Introduction

Trees come in all shapes and sizes – from incredibly tall and slender, to short with wide, flat crowns1,2,3,4,5,6,7. This incredible diversity in tree crown architecture plays an important role in driving variation in growth, water use and competition among tree species1,4,8,9,10. Moreover, tree crown architecture underpins key emergent properties of woody ecosystems, including their 3D canopy structure, aboveground biomass, primary productivity and hydrology8,11,12,13,14,15,16. Consequently, uncovering the environmental, ecological and evolutionary drivers that shape the crown architecture of the world’s trees is central to better understanding the processes that constrain the structure and function of woody ecosystems. It is also essential for developing more realistic representations of these ecosystems in vegetation models17,18,19,20 and bridging the gap between field and remote sensing observations21,22,23,24 – both of which are needed to track how terrestrial ecosystems are responding to rapid global change.

Differences in crown architecture among trees are the result of species employing a variety of strategies to meet a series of competing physiological, structural, competitive, defensive and reproductive demands (Table 1). Trees expand their crowns vertically and laterally to intercept light, compete with neighbours and disperse seeds, while also needing to maintain water transport to their leaves and mechanical stability4,6,7,10,25,26,27,28,29,30,31,32. The balance different tree species strike between these various priorities depends on their environment, ecological strategy and evolutionary history, and will be reflected in the scaling relationships between different axes of tree size, such as their height, crown width and stem diameter8,9,22,33,34,35,36,37,38,39,40,41. For instance, in arid climates woody biomass allocation tends to shift away from height growth and towards crown expansion to limit the risk of hydraulic failure and maximise energy capture, resulting in trees that are shorter for a given diameter and have wider crown profiles2,8,33,34,38,42,43,44. Conversely, when water and nutrients are non-limiting to photosynthesis, strong competition for light leads to greater investment in height growth and relative allocation of carbon to woody tissues44, pushing trees closer to their structural and hydraulic safety margins7,8,16,26,27,32,33. Similarly, tree species have also adapted the size and shape of their crowns to minimise the risk of damage from wind, fire, snow and browsing40,45,46,47,48,49,50. Yet despite clear evidence that crown allometric scaling relationships can vary considerably among tree species, we lack a unified picture of how and why they do so. Nor do we understand how different axes of crown size and shape covary with one another, how they relate to other key plant functional traits, or how they vary among plant lineages.

Table 1 Hypothesised drivers of variation in crown architecture among tree species
Full size table

Here, we assemble a global dataset capturing information on the stem diameter (D), height (H), crown diameter (CD) and crown aspect ratio (CAR, defined as CD/H) for over half a million trees (Fig. 1). Using these data, we develop an approach for modelling variation in H–D, CD–D and CAR–D scaling relationships among species that allows us to compare their crown sizes and shapes while explicitly controlling for differences in their stem sizes. We apply this method to 1914 well-sampled tree species that span all major woody biomes and clades and use it to: (1) characterise the full spectrum of crown architectural types observed across the world’s tree species and biomes; (2) explore whether crown architectural traits are phylogenetically constrained and identify which clades have crown sizes and shapes that are particularly extreme; and (3) test a series of predictions about how H–D, CD–D and CAR–D scaling relationships vary in relation to climate, competition, disturbance and other functional traits related to plant metabolism, hydraulics, structural stability and dispersal (Table 1). We show that tree species span a broad range of crown architectural types, encompassing everything from slender to stout stems and narrow to broad crowns. Variation in H–D scaling relationships is primarily controlled by water availability and light competition, with tropical forests in Southeast Asia home to disproportionately high concentrations of species with tall, slender growth forms. Conversely, crown width is predominantly shaped by exposure to wind and fire, with legume trees in African savannas achieving some of the widest crowns for their stem diameters.

Fig. 1: Overview of the global tree allometry database.
figure 1

a Geographic distribution of the allometric data (n = 374,888 individual trees belonging to 1914 species). Individual tree records were aggregated in 200 × 200 km grid cells (mean number of trees per grid cell = 742). The map was obtained from the Natural Earth database (https://www.naturalearthdata.com) and is displayed using a Robinson projection (EPSG:54030). Relationships between each tree’s stem diameter and its b height (H), c crown diameter (CD) and d crown aspect ratio (CAR) are shown on a logarithmic scale. CAR is defined as the ratio between CD and H, with values lower than 1 indicating a vertical crown profile (H > CD) while values greater than 1 corresponding to a horizontal crown profile (CD > H). Points are coloured according to the aridity index value assigned to each tree based its geographic coordinates, with larger values corresponding to drier conditions (shown in red). Graphical illustration of the approach used to generate size-standardized estimates of e tree height (HRESID), f crown diameter (CDRESID) and g crown aspect ratio (CARRESID) for each tree species. Regression lines are predicted values obtained by fitting a linear model to the entire dataset (grey points). By comparing predicted and observed value of H, CD and CAR, we quantified how much each species departs, on average, from this general trend and identified ones with greater (blue points) or smaller H, CD and CAR values (red points) than expected given their stem diameters. This approach is conceptually similar to generating species-level predictions of H, CD and CAR at a fixed size (e.g., D = 30 cm), but avoids the need to arbitrarily select a size at which to compare species.

Full size image

Results

Global variation in crown architectural types

We used data from 374,888 individual trees to generate size-standardised estimates of H, CD and CAR for 1914 well-sampled species using two complimentary statistical approaches (see Methods for details). The first involved using linear mixed effects models to estimate the height, crown diameter and crown aspect ratio of each species for a tree with a stem diameter of 30 cm (hereafter HD = 30, CDD = 30 and CARD = 30). The second approach is conceptually and quantitatively similar to the first, but avoids the need to choose an arbitrary stem size at which to compare species (Supplementary Fig. 1). It uses the residuals of a linear regression model to determine if a species has H, CD and CAR values that are – on average – larger (positive residuals) or smaller (negative residuals) than expected for their range of stem sizes (hereafter HRESID, CDRESID and CARRESID; see Fig. 1e–g for a graphical representation of the method).

