Global dependence of field-observed leaf area index in woody species on climate: a systematic review

Authors


  • Editor: Ian Wright

Abstract

Aim

Leaf area index (LAI) is one of the key variables related to carbon, water and nutrient cycles in terrestrial ecosystems, but its global distribution patterns remain poorly understood. We evaluated the dependence of LAI on mean annual temperature (MAT) and wetness index (WI; a ratio of annual precipitation to potential evapotranspiration) for three plant functional types (PFTs: deciduous broadleaf, DB; evergreen conifer, EC; evergreen broadleaf, EB) at the global scale.

Location

Global.

Methods

We developed a new global database of unprecedented size (2606 published values) of field-observed LAI (site-specific maximum) values for vegetation of woody species. To maximize the generic applicability of our analysis, we standardized the definition of LAI, and corrected or excluded potentially erroneous data obtained from indirect optical methods.

Results

The global dependence of LAI on MAT showed a reverse S-shaped pattern, in which LAI peaked at around 8.9 and 25.0°C and was lowest at around −10.0 and 18.8°C. The dependence on WI followed a saturation curve levelling off at around log WI = 0.30. LAI for evergreen forests increased linearly with increasing WI, but that for DB showed a curvilinear pattern saturating at log WI = 0.03. EC forests had higher LAI values than those of DB forests under cool conditions (MAT ≤ 8.9°C), but similar values under temperate conditions (MAT = 8.9–18.8°C).

Main conclusions

This analysis of global LAI−climate relationships supports the general belief that temperature limits LAI under cool conditions whereas water availability plays a predominant role under other conditions. We also found that these patterns differed significantly between PFTs, suggesting that the LAI of different PFTs may respond differently to climate change. Our study provides a broad empirical basis for predicting the global distribution of LAI and for analysing the effects of global climate change on vegetation structure and function.

Introduction

Leaf area index (LAI), the amount of leaf area per unit ground area, is one of the dominant factors controlling plant productivity, expressed as, for example, gross or net primary production (e.g. Kira, 1991; Luo et al., 2004), because it is a major determinant of the amount of light intercepted by a canopy and hence of photosynthesis. LAI also strongly influences the water and nutrient cycles in forest ecosystems (e.g. Granier et al., 2000; Erisman & Draaijers, 2003; Schulze, 2006). Consequently, most ecosystem process-based models that simulate carbon, water and nutrient cycles include LAI as a key parameter (e.g. Running & Coughlan, 1988; Sellers et al., 1996; Liu et al., 1997; Ito & Oikawa, 2002).

Although it is widely recognized that LAI is sensitive to climate variables, there are only a limited number of studies investigating LAI−climate relationships under field conditions (e.g. Grier & Running, 1977; Battaglia et al., 1998; Luo et al., 2004; Ladd et al., 2009; Schleppi et al., 2010). Furthermore, there is considerable variation between the LAI−climate relationships reported in different studies. For example, Grier & Running, (1977) showed that LAI was linearly dependent on mean annual precipitation (MAP), while Schleppi et al. (2010) found a saturation effect. Since these studies were conducted on a regional scale, differences can probably be ascribed to limited numbers of observations, shallow gradients for environmental variables and/or site-specific effects such as soil nutrient content and geography. The global dependence of LAI on climate variables thus remains unclear. Quantifying this dependence could make a large contribution to a better understanding of the factors determining the global distribution of LAI, and thus to reliable predictions of carbon and water cycles, vegetation–climate feedback effects and climate change.

LAI may be different for different plant functional types (PFTs). For example, many authors have stated that evergreen forests have higher LAIs than deciduous forests (e.g. Gower et al., 1993; Asner et al., 2003; Larcher, 2003). This has been ascribed to clumping of foliage, which allows deeper penetration of light (Niinemets & Anten, 2009), and a longer leaf life span (Gower et al., 1993; Niinemets, 2010). However, previous studies have not analysed between-PFT variation in LAI in relation to climate variables. The extent to which these putative between-PFT differences are consistent across the globe therefore remains unclear.

