History of functional-type classifications of vegetation
History of remote sensing of vegetation
New sensors and perspectives
Measuring detailed canopy structure
The emerging hypothesis of ‘optical types’
Conceptually, plant functional types represent a classification scheme between species and broad vegetation types. Historically, these were based on physiological, structural and/or phenological properties, whereas recently, they have reflected plant responses to resources or environmental conditions. Often, an underlying assumption, based on an economic analogy, is that the functional role of vegetation can be identified by linked sets of morphological and physiological traits constrained by resources, based on the hypothesis of functional convergence. Using these concepts, ecologists have defined a variety of functional traits that are often context dependent, and the diversity of proposed traits demonstrates the lack of agreement on universal categories. Historically, remotely sensed data have been interpreted in ways that parallel these observations, often focused on the categorization of vegetation into discrete types, often dependent on the sampling scale. At the same time, current thinking in both ecology and remote sensing has moved towards viewing vegetation as a continuum rather than as discrete classes. The capabilities of new remote sensing instruments have led us to propose a new concept of optically distinguishable functional types (‘optical types’) as a unique way to address the scale dependence of this problem. This would ensure more direct relationships between ecological information and remote sensing observations.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
‘What’s the use of their having names,’ the Gnat said, ‘if they won’t answer to them?’ ‘No use to them,’ said Alice, ‘but it’s useful to the people that name them, I suppose. If not, why do they have names at all?’
Lewis Carroll: Through the Looking Glass
Ecologists and plant geographers have long described natural patterns and relationships between plants and their environment by organizing groups of plant species into functional vegetation categories. Recently, ecological research has been driven by the need to develop a predictive capability to understand how ecosystems will respond to increasing anthropogenic impacts. Predicting changes in community composition and ecosystem dynamics requires a better understanding of the ecological importance of structural, physiological and phenological characteristics of today’s vegetation (Chapin et al., 1997). Changes in these key properties affect site productivity and biogeochemical cycling, and can cascade through communities and ecosystems, often with unpredictable consequences. A recurring theme of ecosystem ecology is that a generalization of organisms into functional groups is needed to simplify ecosystem processes so that they can be understood and tested, in part to enable the prediction of environmental changes (Chapin et al., 1996, 1997; Smith et al., 1997). A functional classification system that provides sufficient detail to capture component processes, but with sufficient abstraction to develop regionally and globally useful models, would meet this need (Pacala & Kinzig, 2002).
The term functional type, often used synonymously with functional group, is similar to several others that have been used in ecology in recent years, including life form (Raunkiaer, 1934), guild (Root, 1967) and strategy (Grime, 2001). The concept of plant functional type (PFT) uses structural, physiological and/or phenological features to group species in response to environmental conditions or according to their impacts on ecosystems (Gitay & Nobel, 1997; Lavorel et al., 1997; Shugart, 1997; Tilman et al., 1997). Implicit in this definition is the idea that functional types provide a useful means of predicting the assemblage of species’ responses within an ecosystem to environmental conditions or disturbance (Hooper & Vitousek, 1997). Despite this extended history of development, ecologists have not agreed on a definition of functional types or the key traits that define them. In addition, the characteristics that form the foundation for these groupings often are not fully understood, and there is no universal agreement on what traits should be included in a classifier and what traits should be omitted. As a result, the term has often been applied ad hoc depending on the purpose at hand. For example, some classifications emphasize clearly visible indicators of phenology or life history (e.g. evergreen vs deciduous, annual vs perennial), whereas others focus on more hidden physiological traits, such as nitrogen fixation (Asner et al., 2008b) or alternative photosynthetic pathways (C3, C4 or crassulacean acid metabolism) (Davidson & Csillag, 2003; Anderson et al., 2008). A central tenet in the concept of PFT is that morphological and physiological adaptations are linked in predictable ways by resource limitations, responses to disturbance, biotic factors or other aspects of the environment. The extent to which such linkages are generalized will determine our ability to detect functional types with remote sensing.
In this review, we first consider the evolution of the ecological concept of functional groupings. We next consider the history of remote sensing that has enabled this emerging technology to be used in the classification of vegetation from the local region to the globe. As remote sensing technologies have evolved, so has our ability to detect greater detail in functional classes. In parallel with these technical advances, the terminology and concepts regarding functional types have also evolved, leading to the view that, at some scales, vegetation properties exist along a continuum of traits and environmental conditions rather than as distinct classes. In fact, the history of both vegetation classification and remote sensing illustrates the tension between discrete classes and continuous traits.
Current remote measurement capabilities provide an opportunity for fundamentally new approaches to the classification of functional types, which could even lead to a redefinition of the term itself. We initiate this discussion by suggesting a new concept for approaching this problem by developing an ‘optical-type’ definition built on basic ecological principles, and characterized by optical properties detectable from remote measurements (Gamon, 2008). By optical properties, we mean the measurement of the absorption and scattering properties of plant canopies based on principles of spectroscopy across the electromagnetic spectrum from visible to microwave bands. We address the applications and limitations of such an approach and outline remaining challenges to the concept of optically distinguishable functional types. Finally, we provide arguments for the greater integration between remote sensing and ecology (ecological remote sensing) that links ecological theory with physical principles of spectroscopy and radiative transfer. This could not only help to address the definition of functional type and improve scaling from local to global data, but could also be applied to the challenge of monitoring biodiversity from remote sensing.