Across the 1914 tree species considered in our analysis, we found enormous variation in size-standardized estimates of tree height, crown diameter and crown aspect ratio (Fig. 2). Specifically, HD=30 varied 12.1-fold across species, ranging from <4 m in species like Juniperus osteosperma and Maerua crassifolia to >30 m in several species of the genera Shorea, Parashorea, Hopea and Vatica (all Dipterocarpaceae) and as much as 43.2 m in Eucalyptus regnans. By contrast, CDD=30 was less than half as variable among species, ranging 5.4-fold from <4 m in several species of the genus Picea to >14 m in ones like Brachystegia wangermeeana and Pterocarpus tinctorius in the Fabaceae. As for crown profile shape, CARD=30 ranged 10.5-fold across species. At one end of the spectrum, species like Abies sibirica and Vatica dulitensis were more than five times as tall as their crowns are wide (CARD = 30 < 0.2), whilst several species of the genera Acacia, Vachellia and Senegalia (all Fabaceae) had crowns that are noticeably wider than they are tall (CARD = 30 > 1.2).

Fig. 2: Global spectrum of tree crown architecture.
figure 2

Tree crown architectural types and their distribution across biomes for the 1309 tree species for which both height and crown size were measured. a Tree species were grouped into one of nine architectural types based on their size-standardized height (HRESID) and crown diameter values (CDRESID). The vertical and horizontal lines mark the 25th and 75th percentile of the data and the size of each circle reflects the crown aspect ratio (CARRESID). Examples of tree species that occupy different areas of this crown architectural spectrum are highlighted. b Proportion of species belonging to the nine architectural types for each biome. See Supplementary Table 4 for pairwise comparisons of HRESID, CDRESID and CARRESID values among biomes, and Supplementary Fig. 5 for a breakdown of the nine architectural types among angiosperms and gymnosperms.

Full size image

For the 1309 species for which we were able to estimate both HRESID and CDRESID, we found that these two axes of crown architecture were positively correlated with one another (Pearson correlation coefficient, ρ = 0.26, P < 0.001). However, despite a general trend of taller trees also having wider crowns, the relationship between HRESID and CDRESID was relatively weak. Species occupied all possible combinations of the tree height vs crown diameter spectrum (Fig. 2a), including short and narrow (9.3% of species), short and wide (5.4%), tall and wide (7.9%), and tall and narrow (1.0%).

Where species fell within this spectrum depended, at least in part, on their biome association (Fig. 2b). For example, a high proportion of species found in drylands, temperate woodlands, tropical savannas and tropical dry forests had either short and/or wide crowns (59.6–100% of species, depending on the biome), but almost none were tall (0–8.8%). By contrast, boreal, temperate and tropical rainforests had a considerably higher proportion of species with tall and/or narrow crowns (28.0–60.5%). Overall, differences among biomes were more pronounced for height and crown aspect ratio than for crown width, with biome association explaining 33%, 39% and 5% of the variation in HRESID, CARRESID, and CDRESID among species, respectively (see Supplementary Table 4 for pairwise comparisons among biomes based on ANOVAs). However, our results also revealed considerable variation in tree architectural types within biomes (Fig. 2b), highlighting how tree species with very different crown architectures can be found in similar environments.

Fingerprint of evolution history on crown architecture

From a macroevolutionary perspective, angiosperms and gymnosperms had very similar mean values of HD = 30 (17.8 m and 17.3 m, respectively). However, angiosperms spanned a considerably larger range of heights (3.9–42.3 m), particularly at the tall end of the spectrum, where they included 99 of the 100 species with the highest HD=30 values. By contrast, both crown diameter and aspect ratio were noticeably larger in angiosperms compared to gymnosperms (CDD=30 = 6.4 m vs 5.3 m; CARD=30 = 0.43 vs 0.36). When we placed estimates of HRESID, CDRESID, and CARRESID onto a time-calibrated phylogeny of seed plants, we found that all three exhibited a significant degree of phylogenetic signal (Fig. 3), with Pagel’s λ values of 0.70, 0.54, and 0.63, respectively (P < 0.001 in all cases).

Fig. 3: Variation in tree crown architecture across the tree of life.
figure 3

Size-standardized estimates of a tree height (HRESID, n = 1225 species), b crown diameter (CDRESID, n = 870 species) and c crown aspect ratio (CARRESID, n = 868 species) are mapped onto a time-calibrated phylogeny of seed plants116. Low values (red) indicate species that are shorter, with narrower crowns and smaller crown aspect ratios than expected given the size of their stem, while high values (blue) indicate the opposite. HRESID, CDRESID and CARRESID all exhibited phylogenetic signal, with Pagel’s λ values of 0.70, 0.54, and 0.63, respectively (P < 0.001 in all three cases based on a likelihood ratio test). Plant families and genera with mean HRESID, CDRESID and CARRESID values that are significantly lower (red lines) or higher (blue lines) than zero are highlighted on each phylogenetic tree (see Supplementary Tables 5–6 for full details). Note that only species that were a direct match to those in the phylogeny were used for the phylogenetic analysis.