Recent advances in satellite-based remote sensing enable us to use globally continuous LAI datasets. However, these remote sensing techniques suffer from serious uncertainties that arise from difficulties in both reflectance measurements (e.g. saturation in dense canopies) and empirical assumptions made during data processing, especially with respect to cloud contamination, canopy architecture and vegetation classification (Garrigues et al., 2008). Comprehensive analysis of field-observed LAI is therefore necessary for a reliable evaluation of the dependence of LAI on climate variables. However, we are currently aware of only one previous study that has analysed field-observed LAI on a global scale, that of Asner et al. (2003).

Asner et al. (2003) collected data from 1008 stands and showed that LAI varies greatly between biomes and PFTs. However, they did not study the dependence of LAI on climate variables, so their results are of limited use for predicting the responses of LAI to climate change. Furthermore, they did not use a standardized definition of LAI. This lack of a clear definition is particularly problematic for species like conifers that have non-flat leaves, because in these cases the LAIs can vary more than three-fold depending on the definitions used (Barclay, 1998). Standardization of the definition of LAI is therefore necessary for reliable analysis.

Another difficulty facing those attempting comprehensive analysis of LAI is that the accuracy of estimates of LAI is strongly dependent on the method of measurement used (Gower et al., 1999; Bréda, 2003; Jonckheere et al., 2004). Indirect optical (IR) methods, which estimate LAI from measurements of the interception of light by the canopy, are generally considered to underestimate LAI because they cannot detect the effect of foliage clumping (which is expressed as the clumping index, Ω). Thus, detailed classification of measurement methods and correcting the bias of IR estimates of LAI (by applying a typical Ω value from standard literature) is necessary before analysis, but this was not done in the study of Asner et al. (2003).

In this study, we developed a new global database of LAI for woody species of unprecedented size (about 2600 observations). In it, we used a standardized definition of LAI and corrected the bias in IR estimates of LAI. Our primary objectives were to characterize the global dependence of LAI on climate variables and to determine the extent to which this relationship differs between PFTs.

Methods

Data compilation

A list of abbreviations used here for terms and variables is shown in Table 1. We obtained a total of 2606 observations of field LAI from 554 literature sources (Appendices S1 and S2 in Supporting Information), all published between 1932 and 2011 (Fig. 1). Most of the data were plot-based. As far as possible, we chose site-specific maximum LAI values, and therefore excluded LAI data points with low values that were obviously due to dry season measurements, disturbances or stands being immature or old and declining. In addition, we excluded LAI data affected by artificial treatments such as CO2 enrichment, continuous fertilization and irrigation. We included data from plantations, and divided the vegetation status for records in the dataset into three subgroups: ‘plantation’ (trees had been planted), ‘natural forest’ (trees had grown naturally) and ‘not described’. Plot-related information, such as the name of the dominant species, latitude, longitude, elevation and climate variables, was also included. PFTs of species were classified into six groups based on leaf phenology and leaf shape (Table 2, Fig. 2c,d). The biome type of each plot was derived from a gridded global dataset of the Köppen–Geiger classification (Peel et al., 2007).

Figure 1.

(a) Global distribution of field plots included in the leaf area index (LAI) database. (b) The relationship between mean annual wetness index (WI) and temperature (MAT), classified by mean annual precipitation (MAP).

Figure 2.

Dependence of leaf area index (LAI) on mean annual temperature (MAT) and on wetness index (WI): (a), (b) all data pooled; (c), (d) classified by plant functional type (PFT); (e), (f) the three major PFTs, with averaged data (1 °C and 0.1 intervals for MAT and WI, respectively) for each PFT shown as a thick dark symbol. The average data were removed if the number of data points was less than 10. The lines indicate the best result of the bent-line model. Abbreviations: DB, deciduous broad leaf; EB, evergreen broadleaf; EC, evergreen conifer; DC, deciduous conifer; sDB, semi (dry)-deciduous; Mix, mixture of various functional types.

