 55 Remote sensing products can be used to study these effects, offering broad spatial and temporal 56 coverage. With the availability of nearly four decades of Landsat data (30m spatial resolution), it is now 57 possible to track the fate of deforested areas over time, which includes the changing demography of SF 58 across Amazonia 10,18 . According to satellite-based analysis, SF are typically part of a 5-10 year cycle of 59 clearance and abandonment since they are currently not protected by national policies aimed at curbing 60 deforestation 19,20 . These repeated deforestations are expected to decrease the carbon sink of future 61 regrowth forests. Deforestation of SF amounted to ~70% of total Amazonian forest loss between 2008 62 and 2014 21 . However, the relationship between SF regrowth and environmental and disturbance drivers 63 has never been explored spatially-explicitly using global remote sensing products. 64 Here we aim to produce unique estimates of SF regrowth by constructing spatially explicit models based 65 on multi-satellite products to quantify the carbon sequestration potential of Amazonian SF exposed to 66 multiple environmental and anthropogenic disturbance drivers. We use a novel approach to map SF 67 annually from 1985 to 2017 and determine their ages 10,18 , and provide the first applications of these 68 maps to analyse SF regrowth in terms of Aboveground Carbon (AGC) 22-24 . We present a map of 69 Amazonian SF regrowth rates with the quantification of the contemporary SF carbon sink considering 70 the impact of different drivers on AGC accumulation. We use this to model the future carbon 71 sequestration potential of SF relative to the Brazilian NDC targets. 72

Impact of drivers on Secondary Forest regrowth 74
We used the land cover product MapBiomas (  Tables 2 -7). After forest age, SW radiation is the most 84 important variable influencing AGC (Figure 1g). In areas of very low annual SW radiation (<170 Wm -85 2 ), the overall regrowth rate is almost three times greater compared to areas of high SW radiation (>187 86 Wm -2 ), ~3.4±0.6 and ~1.3±0.4 Mg C ha -1 yr -1 , respectively. MCWD was the second most important 87 driver, where areas with very low MCWD (> -180 mm yr -1 ) assimilate almost double the carbon 88 compared to areas with very high MCWD (< -350mm yr -1 ) in the first 20 years of regrowth (~2.7±0.7 89 Mg C ha -1 yr -1 and ~1.5±0.2 Mg C ha -1 yr -1 , respectively). Similar differences in the regrowth rates can 90 be observed under conditions of low mean annual precipitation (<1920 mm yr -1 ) compared to moderate 91 and high conditions (1920-2210 mm yr -1 and >2210 mm yr -1 , respectively). There was no statistical 92 difference in carbon accumulation under different SCC conditions, furthermore the expected trend, 93 increased carbon accumulation with increased soil fertility, is reversed, probably due to the dominant 94 effect of the other environmental drivers 31,32 (Figure 1d; Supplementary Table 4). 95 For most of our modelled regrowth curves, SF were able to reach AGC levels equivalent to those of 96 old-growth forests, however the time taken to reach these levels is generally more than a century 97 (Supplementary Table 8). Our results also show that in areas of anthropogenic disturbance such as fires 98 and repeated deforestations, the carbon accumulation was up to 3.8 times slower and even plateaued 99 within 20 to 40 years, thus potentially never recovering to old-growth forest AGC values (Figure 1e  were the least important drivers for modelling AGC regrowth across the entire biome. This is in part an 102 artefact of the small number of SF plot being exposed to multiple fires (28.2%) and repeated 103 deforestations (11.3%) (Supplementary Figure 8). 104

