全球变化知识资源中心

Model and observational data.

We used monthly data from transient coupled simulations carried out with 36 GCMs participating in CMIP5 (Supplementary Table 1). The whole period analysed spans from 1960 to 2099, including a single historical run (1960–2005) per model and the corresponding twenty-first century projection (2006–2099) following Representative Concentration Pathway 8.5 (RCP8.5, see Supplementary Information).

In addition to precipitation and sea-level pressure (p_SL), we used three-dimensional wind (u = {u, v}) and specific humidity (q) data for the computation of a vertically integrated water vapour flux (Q), which is derived as:

where g₀ and p_S are the gravitational acceleration and pressure at the surface, respectively. The top pressure level p_T is set to 100 hPa. A dynamic (Q^(ω)) and a thermodynamic (Q^(θ)) component of Q are derived as in equation (1), but prescribing the climatological mean (1960–1999) profiles of q ( ) and u ( ), respectively:

We use a trapezoidal rule to estimate the vertical integrals in the equations (1)–(3) from the pressure levels available in the GCMs’ atmospheric fields. The water vapour flux divergence terms (∇ ⋅ Q, ∇ ⋅ Q^(ω) and ∇ ⋅ Q^(θ)) further used for the analysis depicted in Fig. 1 were computed using central finite-difference approximations in spherical coordinates.

The observational data set includes four gridded products of land precipitation from the Global Precipitation Climatology Centre (GPCC), the National Centers for Environmental Prediction (PREC/L), the Climatic Research Unit (CRU) and from the University of Delaware (UDEL). The monthly data from the Hadley Centre Sea Level Pressure data set³⁰ (HadSLP2) were adopted for p_SL. Further details, references, and an evaluation of these products are provided as Supplementary Information.

The Amazon Basin-wide average of a given variable is calculated as the spatial area-weighted mean over a region of ~7 million km² (Fig. 3). To have a consistent domain within the multiple GCM and observational data sets, all fields were previously interpolated and analysed in a common rectangular grid of 2.0° latitude–longitude.

The P_A responses to ACF are derived as differences between climatologies computed at the end of the twentieth (1960–1999) and twenty-first (2060–2099) centuries. We note that part of these differences could be affected by multi-decadal internal variability simulated in GCMs. The noise induced by these stochastic variations is offset when averages are applied to ensembles (see, for example, Fig. 2), but the standard deviation values should partially reflect uncertainties due to variability.

The dry-season length (DSL) is defined as the number of months per year with precipitation rates below 2 mm d⁻¹. A minimum DSL of three months is used to define a rainfall regime (DSL₃₊) where savannahs/treeless biomes prevail and, therefore, where rainforest is less resilient to permanent changes in precipitation. Vulnerable areas within the Amazonian domain are computed yearly with the pixels satisfying DSL₃₊. The present-day vegetation partitioning is based on the MODIS land-cover product MCD12C1 (ref. 29) averaged from 2001 to 2010. We note the manifest problem of using monthly precipitation data to properly define dry seasons. However, for the purpose of this study, this disadvantage has a lesser impact because we compute time averages and regional statistics to derive, respectively, DSL climatologies (from which fractional values are obtained, Fig. 4a) and yearly DSL₃₊ areas (Fig. 4b).

Empirical models of Amazonian precipitation.

A multivariate regression analysis was performed to derive the link between patterns of precipitation in Amazonia and p_SL of different regions of the globe. The choice of p_SL as an indicator of the large-scale motion relies in part on the availability of historical reconstructions of good quality³⁰. The analysis accounts for the rainfall distribution across the basin. For this purpose, the leading modes of variability were extracted using a standard empirical orthogonal function decomposition. This approach allows us to describe the spatiotemporal variability of P_A with only a few time series (we used the first 10 principal components), hence avoiding multiple analyses at the pixel level. This approach also ensures a spatial coherency on the reconstructed P_A fields.

Several regression models were derived with the historical data (1960–2012) both from the observational products (used to compute the constrained rainfall projections) and from the GCM simulations (used for the method evaluation; see Supplementary Fig. 4). The analysis was done separately for each month of the year. However, information from the months preceding and following the one assessed was also included in the calibrating data, allowing us to increase by a factor of 3 the record length used in the analyses, thereby providing robustness in regression-parameter computation. Hence, both the interannual variability and a small part of the seasonal cycle are accounted for in the models’ calibration.

The explanatory variables correspond to normalized monthly p_SL anomalies averaged onto a rectangular grid of 20° latitude–longitude. Although coarse, this grid defines worldwide a large set of potential predictors (see Supplementary Fig. 3) compared with the record length used in the regression computation (53 × 3 months). We therefore combined a predictor selection and an ensemble technique to manage multi-collinearity/overfitting issues and, hence, to enhance the predictive performance of the models (see Supplementary Information).