Table 1. A list of abbreviations of terms and variables
Term or variableExplanationUnit
DDirect method 
IRIndirect optical method assuming leaf randomness 
ICIndirect optical method taking into account leaf clumping 
OOther methods 
PFTPlant functional type 
DBDeciduous broad leaf 
DCDeciduous conifer 
EBEvergreen broadleaf 
ECEvergreen conifer 
sDBSemi (dry)-deciduous 
MixMixture of various functional types 
HSAHalf of the total leaf surface area 
GLMMGeneral linear mixed model 
PRCPartial regression coefficient 
LAILeaf area indexm2 m−2
LAIeEffective leaf area index (i.e. IR estimates of LAI)m2 m−2
ΩClumping index 
ΩeElement clumping index 
γeWithin-shoot (needle) clumping index 
WIWetness index 
MATMean annual temperature°C
MAPMean annual precipitationmm
MGTMean temperature during growing season°C
MGWIMean wetness index during growing season 
PETPotential (reference) evapotranspirationmm
VPDVapour pressure deficitkPa
RnNet radiationMJ m−2 day−1
Table 2. Statistical metrics of leaf area index for all pooled data, biomes, plant functional types and vegetation states (see Table 1 for abbreviations)
CategoryMeanSDMax.Min.nErroneous*OutlierOriginal
Total
  1. Significant differences (α = 0.05) between mean values are indicated by different letters.
  2. *Potentially erroneous estimates of LAI by the indirect optical methods were excluded, on the basis of the results of general linear mixed model analysis (Appendices S8 and S9).
All4.211.8423.500.6923931041092606
Biome        
Tropical4.13 a1.9022.400.70234289271
Dry2.71 b2.1412.670.711071450171
Temperate4.70 a1.8323.500.69910627979
Continental4.04 a1.7616.660.7211170361153
Polar3.17 b2.179.680.72250732
PFT        
EC4.93 a1.8223.500.75902012914
DB4.09 ab1.6512.100.72809011820
EB3.48 b1.9922.400.6951710472693
DC2.77 c1.8510.811.18240630
sDB3.95 ab2.3117.400.72570259
Mix3.80 b1.7610.310.75840690
Vegetation state        
Plantation4.87 a1.7620.130.72442190461
Natural3.92 b1.9123.500.69129843901431
Mix6.61 c1.459.501.33402143
Not described4.26 ab1.7118.500.756134018671

We standardized the definition of LAI as half of the total leaf surface area (HSA) per unit ground area (Chen & Black, 1992). For conversion from the other definitions to HSA, we employed the factors published by Cannell (1982); the ratio of HSA to projected leaf area is 1.4 for pine, 1.15 for other conifers and 1.0 for broadleaf species. When the definition was not provided in the source literature, we assumed that: (1) LAI for flat broadleaf species was defined on a projected leaf area basis, because other definitions are rarely employed for this leaf type, and (2) LAI for conifer species measured with indirect optical instruments was defined as HSA (Stenberg et al., 1994). The data that did not fit these assumptions were excluded.

Measurement methods for LAI were separated into four groups: (1) direct methods (D, e.g. litter trapping, allometric approaches); (2) indirect optical methods that assumed foliage randomness (IR, i.e. measurements of canopy gap fraction using a plant canopy analyser, fish-eye cameras or quantum sensors); (3) indirect optical methods that allowed for non-randomness of foliage, i.e. clumping (IC, combining measurements of gap fraction and of gap-size distribution in the canopy or direct calibration against D estimates of LAI); and (4) other methods (OT, e.g. point quadrat sampling, litter trapping for evergreen species and other empirical approaches). Details of each methodology and the theory underlying it can be found elsewhere (e.g. Gower et al., 1999; Bréda, 2003; Jonckheere et al., 2004).

We corrected the potential bias associated with IR methods in order to increase the reliability and general applicability of our analysis. Although the source of the bias is mainly attributable to clumping of foliage in the canopy and to interception of radiation by woody elements, we assumed that the contribution of the latter factor to LAI estimates was relatively small (Kucharik et al., 1997). Consequently, each LAI value obtained from an IR method (effective leaf area index, LAIe) was converted to ‘true’ LAI using the formula:

display math(1)

where Ω is the total clumping index. Ω can be expressed in terms of the element clumping index (Ωe; leaf and shoot element for broadleaf species and conifers, respectively) and the within-shoot (i.e. needle) clumping index (γe):

display math(2)

For flat broadleaves, γe is equal to 1.0. These clumping indices were derived from the global database compiled by Pisek et al. (2011). Since we found that the effects of climate on clumping indices were generally marginal (i.e. correlations were weak; Appendix S3), we used the global mean values of Ωe and γe (0.85 and 1.52, respectively) for correcting LAIe.

