Main

Over half of the world’s tropical forests are not old-growth, but naturally regenerating forests5. We focus here on secondary forests that regrow after nearly complete removal of forest cover for agricultural use4. Despite their crucial role in human-modified tropical landscapes, the rate at which these forests will recover, and the extent to which they can provide equivalent levels of ecosystem services as the forests they replaced, remains uncertain. Several studies suggest that, once a certain threshold of disturbance has been reached, tropical forests may collapse6 or switch to an alternative stable state7. However, in general, secondary tropical forests have rapid rates of carbon sequestration, with potentially large consequences for the global carbon cycle3. The magnitude of the tropical carbon sink is relatively well known for old-growth forest8,9, but current global estimates on carbon uptake and storage of secondary forests are rough or biased3,8,9 and we urgently need better-founded biomass recovery rates and resilience estimates.

Secondary forests deliver a suite of ecosystem services that are closely linked to the biomass resilience of these forests10. We define biomass resilience as the ability to (1) recover quickly after disturbance (that is, the absolute recovery) and (2) return to its original (pre-deforestation) state (that is, the relative recovery). Few studies have quantified forest recovery directly by monitoring secondary forest plots over time, and they typically cover only a few years (but see ref. 11). Chronosequence studies have evaluated different plots at different times after abandonment12 and shown that recovery rates vary dramatically with climate13, soils14, and management intensity15. However, the meta analyses so far13,16,17,18 did not consider Neotropical-wide environmental gradients and were not based on original data, making direct comparisons impossible because of variation in methodology and allometric biomass equations used.

Here we quantify rates of recovery of aboveground biomass (AGB) stocks in secondary forests, and assess to what extent recovery rates are driven by abiotic site conditions, landscape forest cover, and previous land use. We hypothesize that recovery rates increase with resource availability (high rainfall13 and fertile soils) and decrease with degree of forest loss in the surrounding landscape matrix (implying lower seed availability), as well as with intensity of prior land use. We test these hypotheses using an unprecedented multi-site analysis including 45 sites, 1,468 plots, and >168,000 stems, spanning the major environmental and latitudinal gradients in the Neotropics, and we present a biomass recovery map to support policy decisions for the Neotropics.

We compiled data from 45 chronosequence studies (Extended Data Table 1); we also included old-growth plots for 28 sites with no signs of recent human disturbance. In each plot all stems ≥5 cm diameter were measured and identified, their biomass estimated using allometric equations, and summed across all trees to obtain live plot AGB.

For each site, AGB was modelled as a function of time elapsed since abandonment: that is, since the main human activity (crop production, cattle ranching) ceased at the site. Biomass resilience was quantified in two ways. For absolute recovery rate, the site-specific models (N = 44; one outlier excluded) were used to predict AGB at 20 years, a representative age for secondary forests in the Neotropics. For relative recovery, AGB was expressed as a percentage of the median AGB of old-growth forest plots in the area. AGB recovery was related to climatic water availability, soil cation exchange capacity (CEC), forest cover, and prior land use using multiple regression analysis. The best-supported equation was used to produce a biomass recovery map of the Neotropics.

Recovery of AGB showed a saturating relationship with stand age for the individual sites. After 20 years of secondary succession, AGB accumulation varied 11.3-fold (from 20 to 225 Mg ha−1, average 121.8 ± 7.5 s.e.m.) across sites (Fig. 1a). Biomass recovery after 20 years, relative to old-growth forest AGB at the same site, varied from 25 to 85% across sites (Extended Data Fig. 1). Across sites, AGB after 20 years increased with annual precipitation (P < 0.001), climatic water deficit (CWD; P < 0.001, a higher value here means a lower deficit), and rainfall seasonality (R2 = 0.59, P < 0.0001; Fig. 2, Extended Data Fig. 2a and Extended Data Table 2). There was considerable variation at a given level of water availability, which is also reflected in the root mean squared error of 31.7 Mg ha−1. CEC, forest cover, and prior land use were not significant predictors of AGB accumulation across sites (Extended Data Fig. 2). For relative recovery, not only rainfall (P = 0.040) but also CEC (P = 0.027) had a significant positive effect on biomass recovery (Extended Data Table 2 and Fig. 1). The biomass recovery map (Fig. 3, see Extended Data Fig. 3 for an uncertainty map) shows high variation in potential AGB accumulation rates across the Neotropics, with low rates in dry forests in Mexico and high rates in humid forests in southern Mexico, Costa Rica, and large parts of the Amazon.