Full size image

In particular, we found several plant genera and families that stand out based on their tree height, crown size, and shape (Fig. 3 and Supplementary Tables 5–6). For tree height, 25 out of 63 families and 31 out of 86 genera that we tested had mean HRESID values that were significantly different from zero. For angiosperms, species in the Dipterocarpaceae (HD = 30 = 24.1 m), Myristicaceae (HD = 30 = 24.4 m), Burseraceae (HD=30 = 21.3 m), Annonaceae (HD=30 = 21.3 m) and Betulaceae (HD = 30 = 19.6 m) were particularly tall for their stem diameters (Fig. 3a). From a biogeographic standpoint, we found that Southeast Asia was home to an especially high concentration of species with tall and slender growth forms, with nine of the 10 species with the highest HRESID values native to this region, including species in the genera Shorea, Parashorea and Vatica (Dipterocarpaceae) and Knema (Myristicaceae). At the opposite end of the spectrum, species in the Ericaceae (HD = 30 = 11.4 m), Combretaceae (HD = 30 = 13.3 m), Fagaceae (HD = 30 = 15.1 m) and Fabaceae (HD = 30 = 16.2 m) were significantly shorter that average for a given stem diameter.

The picture within gymnosperms was equally varied. Cupressaceae were generally shorter than expected (HD = 30 = 14.8 m), despite including species like Sequoia sempervirens which can grow incredibly tall in absolute terms. Conversely, within the Pinaceae we found a clear divide between species in the genus Pinus which are shorter than average (HD = 30 = 16.2 m) and those belonging to Larix and Picea that are taller (HD = 30 = 22.2 m and 20.6 m, respectively).

In terms of crown size and shape, we found that 21/56 families and 26/60 genera (CDRESID) and 12/56 families and 16/60 genera (CARRESID) had values that departed significantly from zero (Supplementary Tables 5–6). One clade that stood out in particular is the Fabaceae (Fig. 3b, c). Of the top 10 species with the highest CDRESID and CARRESID values, four and six were Fabaceae, respectively. Fabaceae had crowns that are both much wider than average (CDD = 30 = 7.6 m) and more horizontal in their aspect ratio (CARD = 30 = 0.58). This trend was predominantly driven by species that occupy savannas in Africa and the Americas, including ones in the genera Senegalia, Acacia, Brachystegia, Vachellia and Pterocarpus (mean CDD = 30 and CARD=30 across 40 tropical savanna specialists = 8.6 m and 0.88, respectively). By contrast, Fabaceae from tropical rainforests (CDD = 30 = 7.2 m; CDD = 30 = 0.39) and temperate forests (CDD = 30 = 6.5 m; CDD = 30 = 0.37) had crown sizes and profiles that were very similar to other species in these biomes.

Within gymnosperms, species in the Podocarpaceae (CDD = 30 =5.2 m), Cupressaceae (CDD = 30 = 5.3 m) and Pinaceae (CDD = 30 = 5.3 m) all exhibited narrower than average crowns (Fig. 3b). In the Pinaceae this effect was far stronger than any variation observed in tree height, resulting in crown aspect ratios that are also much smaller than average (CARD = 30 = 0.34; Fig. 3c). A similar trend emerged for the Rubiaceae and Lauraceae, where species were generally both tall and with narrow crowns, resulting in low crown aspect ratios (CARD = 30 = 0.35 and 0.37, respectively). The exact opposite was true for species in the Combretaceae, Ulmaceae and Fagaceae, which, due to their relatively short stature and wide crowns had particularly large crown aspect ratios (CARD = 30 = 0.73, 0.55, and 0.48, respectively).

Drivers crown architectural variation in the world’s trees

Phylogenetic generalised least squares models relating variation in tree crown architecture among species to climate, tree cover, disturbance, and functional traits explained 55%, 28% and 34% of the variation in HRESID, CDRESID and CARRESID, respectively (Fig. 4). Differences in height among species were predominantly controlled by water availability (aridity, and to a lesser extent rainfall seasonality) and tree cover (Fig. 4a), with HRESID decreasing rapidly with rising aridity and increasing steadily as tree cover increased (Fig. 5a, b). On average, species growing where potential evapotranspiration was equal to or less than mean annual precipitation (aridity index ≤ 1) were almost twice as tall for a given stem diameter (HD = 30 = 20.1 m) as those where the aridity index was ≥2 (HD=30 = 11.1 m). Similarly, HD=30 increased from 11.5 m to 21.4 m when comparing species growing where tree cover was ≤20% and ≥ 80% (Fig. 5a). However, we also found that when water was non-limiting (aridity index ≤ 1), trees could vary hugely in their investment in height growth, with HD=30 ranging anywhere between 6.4 m and 43.2 m (Fig. 5b). In contrast to aridity and tree cover, we only observed a modest positive association between HRESID and mean annual temperature (MAT). Temperature did however indirectly influence HRESID through its interaction with aridity (Supplementary Fig. 6). Specifically, we found that HRESID declines with increasing aridity was much less pronounced for trees growing in cold climates (MAT < 10 °C) compared to warm ones (MAT > 20 °C).

Fig. 4: Drivers of variation in crown architecture among tree species.
figure 4

Standardized model coefficients for each predictor variable were obtained by fitting phylogenetic generalised least squares regressions to size-standardized estimates of a tree height (HRESID, n = 1910 species), b crown diameter (CDRESID, n = 1313 species), and c crown aspect ratio (CARRESID, n = 1309 species). Error bars show both standard errors (thick lines) and 95% confidence intervals (thin lines) of the model coefficients. Significantly positive and negative coefficients are shown in blue and red, respectively, while those for which the 95% confidence intervals overlap with zero are shown in grey.