Climate variables

Mean annual temperature (MAT; °C) and MAP (mm) were derived from the gridded global dataset CRU CL 2.0 (10′ × 10′), containing monthly means (long-term average, 1961–90; New et al., 2002). The net radiation (Rn; MJ m−2 day−1), vapour pressure deficit (VPD; kPa) and Penman–Monteith potential evapotranspiration (PET; mm) were also calculated from global climate variables derived from CRU CL 1.0 (0.5° × 0.5°) and 2.0 according to FAO-56 guidelines (Allen et al., 1998). We also calculated the ratio of MAP to PET to give the wetness index (WI), which represents plot water availability. The climate variables for each plot were summed (precipitation, PET) or averaged (other variables) across all months of the year. When MAT and MAP were given in the source literature we used those values. In addition, we calculated the mean temperature and wetness index during the growing season (MGT and MGW, respectively; the definition of the growing season is given in Appendix S4).

Data analysis

LAI was log10-transformed to meet assumptions of normality and homogeneity for regression analysis, as is common practice in global meta-analysis of leaf traits (e.g. Wright et al., 2004). In addition, since the dependences of LAI on WI showed curvilinear relationships (data not shown), WI values were also log10-transformed to linearize the relationships. To avoid misinterpretation of the results, we determined outliers by applying an interquartile range approach (Appendix S4) to data pooled for all PFTs, and removed these outliers from all statistical analyses presented here. Mean LAI values were compared among biomes, PFTs, vegetation states and measurement methods using one-way ANOVA, followed by Tukey's post hoc comparisons when effects were significant.

Given the significant correlations among climate variables (Appendix S5), we characterized climate on the basis of two variables, MAT and WI, which are critical in shaping vegetation distribution and structure. Since the global dependences of LAI on MAT and WI were nonlinear (Fig. 2), we divided each relationship into groups according to the positions of the points of inflection between groups. The inflection points were objectively determined using a bent-line model (i.e. by fitting a bent line with segment regression; Chappell, 1989). For pooled data for all PFTs, the LAI−MAT relationship was divided into three temperature groups with the inflection points at 8.9 and 18.8 °C (Fig. 2a). Going from low to high temperature, the groups were denoted ‘Cool’, ‘Temperate’ and ‘Warm’, respectively. The LAI−WI relationship was divided into two groups with the inflection point at log WI = 0.30. The groups below and above the boundary were denoted ‘Dry’ and ‘Wet’, respectively. The LAI−MAT relationship for deciduous broadleaves (DB) and evergreen conifers (EC) was also divided into two groups with the points of inflection at 8.6 and 9.7 °C, respectively, but that for evergreen broadleaves (EB) was not divided into groups (Fig. 2e). The LAI−WI relationship for DB was divided into two groups with the inflection point at log WI = 0.03, but those for EB and EC were not divided into groups (Fig. 2f). The nomenclature used for the climate groups for each individual PFT followed that used for the all-PFT pooled data.

The LAI−climate relationship within each climate and PFT group was quantified using a general linear mixed model (GLMM):

display math(3)

where the parameters a and b are model coefficients, c is the intercept and ε is the random effect. We employed ‘site’, which was defined as a data group within which the geographical coordinates (spatial resolution of 0.1°) were the same, as a random effect to incorporate into the analysis the fact that plots within a site may be non-independent due to a similarity of growth environments. MAT and WI were significantly correlated in most of the data groups, but the correlations were generally weak (r2 < 0.240; Appendix S5) irrespective of climate group and PFT. Thus, the problem of collinearity is marginal in our analyses. After analysing the effects of MAT and WI, we evaluated whether or not other climate variables, such as VPD, Rn, MGT and MGW, improve the predictive power of the GLMM, based on Akaike information criteria (AIC; Appendix S6).