105
Mapping the spatial patterns of regrowth 106 To analyse the spatial variation of regrowth rates in our models, we identified different regions of 107 Amazonia according to the three most important environmental drivers influencing carbon 108  Table 9). In contrast, the eastern and southern parts have slower overall regrowth rates (1.3±0.3 -117 1.8±0.3 Mg C ha -1 yr -1 in the first 20 years) with fire and deforestation disturbances reducing their 118 regrowth by around 50% to as low as 0.6 Mg C ha -1 yr -1 in the first 20 years (Figure 3b d). In the 119 North-East and South-Western regions fire disturbance is the third and second most important driver 120 respectively to influence the AGC (Figure 2c and d). 121 We validated our models with field AGC estimates of SF collected across Amazonia (284 samples 122 across 33 locations) and found that our AGC estimates are statistically similar (p > 0.01) within the four 123 regions identified in Figure 2a (Supplementary Figure 10). We also compared the regional models with 124 basin-wide models used in previous studies and within the Brazilian Greenhouse Gas Inventory, which 125 do not consider different environmental or anthropogenic disturbance drivers 21,33,34 . In the western 126 regions, during the first 10 years of growth, our models of 'no disturbance' were visually very similar 127 to the other models (Supplementary Figure 11). We found no significant difference to AGC estimates 128 from the model used in previous research (p > 0.01; Supplementary Table 11). Estimates using the 129 equation from the Brazilian Greenhouse Gas inventory were significantly higher across the 40 years 130 modelled in all four regions with disturbed areas having significantly lower regrowth rates (p < 0.01; 131 the two years to apply the relevant regrowth model seen in Figure 3. Our results show that new 138 regenerating forests and existing SF combined resulted in a carbon sink of 28 Tg C yr -1 , at the expense 139 of 16.1 Tg C yr -1 emitted from SF loss, resulting in a net SF carbon sink of ~12 Tg C yr -1 (Figure 4). 140 We find the total carbon stored in all Amazonian SF in 2017 to be approximately 294Tg C (Figure 4d). 141 We also estimate that the potential carbon stock if all SF had regrown without experiencing any 142 disturbances, namely fire and repeated deforestations, could have reached 320 Tg C in 2017. 143  Table 9 for quantitative interpretations of the qualitative definitions given here, for example "Low precipitation".
Finally, to quantify the potential of the existing 2017 SF to contribute to reducing future net carbon 144 emissions according to Brazil's NDC, we model future potential stocks and annual carbon sink for the 145 decade ahead by considering various levels of preservation ( Figure 5). In 2025 we project an 82% 146 difference in carbon accumulation between the most ambitious preservation plan (preserving all 13.8 147 Mha of SF) and the least ambitious plan (preserving 2.2 Mha including only SF older than 20 years in 148 2017; Figure 5b). 149  In this study, we quantified the impact of environmental and anthropogenic disturbance drivers on 151 carbon accumulation in Amazonian SF. SW radiation was the most important driving variable, with 152 low SW radiation observed in western Amazonia (~163.6 Wm -2 ) (Supplementary Figure 7) having the 153 highest regrowth rates ranging between 2.4±0.8 to 3.0±1.0 Mg C ha -1 yr -1 . These estimates are similar 154 to the previous estimates of 2.95± 0.4 and 3.05± 0.5 Mg C ha -1 yr -1 11,12 . The higher estimated regrowth 155 rates in areas of lower SW radiation is likely linked to higher cloud cover resulting in more diffuse 156 radiation and lower vapour pressure deficit (Figure 1 and 3a). Diffuse radiation can penetrate deeper 157 into closed forest canopies than direct shortwave radiation and enhance productivity and thereby carbon 158 sequestration 13, 35 , whilst a lower vapour pressure deficit encourages leaf stomata to remain open, 159 maximising productivity and thereby regrowth 36 . 160 Additionally, there are synergies between the drivers that influence the regrowth of SF (Figure 1). For 161 example, in the South-East and Northern regions, regrowth rate is approximately 50% lower compared 162 to western regions, likely due to the hydro-climatic conditions which reduce growth (low precipitation, 163 ~1913 mm yr -1 ; very high MCWD, ~-325.5 mm yr -1 ; moderate SW radiation, ~181.7 Wm -2 ). In turn, 164 this results in an environment that is drier and more susceptible to burning, reducing regrowth rates 165 even further (Figure 3d). Previous field-based studies have estimated the reduction in regrowth due to 166 fire to be 50% (reducing from 3.2 to 1.7 Mg C ha -1 yr -1 ) 16 , which is similar to the average reduction 167 estimated in our study (40%). With our method we were able to provide additional information 168 disaggregated by regions, showing that the regrowth rate in the North-Western and South-Western 169 regions SF exposed to fire were 20% and 60% lower, respectively, compared to non-disturbed SF 170 (Supplementary Table 9). The interactions between the drivers and the impact this has on the regrowth 171 rates has never been spatially quantified until now. 172 Across Amazonia, fire and repeated deforestations were evaluated as the least important drivers ( Figure  173 1g). Nonetheless, several other studies have shown that the importance of these drivers is not 174 negligible 16,37,38 , and that the perceived lack of importance may be a local-scale artefact which our model 175 cannot account for in the large environmental regions identified in this study (Figure 2a). Environmental 176 drivers act on regional scales and influence forest type and species physiology. Fire and deforestation 177 act on the local scale by reducing the seed bank, natural biodiversity, soil nutrient, and water 178 availability, which can cause arrested succession (a disturbance preventing the natural successional 179 growth) 39 . Indeed, we see evidence of arrested succession in the slow growth (up to 80% lower) and 180 early plateau in AGC (12 -25 years) in some regions that experienced successive disturbance and sub-181 optimal environmental conditions (Figure 3b and d; Supplementary Table 9). Regions subjected to 182 burning and repeated deforestations that do not reach AGC levels equivalent to those of old-growth 183 forests, highlight that both drivers are much more influential than our model can infer. Additionally, 184 the spatial extent of fire disturbance is likely to be more widespread than presented in our study, as the 185 remote sensing product, based on automatic detection, used in this study underestimates burnt area by 186 ~25% compared to manual photointerpretation methods 40 . 187 Given that our study consisted of 32 years of secondary forest data and one year of AGC data, each of 188 which has associated uncertainties, we take caution with the regrowth rates modelled much beyond this 189 period ( times more rapidly, we estimated the minimum time taken to reach old-growth forest AGC to be ~100 194 years (Supplementary Tables 8 and 9). SF will therefore never replace old-growth forests on policy-195 relevant timescales, stressing the continued need to conserve existing old-growth forests 196 (Supplementary Table 9) 43 . 197 Furthermore, the threat of forest water deficit and, consequently, drought-induced fire disturbances are 198 predicted to increase into the 21 st century due to ongoing climate change 44 . If this kind of climate-199 scenario arises, the reduced regrowth rate of the secondary forest as seen in the South-East region in 200 our analysis is likely to be more widespread and severe (Figures 1-3). This would threaten the 201 permanence of the carbon sequestration potential of SF as we have calculated in this study 17 . Given that 202 some degree of 21 st century climate change is now already out of human control, it is imperative to limit 203 anthropogenic disturbances, such as fire and deforestation. Overall, we estimate these disturbances to 204 have contributed to an 8% reduction in the total potential 2017 carbon stock (Figure 4d), with the highest 205 relative reduction (11%) in North-Eastern Amazonia (Supplementary Figure 12). This has important 206 implications for policies concerning human-induced burning regimes as well as the deforestation of 207 secondary forests. Our analysis has shown that avoiding these actions increases the regrowth potential 208 of SF and will ultimately help Brazil to achieve its NDC goals of reducing net national emissions by 209 37% in 2025 and 43% in 2030 compared to 2005 levels 7 . This amount is equivalent to net emissions of 210 354Tg C yr -1 (1.3GtCO2e yr -1 ) and 327Tg C yr -1 (1.2GtCO2e yr -1 ), respectively. We model the future 211 carbon sequestration rate by preserving all standing SF and find that the annual carbon accumulation 212 would be equivalent to providing an additional 5±1% reduction to the 2025/2030 emissions target 213 (Figure 5b). Conversely, if only SF older than 20 years in 2017 were preserved, the additional mitigation 214 potential would reduce to less than 1% (Figure 5b). The modelling shows that various levels of SF 215 preservation can contribute significantly to Brazil reaching its NDC targets. However, these estimates 216 assume that future rates of deforestation in SF and old-growth forests remain sustainable.

