Basin-wide variation in tree hydraulic safety margins predicts the carbon balance of Amazon forests

Abstract

Tropical forests face increasing climate risk^1,2, yet our ability to predict their response to climate change is limited by poor understanding of their resistance to water stress. Although xylem embolism resistance thresholds (for example, [Formula: see text]₅₀) and hydraulic safety margins (for example, HSM₅₀) are important predictors of drought-induced mortality risk^3-5, little is known about how these vary across Earth's largest tropical forest. Here, we present a pan-Amazon, fully standardized hydraulic traits dataset and use it to assess regional variation in drought sensitivity and hydraulic trait ability to predict species distributions and long-term forest biomass accumulation. Parameters [Formula: see text]₅₀ and HSM₅₀ vary markedly across the Amazon and are related to average long-term rainfall characteristics. Both [Formula: see text]₅₀ and HSM₅₀ influence the biogeographical distribution of Amazon tree species. However, HSM₅₀ was the only significant predictor of observed decadal-scale changes in forest biomass. Old-growth forests with wide HSM₅₀ are gaining more biomass than are low HSM₅₀ forests. We propose that this may be associated with a growth-mortality trade-off whereby trees in forests consisting of fast-growing species take greater hydraulic risks and face greater mortality risk. Moreover, in regions of more pronounced climatic change, we find evidence that forests are losing biomass, suggesting that species in these regions may be operating beyond their hydraulic limits. Continued climate change is likely to further reduce HSM₅₀ in the Amazon^6,7, with strong implications for the Amazon carbon sink.

Fig. 1. Sampled sites: spatial distribution and climatological variation.

The map depicts long-term climatical water deficit (CWD) obtained from ref. (2.5 arcsec resolution). Bar graphs show mean precipitation per month (1998–2016) per site. The red lines at 100 mm represent the definition of dry season, where the monthly precipitation is below 100 mm. Precipitation data were obtained from TRMM (the Tropical Rainfall Measuring Mission—TMPA/3B43 v.7) at 0.25^° spatial resolution. Aseasonal ever-wet sites (blue bars): Sucusari (SUC) and Allpahuayo (ALP-1 and ALP-2). Intermediate DSL sites (green bars): Acre (FEC), Caxiuanã (CAX), Manaus (MAN), Tambopata (TAM) and Tapajós (TAP). Ecotonal long DSL sites (brown bars): Kenia (KEN-1 and KEN-2) and Nova Xavantina (NVX).

Fig. 2. Hydraulic trait variation across and within Amazon forest sites.

a,d, Xylem water potential at which 50% of the conductance is lost ($\Psi$₅₀). b,e, HSMs related to $\Psi$₅₀ (HSM₅₀ = $\Psi$_dry − $\Psi$₅₀). c,f, In situ dry season leaf water potential ($\Psi$_dry). d–f, Show hydraulic trait variation within intermediate DSL forests, subsetted according to Amazon region (central eastern Amazon: TAP, CAX and MAN; western Amazon: FEC and TAM). Dashed lines denote the mean value of each trait across all tree taxa in the dataset whereas the red line indicates HSMs equal to zero. Boxplots show the 25th percentile, median and 75th percentile. The vertical bars show the interquartile range ×1.5 and datapoints beyond these bars are outliers. Sites are sorted according to increasing water availability. Red, green and blue colours represent sites from ecotonal long DSL, intermediate DSL and aseasonal ever-wet forests, respectively. Each point represents one species per site (N_total = 170 species). Significant differences at P < 0.05 are shown on the figure (Wilcoxon rank sum tests).

Fig. 3. Relationship between WDA and HTs across western Amazon tree species.

a–c, Embolism resistance $\Psi$₅₀ (a), hydraulic safety margin HSM₅₀ (b) and minimum leaf water potential observed in the dry season $\Psi$_dry (c). Individual points indicate species mean trait values (n = 87). Less negative WDA values denote wet-affiliated species and more negative WDA denote dry-affiliated species. Species-level WDA data were obtained from ref. . SMA regressions are shown by solid lines. The grey shaded areas represent the 95% bootstrapped confidence intervals for the slopes and intercepts. The R² of each regression is shown on the figure. For this analysis, we subset our dataset to include only species collected in the western Amazon as done by ref. .