The effects of measurement methods and PFTs on the climate−LAI relationships were also analysed with a GLMM:

display math(4)

where FAC is a categorical factor, i.e. either measurement method (D versus IR after application of Ω correction) or PFT (DB, EB or EC), parameters a to e are model coefficients and f is the intercept. The effect of the categorical factors was evaluated with a null hypothesis test (t-test) of parameters c to e which represent the effect of the categorical on the intercept, MAT response and WI response, respectively. We first examined the effect of Ω correction on IR estimates of LAI. If the LAI estimates from corrected IR differed significantly from the D estimates, we regarded them as potentially erroneous and excluded them from all other analyses. We also examined the effect of vegetation state (‘plantation’ versus ‘natural forest’) but found that differences between them were mostly marginal (Appendix S7). All statistical analyses were performed using spss software (IBM spss 19.0, Armonk, NY, USA) and the R software package (R Core Team, 2012). Further details of the methodology are described in Appendix S4.

Results

To exclude the potential bias introduced by indirect optical methods that assume foliage randomness (IR), we first examined the effect of clumping correction (i.e. applying a typical clumping index, Ω, to IR estimates; D versus corrected IR) on the LAI−climate relationships for individual PFTs (Appendix S8). Since the IR estimates of LAI were generally lower than the D estimates, applying Ω correction lessened the bias in most data groups. However, it still caused significant underestimation of LAI for EB forests in the MAT range 10−20 °C (Appendix S9). Although significant differences in the slope of the LAI−MAT relationship were also found for EC under cool conditions, the LAI difference for a given MAT was marginal because the two regression lines crossed at the midpoint of the MAT range (Fig. e in Appendix S8). We therefore excluded the IR estimates of LAI for EB in the MAT range 10–20 °C from all subsequent analyses as being erroneous. In the other data groups, the corrected IR estimates of LAI were included in the analyses.

The global mean LAI value in the database was 4.21 (Table 2). LAI differed significantly among biomes, with ranking (high to low) being temperate ≈ tropical ≈ continental > polar ≈ dry climate. Although we classified the dataset into six PFTs, more than 93% of the total records were obtained from three PFTs: EC, DB and EB. The mean LAI was lower for EB than for DB and EC, while the standard deviation was markedly larger for EB than for the other PFTs. Plantations had significantly higher LAIs than natural vegetation.

The global dependence of LAI on MAT showed a reverse S-shaped pattern which was highest at around 8.9 and 25.0 °C and lowest at around −10.0 and 18.8 °C (Fig. 2a). The dependence on WI was curvilinear, saturating at around log WI = 0.30 (Fig. 2b). These dependences were quantified using a GLMM applied to individual climate groups (equation (3), Fig. 3). Although LAI showed strong (P < 0.01) and positive correlations with MAT under cool dry conditions, the opposite trend was found under temperate dry conditions. Under warm dry conditions, LAI was weakly correlated with MAT. The dependences of LAI on MAT were generally marginal under wet conditions, irrespective of temperature group. With respect to WI, significantly positive dependences of LAI were found under dry conditions irrespective of temperature group (Fig. 3b). By contrast, LAI was negatively correlated with WI under warm wet conditions.

Figure 3.

Result of general linear mixed model analysis (equation (3)) for the dependence of leaf area index (LAI) on climate variables. (a) Partial regression coefficient (PRC) of the mean annual temperature (MAT) term. (b) PRC of the wetness index (WI) term. ns, P ≥ 0.05; *P < 0.05; **P < 0.01; ***P < 0.001. Other parameters are shown in Appendix S11.

The dependence of LAI on climate variables differed among the three major PFTs (Fig. 2e,f). Although the LAI−MAT relationship for DB and EC showed a bell-shaped pattern peaking at MAT = 9.8 and 7.9 °C, respectively, no correlation was found for EB. With respect to WI, the LAI for evergreen forests increased linearly with increasing WI, but that for DB showed a curvilinear pattern, saturating at around log WI = 0.03. To examine these differences in more detail, the LAI−climate relationships for individual PFTs were compared for the same climate groups (Fig. 4, Table 3). EB data under cool conditions (MAT < 8°C) were excluded from the analysis due to the small number of data points available. Under cool wet conditions, the slope of the LAI−WI relationship was significantly steeper for EC than for DB (Fig. 4b,d). Consequently, LAI in a given climate condition was generally higher for EC than for DB. However, such differences disappeared under temperate conditions, due mainly to an increase in LAI for DB relative to EC in a given WI (Fig. 4d,e). The slope of the LAI−WI relationship was significantly greater for EC than for EB. Consequently, LAI in a given climate variable was higher for EC than for EB (Fig. 4c,f). LAI for DB also tended to be higher than that for EB, though GLMM analysis showed that the differences were generally marginal (Fig. 4a,e).