Figure 1: Relationship between forest biomass and stand age using chronosequence studies in Neotropical secondary forest sites.
figure 1

a, AGB (N = 44); b, AGB recovery (N = 28). Each line represents a different chronosequence. The original plots on which the regression lines are based are indicated in grey (N = 1,364 for AGB, N = 995 for AGB recovery). AGB recovery is defined as the AGB of the secondary forest plot compared with the median AGB of old-growth forest plots in the area, multiplied by 100. Significant relations (two-sided P ≤ 0.05) are indicated by continuous lines; non-significant relationships (two-sided P > 0.05) are indicated by broken lines. Plots of 100 years old are also second-growth. See Extended Data Fig. 4 for the same figure with plots colour-coded by forest type.

PowerPoint slide

Full size image
Figure 2: AGB after 20 years.
figure 2

a, In relation to annual rainfall; b, in relation to CWD for Neotropical forest sites. Lines indicate predicted AGB at 20 years based on a multiple regression including 1/rainfall, CWD, and rainfall seasonality (R2 = 0.59). Other variables were kept constant at the mean across sites (two-sided P < 0.005 for 1/rainfall; P = 0.03 for CWD). The third, less significant factor (rainfall seasonality) is shown in Extended Data Fig. 2. N = 43 sites (one site was excluded because no climatic data were available).

PowerPoint slide

Full size image
Figure 3: Potential biomass recovery map of Neotropical secondary forests.
figure 3

The total potential AGB accumulation over 20 years of lowland secondary forest growth was calculated on the basis of a regression equation relating AGB with annual rainfall (AGB = 135.17 − 103,950 × 1/rainfall + 1.522 × rainfall seasonality + 0.1148 × CWD; see Methods). The colour indicates the amount of forest cover recovery (red is low recovery; green is high recovery). The 44 study sites are indicated by circles (the symbol on the ocean belongs to an island) and the size of the symbols scales with the AGB attained after 20 years. The grey areas do not belong to the tropical forest biome. The map focuses on lowland tropical forest (altitude <1,000 m). See Extended Data Fig. 5 for a colour-blind-friendly map.

PowerPoint slide

Full size image

We show that AGB of 20-year-old secondary forest varies more than 11-fold across sites, with potentially large consequences for the carbon storage and mitigation potential of tropical forests (Supplementary Information S1). The carbon sequestration and mitigation potential of secondary forests is immense. Average biomass after 20 years corresponds to an annual net carbon uptake of 3.05 Mg C ha−1 yr−1, 11 times the uptake rates of Amazonian old-growth forests in 2010 (0.28 Mg C ha−1 yr−1 (ref. 19)) and 2.3 times the uptake rates of selectively logged Amazonian forests (1.33 Mg C ha−1 yr−1 (ref. 20)) where reduced impact logging techniques have been applied. Clearly, anthropogenic disturbances open up the canopy, enhance light availability, and set the system back to an earlier successional stage, leading to lower standing biomass but faster growth rates of the remaining forest stand. Although second-growth forests have substantially lower carbon stocks and biodiversity than the old-growth forests they replace, their carbon sequestering potential is high. Therefore, we urge halting deforestation and advocate actions that promote natural regeneration in deforested areas. Recognition of the significant carbon mitigation potential and other important services and values delivered by second-growth forests21 can be an important motivating factor to reach national targets for forest restoration inspired by the Convention on Biological Diversity Aichi Targets, the Bonn Challenge, and the New York Declaration on Forests.