Conclusions 224
Our model results have the potential to benefit both the carbon modelling and carbon-policy 225 communities to help understand the regional variations of regrowth under different drivers. The carbon 226 modelling community will benefit from the ability to spatially monitor carbon dynamics, which can be 227 incorporated into models and scenarios of land cover and climate change. Additionally, the methods 228 used in this study can be developed further to include other important variables that influence regrowth. 229 This includes variables such as the type of previous land use practises (livestock, agriculture, and 230 forestry) and the period of active land use before abandonment. For instance, SF regrow 38% faster on 231 land used for agriculture than those for cattle pastures 37,46 . Our models will benefit carbon-policy 232 communities by helping to assess locations for restoring and reforesting 12Mha of forests, as proposed 233 by Brazil's NDC, that would maximise regrowth and thereby be most beneficial to mitigating climate 234 change. This includes areas with limited anthropogenic disturbances, which will minimise forest 235 restoration and thereby costs of implementation and conservation. Additionally, the results can be used 236 to improve monitoring under the Reducing Emissions from Deforestation and Degradation (REDD+) 237 scheme. This approach would not be limited to Amazonia and could be applied in other countries where 238 field data may be limited. 239 A wide range of remote sensing products can be used to monitor SF change, and more are in 240 development. Large-scale single-date AGC products, such as the ESA-CCI, allows us to apply space-241 for-time substitution techniques and improve our understanding of forest growth and potential. 242 However, the application of these methods for predicting future carbon stocks will bring large 243 uncertainties without the ability to validate model results against temporal products. This could be 244 achieved in future research using high spatial and temporal resolution orbital LiDAR data derived from 245 GEDI (Global Ecosystem Dynamics) or IceSat-2 (Ice, Cloud and land Elevation Satellite) 47 as well as 246 the continuous production of the ESA-CCI product used in this study. With the use of temporal 247 products, we can better understand and monitor the current and future role of these forests in the carbon 248 cycle and as climate mitigation strategies on potentially a global scale. 249 Our study has quantified the varied and complicated regrowth rates of SF across Amazonia influenced 250 by multiple drivers across Amazonia. Given the uncertain and potentially threatened status of old-251 growth forest sinks due to ongoing climate change 4 , it is imperative to limit human-induced fire and 252 deforestation disturbance in both old-growth and SF. By preserving the remaining old-growth forest 253 stock and sustainability managing SF we can maintain and increase the carbon sink of this globally 254 important biome and help it to achieve its climate mitigation potential. 255