Full size image
Fig. 5: Variation in the crown architecture along gradients of tree cover, aridity, and wind gust speed.
figure 5

Points are species-level estimates of size-standardized a–c tree height (HRESID, n = 1910 species), d–f crown diameter (CDRESID, n = 1313 species) and g–i crown aspect ratio (CARRESID, n = 1309 species). Fitted lines correspond to phylogenetic generalised least squares model predictions generated by keeping all other predictors fixed at their mean values. Negative values of HRESID, CDRESID, and CARRESID indicate species that are shorter, with narrower crowns and smaller crown aspect ratios than expected given the size of their stem, while positive values denote the opposite. Values of tree cover, aridity index and wind gust speed represent are means calculated across all individual trees of a given species. Note that the aridity index was log-transformed and that larger values correspond to drier conditions.

Full size image

Water availability and tree cover also emerged as significant predictors of CDRESID, with species generally having narrower crowns in more arid and seasonal climates, and where tree cover was higher (Fig. 4b). But these effects were much less pronounced than for tree height (Fig. 5d, e). Consequently, we found that variation in CARRESID along aridity and tree cover gradients was mostly driven by changes in HRESID, with crown profiles becoming markedly more vertical as aridity decreased and tree cover increased (Figs. 4c, 5g, h). For example, the average CARD=30 of species growing where tree cover was ≤20% (0.77) was more than twice that of those where tree cover was ≥ 80% (0.36). Conversely, crown diameter was much more strongly influenced by disturbances such as wind and fire (Fig. 4b). In particular, CDRESID decreased markedly as maximum wind gust speeds increased (Fig. 5f). Species also tended to be shorter for a given stem diameter where wind speeds were higher (Fig. 5c), but this effect was more subtle than for crown diameter, meaning that overall crown profiles became significantly narrower as wind speeds increased (Fig. 5i). By contrast, areas with higher frequencies of wildfires harboured species with wider crowns and higher crown aspect ratios (Fig. 4b, c), but with similar HRESID values.

While variation in HRESID, CDRESID, and CARRESID among tree species was predominantly associated with climate, tree cover and risk of disturbance, we also found that these crown architectural traits covaried with other plant functional traits (Fig. 4). After accounting for environmental effects, we found that both HRESID and CDRESID were significantly greater in species with higher leaf nitrogen content, with HRESID also positively correlated to specific leaf area. Additionally, while we observed no clear relationships between HRESID and wood density, we did find that species with wider crowns and more horizontal crown profiles had denser wood. By contrast, none of the three crown architectural traits exhibited any relationship with seed mass.

Discussion

Using allometric data from hundreds of thousands of trees across the world, our study provides a global picture of how tree species vary in their crown architecture (Fig. 2) and what drives this variation (Fig. 4). While ecologists have long been aware that trees can differ in their crown size and shape1,2,5,7,10, we have lacked a quantitative understanding of where the boundaries of this crown architectural spectrum lie. Not only did we show that tree species can vary considerably in the scaling relationships between their height, crown width, and stem diameters, but we also found that size-standardised estimates of tree height and crown diameter were largely decoupled, forming two independent axes of variation in crown architecture (Fig. 2a).

Where species fell within this crown architectural spectrum depended largely on their environment, their evolutionary history, and ecological strategy. Like previous studies, we found that species confined to more arid and seasonal biomes were generally much shorter for a given stem diameter than those growing in more humid climates2,8,38,44,51. We also found that the range of crown architectural types was much greater in biomes where water and temperature were non-limiting and angiosperms dominate the flora, such as tropical rainforests and temperate forests. In these environments, tree species living side by side can have incredibly different crown forms depending on their ecological strategy and evolutionary history8,36,52,53,54,55, whereas where conditions for growth are harsher, there is less flexibility in the range of crown sizes and shapes that species can assume.

From a macroevolutionary standpoint, when growing in similar environments angiosperm were only marginally taller for a given diameter than gymnosperms. However, angiosperms consistently had wider crowns. This pattern reflects a fundamental difference in the growth strategy of the two major clades, with gymnosperms investing less in lateral crown expansion due to strong apical dominance and control56, resulting in crown profiles that are generally more vertical than those of angiosperms22,33. When exploring how evolutionary history has shaped variation in crown architecture in more detail, we found that HRESID, CDRESID, and CARRESID all exhibited a clear phylogenetic fingerprint (Fig. 3), with several plant linages standing out. For instance, we found that dipterocarps – and trees native to Southeast Asia more generally – achieve remarkable heights for a given stem diameter39,40,57. In terms of crown size, we showed that several species of Fabaceae that grow in tropical savannas in Africa and the Americas had exceptionally wide crowns. This helps explain previous observations that trees in these regions have larger crowns than those of Australia8,34,40, where savannas are largely dominated by smaller-crowned eucalypts.

Aridity and tree cover emerged as the strongest predictors of tree height scaling relationships, which is consistent with previous empirical and theoretical research showing how investment in height growth is modulated by risk of hydraulic failure2,17,30,42,43,44,58 and competition for light7,16,32,33,59. Interestingly, species that were tallest for a given stem diameter (HD = 30 > 30 m, top 1% of species) occurred within a narrow band of aridity values (0.65–0.81) where rainfall only slightly exceeded potential evapotranspiration (Fig. 5b). Once the aridity index decreased below 0.5, species with very high HRESID values disappeared, possibly due to a combination of waterlogging from excessive rainfall and/or growth limitations linked to lower temperatures and high cloud cover60,61,62.

CDRESID also decreased with aridity, suggesting trees limit the size of their crowns (and by proxy their total leaf area) in environments where water is scarce to reduce transpiration and minimise risk of hydraulic stress8,33,34,63. As expected, we also found that CDRESID decreased with tree cover, which is consistent with trees prioritizing height growth over crown expansion in response to increasing competition for light7,16,32,33,34. However, the effects of both aridity and tree cover were much weaker for crown diameter than for height. One possible reason for this is that tree cover and aridity are generally negatively correlated, leading to their effects cancelling each other out: in humid climates where tree cover also tends to be high, trees would be able to support wider crowns, were it not for increased competition for space. The net result is that investment in crown expansion relative to height growth (as captured by CARRESID) increased progressively in drier and more open habitats (Fig. 5g, h), a growth strategy that maximises energy capture and hydraulic safety44.

