Introduction
Global change is expected to cause extensive changes in terrestrial ecosystems, driving unprecedented redistribution of plant species and their associated traits (Pecl et al., 2017, Diaz and Cabido, 2001). Plant functional traits are involved in key ecosystem processes from local community assembly (McGill et al., 2006, Sterck et al., 2011) to global biogeochemical cycles, and these processes are interconnected across scales (Reichstein et al., 2014, Peaucelle et al., 2019). Relationships between traits, such as the leaf economic spectrum (LES)(Wright et al., 2005) , reveal information about biological constraints in leaf mass allocation that impact plant ecophysiology and have the potential to improve ecosystem models (Fisher et al, 2015). Biotic interactions, micro-climate, and soil conditions can affect species co-occurrence and influence local trait distributions (Bruelheide et al., 2018, Simpson et al., 2016), and variation in climate within species ranges can affect realized niches and drive trait responses (Chave, 2013). Given their central role across levels of organization, understanding how traits vary within and among species across scales and environments is essential for conserving present and future ecosystem function (Violle et al., 2014).
Plant traits vary geographically through a combination of interspecific shifts in species abundances and intraspecific trait variation (Leps et al., 2011, Laughlin et al., 2012, Valladares et al., 2014, Münzbergová et al., 2017). Understanding how traits respond to the environment across wide geographic areas requires approaches integrating both sources of trait variation. This is challenging because individual level trait data are geographically and taxonomically limited, making it hard for traditional methods to identify the relative importance of inter- and intraspecific variation at large scales (Henn et al., 2018). In addition, effectively predicting broad-scale geographic patterns requires making predictions for species not included in trait datasets but widely distributed across the continent. Most current approaches to understanding and predicting trait variation use community weighted mean approaches (CWM), which circumvent these limitations by either focusing directly on trait–environment relationships (based on direct relationships between environment and community trait averages while ignoring species) or by estimating traits from species averages (using the environment only for predicting species assembly but ignoring the effect of intraspecific variation) (Miller & Ives, 2019).
CWMs relying on direct trait-environment relationships offer the advantage of predicting large scale community trait distributions without requiring field surveys or estimates of species assembly. However, by ignoring species identity (e.g., Ordonez et al., 2009) this approach ignores known phylogenetic signals in trait variation driven by biological, physical, and historical constraints (Wright et al., 2004, Anderegg et al., 2018), and assumes that the environment implicitly captures relevant changes in species distribution and abundance. Since the information about species identity is missing from predictions, these models cannot be used to explicitly identify the relative contributions of inter- and intraspecific variation (but see Moles et al. 2014) and could potentially yield worse predictions for assemblage level trait values (when abundance data is available) unless the environmental model fully captures relevant shifts in species abundance.
Due to these limitations, it has been suggested that predicting traits directly from species’ average values (Swenson 2010, Clark 2016, Wieczynski et al., 2019, Swenson 2017, Stahl et al., 2014) offers better estimates of large-scale trait distributions. These approaches assume that environmental drivers affect trait distributions indirectly by shaping community structure and species abundance, implying that species distributions are the best predictor of traits and associated ecosystem function. These models can be used to make predictions for traits over large areas by training on forest inventories, leveraging species distribution models to forecast future shifts in species and consequently trait distributions (Swenson & Weiser, 2010, Clark 2016). Yet, this approach ignores intraspecific variation, which can be larger than interspecific variation for broadly distributed species (Niinemets, 2015, Messier et al., 2017), and overlooks that species averages within regional communities may diverge from their global averages (Hulshof & Swenson, 2010).
Both these approaches contribute to our understanding of trait variation and allow for predicting community level trait distributions without requiring extensive field surveys of traits. However, neither approach is designed to predict trait variation at the individual level, nor allow for estimating intra-species variation. Also, they often fail to account for the effect of phylogenetic signal on traits of closely related species (but see Swenson et al., 2017), which can be potentially important for generating more robust trait predictions for species that are sparsely sampled (or not sampled at all) across vast geographic areas (Blomberg et al. 2003, Swenson 2013, Swenson et al., 2014). These limitations prevent assessment of the relative importance of inter- and intraspecific variation on trait distributions across a continuum of geographical scales, reducing our ability to generalize and understand mechanisms driving trait distributions.
To address these limitations, we developed a model that combines species, their phylogenetic relationships (from the Tree of Life; Hinchliff et al., 2015) and environmental drivers (climate, elevation, slope, terrain aspect) with large scale leaf trait data from the National Ecological Observatory Network (NEON) (National Ecological Observatory Network, 2020). While this method can only predict trait distributions for locations with species abundance data, it makes it possible to estimate the relative contribution of intra and interspecies variation on trait distributions across wide geographic areas. This allows us to address whether changes in environmental conditions have a direct effect on the LES, within and across species. We jointly modeled eight leaf traits: nitrogen (N%), carbon (C%), chlorophyllA (ChlA%), chlorophyllB (ChlB%), carotenoids (Crt%), leaf mass per area (LMA, g m-2), lignin (%) and cellulose (%). We compared this combined model to models based on only environmental drivers or only species and phylogeny information. We integrated the combined model with US Forest Inventory and Analysis (FIA; USDA Forest Service, 2001, Smith et al., 2002) and Daymet data (Thornton et al., 2018) to make trait predictions for ~1.2 million trees across the eastern US. We compared these predictions to the other two approaches to assess the influence of model differences on large scale prediction, analyzed the relative contribution of environmental factors and phylogeny to leaf trait variation across ecoregions, and demonstrate the potential use of this data for understanding the processes structuring ecological systems at scale.