Secondary forests differ dramatically in their biomass resilience, driven mainly by variation in water availability (Fig. 2, cf. ref. 13), although other factors play a role as well. Both higher rainfall and lower CWD increase plant water availability and extend the growing season, thereby increasing biomass growth of trees and forest stands22. Climate change scenarios predict less and/or more variable rainfall for several regions in the tropics23, which may potentially hamper biomass recovery and forest resilience in these regions.

Soil fertility (CEC) had a significant positive effect on relative biomass recovery, but no effect on absolute recovery. Perhaps the effect of CEC was weak because for many sites CEC was obtained from a global database rather than measured in situ, because large differences in macroclimate may overrule more subtle differences in soil fertility, or because forest productivity is often limited by phosphorus24 or nitrogen25, which we did not measure. Forest cover in the surrounding matrix and prior land use did not affect biomass recovery (Extended Data Figs 1 and 2), perhaps because these were estimated. Forest recovery rate may slow down in isolated secondary forest patches because of reduced plant colonization and survival, harsher environmental conditions, and frequent disturbances. Biomass recovery did not differ between abandoned pastures and agricultural fields, possibly because of high within-category variation in land-use intensity (such as use of fire and time under land use and frequency of cultivation cycles), which may strongly affect forest recovery15,16,17.

Dry and moist forests differ substantially in old-growth forest biomass, and relative recovery may therefore be a better indicator of forest resilience than absolute recovery rates. It takes a median time of 66 years to recover to 90% of old-growth values. However, several mid-successional sites of 40–100 years old attained higher biomass than old-growth forest (Fig. 1b), which has also been found by modelling studies26. This is most probably because of a high abundance of old trees of long-lived pioneer species in mid-successional sites, which tend to be large and store large amounts of carbon27 before they die.

Several studies suggest that, once a certain threshold of disturbance is reached, tropical forests may collapse6 or switch to an alternative stable state7. Our study shows that forests can be resilient, and that their biomass resilience strongly depends on water availability. By mapping potential for biomass recovery across the Neotropics, attention can be focused on particular areas that should be conserved or treated with extra caution because they are more difficult to restore (slow recovery). The recovery map can also be used to identify areas with high carbon sequestration potential for United Nations collaborative initiative on Reducing Emissions from Deforestation and forest Degradation (REDD+) programmes, or where (assisted) natural regeneration, or restoration and afforestation activities, may have the highest chance of success28,29 (high recovery). Such a spatially explicit, resource-based approach paves the road towards a more sustainable design and management of human-modified tropical landscapes.

Methods

Site description

We compiled the largest data set of chronosequence studies in the lowland Neotropics so far, including 1,468 plots established in 45 forest sites (Extended Data Table 1), containing 168,210 tree stems. The sites were located in eight countries, covering the full latitudinal range in the tropics varying from 20° N in Mexico to 22° S in Brazil. We focused on lowland sites, generally below 1,000 m altitude. Annual rainfall varied fivefold across sites (from 750 to 4,000 mm yr−1), thus all major lowland forest types were included. Soil CEC varied 33-fold across sites (from 2 to 65 cmol(+) kg−1). Forest cover in the landscape ranged from 31 to 100% across sites.