Figure 4.

Dependence of leaf area index (LAI) on mean annual temperature (MAT) and on wetness index (WI), classified by climate group and plant functional type (PFT). When comparing deciduous broadleaf (DB) with evergreen conifer (EC) and broadleaf (EB), dry and wet conditions for evergreen species were defined as log WI ≤ 0.1 and > −0.1, respectively, and cool conditions for EB were defined as 8 °C < MAT ≤ 22 °C. Data were fitted by a general linear mixed model (y = ax + b; the regression parameters are not shown).

Table 3. Result of general linear mixed model analysis (equation (4)) for the effect of plant functional types (PFTs) on leaf area index−climate relationships
Data groupModel termn
ClimatePFTMAT × PFTWI × PFTPFTDBECEB
  1. When comparing deciduous broadleaf (DB) with evergreen conifer (EC) and broadleaf (EB), dry and wet conditions for evergreen species were defined as log WI ≤ 0.1 and > −0.1, respectively, and cool conditions for EB were defined as 8 °C < MAT ≤ 22 °C.
  2. Significance for each model term is shown as follows: ns, P ≥ 0.05; *P < 0.05; **P < 0.01; ***P < 0.001.
Cool dryEC vs. DBnsnsns198352 
Cool wetEC vs. DBns***ns367506 
Temperate dryDB vs. ECnsnsns96123 
Temperate wetDB vs. EBnsns*96 283
Temperate dryDB vs. ECnsnsns148257 
Temperate wetDB vs. EBnsnsns148 261
TemperateEC vs. EB******** 293430

When VPD and Rn were incorporated into the model, and when MGT and MGWI, respectively, were employed instead of MAT and WI, there was no improvement in the predictive power for most of the climate groups (Appendix S6), indicating that MAT and WI explained most of the variance in LAI.

Discussion

The LAI−climate relationship

The global LAI−WI relationship showed a saturation pattern, with the inflection point at around log WI = 0.30 (Fig. 2b). Since irrigation of forests in dry regions increases the LAI (Dobbertin et al., 2010), the significant positive LAI−WI relationship observed under dry conditions indicates that the LAIs are limited by site water availability. Although a decrease in LAI under water-limited conditions reduces light harvesting by the canopy, it also reduces water loss from the leaves, thus allowing higher stomatal conductance and photosynthesis per unit leaf area. Optimality modelling of forest carbon exchange indicates that this strategy may increase carbon gain under water-limited conditions (McMurtrie et al., 2008; Sterck et al., 2011). Under wet conditions, the slope of the LAI−WI relationship was markedly lower than that under dry conditions (Fig. 3b). This may indicate that resources other than water limit biomass production and thus LAI. However, under warm wet conditions, LAI was negatively correlated with WI. This is probably associated with the negative effect of heavy precipitation on LAI (Santiago et al., 2000; Schuur et al., 2001), which most likely results from indirect effects such as low irradiance due to frequent cloud and fog and reductions in plant nutrient uptake due to nutrient leaching from the soil and to hypoxic limitation of litter decomposition and root activity.

The present study found that the global relationship between LAI and MAT shows a reverse S-shaped curve with the highest points at 8.9 and 25.0 °C and the lowest at −10.0 and 18.8 °C (Fig. 2a). Such a pattern has not been observed in any previous study; for example, a bell-shaped pattern was reported for LAI−MAT relationships by Battaglia et al. (1998 and Luo et al. (2004). This discrepancy probably lies at least partly in the fact that the previous studies were made at a regional scale and for limited ranges of MAT. Our dataset suggests that the pattern of global LAI−climate relationships is more complicated than has been believed up to now.