II. History of functional-type classifications of vegetation
Meaningful descriptions of natural functional groupings of plants were driven by the exploration of the New World. The explorations of Alexander von Humboldt (1807) in the late 1700s in Europe, Africa and the Americas led to his recognition of at least 16 species-based structural classes having different physiognomies or plant growth forms, which were not dependent on taxonomic affiliation (von Humboldt, 1849). Although von Humboldt did not explicitly define functional relationships, he did relate physiognomic forms to their physical environment, laying the foundation for later functional classifications. Andreas F. W. Schimper (1903) developed the first detailed understanding of the relationships between environmental factors and the geographical distributions of physiological functions, plant growth forms and life history traits. These ideas remain active and classifications based on Schimper’s work are still used in ecosystem (Prentice et al., 1992) and PFT (Box, 1995) models. Relationships between functional growth forms and the environment were extended by Eugen Warming (1909) and his student, Christen Raunkiaer, (see Raunkiaer, 1934 for a collection of papers), who developed these ideas into a formal growth form-based classification system.
In parallel with the geographical approach, Clausen et al. (1948) demonstrated the interactions between climatic and genetic controls on the distribution of plant growth forms or ecotypes in transplant gardens at Stanford University and other sites in California (USA). These groundbreaking experiments formed a foundation for further ecophysiological research, which improved our understanding of physiological responses of plants to environmental conditions and how these relate to plant distribution (e.g. Berry & Björkman, 1980; Pearcy & Ehleringer, 1984; Sage et al., 1999).
The study of form and function further expanded in the 1970s (Wiens, 1976) and 1980s (Urban et al., 1987) in parallel with advances in the understanding of the physiological characteristics of plant species and relationships to plant form (Givnish, 1986). During this period, the realization of the scope of environmental changes at the global scale from human impacts on land use and land conversion coincided with emerging concerns about environmental sustainability, global habitability and climate change, which led to increasing emphasis on ecosystem science and the emergence of global ecology in the 1990s.
The convergence of structure, physiology and phenology is thought to be driven by adaptations to similar climatic, topographic and edaphic conditions arising in different regions of the earth (e.g. Walter, 1931, 1979; Whittaker, 1956, 1962, 1975). A more recent restatement of the concept of functional convergence explains how resource constraints define plant traits related to resource acquisition. A body of ecological theory, grounded in economic arguments (Mooney & Gulmon, 1979; Bloom et al., 1985), states that plant form and function are constrained by resource limitations (e.g. Field et al., 1992; West et al., 1997; Wright et al., 2004). One consequence of this is that a plant’s allocation of energy to structural and physiological components should reflect the integration of multiple stresses and resource availability in predictable ways (Fig. 1).
Examples of structural responses to resource limitations (or ‘stress’) include patterns of above-ground growth or foliar mass, root to shoot ratios, specific leaf area (SLA), leaf longevity, leaf orientation and elemental composition (Bloom et al., 1985; Field & Mooney, 1986; Ehleringer & Werk, 1986; Givnish, 1986). Similarly, physiological investments in photosynthesis, light harvesting and water or nutrient acquisition should be resource constrained, and should also be tied to structural investments (Field et al., 1992). This framework has sometimes been called the ‘functional convergence hypothesis’ (Field, 1991) Although, in its original expression, it primarily addressed the biochemical capacity for photosynthesis, this hypothesis can also be extended to canopy structure (Field, 1991). Here, we apply this concept in a broad sense to indicate the convergence of form and function defined by resource limitations and evolutionary constraints. This hypothesis provides a useful paradigm for understanding PFTs, because it tells us that readily apparent morphological and phenological traits should be linked to underlying physiological traits.
At the leaf scale, functional convergence is seen in the significant relationship between maximum assimilation (Amax) and leaf nitrogen content (Field & Mooney, 1986), or in the tendency of SLA to modulate this relationship (Reich et al., 1998). In an analysis of 2548 species from 175 sites around the world, Wright et al. (2004) demonstrated that c. 74% of the variability in leaf traits can be explained by a single axis (principal component) derived from leaf longevity, leaf mass per area, leaf nitrogen content, leaf phosphorus content, photosynthetic capacity and dark respiration. For the majority of species, representing different growth forms, biomes, climates and functional types, photosynthesis and respiration increase proportionally with leaf nitrogen, leaf thinness and shortness of leaf lifespan. The convergence of so many plant species along a single axis of linked physiological and structural traits, named the ‘worldwide leaf economics spectrum’ (Wright et al., 2004), is a powerful restatement of the functional convergence hypothesis. Much of the variation along major and minor axes represents co-existing species that could be represented in principal components space as higher dimensional axes (such as shown in Fig. 1), and which could express different resource constraints.
At stand scales, functional convergence argues that the investment in canopy structure and physiology matches resource availability; consequently, there should be predictable relationships between vegetation structural attributes detectable by remote sensing and physiological behavior characteristic of functional types (Field, 1991; Field et al., 1992). Fig. 1 (see also Fig. 1 in Wright et al., 2005) illustrates how functional traits can be distributed among species along a continuum of two limiting resources. Variability along the line indicates that additional resource axes may exist, which, in ecological terms, would represent niche differentiation among co-occurring species. At the local scale of communities or plant stands, where alpha diversity is not high, each individual species is located in a discrete region or trait space, and is therefore distinguishable, whereas, at a larger scale, distributions become more continuous because of the greater diversity. This allows us to explore the possibility of regionally distinct functional categories (hence spreading away from the main axis) within a continuous spectrum of traits.
There is a significant difference between all of the functional types in an ecosystem contributing to its biodiversity and the functional types that exert the greatest influence on ecosystem properties such as productivity, carbon and water budgets, and biogeochemical cycling. These properties are primarily affected by the dominant plant species, which tend to be the most abundant overstory species with the largest biomass, as expressed in the mass ratio hypothesis of Grime (1998) and widely demonstrated in the ecological literature (e.g. Diaz et al., 2004; Garnier et al., 2004; Hooper et al., 2005). In this view, at the ecosystem level, biodiversity is represented as having complementary functional traits (Loreau et al., 2001; Hooper et al., 2005). This distinction is important as most remote sensing measurements, with the exception of light detection and ranging (LiDAR) and radio detection and ranging (RADAR) systems, primarily extract information about the upper canopy and not the vegetation in the understory, thus much of the heterogeneity in functional traits for understory species is not readily observed.
