Supporting Information

Appendix S1: projected climate change by the three GCMs

We conducted cell-by-cell temporal trend analysis on the major climate variables

from the bias-corrected results of the three GCMs using the non-parametric Mann-

Kendall test. The results show considerable differences between the projected

climatologies by the three GCMs (Fig. S1). All three predict significant (P < 0.01),

region-wide warming trends over the 21st century (Fig. S1a); however, the magnitudes of

warming trends differ: HadCM3 has the strongest warming trend followed by CCSM3

and PCM. Although the magnitudes of warming trends in CCSM3 and HadCM3 differ,

both models predict that the eastern and southern Amazon will experience the highest

rates of warming during the 21st century (Fig. S1a), and that these regions will also

experience significant upward trends in the vapor pressure deficit (VPD) (Fig. S1b). In

contrast, PCM predicts that most of the basin will experience relatively mild changes in

temperature, and little or no changes in VPD.

With regard to rainfall, the HadCM3 projection indicates that approximately half

(53%) of the region, mainly the eastern and southeastern Amazon, will suffer significant

reductions in precipitation during the 21st century (Fig. S1c). In contrast, PCM and

CCSM3 predict that significant portions of the basin (47% and 62% of the region,

respectively), located mainly in southern and western portion of the basin, will

experience significant increases in precipitation. Comparison of these predictions against

the nineteen GCM projections examined by Malhi et al. (2009) indicate that they span the

range of precipitation predictions for the Amazon region: the PCM projection presents a

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

slightly warmer but wetter future climate, while the HadCM3 projection represents an

extremely hot and dry scenario and the CCSM3 projection falls in between.

Fig. S1. Maps of temporal trends from 2009 to 2100 in (a) annual air temperature, (b)

vapor pressure deficit and (c) precipitation from the bias-corrected projections of three

GCMs (i.e. PCM, CCSM3, and HadCM3); grey areas denote non-significant trends with

90% confidence.

23

24

25

26

27

28

29

Appendix S2: spatial patterns of the Business-As-Usual and Governance land-use

scenarios in Amazonia

The Global Land-Use dataset (GLU) incorporates the SAGE-HYDE 3.3.1 dataset

and provides global land-use transitions on a 1° grid from 1700 to 1999 (Hurtt et al.,

2006). The Business-As-Usual (BAU) and Governance (GOV) datasets provide yearly 1-

km horizontal resolution land-use maps from 2002 to 2050. Consistent land-use transition

data for the entire period (1700–2100) were produced by following the procedure: (1) the

1-km BAU and GOV land-use maps were used to calculate the fractions of the three

land-use types for each 1° grid cell; (2) linear interpolation was used to connect the 1999

GLU land-use map to the 2008 GOV land-use map to produce continuous land-use maps

for the 2000–2008 period; (3) the land-use transition rates between 2000 and 2008 were

computed from the interpolated land-use maps; (4) the BAU and GOV land-use maps for

the 2009–2050 period were converted to their respective land-use transition rates for the

same period; and (5) BAU and GOV land-use transition rates were extrapolated from

2050 to 2100 by assuming the continuation of the same transition rates in 2050 until no

forest left in the grid cells.

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

Fig. S2. Spatial patterns of land-use composition in 2100 under the (a) GOV and (b)

BAU scenarios; the projected rates of land-use transformation were derived and extended

from Soares-Filho et al. (2006).

46

47

48

49

Appendix S3: evaluation of the biosphere models' performance

In this study, we assessed the three biosphere models' predictive capabilities by

comparing modeled above-ground live biomass (AGB) and percent tree cover against

two sets of satellite remote sensing (RS) derived AGB estimates (Baccini et al., 2012,

Saatchi et al., 2011) and the MODIS Collection 5 MOD44B percent tree cover product

(DiMiceli et al., 2011), respectively. The Baccini et al. (2012), Saatchi et al. (2011), and

MOD44B data have nominal spatial resolutions of 500m, 1km, 500m, respectively. We

first aggregated these data to 1-degree and then compared them with our model results.

Both of the two RS biomass products were produced using a combination of data from

spaceborne LiDAR, optical and microwave imagery, and in situ inventory plots. An

evaluation of models' performance against site-level measurements of carbon fluxes and

aboveground biomass dynamics can be found in an ongoing study (Levine et al., The role

of short-term climate variability in governing Amazonian biomass dynamics, in

preparation) and Powell et al. (2013).

Although there are non-negligible discrepancies between the two RS products,

across the models, and between the models and RS products, the model predictions of

AGB show a similar spatial gradient to the satellite-derived estimates of regional AGB,

with AGB increasing from the southern and southeastern dry savanna zones to the

western and northeastern dense, moist forest regions (Fig. S3). ED2 AGB agrees well

with the RS AGB in these high biomass regions, but tends to underestimate AGB in low

biomass regions (Fig. S3a,d-f). IBIS AGB estimates are systematically lower than the RS

estimates in areas of high biomass but higher in areas of low biomass (Fig. S3b,d-f),

while JULES AGB is systematically higher than the RS AGB (Fig. S3c-f). A recent study

5

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66

67

68

69

70

71

72

by Levine et al. (2014) compared site-level AGB estimates from the ED2 model and the

above two RS biomass products with the measurements of the RAINFOR network(Malhi

et al., 2002), and found that these model and RS estimates are qualitatively consistent

with that observed in the RAINFOR plots.