Fig. 4. Relationship between relative ΔAGB and basal area weighted mean vegetation traits and climatic factors across clusters of the Amazonian forests.

a, Variance explained by individual predictors when using SMA models to predict plot-level relative ΔAGB, with ΔAGB calculated as (AGB_end − AGB_start)/period of monitoring length/standing woody biomass. Climatic data (MAT, MAP and MCWD), HTs ($\Psi$₅₀, $\Psi$_dry and HSM₅₀, defined as the difference between $\Psi$_dry and $\Psi$₅₀) and other plant traits (stem and branch wood density and LMA) are indicated as red, blue, brown and green bars, respectively. Asterisk denotes statistically significant bivariate relationships after correcting for multiple hypothesis testing, using Bonferroni-corrected P < 0.05. Stem wood density values were extracted from the Global Wood Density database^,. Bivariate plots and statistics for all predictor variables considered are shown in Extended Data Fig. 6 and Supplementary Table 3. b, Relationship between basal area weighted mean HSM₅₀ and plot-level relative ΔAGB. We computed relative ΔAGB due to high standing AGB variance across plots. However, we also repeated B regression by considering absolute ΔAGB and this result was independent of whether absolute or relative ΔAGB were used in the bivariate regressions (Extended Data Fig. 7d). c, Relationship between basal area weighted mean HSM₈₈ and annual instantaneous stem mortality rate (equation (4); ref. ) across forest plots. The solid line is the best fit line of the SMA model and the shaded area represents the 95% bootstrapped confidence interval. KEN plots were excluded from all forest dynamics analyses because of a fire event that occurred in the region in 2004 and may still be affecting biomass accrual.

Extended Data Fig. 1. Precipitation regimes of sampled sites.

Precipitation data were obtained from TRMM (the Tropical Rainfall Measuring Mission—TMPA/3B43 version 7) at 0.25^o spatial resolution from 1998–2016. Maximum cumulative water deficit (MCWD) was computed following Aragão et al. (2007) but replacing universal ET values with site-specific values derived from the ERA-5 re-analysis product. MCWD is defined as the maximum climatologically-induced water deficit (see equation 1 in Methods). Sites in which MCWD~0 do not experience seasonality (dry season length (DSL) = 0), while sites with very negative MCWD values are strongly seasonally water-stressed. Sites are colour-coded by forest types, based on their seasonal rainfall patterns: aseasonal (blue), intermediate DSL (green) and long DSL (brown).

Extended Data Fig. 2. Climatological data corresponding to Ψ_dry (in situ dry season leaf water potential) sampling at each site.

Grey bars and error bars show the mean and standard deviation of monthly precipitation from 1991 to 2018 (CRU data ts.4.0338). The blue dashed lines represent the year of sampling, while the brown points show the months at which $\Psi$_dry was measured. Hydraulic traits and climatic data for TAP were obtained from Brum et al.. We display CRU data in this figure due to no availability of TRMM (Tropical Rainfall Measuring Mission) data beyond 2016. Sampling years can be found in Supplementary Table 8.

Extended Data Fig. 3. Functional trait range of species sampled in this study.

Histograms of life-history related traits of the sampled species (red) in relation to comprehensive histograms of the broader Amazon tree flora (grey). A) Mean wood density (g cm⁻³)^,; Potential size, calculated as the 95th percentile of diameter distribution (cm); C) Maximum growth, calculated as the 95th percentile of individual growth rates available for a given species (cm yr⁻¹); D) Mean growth rate (cm yr⁻¹); E) Mean mortality rate (% yr⁻¹). All trait data shown in this figure were extracted from Coelho de Souza et al..

Extended Data Fig. 4. Hydraulic traits variation at species and community level.

Top panels: variation in hydraulic traits across Amazon forest types; long DSL (brown), intermediate DSL (green) and aseasonal (blue). A) xylem water potential at which 50% ($\Psi$₅₀) of the conductance is lost. B) hydraulic safety margins related to $\Psi$₅₀ (HSM₅₀ = $\Psi$_dry – $\Psi$₅₀). C) minimum leaf water potential observed in the dry season ($\Psi$_dry). Dashed lines show the mean value of each trait across all tree taxa. Red lines denote hydraulic safety margins equal to zero. Significant differences at p < 0.05 are shown by letters above each boxplot (Kruskal–Wallis followed by post hoc Mann–Whitney–Wilcoxon Rank Sum test). Each point represents one species per site. Long DSL, intermediate DSL and aseasonal forests encompass 3, 5 and 3 forest sites, respectively. Boxplots display the 25^th percentile, median and 75^th percentile. The vertical bars show the interquartile range times 1.5 and datapoints beyond these bars are outliers. Bottom panels: Relationship between tree basal-area weighted mean hydraulic traits and maximum cumulative water deficit (MCWD). D) Basal area-weighted mean $\psi$₅₀ (xylem water potential at which 50% of the conductance is lost); E) Basal area-weighted mean hydraulic safety margin (HSM₅₀); F) Basal area-weighted mean minimum leaf water potential observed in the dry season ($\Psi$_dry); n = 11 sites. Significant linear relations are shown by regression lines and 95% confidence intervals, by shaded areas. See methods for MCWD calculations.