Identifying areas of Secondary forest and their ages 257
The underlying product for this research was the land-use and land-cover product (MapBiomas 258 Collection 3.1), available for the whole of Brazil for the years 1985 to 2017 24 . The dataset is based on 259 Landsat image classification, mapping annual land-use and land-cover at 30m spatial resolution. We 260 follow a very similar methodology applied by Nunes et al. 10 and Silva Junior et al. 18 . to identify areas 261 of secondary forest (SF) and determine their respective ages. We reclassified forest land and all land 262 under human use to values of 1 and 0, respectively and tracked, when a conversion from anthropogenic 263 (0) to forest land (1) took place. Consecutive years following this transition in which a forest remained 264 forest, were considered to be SF and used to estimate their respective ages (in years). Ages ranged from 265 1 to 32 years since the MapBiomas product (v3.1) is available for the period 1985 to 2017. Any forest 266 land pixels that did not undergo a transition during this period were considered an old-growth forest. A 267 limitation is therefore that this method cannot classify forests as secondary forests that were deforested 268 and regrew before 1985. If an area of SF was deforested during the period of analysis, we disregarded 269 the area as SF and only began calculating the age again if a conversion from 0 to 1 took place. From 270 this we also calculated the number of times an area of SF was deforested during the period 1986 to 271

272
Previous research has shown that the MapBiomas product misclassifies perennial crops such as oil palm 273 plantations 10 and other plantation forests as natural forests (Supplementary Figure 2). To remove 274 misclassified areas, we used the latest land cover data of another, widely used Brazilian land cover 275 product, TerraClass-2014 9 . Finally, we excluded areas of SF (within a 3km radius) that overlay field 276 inventory sites of SF for cross validation of our method (Supplementary Figure 10; Supplementary 277 Table 10). 278

Modelling carbon sequestration with different drivers 279
To model the regrowth of SF we applied a space-for-time substitution method. Instead of tracking the 280 associated Aboveground Carbon (AGC) regrowth over time, the regrowth was estimated by considering 281 the available ages of the standing SF in 2017 and the associated AGC at the same time. Here we explain 282 the methods used to determine SF AGC using the ESA-CCI Aboveground Biomass (AGB) product 283 (100-m) for the year 2017 22 (see Supplementary Discussion for further details). All analysis was carried 284 out in the units of the original product (AGB) but expressed as AGC by assuming a 2:1 ratio of biomass 285 to carbon 23 . The ESA-CCI AGB product was only released in late 2019 and was in its early phases of 286 development at the time of use. However, given that its spatial resolution was high enough to separate 287 areas of only SF and its recent acquisition warranted its use for this research. Only areas of SF greater 288 than 9,000m 2 were considered for further analysis, an area approximately equal to 1 pixel of the ESA-289 CCI product. The SF map was laid over the AGC data and the modal AGC was extracted for each SF 290 patch. We then aggregated the AGC values by the age of SF and used the median AGC value for each 291 age in further analysis. We applied a bias correction to the median AGC values, subtracting the smallest 292 median value from all values to shift the data to begin at or near 0 Mg C ha -1 AGC for a 1-year old SF. 293 Following this, we used six remote sensing products of driving variables widely accepted to influence 294 regrowth of forests. The data products included four environmental drivers (1 -4) Table 1) and so had to be resampled to the size of SF pixels 301 (30-m spatial resolution) using the "resample" package in the Geographic Information System 302 programme, ArcMap10.6. We calculated the key zonal statistics of these variables such as the mean 303 value of the driver affecting a specific area of SF. 304 The drivers were then grouped according to numerical limits, such as the 25 th , 50 th and 75 th percentiles. 305 We then modelled the AGC for the age of SF under these groupings using the commonly used 306 Chapman-Richard model for regrowth 49 : 307 where Yt refers to the AGC at age t; A is the AGC asymptote or the AGC of the old-growth forest; k is 309 a growth rate coefficient of Y as a function of age; c is a coefficient that determines the shape the growth 310 curve; and ε is an error term. We assumed that after a given amount of time, the AGC could return to 311 levels equivalent to old-growth forests, and reach a precalculated asymptote. As such, we extracted the 312 median, bias-corrected AGC value of old-growth forests under each variable condition from the ESA-313 CCI AGC product to represent the value of the asymptote. From this, we could also determine if and 314 when the SF AGC regrowth models would reach those equivalent to old-growth forest levels. Forcing 315 the models to "fit" to an expected value for the asymptote value naturally increases the error of our 316 model, partly due to heterogeneity in old-growth forest values within each variable condition. 317