In contrast to aridity and tree cover, temperature played a secondary role shaping variation in tree architecture (Fig. 4). However, we did find that trees tended to be taller for a given diameter in warmer climates, which is consistent with observations that the world’s tallest trees inhabit mild and warm climates with little seasonality62. This also fits our understanding of how cold temperatures impact tree height growth64 and the fact that trees in cold climates generally have small vessels to minimise risk of embolisms under freezing conditions, which limits their ability to grow tall42. Moreover, we found that temperature indirectly influenced tree height scaling relationships by exacerbating the effects of aridity (Supplementary Fig. 6). Warmer temperatures are associated with higher vapour pressure deficits, requiring trees to have more soil water to meet higher evaporative demands65,66,67. This suggests that even small decreases in rainfall and/or increases in temperature in warm climates could lead to disproportionately large impacts on forest structure2,68,69.

Risk of disturbance by wind and fire emerged as stronger predictors of crown diameter scaling relationships than climate (Fig. 4). Species adapted to windy conditions were shorter, had narrower crowns, and more vertical crown profiles for a given stem diameter than those where risk of exposure to high wind speeds was low. These adaptations would make them less prone to uprooting and snapping in high winds, as the risk of both is proportional to total crown surface area27,31,45,46. Our findings are also consistent with observations that some of the world’s tallest tropical trees grow where the risk of wind disturbance is low, such as the Guiana Shield and in Borneo27,31,39,70.

Crown width was also positively associated with burned area fraction, indicating that trees in fire-prone environments generally allocate more resources to lateral crown expansion. This could partly be a result of lower competition for light in more open environments where fire is frequent, which would allow trees to maximise light interception by spreading their crowns laterally40. It could also reflect a more direct response to fire, with trees developing wide crowns to shade out grasses and limit fuel loads71. Additionally, adaptations to fire may be confounded with those associated with herbivory, which also plays an important role in shaping tree architecture in savannas49,50, with wider crowns serving as a protective strategy against browsing40,72. In terms of how fire might affect height scaling relationships, our expectation was that trees would generally invest more in height growth to escape fire (and herbivory). By contrast, we found no relationship between HRESID and burned area fraction, which could be because some species adopt the alternative strategy of fire resistance through the growth of thicker stems and bark40,49.

As for other disturbance agents, such as snow accumulation, we would expect trees exposed to high snow loads to have narrower crowns and more slender profiles47,48. While we did not test this directly, we did find that CARRESID was generally lower in colder climates. Given that snow cover duration and mean annual temperature were highly correlated (Supplementary Fig. 3), this temperature response may in part reflect an adaptation to minimising the risk of stem breakage from snow accumulation in the crown in cold climates.

Crown architectural traits covaried with several other plant functional traits related to photosynthesis and structural integrity. For instance, we found that CDRESID and CARRESID were positively associated with wood density, which is thought to confer the mechanical strength and resistance needed for trees to grow large branches and wide crowns9,36,37. However, while our results support this hypothesis, we also found that numerous species had wide, horizontal crowns despite having relatively low wood densities. This highlights how other properties, such as branching architecture, may be just as important in determining a tree’s structural integrity27.

Species that were taller and with wider crowns for a given stem diameter also generally had higher concentrations of nitrogen in their leaves. Leaf nitrogen content is a cornerstone of the ‘fast-slow’ plant economic spectrum73,74, with species that have high leaf nitrogen generally capable of rapid growth, but also less able to tolerate shade due to higher metabolic and respiration rates75,76. Based on this, we would expect species with higher leaf nitrogen to invest more in both height growth and crown expansion to allow them to intercept more light, limit self-shading, and optimise the distribution of leaves across their crowns. Moreover, higher leaf nitrogen content should lead to greater photosynthetic rates and investment in height growth, as found previously along a rainfall gradient in Australia44. This is also consistent with work from the U.S. showing that taller and more slender trees tend to be less shade tolerant38. Similarly, early successional species in tropical forests that are adapted to grow rapidly in height to take advantage of gap openings generally have high leaf nitrogen content52,77.

By contrast, we found no evidence that large-seeded species were architecturally any different to those with small seeds. Seed mass has previously been shown to correlate positively with maximum plant height and canopy area, which some have proposed is the result of species with longer life spans and greater adult sizes being more likely to invest in large seeds78,79,80. Alternatively, seed mass may correlate with crown architecture through its association with seed dispersal28,81. For example, species with light, wind-dispersed seeds might profit from being taller to increase dispersal range, while species that have ballistic seeds or ones that simply drop to the ground might benefit from wider branches to increase distance from parent trees52,81. However, we found no support for either of these hypotheses when relating variation in seed mass to HRESID and CDRESID. If these processes are at play, they may well be better captured by other facets of crown architecture (e.g., maximum tree height and crown width).

Our findings highlight several fruitful avenues for future research. An obvious next step would be to expand the spectrum of crown architectural traits to other axes of crown size and shape, such as crown depth, surface area, and volume8,16,34. This would allow us to test long-standing predictions about how crown size and shape reflect a compromise (in terms of carbon gains) between greater light interception and higher maintenance costs6, and explore how the outcome of these trade-offs varies with water and light availability8,82. In this regard, efforts to better characterise crown architecture are likely to benefit from growing access to technologies such as terrestrial laser scanning (TLS). These can provide a much richer picture of a tree’s crown and local surroundings, including reconstructing its branching structure, accurate 3D volumes and within-crown distribution of leaves81,83,84. Extracting these measurements from TLS point clouds remains a challenge, but access to data and automated processing pipelines are continuously improving1,85.