Site conditions

We assessed whether biomass recovery rates were related to abiotic site conditions (rainfall, soil fertility), and land-use intensity (prior land use and percentage forest cover in the surrounding matrix). For each site, mean annual rainfall was obtained from the nearest climatological station. Biomass recovery should ultimately respond to plant water availability, which is a function of annual rainfall and its seasonal distribution, topography, soil depth, and the water holding capacity of the soil. Rainfall seasonality (expressed as a coefficient of variation of rainfall) was obtained from WorldClim30 and CWD (in millimetres per year) from http://chave.ups-tlse.fr/pantropical_allometry.htm. CWD is the amount of water lost during dry months (defined as months where evapotranspiration exceeds rainfall), and is calculated as the total rainfall minus evapotranspiration during dry months. This number is by definition negative, and sites with CWD of 0 are not seasonally water stressed. CEC (in centimoles of positive charge per kilogram of soil) was used as an indicator of soil fertility of the site. Ideally, local soil data should be used, but these were only available for 12 sites. For the other sites, CEC data were obtained from the Harmonized World Soil Database (HWSD)31. For those 12 sites, the local CEC and HWSD data are indeed strongly positively correlated (r = 0.84, P < 0.0006, n = 12), indicating that the HWSD ranks the sites well in terms of CEC. When local soil fertility data were available, the CEC of the old-growth forest was used, as this indicates potential site fertility, which is also included in the world soil database. The HWSD database did not contain data on soil nitrogen and phosphorus. Phosphorus is thought to limit plant growth in highly weathered and leached tropical soils and correlates well with the productivity of old-growth Amazonian forests24. Nitrogen may be more limiting during the first decades of secondary succession because a large part of vegetation nitrogen is volatilized when the slashed vegetation is burned25, while soil preparation, crop cultivation, and nutrient leaching may further deplete soil nutrients. Indeed, early successional forests have relatively many nitrogen-fixing trees. To describe the landscape matrix, we estimated per site, for all plots combined, the average cover in the following five land use classes in a 1 km radius around the plots: mature forest, secondary forest, plantations/agroforestry systems, agriculture/pasture, other. This estimate is based on the long-term field experience of the authors at the site where they worked. We then calculated per site the average percentage total forest cover (of mature and secondary forest) in a 1 km radius around the plots. We used total forest cover because forests improve microclimatic conditions at the landscape level and act as important sources of seeds and biotic agents (mammal and bird dispersal agents, mycorrhizae) that may help to speed up the rate of secondary succession. We classified the prior land use of each site as cattle ranching or shifting cultivation (or a combination of these two), which are the two major land-use types where forest regeneration occurs in the Neotropics. We acknowledge that different land use practices, such as fire frequency16 and length of cultivation15, could affect biomass recovery rates of abandoned fields, but it is very difficult to obtain that information in a consistent and quantitative way across all sites.

Plots

Plot size varied from 0.01 to 1 ha. Most plots were rectangular, with a length varying from 20 to 200 m and width ranging from 1 to 50 m. The mean plot area across all ~1,468 plots was 947 m2, and the mean plot area across all 45 sites was 1,273 m2. On average 33 (5–291) plots were established per chronosequence. Chronosequence sites varied in the range of stand ages across the included secondary forest plots (Extended Data Table 1). For 28 out of the 45 sites, old-growth plots were included. Old-growth forests did not have signs of recent human disturbance, and were at least 100 years old. In most cases, the sampling design was the same as for secondary forest plots; however, for five of the sites, old-growth plots were slightly larger, for one site smaller. In each plot, all woody trees, shrubs, and palms ≥5 cm stem diameter at breast height (at 1.3 m height) were measured for their diameter and identified to species, with the exception of three sites for which only data for trees ≥10 cm diameter at breast height were available.

Tree biomass

We first evaluated how estimates of live AGB varied on the basis of the application of three frequently used sets of allometric equations. We calculated live AGB for each tree using equations from refs 32, 33, 34 (Supplementary Information S2). The equations in ref. 32 are based on stem diameter only, and the equations in refs 33, 34 are based on a combination of stem diameter and wood density. Palms were treated in the same way as trees35, which may slightly overestimate palm biomass. Biomass was summed across all trees to obtain live plot AGB. The three different equations from refs 32, 33, 34 resulted in small differences in absolute plot AGB estimates for dry forests (average AGB after 20 years ranged from 68 to 85 Mg ha−1 among the three equations), but yielded larger differences in AGB estimates for moist forest (precipitation 1,500–2,500 mm yr−1, AGB after 20 years ranged from 142 to 179 Mg ha−1; Extended Data Fig. 6). For 20-year-old forest, the earlier widely used equations32,33 predicted about 21–26% more biomass than the more recently developed equation (Extended Data Fig. 6). These differences become even greater with increasing forest age, with large consequences for assessing the magnitude of carbon uptake and storage in tropical forests (see, for example, ref. 19), as moist forests contribute most to the current forested area in the Neotropics. For further analyses we used the equation in ref. 34 because it provides an allometric equation based on the largest harvested tree data set analysed so far (and includes and extends the equation of ref. 33), including many secondary forest species for trees from wet and dry forests, for trees of most tropical continents. It is ecologically very relevant and methodologically elegant, as it uses a correction factor E. This bioclimatic stress variable takes temperature variability, precipitation variability, and drought intensity into account, and therefore continuous variation in site aridity. The equation in ref. 34 also takes variation in wood density into account, which is an important source of plot AGB variation across the Neotropics36. For one site (Providencia Island) for which no E value was available, we estimated E from a regression equation relating E to annual rainfall.