Extended Data Fig. 5. Leaf habit information of the sampled species per plot (top panels), hydraulic trait variation across leaf habit groups (middle panels) and relationship between basal area-weighted mean hydraulic traits and maximum cumulative water deficit when excluding deciduous and semideciduous species.

A) and D) xylem water potential at which 50% of the conductance is lost ($\Psi$₅₀); B) and E) Hydraulic safety margins related to $\Psi$₅₀ (HSM₅₀ = $\Psi$_dry – $\Psi$₅₀ and C) and F) Minimum in situ leaf water potential observed in the dry season ($\Psi$ dry). Dashed lines denote the mean value of each trait across all tree taxa in the dataset. Red line, the hydraulic safety margins equal to zero. Boxplots display the 25^th percentile, median and 75^th percentile. The vertical bars show the interquartile range times 1.5 and datapoints beyond these bars are outliers. Sites are sorted by increasing water availability. Each point represents one species per site (N = 170 species) in the top panels and species mean ($\Psi$_50, HSM₅₀) and species minimum $\Psi$_dry per leaf habit in the bottom panels (N = 136 species). Deciduous, semideciduous and evergreen species are represented by red, blue and green points, respectively. Grey points or NA represent species for which leaf habit information was not available. Significant differences at p < 0.05 are represented by different letters above each boxplot (Kruskal–Wallis followed by post hoc Mann–Whitney–Wilcoxon Rank Sum test). G) $\psi$₅₀ (xylem water potential on which 50% of the conductance is lost; H) Hydraulic safety margin (HSM₅₀ = $\Psi$_dry – $\Psi$₅₀); I) In situ leaf water potential observed in the dry season ($\Psi$ dry); n = 10 sites. Significant linear relations are shown by regression lines. The shaded area represents the 95% confidence interval of the regression slope. Further leaf habit information of sampled species is provided in Supplementary Table 10.

Extended Data Fig. 6. Bivariate relationship between plot relative aboveground net biomass change (ΔAGB) and basal-area-weighted mean vegetation traits (A–E) and climatic variables (F–I) across Amazonian forest plots.

Due to high standing aboveground biomass variability across sites, we computed relative ΔAGB values as: ((AGB _end – AGB _start)/census length)/AGB (time-weighted mean standing woody biomass). Each point indicates a forest plot. Information about each plot and the observation period used can be found in Supplementary Table 5. Stem wood density data were extracted from the Global Wood Density database^,. KEN plots were excluded from all forest dynamics analyses because of a fire event that occurred in the region in 2004 and may still be affecting biomass accrual. Regression lines show significant relationships using standard major axis (SMA) models after Bonferroni correction for multiples hypothesis testing. Supplementary Table 3 shows the results of the SMA models.

Extended Data Fig. 7. Plot-level analyses: bivariate relationships between basal area-weighted mean hydraulic safety margin and forest dynamics across Amazonian forest plots.

Basal area-weighted mean HSM₅₀ in relation to relative (A-C) and absolute (D-H) forest dynamics values. A) Relative annual aboveground biomass net change (ΔAGB/AGB), where: ΔAGB is the difference in aboveground biomass between the final and initial censuses (AGB_{final census} – AGB_{initial census}) divided by total monitoring length for that plot (Date_{final census} – Date_{initial census}) in years and AGB is the time-weighted mean standing woody biomass across censuses per plot; B) Relative annual mortality in terms of biomass: (AGB_MORT/AGB). Where AGB_MORT is the sum of the AGB of all dead trees and the unobserved components (see methods), divided by the census interval length;C) Relative annual AGB wood productivity: (AGWP/AGB), where AGWP is defined as the sum of the biomass growth of surviving trees >10 cm DBH, new recruits > 10 cm DBH and the unobserved components (see methods), within a plot in a given census interval, divided by the census interval length; D) Annual aboveground biomass net change (ΔAGB); E) Annual aboveground biomass mortality (AGB_MORT); F) Annual aboveground wood productivity (AGWP), G) Annual instantaneous stem mortality rate (See methods Equation 4); H) Residence time of woody biomass calculated as the ratio of mean standing biomass to mean biomass mortality rate. All these parameters were calculated for each census interval and we calculated time-weighted mean to have one value per plot. Each point indicates one forest plot. Significant linear relations are shown by regression lines (Standard major axis models). The shade area represents the 95% bootstrapped confidence interval. Information about each plot and the observation period is available on Supplementary Table 5. Supplementary Table 4 shows the results of the SMA models.