In addition to extending the range of crown architectural traits, another important direction for future research would be to explore how crown allometric scaling relationships vary within species. Our study focused exclusively on species-level differences in crown architecture, as for most species we had insufficient data to robustly incorporate intraspecific variation. However, it is well known that trees can exhibit considerable plasticity in their crown shapes and sizes, shifting their allocations to vertical and horizontal growth as they age and in response to both competition and climate2,4,16,33. What remains less clear is the extent to which shifts in crown architecture along environmental gradients observed across species are mirrored by ones occurring within species. A previous study of 342 widely-distributed species suggests that shifts in H–D scaling relationships along aridity gradients reflect both turnover in species composition and intraspecific plasticity2. However, it also revealed that patterns across and within species were only consistent in 70% of cases, and that intraspecific plasticity was of secondary importance (relative to species turnover) in driving shifts in H–D scaling relationships along aridity gradients2. Moreover, this analysis focused exclusively on tree height, and much less is known about plasticity in crown size and shape outside of a few temperate forest species16,33. A big question concerns the relative importance of ontogeny, competition, and the environment in driving plasticity in crown architecture. Similarly, it is unclear whether certain plant lineages show greater ability to plastically adapt the size and shape of their crowns compared to others, and if so why.

While there is clearly plenty of scope to build on our findings, our study provides a key starting point for characterising the crown architectural spectrum of the world’s trees. Not only do we capture the range of possible crown architectural types, we also take an important step towards explaining what drives this immense variation. Our results highlight how crown architecture is jointly constrained by a range of processes related to a tree’s environment, ecological strategy, and evolutionary history. This understanding will underpin ongoing efforts to leverage remote sensing technologies to track tree carbon stocks and dynamics at scale21,22,23,24. It is also critical for developing the next generation of Earth System Models that accurately simulate variation in vegetation structure and dynamics by incorporating more realistic representations of how tree crowns vary among biomes, plant functional types, and in coordination with other traits17,19,20. All of this is essential to better understanding the processes that shape the structure and function of woody biomes and tracking how these are responding to rapid global change.

Methods

Individual tree height and crown size data

We compiled 528,311 georeferenced records of individual trees for which stem diameter (D, cm), height (H, m) and/or crown diameter (CD, m) were measured (Fig. 1a). For trees where both H and CD were measured (n = 340,221; 64.4%), we also calculated their crown aspect ratio (CAR) as CD/H, where CAR < 1 denotes a vertical crown profile and CAR > 1 a flat or horizontal profile1,86. These data were obtained from 62,435 globally distributed sites which encompass all major terrestrial biomes and span a gradient in mean annual temperature of −15.1–30.1 °C and 143–7157 mm yr-1 in rainfall. Sampled trees span multiple orders of magnitude in size and crown shape (Fig. 1b–d) and represent 5161 tree species from 1451 genera and 187 plant families.

Most of the data (94.4% of records) were sourced from the Tallo database2. Additionally, we also obtained data from Alberta’s Permanent Sample Plots network in Canada (n = 12,171 trees) and the ICP Forests network in Europe (n = 17,540 trees). Allometric data were quality controlled following the protocols of the Tallo database2. Briefly, we first used Mahalanobis distance to identify and remove possible data entry errors by screening for trees with unrealistically large or small H and CD values for a given stem diameter. Species names were then standardized against those of The Plant List (TPL) using the taxonstand package87 in R (version 4.2.2)88. Lastly, we excluded records from species that did not meet our working definition of trees: perennial woody seed plants with a single dominant stem that are self-supporting and undergo secondary growth (i.e., excluding ferns, palms, short multi-stemmed shrubs, and lianas).

Species level, size-standardized estimates of tree height, crown diameter, and crown aspect ratio

Tree species can differ considerably in their maximum size and developmental strategies, so to directly compare their crown architecture, we used two complimentary approaches to generate size-standardized estimates of H, CD, and CAR at the species level. For these and all subsequent analyses, we focused on species with at least 10 trees sampled within the same biome and spanning a minimum D range of 20 cm between the smallest and largest measured tree (see Environmental data for details on how trees were assigned to biomes). In total, 1914 species represented by 374,888 individual trees met these criteria for at least one of the three axes of crown size and shape and 1309 species represented by 251,733 trees had sufficient data for all three (1910 species for H, 1313 for CD and 1309 for CAR; see Supplementary Table 1 for details). These 1914 species cover 755 genera and 131 plant families.

The first approach to comparing species’ crown architectures involved generating estimates of H, CD, and CAR for a tree of fixed size (D = 30 cm) for each species33,52. To do this, we modelled variation in H, CD and CAR among individual trees as a power-law function of D by fitting linear mixed-effects regressions to log-log transformed data8,33,34:

$$\log \left(Y\right)={\alpha }_{f/g/s}+{\beta }_{f/g/s}\log \left(D\right)$$
(1)

where Y denotes either H, CD or CAR, α is the intercept (or normalization constant) and β is the slope (or scaling exponent). Models were fit using the lme4 package89 and both the intercept and slope of the regressions were allowed to vary among tree species, genera, and families with a nested random effects structure (denoted by the f/g/s subscripts in the equation above). R2 values accounting for both fixed and random effects components of the models were 0.82, 0.74, and 0.45 for H, C,D and CAR, respectively90. The fitted models were then used to predict H, CD, and CAR for each specie,s assuming a fixed stem size of 30 cm (hereafter HD = 30, CDD=30 and CARD = 30). Models were fit using least squares regression (as opposed to approaches such as major axis regression) as we were not interested in comparing scaling coefficients among species, but in using the fitted models to generate predictions91. For the H–D model, we compared the power-law with a saturating Michaelis-Menten function, as previous work has suggested that the latter might fit the data better37. We found that both approaches yielded very similar estimates of HD = 30 (ρ = 0.93), but that overall the power-law provided a better fit to the data (Supplementary Fig. 2).