Data about species-specific wood density (in grams per cubic centimetre) came from the local sites, or from the Neotropical data of a global wood density database37,38 (http://datadryad.org/handle/10255/dryad.235), henceforth referred to as ‘Dryad’. For all estimates of wood density, we used the data source (local or Dryad) that had the highest level of taxonomic resolution. In the case of the same level of taxonomic resolution, we used the local data source. If no wood density information was available at the species level, then genus- or family-level wood density values were used, as wood density is phylogenetically strongly conserved39. Across all sites except Surutu 1 and 2, Cayey, and Bolpebra, 94.0% of the trees of the secondary and old-growth plots had been identified to species (range 74.4–100%), 97.9% to genus (range 90.8–100%), and 98.9% to family (range 94.9–100%). For the other four sites, 36.0–55.6% of the trees had been identified to species, 61.6–86% to genus, and 61.6–86% to family.

Plot structure

Biomass was summed across all trees in a plot to obtain live plot AGB and expressed on a per hectare basis (AGB, in megagrams per hectare). For one site (San Carlos) the original tree inventory data were not available anymore, so for this site we used the plot biomass that the author had estimated using locally derived allometric equations. For plots with a nested design, we calculated AGB per hectare by extrapolating the biomass of trees in smaller size classes that were measured in a part of the plot only to the total area measured, thus assuming that forest structure within the plot did not vary. For four sites, only trees ≥10 cm diameter at breast height were inventoried, and for these sites a correction factor was used to obtain the estimated plot AGB when trees between 5 and 10 cm diameter at breast height would have been included. Plot AGB was therefore multiplied by 1.18 for dry forest and by 1.03 for moist forest (compare ref. 40). Some plots had extremely high AGB values (>500 Mg ha−1), because of the occurrence of a single large tree, in combination with a small plot size. These outliers (2.5% of the total number of plots) were excluded from the analysis. One site (Marques de Comillas) had extremely high average AGB (>100 Mg ha−1) 2 years after land abandonment, because of a combination of many large remnant forest trees being left during land use, and a small plot size. This site was excluded from the analysis. Our analysis focused on AGB only, which should be a good indicator of carbon recovery over time, as below-ground biomass in roots scales tightly with AGB of trees41, and soil organic matter (which includes necromass of dead fallen trees and roots) is rather constant with time after abandonment18.

Statistical analyses

No statistical methods were used to predetermine sample size. To evaluate how AGB (estimated with the equation from ref. 34) recovered over time, we modelled AGB in each site as a function of natural-logarithm-transformed age. We included an intercept, as in some cases there was AGB at time zero because remnant trees were left during forest clearance. Biomass resilience was quantified in two complementary ways: the absolute recovery rate and the relative recovery rate. For absolute recovery, the site-specific models (N = 44 sites) were used to predict AGB at 10 and 20 years, as these ages were included in the age range of most sites, and are representative ages for secondary forests. To calculate annual carbon sequestration rates, we used the average AGB attained after 20 years, multiplied it by 0.5 (which is the average carbon value in dry biomass), and divided it by 20. To evaluate how fast AGB of secondary forests recovers towards old-growth forest values, we calculated the relative AGB recovery as the AGB of the secondary forest plot compared with the median AGB of old-growth forest plots in the area (as a percentage). For each site, we estimated the recovery time as the time needed to recover to 90% of old-growth values. We used time to recover to 90%, rather than to recover 100% of old-growth values, as it indicates when the biomass is close to old-growth values. To attain the full 100% may take much more time, as we modelled biomass build-up to show an asymptotic relationship with time. This study aimed to assess the rates of recovery of tree biomass under conditions where regeneration is occurring naturally. Sometimes, succession can be arrested owing to high-intensity land use or seed limitation. We did not target sites where succession was arrested, so it is possible that our ‘average’ conditions could have overestimated rates of recovery when considering all potential areas. However, neither did we specifically select sites that were regenerating particularly fast. Given the high number of sites and plots across the Neotropics that cover the entire range of typical land-use types and intensities, we are confident that we provide realistic rates of biomass accumulation for dominant land-use types and intensities in the Neotropics.