An example of convergence among traits of dominant species at the canopy scale is the consistent relationship between the absorbed radiation (APAR, measured by remote sensing) and plant productivity (e.g. net primary production, NPP), which has been observed for much of the world’s vegetation (Goward et al., 1985). Similarly, a wide range of remotely sensed data have been used to establish close links to total canopy leaf area and NPP. Ollinger et al. (2008) found strong correlations in a globally extensive dataset, between canopy nitrogen, canopy structure and canopy near-infrared (near-IR) reflectance, which is readily detected by remote sensing.
Ecological theory, including the concept of functional convergence, also presents many challenges to the concept of discrete functional types. At first sight, these demonstrations that plant species or vegetation types fall on a continuum described by structural and physiological characteristics linked to resources tend to call into question the concept of distinct categories embodied in the term ‘functional type’. However, it need not be inconsistent because, from this perspective, ‘functional types’ devolve from universally applicable relationships represented in n-dimensional resource space (as illustrated in Fig. 1) to locally defined regions, in much the same way that biomes can be defined as local regions in climate space (Whittaker, 1975) or that niches are locally defined regions in n-dimensional resource space (Hutchinson, 1958). In this context, functional types are clearly context dependent, according to the resources used to define the n-dimensional space (often a consequence of the tools at hand). A good example of this apparent contradiction is found in arctic vegetation, which can be characterized either as discrete patches with contrasting functional characteristics, or by a continuous approach based on a few key resources, including water, nitrogen and temperature (Shaver et al., 2007).
A long history of ecophysiological research illustrates that differences in environmental tolerances often exist within different ecotypes of the same species (Clausen et al., 1948; Clausen & Heisey, 1958), and so, even at this level, the exact meaning of functional type becomes unclear. In addition, we are unable to define a universal functional-type classification system when species with intermediate traits exist, as happens when extending classifiers to large geographical areas (e.g. see Lavorel et al., 1997) or when biodiversity is high, as in the tropics, and the variability is continuous (Westby & Leishman, 1997). As classification systems expand and become more inclusive, more species with intermediate characteristics accumulate, confounding the classifier. Gitay & Nobel (1997), among others, question whether consistent ecologically interpretable groups actually exist, and doubt that a universal functional classification can be defined because of the lack of agreement on traits in space and/or time. Nonetheless, the concept of functional type has emerged as a useful tool to provide balance between abstraction and detail to address large-scale ecosystem processes.
III. History of remote sensing of vegetation
The history of remote sensing mirrors the evolution of ecological concepts described above. Advances in technology have progressively expanded the capability for distinguishing the structure, phenology and physiology of vegetation, yielding new insights into the concept of PFTs. As these data have become more widely accepted by the ecological community, there has been renewed interest in developing functional classification schemes driven by remote measurements.
To understand how the development of remote sensing technologies has shaped our thinking about functional vegetation categories, it is necessary to briefly review this history. The heritage of many vegetation mapping methods stems from their photogrammetry origins. Because different types of plants have characteristic spatial distributions, density and canopy architectures, the color and texture patterns associated with them become recognizable to a photo-interpretation analyst.
By World War II, color-infrared (CIR) film began to be used to detect the state of health of the vegetation. It was possible to detect severe plant stress because, as leaves senesce and foliar density declines, reflectance, i.e. the brightness or the amount of reflected sunlight in the near-IR wavelength band of CIR photography (displayed or printed in red), declines and reflectance in the true red wavelength band increases (displayed or printed in green). This causes the color balance to change from bright red for healthy plants to brown for unhealthy ones (Jackson, 1986). However, apart from detecting obvious stress, aerial photography has limited ability to detect physiological state, thus limiting its utility for PFTs.
The development of digital sensors, concurrent with advances in informatics, has accelerated computer-based image analysis in the last two decades, and produced methods that are rapid, repeatable, objective and can be globally applied. Digital remote sensing technology has made it possible to routinely and systematically monitor ecosystems at a range of spatial scales from submeter to kilometer resolutions, which has facilitated further development of the functional classification of vegetation. The data from these systems are spatially explicit and extensive rather than extrapolated from limited subsamples, as would be the case in any field measurement campaign. Compared with film, digital imagery increases the spectral range that can be measured, particularly in the solar-IR and in the region of thermal emissions. Digital sensors allow greater automation in data processing relative to film-based methods and can be rigorously statistically analyzed. When repeated in time, digital measurements provide detailed information about changes in form, state and abundance of vegetation. All of these features have facilitated the development of the functional classification of vegetation.
Satellite and airborne observations of solar reflectance in the visible and IR regions primarily measure the surface or near-surface conditions. For vegetation communities, this primarily limits observations to the dominant species in the canopy overstory, and thus emphasizes the abundance and distribution of land cover types and ecosystem-based classifications rather than species (e.g. conifer forests rather than individual species). Until recently, the ability of remote sensing to resolve PFTs has been limited by the relatively coarse pixel resolution, few and broad spectral bands, and infrequent temporal coverage of satellite and airborne observing systems. Fig. 2 illustrates how the percentage reflectance of a typical plant spectrum appears in the different spectral resolutions of three current satellites and one airborne system, which impacts their ability to discriminate PFTs. In this figure, the measured reflectance is seen in the dark lines, the width of which illustrates the resolution of the instrument; connecting gray lines are gaps between the location of spectral bands. Gaps in the spectrum at c. 1.25–1.375 μm and 1.75–2.0 μm are regions in which atmospheric water vapor absorbs energy, and therefore the land surface is not detected. The progressive improvements in sensor capabilities over the past 40 yr, including the improved spatial and spectral resolution, and their implications for assessing PFTs, are further outlined below.