To more quantitatively describe the vegetation structure, we calculate the percent

tree cover (f treecover ¿ using the fully projected tree foliage cover by following Kucharik et

al. (2000): f treecover=1−exp (−0.5 ∙ LAI tree), where 0.5 is the canopy extinction coefficient,

and LAI tree is the total leaf area index of all tree plant function types. Percent tree cover

from all models generally compares well with the RS based estimate (Fig. S4). f treecover

shows the similar spatial gradient as AGB in models and RS estimates (Fig. S3,S4); the

higher biomass regions have higher f treecover, while the lower biomass regions have lower

f treecover. The inset quantile-quantile (Q-Q) plot (Fig. S4) shows that IBIS and JULES have

generally higher tree cover fractions than ED2 and the MODIS product. Although ED2

predicts lower tree cover in the lower tree cover region relative to the MODIS product,

ED2's predictions of tree cover fraction are close to the MODIS values in the areas with

middle to high tree cover fractions (Fig. S4).

Despite the uncertainty in the AGB and f treecover estimates of models and remote

sensing products and some discrepancies between the model and remote sensing results,

the overall similar spatial patterns and gradients from the model and remote sensing

results suggest that all three biosphere models are able to reasonably capture the present-

day composition and spatial variability of Amazonian ecosystems.

6

73

74

75

76

77

78

79

80

81

82

83

84

85

86

87

88

89

90

91

92

93

Fig. S3. Spatial patterns of present-day (2000~2008) above-ground biomass across

Amazonia from model estimates of (a) ED2, (b) IBIS, and (c) JULES, and remote

sensing based estimates of (d) Baccini et al. (2012) and (e) Saatchi et al. (2011), and (f)

the quantile-quantile plots of model estimates against remote sensing (RS) based

estimates.

7

94

95

96

97

98

99

Fig. S4. Spatial patterns of present-day (2000~2008) percent tree cover across Amazonia

from (a) ED2, (b) IBIS, (c) JULES, and (d) MODIS collection 5 MOD44B product. The

inset graph shows the quantile-quantile plot of model estimates against remote sensing

based estimates.

8

100

101

102

103

104

Appendix S4: evaluation of association between water stress regime and AGB from

model simulations and remote sensing estimates across the Amazon.

Table S1. Summary of the strength of association between water stress (MCWD) and

AGB from model simulations and remote sensing estimates across the Amazon; the

strength of association is quantified by Pearson's simply linear correlations and Kendall's

Tau (i.e. the rank correlation).

Water Stress Regime

Statistical Test AGBED2 AGBIBIS AGBJULES AGBBaccini AGBSaatchi

MCWDPearson's r 0.75*** 0.67*** 0.61*** 0.65*** 0.63***

Kendall's τ 0.61*** 0.45*** 0.45*** 0.48*** 0.42***

*** P<0.01

References

Baccini A, Goetz SJ, Walker WS et al. (2012) Estimated carbon dioxide emissions from

tropical deforestation improved by carbon-density maps. Nature Climate Change, 2, 182-

185.

Dimiceli CM, Carroll ML, Sohlberg RA, Huang C, Hansen MC, Townshend JRG (2011)

Vegetation Continuous Fields MOD44b. College Park, Maryland, University of

Maryland.

Kucharik CJ, Foley JA, Delire C et al. (2000) Testing the performance of a Dynamic

Global Ecosystem Model: Water balance, carbon balance, and vegetation structure.

Global Biogeochemical Cycles, 14, 795-825.

Levine NM, Zhang K, Longo M et al. (2014) Ecosystem heterogeneity determines the

resilience of the Amazon to climate change. Nature Climate Change, In revision.

9

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

Malhi Y, Aragao LEOC, Galbraith D et al. (2009) Exploring the likelihood and

mechanism of a climate-change-induced dieback of the Amazon rainforest. Proceedings

of the National Academy of Sciences of the United States of America, 106, 20610-20615.

Malhi Y, Phillips OL, Lloyd J et al. (2002) An international network to monitor the

structure, composition and dynamics of Amazonian forests (RAINFOR). Journal of

Vegetation Science, 13, 439-450.

Powell TL, Galbraith DR, Christoffersen BO et al. (2013) Confronting model predictions

of carbon fluxes with measurements of Amazon forests subjected to experimental

drought. New Phytologist, 200, 350-364.

Saatchi SS, Harris NL, Brown S et al. (2011) Benchmark map of forest carbon stocks in

tropical regions across three continents. Proceedings of the National Academy of Sciences

of the United States of America, 108, 9899-9904.

Soares-Filho BS, Nepstad DC, Curran LM et al. (2006) Modelling conservation in the

Amazon basin. Nature, 440, 520-523.

10

125

126

127

128

129

130

131

132

133

134

135

136

137

138