To evaluate how predicted absolute and relative recovery of AGB after 20 years varied with abiotic site conditions and land-use intensity, we used multiple regression analysis. We related AGB recovery to rainfall, rainfall seasonality, CWD, CEC, forest cover in the landscape matrix, and prior land use (pasture, mix of pasture and agriculture, agriculture) for 43 sites (for one island site no WorldClim data were available). AGB showed a saturating relationship with rainfall; therefore we used 1/rainfall as a predictor rather than absolute rainfall or the natural logarithm of rainfall, which improved model fits. We compared models of all possible combinations of predictors based on Akaike’s information criterion adjusted for small sample sizes (AICc), and selected the best-supported model with the lowest AICc. All analyses were performed in R 3.1.2 (ref. 42).

On the basis of our regression analysis, we constructed a potential biomass recovery map of the Neotropics at a resolution of 1 km2. First, a map of world ecoregions was obtained from the Nature Conservancy (http://maps.tnc.org/gis_data.html; last accessed on 26 January 2015). Eight hundred and sixty-seven distinct units were categorized into 14 biomes in 8 biogeographical realms43. Of these 14 biomes, three are relevant to our results: (1) tropical and subtropical moist broadleaf forests, (2) tropical and subtropical dry broadleaf forests, and (3) tropical and subtropical coniferous forests. Our scaled-up study region was then defined as the full spatial extent encompassing these three biomes in Central or South America, the region in which our network of study sites was located. Total annual precipitation was calculated by summing the individual monthly totals provided by WorldClim. Data for mean annual rainfall (defined as the average of 1950–2000) and rainfall seasonality (defined as variable ‘BIO15’ in WorldClim) was obtained at a 30 s resolution (approximately 1 km × 1 km) from WorldClim (http://www.worldclim.org/current), and CWD was obtained from http://chave.ups-tlse.fr/pantropical_allometry.htm. We then calculated the total potential AGB accumulation over 20 years of secondary forest growth (assuming that the initial year 0 condition was a fully cleared area), based on annual rainfall, rainfall seasonality, and CWD, because they were included in the best model explaining absolute recovery (Extended Data Table 2). The regression equation obtained in this study, relating AGB after 20 years with the climatic variables, was AGB at 20 yr = 135.17 − 103,950 × 1/rainfall + 1.521983 × rainfall seasonality + 0.1148 × CWD. Estimated AGB would then indicate the absolute biomass recovery potential. We stress that this is a potential biomass recovery map based on the relationship between recovery and rainfall for our sites; realized local recovery may vary because of differences in local soil conditions, land use history, the surrounding matrix, and availability of seed sources. We also made an uncertainty map of potential biomass based on the confidence interval of the predicted mean AGB at 20 years obtained from the multiple regression. We calculated the confidence interval of the predicted AGB at 20 years for each pixel based on 1/rainfall, CWD, and rainfall seasonality. The relative uncertainty for each pixel was calculated as 100 × (0.5 × 95% confidence interval of the mean)/predicted AGB. For the maps we focused on lowland forest (<1,000 m altitude), between the Tropics of Cancer and Capricorn. To make the maps, we had to extrapolate very little beyond our climatic parameter space; our sites cover a rainfall range from 750 to 4,000 mm yr−1. Sites with rainfall >4,000 mm yr−1 (where we extrapolate with our equation) cover only 0.88% of the lowland Neotropics.