A positive LAI−MAT relationship under cool conditions is to be expected, because it is known that LAI tends to be lower at high latitudes (Van Cleve et al., 1983; Larcher, 2003). This may be explained partly by nutrient availability, since soil nitrogen (N) availability increases with increasing MAT in cool climates (Craine et al., 2009) because the decomposition and mineralization processes are limited by low soil temperature (Robinson, 2002), and LAI increases with increasing nitrogen availability (Anten et al., 1995; Hikosaka, 2003; McMurtrie et al., 2008). Lower Rn at high latitudes may also be associated with the positive LAI−MAT relationship, because Rn decreases with decreasing MAT across our entire dataset (r2 = 0.741, P < 0.001), and because LAI decreases with decreasing light availability (Hikosaka, 2003).

A significant depression in LAI was found at around MAT = 19 °C under dry conditions (Figs 2a & 3). This depression may be ascribed to the fact that many of the world's drier regions that are characterized by low LAI (e.g. open woodland in savanna; warm Mediterranean climates) tend to have an annual MAT of around 19 °C (Fig. 1b). Although a significant negative dependence of LAI on MAT was also found, exclusion of the MAT term from the model (equation (3)) had little influence on the predictive power (AIC increased by only 0.8), indicating that contribution of MAT to the depression is marginal. We thus concluded that the depression of LAI observed at around 19 °C is associated with water availability.

It should be mentioned that our data tend to concentrate on those regions where ecologists have conducted extensive studies, such as the USA, Europe, Australia and Japan (Fig. 1a). Consequently, under temperate and warm conditions, data were relatively sparser from wet conditions than from dry conditions (Fig. 1b), indicating that results from our dataset may overemphasize the depression of LAI at around 19 °C. We thus stress the importance of making additional measurements in warm and wet climates, such as the humid subtropics, which are mainly distributed in the southern part of China and in Brazil and Africa.

Our results support the general ideas underlying previous studies that temperature and water availability play the predominant roles in controlling LAI. However, previous studies characterized global LAI−climate relationship at the biome scale (Asner et al., 2003; Larcher, 2003), which was too coarse a level to predict global distribution of LAI. We characterized the details of the dependence of LAI on climate at the plot scale and our results can be summarized as follows: (1) water availability limits LAI at log WI < 0.30; (2) temperature limits LAI at MAT < 8.9 °C; and (3) heavy precipitation limits LAI in warm and humid climates (MAT > 18.8 °C and WI > 0.30). The LAI−climate relationships investigated here may help us to gain a better understanding of the global distribution of LAI and ecosystem CO2 fluxes and their possible responses to climate change over the next few centuries.

Differences in LAI−climate relationships between PFTs

It has been widely recognized that LAIs of EC forests are higher than those of DB forests within a given biome (Asner et al., 2003; Larcher, 2003) and site (Gower et al., 1993). In this study, LAI under cool conditions was generally higher for EC than for DB (Table 3, Fig. 4), which is consistent with this idea. One explanation for the difference in LAI between PFTs may be the fact that foliage clumping is higher in EC than in DB. Since foliage clumping increases irradiance in deeper canopy layers (Niinemets & Anten, 2009), such plants can maintain high LAIs (Hikosaka & Hirose, 1997). Longer leaf life span in EC than in DB has also been believed to contribute to this difference, because an increase in leaf life span enables plants to support a greater foliage area (Gower et al., 1993; Niinemets, 2010). The fact that LAI was higher for EC than for EB under temperate conditions is also consistent with these assumptions. However, the above-mentioned differences in LAI between DB and EC disappeared under temperate conditions, suggesting that the idea that EC forests have larger LAIs than DB forests is not globally applicable. The lack of between-PFT differences under temperate conditions was mainly associated with an increase in LAI for DB in temperate conditions relative to cool conditions (Fig. 4d,e). Although there is no theory that explains why EC have similar LAI to DB in these conditions, many studies have reported an empirical relationship that more productive canopy tends to have higher LAI (e.g. Fassnacht & Gower, 1997; Luo et al., 2004). A longer growing season in temperate conditions than in cool conditions would particularly benefit DB species, which tend to have higher photosynthetic capacities than EC species. By contrast, for evergreen species, carbon loss due to foliage respiration during the dormant season would increase with increasing growth temperature. These contrasts in physiological and phenological characteristics between PFTs may explain the lack of the differences in LAI between DB and EC. We also found that not only EC but also EB showed similar LAI to that of DB, although leaf life span is generally longer in EB than in DB. This suggests that leaf life span is not one of the principal factors accounting for between-PFT variations in LAI.