Extended Data Fig. 8. Cluster-level analyses: Relationship between basal area weighted mean HSM₅₀ and cluster mean forest dynamics across clusters of forest plots.

Basal area-weighted mean HSM₅₀ in relation to cluster mean relative (A–C) and absolute (D–H) forest dynamics values. A) Relative annual aboveground biomass net change (ΔAGB/AGB), where: ΔAGB is the difference in aboveground biomass between the final and initial censuses (AGB_{final census} – AGB_{initial census}) divided by total monitoring length for that plot (Date_{final census} – Date_{initial census}) in years and AGB is the time-weighted mean standing woody biomass across censuses per plot; B) Relative annual mortality in terms of biomass: (AGB_MORT/AGB). Where AGB_MORT is the sum of the AGB of all dead trees and the unobserved components (see methods), divided by the census interval length;C) Relative annual AGB wood productivity: (AGWP/AGB), where AGWP is defined as the sum of the biomass growth of surviving trees >10 cm DBH, new recruits > 10 cm DBH and the unobserved components (see methods), within a plot in a given census interval, divided by the census interval length; D) Annual aboveground biomass net change (ΔAGB); E) Annual aboveground biomass mortality (AGB_MORT); F) Annual aboveground wood productivity (AGWP), G) Annual instantaneous stem mortality rate (see methods); H) Residence time of woody biomass calculated as the ratio of mean standing biomass to mean biomass mortality rate. All these parameters were calculated for each census interval and we calculated time-weighted mean to have one value per plot. Each point indicates the mean value across clusters of forest plots, which in total encompass 31.37 ha of forest spread across 34 plots. Information about each cluster and the observation period used for each cluster is provided in Supplementary Table 5. The solid line is the best fit line of the standard major axis (SMA) model and the shaded area represents the 95% bootstrapped confidence interval. Supplementary Table 4 shows the results of the SMA models.

Extended Data Fig. 9. Relationship between published pan-Amazonian species-level growth rates and hydraulic safety margins.

We restrict our analysis to the western Amazon, where there is the highest overlap of species between ours and Coelho de Souza et al. dataset and to avoid biases due to soil/forest dynamic differences across Amazonian regions. Hydraulic safety margins for species occurring across multiple sites were averaged to yield one value per species.

Extended Data Fig. 10. HSM₅Plot-level analyses: bivariate relationships between basal area-weighted mean HSM₈₈ and forest dynamics across Amazonian forest plots.

Basal area-weighted mean HSM₈₈ in relation to relative (A-C) and absolute (D-H) forest dynamics values. A) Relative annual aboveground biomass net change (ΔAGB/AGB), where: ΔAGB is the difference in aboveground biomass between the final and initial censuses (AGB_{final census} – AGB_{initial census}) divided by total monitoring length for that plot (Date_{final census} – Date_{initial census}) in years and AGB is the time-weighted mean standing woody biomass across censuses per plot; B) Relative annual mortality in terms of biomass: (AGB_MORT/AGB). Where AGB_MORT is the sum of the AGB of all dead trees and the unobserved components (see methods), divided by the census interval length;C) Relative annual AGB wood productivity: (AGWP/AGB), where AGWP is defined as the sum of the biomass growth of surviving trees >10 cm DBH, new recruits > 10 cm DBH and the unobserved components (see methods), within a plot in a given census interval, divided by the census interval length; D) Annual aboveground biomass net change (ΔAGB); E) Annual aboveground biomass mortality (AGB_MORT); F) Annual aboveground wood productivity (AGWP), G) Annual instantaneous stem mortality rate (see methods Equation 4); H) Residence time of woody biomass calculated as the ratio of mean standing biomass to mean biomass mortality rate. All these parameters were calculated for each census interval and we calculated time-weighted mean to have one value per plot. Each point indicates one forest plot. Significant linear relations are shown by regression lines (Standard major axis models). The shade area represents the 95% bootstrapped confidence interval. Information about each plot and the observation period is available on Supplementary Table 5. Supplementary Table 4 shows the results of the SMA models.