Table 1. Summary of Amazon forest simulator results.

Intermediate-scale disturbances	Large-scale blow-downs25,26
	None	Central Amazon	All Amazon region
Lidar data24 from terra firme (gaps age30 ~1-year old)
dM/dt* (Mg C ha−1 y−1)	—	0.85	—
σ* (Mg C ha−1 y−1)	—	4.40	—
tobs (N=135)	—	2.24	—
tobs (N=1,545)	—	7.59	—
Lidar data24 from terra firme (gaps age30 ~3.6-year old)
dM/dt* (Mg C ha−1 y−1)	0.94	0.94	0.94
σ* (Mg C ha−1 y−1)	2.19	3.77	12.4
tobs (N=135)	4.99	2.90	0.88
tobs (N=1,545)	16.9	9.80	2.98

Mean and statistical significance of simulated AGB gains for a range of scenarios. We vary occurrence of large-disturbance blow-downs25,26, that is, the large-end tail of the disturbance frequency distribution, and age of intermediate-range disturbances. For blow-downs we distinguish three cases: (i) no large-disturbance blow-downs25,26, (ii) large blow-downs as observed only in central Amazon (~20% of the Amazon region), (iii) everywhere in the Amazon with the same frequency of events as in the central Amazon (that is, with five times more large-area events than observed). For intermediate-range disturbances we distinguish disturbances occurring across the entire Amazon region distributed according to lidar surveys24 (plots 1, 4, 5 and 12) of erosional terra firme (ETF) forests (33,196 ha) with either a mean gap age of 1 or 3.6 years based on gap closure observations of a 50 ha plot on Barro Colorado Island30. We assumed an annual mean mass gain (G) (live tree mass gains plus mass gains due to recruitment8,10,11) of 2.5 Mg C ha⁻¹ y⁻¹ in areas of terra firme forests. The simulator of forest mortality (D) is based on the frequency distribution of disturbance area. To convert area losses to biomass losses we assumed a forest mass density of 170 Mg C ha⁻¹ for all simulations, a high value and ~50% greater than the actual biomass density in the lidar landscape in southern Peru used to estimate intermediate disturbance dynamics8,11. Assessment of each scenario is based on a set of 10⁹ annual equivalent samples. The most credible results are in highlighted bold.

^*Significance is assessed with a t-test considering t_sim=(dM/dt)/(σ/sqrt(N)) where dM/dt is ensemble mean mass gain (Mg C ha⁻¹ y⁻¹), σ the s.d. of the mass gain distribution and N the number of observations.

For N we use the RAINFOR sample published in 2009, either conservatively N=135, the total number of observational plots or N=1,545, the total number of plot census years, reflecting the stochastic nature of disturbance and therefore the near independence of plot results from year-to-year. Net gain results are statistically significant at the 95% level if t_sim≥t_{(0.975,N=135)}≈t_{(0.975,N=1,545)}=1.96 and at the 99% level if t_sim≥t_{(0.995,N=135)}≈t_{(0.995,N=1,545)}=2.58.