The era of global digital imagery began in 1972 when, for the first time, repeated opportunities for synoptic views of entire continents became possible. Since that time, three categories of satellite instruments have had widespread use for land cover mapping (National Research Council of the National Academies, 2008). The first group includes the ‘environmental monitoring’ satellites, such as Landsat Thematic Mapper, which make measurements twice monthly, but at resolutions of tens of meters, allowing the determination of physiognomy, cover type and fractional canopy cover. These have had extensive use for land cover classification at regional scales (Ustin et al., 1986; Smith et al., 1990a,b). The second group includes the global weather satellites (e.g. AVHRR and MODIS) which monitor vegetation structure and phenology at daily or near-daily intervals at the kilometer scale designed for monitoring climate responses (Tucker et al., 1985). Lastly, the third group represents recent additions to the suite of civilian satellites (e.g. GeoEye, Ikonos, Quickbird, TerraSAR-X) and provide very high spatial resolution measurements, from submeter to a few meters. These latter sensors provide vegetation measurements at scales that overlap traditional field experiments, but the images cover very limited spatial extents.
In addition to these satellite sensors, a wide range of powerful aircraft and field sensors now facilitate the interpretation of satellite data. As further discussed below, airborne imaging spectrometers and field spectrometers are making novel contributions to our understanding of functional vegetation categories (Gamon et al., 2004; Ustin et al., 2004), and LiDAR instruments are contributing new forms of explicit three-dimensional structural information (Lefsky et al., 1999; Gillespie et al., 2004; Vierling et al., 2008). However, because these sensors represent newer technologies not yet fully deployed to collect systematic global data, most land cover analyses have been restricted to one of the first three types of instruments operating at coarser spectral and spatial scales. Consequently, most remotely sensed global vegetation classifications to date have been limited to general categories of growth form, fractional cover and phenology.
The first land observing satellite, the Earth Resources Technology Satellite (ERTS 1), later renamed Landsat, led to the development of a land cover classification system for North America that is still in wide use and based on plant physiognomic types (Anderson et al., 1976) that generally correspond to eight of the 16 growth forms defined by von Humboldt (1807). Although this level of class resolution is coarse, land cover classes were accurately mapped (Anderson et al., 1976). These early classifications often succeeded in distinguishing broad functional types, in part because the biochemical composition and morphological structure of foliage are accentuated by differences in growth form, combined to produce characteristic reflectance patterns (e.g. very low reflectance in the near-IR in conifer and wetland ecosystems compared with grasslands and crops). To put this in an ecological context, the convergence between form and function makes it possible to infer physiology from physiognomy and phenology.
Beginning in the mid-1980s, the focus of research switched from local scale mapping to the development of predictive models to assess the consequences of land use change and climate change over the whole globe, which renewed interest in the development and refining of the concepts of vegetation functional types applicable to this scale. Until this time, most global vegetation maps were based on potential vegetation predicted from climate space (Holdridge, 1947; Walter, 1973), and not on actually measured land cover maps. For the first time, the global distribution of land cover types could be developed by compiling monthly AVHRR weather satellite data to classify global land cover. Initially performed for the African continent, Tucker et al. (1985) showed that it was possible to use a ‘greenness’ index, specifically the Normalized Difference Vegetation Index (NDVI), from weather satellite data to develop a consistent continental-scale map of actual vegetation, rather than a potential vegetation map based on climate. Continuing this approach, DeFries et al. (1999) used AVHRR to produce a consistent land cover map with 12 vegetation categories based on life form (herbaceous, woody, open or closed canopy), leaf type (needleleaf or broadleaf) and leaf duration (evergreen or deciduous). The fact that this many cover types could be identified from the two AVHRR bands (red and near-IR) testifies to the strong correlation between light absorption by photosynthetic pigments and structural features indicative of functionally distinct types, a further demonstration of functional convergence measurable by remote sensing systems.
DeFries et al. (1995, 1999) also developed an alternative to the classification of distinct classes that describes vegetation properties in terms of ‘continuous fields’ of functional attributes, such as growth form (tree, shrub, herbaceous), life history (annual, perennial), seasonality (evergreen, deciduous) and photosynthetic pathway (C3, C4, crassulacean acid metabolism). This, as in the previous example, was performed by capturing the ‘greenness’ of the land cover at monthly or bimonthly intervals from the global coverage of the AVHRR weather satellite, explicitly introducing phenology into the classification scheme. The idea for defining proportional estimates of plant properties on a pixel basis, rather than assigning the pixel to a ‘type’, was an extension of the mixture analysis methods that produce maps of the fractional composition of the scene components. These methods were developed to address the classification problems that arise when more than one vegetation type occurs in a pixel. The technique, derived from analytical chemistry and widely used in many environmental applications, was initially developed for Landsat data in the 1980s (e.g. Ustin et al., 1986; Smith et al., 1990a,b). The early spectral mixture analysis methods applied to single-date Landsat and other mid-resolution data used multiple spectral bands to identify different land cover components, such as the fraction of plant cover, the proportion of foliage, plant litter (including woody debris), soil, water, etc. in each pixel (discussed further in Section IV ‘New sensors and perspectives’).
Because the AVHRR has only two spectral bands (red, near-IR) at coarse spatial resolution (8 km × 8 km pixels), it was not directly able to differentiate vegetation types from remote sensing data alone, but had to include information on geographical location and temporal pattern (Fig. 3). In Fig. 3, differences in the maps come from assigning a PFT to a biome type (top) vs estimating the PFT from AVHRR data (bottom), which is then proportionally consolidated with up to four PFTs per grid cell. Biomes that include multiple PFTs are poorly represented in the biome-based map. Bonan et al. (2002) concluded that the use of spatially continuous distributions of coexisting PFTs, as shown in the bottom panel, is necessary to link climate and ecosystem models.