The second approach we developed to capture variation in crown architecture among species is conceptually similar to the first, but avoids the need to choose an arbitrary stem size at which to compare species8. As for the previous method, we began by using the tree-level data to model variation in H, CD and CAR as a power-law function of D. However, in this case we explicitly ignored differences among species and simply estimated the overall scaling relationships between H–D, CD–D and CAR–D across the whole dataset by fitting ordinary linear regressions to log-log transformed data. Using the residuals of the models (i.e., the difference between observed and predicted values of H, CD and CAR on a log-log scale), we then determined post hoc whether a given species has H, CD and CAR values that are – on average – larger (positive residuals) or smaller (negative residuals) than expected after accounting for differences in stem size between trees (see Fig. 1e–g for a graphical representation of this approach). Because sample sizes varied considerably among species (n = 10–22,835 trees per species), we subset the data by randomly selecting 10 trees per species prior to model fitting. Without this step, well-sampled species would dominate the signal of the regression and skew the values of the residuals. This randomization step was repeated 100 times, and for each species we then calculated the mean value of the residuals across all model runs as a measure of size-standardized H, CD, and CAR (hereafter HRESID, CDRESID, and CARRESID).

Quantitatively, the two approaches gave very similar results (Supplementary Fig. 1). However, the second method based on model residuals is better suited to comparing species that exhibit contrasting growth trajectories or vary in their size at maturity, as it integrates data across all observed tree sizes instead of focusing on a single point of comparison (e.g., D = 30 cm, which could correspond to a small tree for some species and a very large one for others). For subsequent analyses, we therefore focus on comparing values of HRESID, CDRESID, and CARRESID across species, but to aid the interpretation of results, we also report values of HD = 30, CDD = 30 and CARD=30. Note that both of these approaches overlook plasticity in allometric scaling relationships within species. This is partly because previous work based on the Tallo database has shown that allometric variation along environmental gradients is twice as pronounced among species as it is within them2 – prompting us to focus on identifying drivers of species-level differences in crown architecture. But it also reflects the fact that for many species in our analysis we have insufficient data to appropriately model intraspecific variability (314 species were measured at a single site and <20% were sampled widely enough to characterise their architectural plasticity along environmental gradients2).

Environmental data

To understand how environmental conditions shape variation in HRESID, CDRESID and CARRESID among tree species, we used the geographic coordinates of individual trees to assign attributes related to climate, competition, and disturbance (see Supplementary Table 2 for full details on sources of environmental data). These environmental predictors were chosen based on previous work suggesting they play an important role in shaping tree crown allometry by constraining plant hydraulics, growth and competition (Table 1). Importantly, they were also selected as they were not strongly correlated with one another (Supplementary Fig. 3), allowing their effects to be teased apart in subsequent analyses. In addition to the environmental predictors described below, trees were also assigned to one of seven biome classes based on the classification used by the Terrestrial Ecoregions of the World database92.

For climate, we focused on the effects of mean annual temperature (MAT, °C), precipitation seasonality (mm,) and aridity (unitless index). MAT and precipitation seasonality were obtained from the WorldClim2 database at a resolution of 30 arc-seconds93. Aridity was instead calculated as the ratio between potential evapotranspiration (PET, mm) and mean annual precipitation (MAP, mm), where MAP was obtained from WorldClim2 while PET was derived from the Global Aridity Index and Potential Evapotranspiration Climate Database at 30 arc-second resolution94. This is the inverse of how aridity is often expressed94, but has the advantage of being easier to interpret as larger values of the aridity index correspond to drier conditions95.

As a proxy for local competitive environment, we used estimates of tree cover derived from MODIS at 15-arc second resolution for the year 200896, which broadly overlaps with the period when the majority of the allometric data were collected. We chose this approach as for most trees we lacked information on stand-level attributes commonly used to characterise competition, such as basal area or stem density16,33. However, for a subset of sites across which these field data were available, we found good agreement between MODIS-derived estimates of tree cover and stand basal area (Supplementary Fig. 4), suggesting satellite estimates of tree cover provide a reliable indicator of competitive environment.

As indicators of disturbance, we focused on wind speed and fire risk. Specifically, we used the ERA5-Land data to calculate the maximum wind gust speed experienced by each tree between 2010–2020. To quantify exposure to fire, we calculated the mean burned area fraction between 2001–2010, as estimated from MODIS in the Global Fire Emissions Database97. Note that we also tried to assess the impacts of snow accumulation on tree crowns, but found that estimates of snow cover duration derived from MODIS were strongly correlated with MAT (ρ = −0.80; Supplementary Fig. 3), and were therefore not considered in subsequent analyses.

For each of the 1914 species that met the minimum sampling criteria described previously, we then calculated mean values of aridity, MAT, precipitation seasonality, tree cover, maximum wind gust speed and burned area fraction across all sampled trees. Each species was also assigned to a unique biome based on the terrestrial ecoregion in which they were recorded most frequently98.

Functional trait data

To test how variation in HRESID, CDRESID and CARRESID among tree species relates to other key plant functional traits, we compiled data on wood density (g cm-3), leaf nitrogen content (mg g−1), specific leaf area (SLA, mm2 mg−1) and seed mass (g) from multiple sources (Supplementary Table 3). This includes the TRY plant trait database99, the Botanical Information and Ecology Network (BIEN) database100, the global wood density database101, the Royal Botanic Gardens Kew seed information database, the AusTraits database102, the China plant trait database103, the Terrestrial Ecosystem Research Network (TERN), as well as selected publications36,77,104,105,106,107. These four functional traits were chosen as previous work suggests they may covary with crown size and shape through their influence on whole-plant growth, size, hydraulics, and mechanical stability33,36,37,75,76,79 and have been measured for numerous tree species.