Determining the importance of each driver 318
We used a random forest model to assess which of the drivers used in this research were the most 319 important in influencing the regrowth of SF. To maximise computational speed and to account for any 320 biases in the products used we applied a stratified random sample equating to 2% of the data into the 321 random forest model (n = 50,000). This sample size was more than the minimum number of samples 322 needed (1,000 = 0.04%) to ensure results would be within the 95% confidence interval with a sampling 323 error of 5% using a multinomial function 50 . We carried out all analysis using the conditional random 324 forest model "cforest" available in predictive model package "caret" for the statistical software 325 "R"(v4.0.2) 51,52 . The "cforest" random forest model provides more accurate importance estimates 326 compared to more traditional random forest models such as "randomForest" when the dataset includes 327 both continuous (e.g. precipitation) and categorical data (e.g. burnt, not burnt) data 53 . We used 80% of 328 the sampled data for training the model and the remaining 20% to test the model. From this analysis we 329 estimated the "Permutation importance" also known as the "Mean Decrease in Accuracy" (as a 330 percentage) for each variable, which indicates the most important variable in terms ranking (higher 331 value meaning more importance to the determine the value of AGC). We show the variable importance 332 relative to SF age, identified as the most important variable in influencing AGC. The interpretation of 333 the results should be limited to the rankings and not the absolute values of the percentages 54 . 334

Representing spatial patterns of secondary forest regrowth 335
We created a regional classification based on the three most important environmental variables driving 336 regrowth. We used an unsupervised K-means cluster analysis to group Amazonia into regions based on 337 similarities between the SF in terms of the drivers' variability. We then subclassified each region based 338 on the type of disturbance (fire and/or deforestation) experienced by the SF. The aim of this was to 339 show areas of SF that experience similar conditions and the effect this has on regrowth in a spatially 340 explicit manner. We developed 16 regional-models of regrowth and included the median, bias-corrected 341 AGC value for old-growth forest in each of the regions as the asymptote of the models. Using the 342 random forest model, we again determined the importance of drivers for each region, as described in 343 the previous section. 344 Estimating 2017 carbon stock and future carbon sinks 345 We estimated the 2017 carbon stock by applying the corresponding regional models to all pixels initially 346 identified as SF with respect to the pixel age, and whether the pixel experienced any disturbances. From 347 this we were able to estimate the carbon stock in 2017 for all SF and the net carbon change from 2016 348 to 2017. We also considered an alternative scenario in which no forest disturbance occurred during 349 regrowth by applying the no-disturbance models to the corresponding regions. In this alternative 350 scenario, we were able to calculate the resulting potential 2017 carbon stock, and associated reduction 351 due to disturbances. Finally, by aging the standing SF in 2017, we modelled the carbon stock and annual 352 carbon accumulation for the next decade considering different scenarios of SF preservation: (1) all 353 forests; (2) forest with ages 5+; (3) forest with ages 10+; (4) forest with ages 15+; (5) forest with ages 354 20+ years. 355

Data and Code Availability 356
The original data used in this study are all publicly available from their sources (see references).

357
Processed data, products and codes produced in this research are available from the corresponding 358 author upon reasonable request. The regrowth models can be built by users using Equation 1 and 359 information provided in Supplementary Tables 8 and 9.