The global mapping methods developed with AVHRR have since been extended to other satellite sensors. Launched in 2000, MODIS was expected to produce a new generation of maps based on actual vegetation types. MODIS has seven bands (Fig. 1c), with six located in wavelength positions similar to the Landsat Thematic Mapper, plus one extra band in the shortwave infrared (SWIR) region, that are measured at 500 m × 500 m pixels. This increase in information content compared with the coarse resolution of AVHRR suggests that plant properties similar to those derived from Landsat data could be measured at the global scale by MODIS.
Currently, the methods to retrieve more detailed maps are still under active research, which reveals that challenges still remain to achieve consistent vegetation mapping at the global scale. Different sensors often yield different classes; consequently, there is no universal agreement between the classifications from different sensors and the comparison of results is difficult because of spectral and spatial resolution differences. For example, the 16-class map of Friedl et al. (2002) showed significant disagreement with a global map based primarily on the Satellite Pour l’Observation de la Terre (SPOT-4) vegetation satellite, after data were aggregated to eight classes (Fritz et al., 2003). Several issues make it difficult to unravel the sources of this misclassification, including technical differences in sensor design and calibration, differences in analytical methods, misregistration of the pixels and mapping errors, such as misclassification of edge pixels, inconsistencies in the base map classifications, classification criteria and errors in the assumed ‘true’ ground validation data. The results from different instruments could be more rigorously compared by simulating each instrument’s performance using data from a spaceborne imaging spectrometer (also called a hyperspectral imager), a new technology which several countries are proposing to fly within the next decade.
Despite the availability of new instruments measuring more spectral bands over a wider range of the electromagnetic spectrum, it has been, and continues to be, difficult to identify more than these few growth form types in individual remote sensing datasets, because of the seemingly inherent high levels of variance within and between classes. This problem relates to the spatial extent of the analysis and the number of statistically distinct spectral classes, which is greatest at the finest spatial scales and decreases as the pixel size increases (Woodcock & Strahler, 1987). Given the increasing ecological interest in large-scale ecosystem processes, approaches that define the continuous variation of relevant functional properties may be a more practical alternative than defining a large number of discrete types (DeFries et al., 1995; Shaver et al., 2007). Obtaining measurements of plant biophysical characteristics that are independent of taxonomic or phylogenetic relationships has created opportunities to explore plant functional properties from entirely new perspectives. The emergence of time-series satellite products has enabled a consistent analysis of seasonal patterns (phenology) that is continuing to expand the concept of vegetation functional types.
IV. New sensors and perspectives
Historically, our ability to apply remote sensing to mapping PFTs has been limited by the spatial, temporal and spectral scales of the technology at hand. As described above, instruments with broad spectral bands yield general vegetation classes, and the addition of more bands in other parts of the electromagnetic spectrum, more viewing angles or the fusion of data types can reveal finer distinctions in vegetation function. Fortunately, the newer remote sensing approaches have the potential to revolutionize observations, particularly imaging spectrometers and LiDAR imagers, which are already providing ‘sharper tools’ for assessing physiology and structure. Essentially, these new tools are expanding the range of detectable structural and physiological properties and, by extension, expanding the resource volume that can be observed (the number of n-dimensional axes, as illustrated for two resource axes in Fig. 1). This effectively increases the number of classes of vegetation that can be distinguished remotely, and can in turn yield new insights into the study of functional types.
Fig. 4 shows several biochemical and structural differences that are observed in plant and soil spectra from the visible and reflected solar-IR spectrum. In the visible region (400–700 nm), the overall shape of the reflectance spectrum of foliage is determined by photosynthetic pigments but, at wavelengths longer than 1500 nm, the absorption of energy by water dominates vegetation reflectance. In the near-IR (between 700 and 1500 nm), reflectance is controlled by scattering between air–water interfaces in the leaf and between leaves in the canopy, thus capturing a component of leaf tissue structure and leaf area index and leaf angle distribution at the canopy scale. Species with more compact leaves and thicker cell walls (lower leaf mass per area) have fewer air spaces, less water and reduced scattering from air–water interfaces, and consequently have lower near-IR reflectance than thin leaves (Lichtenthaler et al., 1981; Abrams & Kubiske, 1990). Typically, the remote sensing of biochemical content is stronger on an area basis (g cm−2) than on a mass basis (g g−1) (e.g. see Grossman et al., 1996; Jacquemoud et al., 1996). Poorter et al. (2009) have reported that leaf mass area varies by a factor of 100 among species (between 30 and 330 g m−2 for most species), with c. 35% attributed to functional groups. Leaf mass area is a key species’ property proportional to the relative growth rate (Poorter et al., 2009) and is a measure of the life history strategy (Westoby, 1998; Grime, 2001; Westoby et al., 2002). Recent reports have indicated that spectral differences related to specific leaf mass can be measured at both leaf and canopy scales in tropical forest vegetation (Asner & Martin, 2008; Sánchez-Azofeifa et al., 2009).
Leaf level differences can be amplified at the canopy level by the three-dimensional canopy architecture (Horn, 1971, 1975) as a result of multiple scattering. Consequently, needleleaf (e.g. pine, Pinus spp.) and schlerophyllous leaf species (California Coast Live Oak, Quercus agrifolia) are much darker (i.e. lower reflectance) in the near-IR than are broadleaf deciduous species such as arroyo willow (Salix lasiolepis) (Fig. 4). As a result, both the leaf and canopy spectra contain information relevant for PFT classification.
Other plant biochemicals provide spectral information on the functional status and characteristics of an ecosystem. Dry plant litter has absorption features caused by cell wall constituents, sugars and nitrogen compounds that become observable after water is removed (Curran, 1989; Elvidge, 1990; Wessman, 1990), for example as seen in Fig. 4 at 1750, 2150 and 2300 nm. The dry needles from a conifer, Sequoia sempervirens (coast redwood) and dry grass foliage in Fig. 4 show the characteristic cellulose absorption in the 2000–2200 nm region. Lastly, a soil spectrum from Jasper Ridge Biological Preserve, Stanford, California shows the typical monotonically increasing soil reflectance from 400 to c. 1800 nm and slightly lower reflectance to 2500 nm, but with some spectral structure between 2300 and 2500 nm.