To obtain species-level mean values for each trait, we first grouped together individual records by site (based on shared geographic coordinates) and then species108. For species where no individual-level records could be sourced, species-level values reported in the literature were used instead if available. Of the 1914 tree species for which we estimated HRESID, CDRESID, and/or CARRESID, we obtained wood density estimates for 1572 species (82%), leaf nitrogen content for 1085 (57%), SLA for 1120 (59%), and seed mass for 1108 (58%).

Mapping the spectrum of crown architectural types and its distribution across woody biomes

To characterise the range of crown forms that tree species can assume and better understand how these vary among woody biomes, we used estimates of HRESID, CDRESID and CARRESID to determine how species cluster into architectural types based on their height, crown size, and shape1. To do this, we first calculated the correlation between HRESID and CDRESID to determine how tightly constrained these two axes of crown architecture are across the 1309 species where both had been measured. A strong positive correlation would indicate that species that are taller for a given stem diameter also tend to have larger crowns and vice versa. Conversely, a weak correlation between HRESID and CDRESID would suggest that when standardized by size, tree species are able to adopt a wide range of crown architectural forms, from tall and narrow to short and wide.

We then grouped species into one of nine crown architectural types: (1) short and narrow, (2) narrow, (3) tall and narrow, (4) short, (5) medium-sized, (6) tall, (7) short and wide, (8) wide, and (9) tall and wide species. Species were assigned to groups based on their HRESID and CDRESID values and whether these fell in the lower quartile, interquartile range, or upper quartile of data (see Fig. 2a for a visual representation). Note that CARRESID was not used to group species, as any differences in CARRESID among species can be directly attributed to ones in HRESID and CDRESID. To determine the degree to which crown forms are adapted and confined to specific environments, we quantified the relative frequency of each architectural type across different biomes. To support this analysis, we also used one-way ANOVAs to compare mean values of HRESID, CDRESID, and CARRESID among species from different biomes.

Evolutionary history and its fingerprint on crown architecture

To determine whether crown architectural traits exhibit phylogenetic signal, we mapped HRESID, CDRESID, and CARRESID onto the Smith & Brown (2018) phylogeny of seed plants and calculated Pagel’s λ as a general test of phylogenetic signal for each crown attribute109,110. A λ value of 0 indicates no phylogenetic signal, while a value of 1 corresponds to a trait that has evolved according to Brownian motion, indicating strong phylogenetic signal109. Pagel’s λ was calculated using the phytools package111, which uses a likelihood ratio test to determine whether λ is significantly different from 0. Because tests of phylogenetic signal are sensitive to errors in the phylogeny, such as those associated with branch lengths110, only species that were a direct match to those in the time-calibrated phylogeny were retained for subsequent analyses (1225 species for HRESID, 870 for CDRESID and 868 for CARRESID; Supplementary Table 1).

To complement λ – which provides a global test of phylogenetic signal across the entire phylogeny – we also explored how HRESID, CDRESID and CARRESID varied among clades within the phylogeny. Specifically, we used one-way ANOVAs fit without an intercept to identify plant families and genera where species’ mean HRESID, CDRESID and CARRESID values are significantly greater or smaller than zero. For this purpose, we only retained families and genera represented by at least five species in our dataset (n = 63 families and 86 genera for HRESID, and n = 56 families and 60 genera for CDRESID and CARRESID).

Effects of climate, competition, disturbance and functional traits on crown architecture

To quantify the effects of climate, competition, and disturbance on tree crown architecture, we modelled variation in HRESID, CDRESID and CARRESID among species as a function of aridity, MAT, precipitation seasonality, tree cover, maximum wind gust speed, and burned area fraction using multiple regression. Models also included an interaction term between aridity and MAT to test whether the effects of low water availability on tree height, crown size, and shape would be strongest in hotter environments65, as well as a binary variable testing for systematic differences in HRESID, CDRESID, and CARRESID between angiosperms and gymnosperms9,33,38.

To determine whether HRESID, CDRESID and CARRESID also vary in relation to species’ functional traits, we then fit separate models in which either wood density, leaf nitrogen content, SLA or seed mass was added to the multiple regression alongside the environmental predictors described above. Note that models including functional traits as predictors were restricted to the subset of species for which trait data were available (see Supplementary Table 1 for details).

Prior to model fitting, both aridity and seed mass were log-transformed to linearise relationships between response and predictor variables. All continuous predictor variables were then centred and scaled by subtracting the mean and dividing by 1 standard deviation, while the binary variable grouping species into major evolutionary clades was coded as –1 for gymnosperms and 1 for angiosperms. This allowed us to directly compare the effect sizes of different predictors both within and across models based on their regression coefficients. To ensure model coefficients were not affected by collinearity among predictors, we calculated variance inflation factors for all models to confirm they were all ≤ 2.

To account for non-independence among species due to shared evolutionary history, regression models were fit using phylogenetic generalised least squares (PGLS)112. PGLS models were fit using the gls function in the nlme package113, where the correlation structure among species was captured using Pagel’s λ as implemented by the corPagel function in the ape package114. A phylogenetic tree capturing evolutionary relationships among species was generated using the V.PhyloMaker package115, which uses a comprehensive time-calibrated phylogeny of 79,881 seed plant species as a backbone116. Model R2 values that account for the phylogenetic structure of the data were calculated using the rr2 package117.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.