PFT-dependent variations in LAI were relatively small compared with the climatic effects, suggesting that climate has a more important influence on LAI than does PFT. However, LAI−climate relationships were apparently different between PFTs, especially for WI (Table 3). For example, LAI values for evergreen species were linearly correlated with WI, while the LAI−WI relationship for DB exhibited a saturation pattern (Fig. 2f). This result suggests that LAIs for individual PFTs may respond differently to climate change. These differential responses may, in turn, have important implications for regional and global patterns of carbon, water and nutrient cycling under climate change as well as for feedback between vegetation and climate.

Uncertainty in our analysis

Although LAI was found to be significantly correlated with MAT, WI and PFTs, there was still a fair degree of unexplained variation in our global LAI data set. This variation could stem from methodological inaccuracies and/or biotic and abiotic factors that are not considered in this analysis. Regarding methodological uncertainties, different experimental designs used in the source studies, such as plot sizes and number of samplings, could have been responsible for some of the unexplained variation in our database. Furthermore, the extensive time period over which the source literature was published (1932−2011) could lead to uncertainty because measurement techniques, especially for the IR approach, have changed considerably over time and because ambient CO2 concentration has increased by more than 50 p.p.m. in the past 50 years, which may have affected LAI (Norby et al., 2005; McCarthy et al., 2006). The fact that only 6% of the total number of records explicitly included understorey LAI (grass/herb/shrub layers) may also cause uncertainty, especially in the case of regions of sparse vegetation such as savanna and tundra. However, existing data showed that understorey LAI was fairly constant at around 1.0 irrespective of overstorey LAI (Appendix S10) and climate (data not shown), probably because of severe environmental limitations on LAI in sparse vegetation. Consequently, inclusion of the understorey LAI in GLMM analysis (i.e. overstorey LAI plus 1.0) did not affect our results. We thus conclude that the uncertainty due to our not having explicitly considered understorey LAI is marginal.

With respect to biotic and abiotic factors, between-site differences in soil nutrient availability, a factor which is known to affect LAI (Anten et al., 1995; Hikosaka, 2003), could also account for some of the variation in LAI in our dataset. Furthermore, although we excluded data where LAI values were obviously low due to forest stands being either immature or old and declining, differences in stand age could have caused the variation in LAI, because in some cases LAIs gradually decrease after a peak age and the rate of decline can differ considerably between species and growth conditions (Ryan et al., 1997). Physiological variability among species within each PFT could also be an important factor contributing to the variability in LAI. Moreover, natural and human disturbance (e.g. change of land use, air pollution, wildfire, typhoon-mediated windfall) in the past few decades could have caused variation in LAI. All of these factors probably contributed to the variation in our LAI analysis, but it is difficult to evaluate these uncertainties. Nevertheless, we obtained strongly statistically significant results (P < 0.01), most of which are consistent with previous ideas. We thus believe that these uncertainties are unlikely to cause significant changes in the observed patterns of global relationships between LAI, climates and PFTs presented in this paper.

Acknowledgements

We thank many researchers for providing additional information about their published studies. We also thank Dr D. Kabeya and Dr K. Nishina for helpful suggestions about statistical analysis. This study was supported by a Grant-in-Aid for Scientific Research on Innovative Areas in Japan ‘Comprehensive studies of plant responses to high CO2 world by an innovative consortium of ecologists and molecular biologists’ (no. 22114513).

Note: Additional references for the LAI dataset (Appendix S1) can be found in Appendix S2 at [http://onlinelibrary.wiley.com/doi/10.1111/geb.12133/suppinfo].

A. Iio is a post-doctoral researcher at the National Laboratory of Environmental Studies (NIES) in Japan. His current research interest is global meta-analysis of key parameters of ecosystem process-based models.

Author contributions: K.H., N.A and A. Ito conceived the study. A. Iio collated data, undertook analysis and wrote the manuscript. Y.N. helped with programming and statistical analysis. K.H., N.A and A. Ito revised the manuscript and contributed to the final version.

Ancillary