Fortuitously, the ability to detect different levels of functionally important compounds is often facilitated by the structural properties of vegetation. Absorption features present at leaf scales are accentuated by the multiple scattering of plant canopies, which enhance their delectability at canopy scales. Roberts et al. (2004) have provided examples of how the spectra of different species change across spatial scales from leaf, branch and canopy, for a large number of species common to the Douglas Fir–western hemlock forests of the Pacific Northwest. Their study illustrates how weak absorption features at the leaf scale are compounded at the branch and canopy scales, enhancing these features and strengthening species’ differences. In addition, different levels of nitrogen and chlorophyll tend to be associated with different patterns of canopy structure, often contributing to distinguish functionally distinct species within a landscape, as recently shown for nitrogen concentration by Ollinger et al. (2008). In another example, Zygielbaum et al. (2009) found that they could detect the early onset of water stress, not directly from changes in the IR absorption features of water, but from systematic changes in the albedo of photosynthetically active radiation.
The spectra in Fig. 4 could represent end members (components that can be spectrally identified in images: soils, water, dry plant materials, different vegetation types) and used in a spectral mixture analysis, as introduced above. A mixture analysis computes the fractional composition of each end member on a pixel-by-pixel basis for all pixels in the image, based on a best-fit criterion. The assumption is that the pixel spectrum is produced by adding the fractional composition of each end member in the pixel times the end member reflectance spectrum for all end members. This is achieved by solving for the spectral composition that matches the measured pixel spectrum (e.g. described in Smith et al., 1990a,b). Fig. 5 shows an example of a mixture analysis for Stanford University’s Jasper Ridge Biological Preserve (Ustin et al., 1998a). In this example, we show an independently produced vegetation map for the preserve and three images from spring, summer and fall that capture seasonal patterns in this ecosystem. The images were derived using a healthy green vegetation spectrum (such as the arroyo willow), a dry grass and a soil spectrum, as shown in Fig. 4. Here, the evolving colors and tones represent changing fractions of the end members in the pixels, yielding accurate vegetation maps. There is an obvious correspondence to the land cover types shown in the vegetation map, and this correspondence becomes most evident by multi-temporal analysis revealing contrasting phenological patterns for different vegetation types. This example illustrates how combining the structural and physiological signals present in imaging spectrometry with phenological information can improve the differentiation of contrasting vegetation types.
Using a modified form of spectral mixture analysis (multiple end member spectral mixture analysis, or MESMA), Dennison & Roberts (2003) mapped the dominant species in chaparral vegetation in the Coast Range of Central California, as shown in Fig. 6. These species have different reproductive and wildfire response strategies, providing an example of the power of high spectral resolution approaches to distinguish between functionally distinct species. For example, the dominant species, Adenostoma fasciculatum (chamise), are evergreen with reduced needle-like leaves; the shrub has the capability to stump sprout after wildfire (McMurray, 1990). Ceanothus megacarpus (bigpod ceanothus) has larger evergreen leaves compared with Adenostoma; after fire, the species regenerates from seed (Schlesinger & Gill, 1978). Arctostaphylos glandulosa (eastwood Manzanita) is also evergreen with sclerophyllous leaves; the shrub regrows from seed and stump sprouts (Howard, 1992). In addition to shrubs, this community includes evergreen, sclerophyllous oaks (e.g. Quercus agrifolia) and grasslands that are dominated by winter-active annual grasses (e.g. Bromus spp.).
The capability of imaging spectrometers to distinguish between functionally distinct types has led to powerful new mapping approaches that offer novel ways of categorizing vegetation function. For example, Xiao et al. (2004) used high spatial resolution imaging spectroscopy to map street trees by PFTs (Fig. 7) and by species (Fig. 8). As these different species have contrasting effects on energy balance and water usage, high-resolution spectrometry provides a new tool for the classification of PFTs in the urban environment.
The spectral and spatial resolution required to map PFTs at the scales relevant for the discrimination of processes and patterns is not fully understood, but is starting to be realized through scaling studies. In Fig. 9, the study of Underwood et al. (2007) is illustrated, in which four map versions were compared of two plant communities along the central California coast with invasive plant species. They compared the classification accuracy from high spectral resolution imaging spectroscopy data (174 bands) acquired at 4 m × 4 m spatial resolution (pixel size) and used these data to simulate the resolution of the six Landsat Thematic Mapper bands at 4 m and 30 m pixels. They found that the map classification accuracy was best (75%, with a kappa statistic of 0.7, or excellent agreement) using the full spectral and spatial resolution data (Fig. 9a). In this panel, the class composition shows little pixelation and classes have mostly continuous spatial patterns. When the spatial resolution was degraded to 30 m but the spectral resolution was retained, the accuracy declined to 58% with a kappa statistic of 0.5, or moderate agreement (Fig. 9b), whereas the spectrally degraded data matching the six Landsat bands only had an accuracy of 42% (kappa statistic of 0.3, poor agreement), even though the spatial resolution was still 4 m. This accuracy declined slightly more, to 37% (kappa statistic of 0.3), for images matching the 30 m spatial and spectral resolution of Landsat. Fig. 9(b–d) all show increasing pixelation of the spatial patterns, indicating that the classification is tenuous and detection is inconsistent. The improved discrimination obtained with higher spatial and spectral resolution is consistent with previous findings (e.g. Gamon et al., 2004), and demonstrates the importance of higher spectral resolution in mapping PFTs, and also shows, for this community, that a spatial resolution better than 30 m is not always required to achieve good classification results.
The improved precision of today’s measurements is a reflection of both improved measurement technology (better sensors) and progress in quantitative analysis. For example, the most widely used leaf radiative transfer model, PROSPECT (Jacquemoud & Baret, 1990), has had 10 yr of development (Jacquemoud et al., 2009). The current version of the model, linked with one of the several radiative transfer canopy models, most commonly a version of SAIL (Verhoef, 1984), can invert reflectance in imaging spectrometry images to predict, on a pixel basis, concentrations of chlorophyll a and b, carotenoids, water, SLA and dry matter (Feret et al., 2008). This capability to accurately measure and model the full reflectance spectrum in the solar domain permits model predictions of the biochemistry from the reflectance of any band combination or band widths. Thus, through radiative transfer modeling, a pathway exists to directly relate the detailed site reflectance data and functional type descriptions (such as leaf area index, foliage biomass, pigment composition, woody debris fraction, etc.) to the continuous fields extracted from spectral attributes measured at the global scale. The ability to directly scale between optical measurements at fine spatial and spectral scales using aircraft or moderate resolution satellites makes it possible to simultaneously measure local variation in functional types and link these to global observations of continuous traits from instruments with few spectral bands and large pixels.
V. Measuring detailed canopy structure
Information about vegetation structure, from the leaf to the entire stand, is an essential component for assessing PFTs. Instruments such as imaging spectrometers do not measure three-dimensional structural properties of vegetation directly, although they can be used to infer structural properties, in part through the shadows cast by plant structure (Greenberg et al., 2005; Leboeuf et al., 2007). Stratifying images by shadow fraction allows some canopy structure to be estimated indirectly. Richardson et al. (1975) first reported the impact of shadows on crop reflectance. This information was incorporated into spectral mixture models (e.g. Smith et al., 1990a) that define a ‘shade fraction’ and used in classifications of land cover, and additionally, when inverted, as a surrogate for quantifying albedo (eg. Roberts et al., 2004). Geometric-optical models (e.g. Li & Strahler, 1986) classify vegetation, particularly woodlands and forests, in terms of the proportion of sunlit and shaded canopies. Recently, multiple view-angle instruments have attracted interest because of their potential to quantify structural information though a more rigorous analysis of the bidirectional reflectance distribution function (BRDF). Multiple view-angle data typically use only one or a few spectral bands, but probe each pixel with measurements from multiple directions. When combined, these data can be modeled to provide information about the three-dimensional surface structure. However, the large pixel sizes of satellite data and inadequate access to higher spatial resolution multi-angle airborne or satellite data have limited the adoption of these methods (Su et al., 2007).
The NASA BOREAS field campaign, spanning the boreal forests of Manitoba and Saskatchewan, Canada, was a multi-agency program that took place throughout most of the 1990s (Sellers et al., 1995). Multi-angle studies in BOREAS found that BRDF provided useful vegetation mapping tools (Lacaze & Roujean, 2001; Lacaze et al., 2002; Gamon et al., 2004). As structural information varies systematically between functional types (e.g. conifers vs deciduous trees) that vary in their patterns of light capture, BRDF measurements provide novel approaches to functional vegetation classification. The SPOT satellite was the first to allow pointing to collect stereo views of particular sites. This satellite was followed by the Multi-angle Imaging SpectroRadiometer (MISR) instrument on the NASA’s Terra platform. Designed primarily for atmospheric measurements, it has also been used for land cover mapping, and collects four spectral bands (blue, green, red and near-IR) at nine view angles for 275 m pixels. The European Space Agency has led the development of BRDF for ecological applications by flying the demonstration multi-view-angle satellite CHRIS (Compact High Resolution Imaging Spectrometer)/Proba (Project for on-board autonomy) since 2001 (Barnsley et al., 2004). Because of the nature of this technology demonstration, the availability of these data has been strictly limited (Guanter et al., 2005). Nonetheless, Chopping et al. (2006) showed that, by using four viewing angles, they were able to accurately estimate the fractional shrub cover against a grassland background (c. 17 m pixel resolution) in the Jornada Experimental Range in New Mexico. Although BRDF methods offer considerable promise for understanding PFTs, these methods must be better integrated with ecological theory, and tools capable of multi-angle observation must become more widely available, before BRDF approaches can make a significant contribution to our understanding of PFTs.
An alternative technology that can measure detailed structural information about land cover and functional types involves ‘active’ sensors that emit signals at specific frequencies. Detailed structural information can be obtained from both RADAR and LiDAR technologies. RADAR emits radiation in the microwave region of the spectrum (c. 2 mm to c. 70 cm) and LiDAR emits radiation at 1064 nm or 1535 nm in the near-IR region for terrestrial applications or in the green region at 550 nm for water applications. Because both systems use internal energy sources, they can be operated at any time of the day or night. Both systems precisely measure the time for the signal (the pulse) to return to the detector, thus locating the position of the scattering object in space with high precision. These instruments can emit and measure sufficient numbers of pulses to characterize the plant canopy surface and the ground (topographic) surface. Typically, only the ‘first’ and ‘last’ LiDAR returns are used for canopy and ground structure, mostly because of the large size of these files. Because some photons pass between leaves of the outer canopy surface, both the canopy depth and an understory, if present, can be mapped.
Fig. 10 shows the three-dimensional data derived from an advanced LiDAR system for which all returned pulses are recorded and analyzed. This type of full waveform data remains experimental, but demonstrates the potential to provide detailed information on the three-dimensional structure of canopies. Such data are particularly valuable to obtain synoptic descriptions of multilayered forests. In their study at this site, Asner et al. (2009b) found significant structural changes in the three-dimensional savanna landscape resulting from herbivory. One advantage of LiDAR over RADAR for assessing structure is the fine horizontal and vertical spatial resolution that can be detected. However, the new commercial TerraSAR satellite has a 1 m resolution X-band RADAR system, so that this advantage may not be an absolute limitation. An advantage of RADAR is that the data can be acquired under all sky conditions. This makes RADAR particularly valuable for tropical and polar research, where frequent cloud cover restricts the use of visible–IR sensors. Although RADARs have been available longer than LiDARs and have been flown more extensively in space [Ice, Cloud and Land Elevation Satellite (ICESat) carries the first spaceborne LiDAR, launched January 2003], the compact size and reliability of airborne lasers, combined with their high spatial resolution, have caused ecologists to prefer LiDAR technology for mapping vegetation structure at the local scale (e.g. Vierling et al., 2008).
Today’s airborne LiDARs can map the three-dimensional profile of the canopy and ground surface, quantifying the depth of the canopy, the height and spacing of individual trees and shrubs, the distribution of gap fractions, presence of a shrub or herbaceous forest understory, and presence of fuel ladders, among other vegetation characteristics (Lefsky et al., 1999; Riaño et al., 2003, 2007). Airborne LiDARs are generally flown at low altitude (because of the limitations on the power of the laser) and, as a result, are usually acquired at very high spatial resolution. Commercial airborne LiDAR instruments are flown today with horizontal resolutions of 0.5 m pixels and up to 10 cm vertical resolution. Among other examples relevant for defining PFTs, LiDAR has been used to estimate aboveground biomass (Nelson et al., 2003; Omasa et al., 2003; Patenaude et al., 2004), individual trees within a stand (Popescu, 2007), leaf area density and crown bulk density (Riaño et al., 2004a,b) and forest gap characteristics (Asner et al., 2008a,b).
Lefsky et al. (2005) showed that even coarse resolution satellite LiDAR data could be used to estimate tree height over large areas employing the Geoscience Laser Altimeter System (GLAS) instrument on ICESat. By processing the GLAS data with topographic information from the Shuttle RADAR Topography Mission (SRTM), they measured the maximum forest height in three ecosystems (tropical broadleaf forests in Brazil, temperate broadleaf forests in Tennessee, and temperate needleleaf forests in Oregon), which were correlated with field estimates of the tree height, equivalent to 59%–68% of the variance in aboveground biomass estimates and equivalent to 73% of the variance for the Brazilian forests.
Clearly, these types of data provide unprecedented levels of detail about canopy structure at a scale that cannot be obtained practically from ground measurements. To the extent that function can be inferred from structure, these methods can assist in functional vegetation mapping, although this potential remains relatively underutilized. Because they are largely limited to detecting the structural (as opposed to the functional) aspects of vegetation, RADAR and LiDAR have not yet been widely used to distinguish functional types. However, the structural power of these sensors, particularly when combined with the ability of imaging spectrometers to detect physiological details (discussed above), may provide innovative approaches to increase our understanding of physiological functioning in ways that transcend the measurement limits of physiological ecology at regional and global scales.
Combined LiDAR and imaging spectrometer datasets outperform either type of individual dataset for mapping dominant tree species and determining their abundance and distribution (Asner et al., 2008b). Hill & Thomson (2005), Sadro et al. (2007) and Anderson et al. (2008) have shown that combined high spatial resolution LiDAR and imaging spectroscopy can accurately map individual woodland, salt marsh and forest species, respectively. Combining data, especially from these two types of instruments, promises to increase our ability to measure PFTs because of the simultaneous acquisition of information about both ecosystem structure and function.
By combining multi-angle imaging spectroscopy data with LiDAR data to stratify vegetation reflectance by shadow fraction, Hilker et al. (2009) and Hall et al. (2008) were able to better isolate subtle features in forest canopy reflectance that were associated with the photoregulatory xanthophyll pigments. The activity of these photoregulatory pigments and their associated spectral features vary between functional types (Gamon et al., 1997), and are often confounded by vegetation structure (Barton & North, 2001; Middleton et al., 2009). Consequently, integrated analysis, made possible by the fusion of LiDAR and imaging spectrometer data, can provide new insights into the spatial patterns of physiological function.
VI. The emerging hypothesis of ‘optical types’
Clearly, remote sensing today offers many innovative tools for assessing PFTs at a range of scales. To fully realize the potential of these technologies, the data must be combined with ecological theory linking structural, physiological and phenological traits based on resource constraints. Remote sensing enables novel approaches to the assessment of PFTs based on optical principles. This linkage leads us directly to the concept of ‘optical type’ and the related hypothesis that functional types can be distinguished largely on the basis of optical properties detectable by remote sensing as conceptualized in Fig. 11. If the resource space (moisture, nutrient, light, temperature, etc.) can be defined in terms of n-multivariate axes, analogous to principal components, then the goal is to orient the axes in the optical space such that they map onto the axes of the resource space.
In essence, this concept builds on the ‘functional convergence’ hypothesis and begins to develop the theoretical basis to explain why remote sensing works as reliably as it does (Field, 1991). To put it another way, plants are essentially solar energy factories with their canopies structured to optimize the capture of light within existing resource constraints (Ehleringer & Werk, 1986). Consequently, by inverting the view and looking down from above, remote sensing directly assesses key plant structural and physiological features that readily reveal resource constraints, and can be used to explore the concept of functional type from a fresh perspective. Because this concept of optical type is based on fundamental physical principles (e.g. radiative transfer theory and principles of spectroscopy), which are linked to ecological theory, it provides a potentially more rapid, uniform, scalable approach to the problem of measuring functional types than is possible from field observations alone.
According to our hypothesis of optical types, the resource space axes that define optically distinguishable functional types (Fig. 1) can be described by structural, biochemical or physiologically significant constituents detectable with new remote sensing technologies. At the leaf level, these biochemical constituents include pigments (chlorophyll, carotenoids and anthocyanins), plant water, nitrogen and structural components (e.g. lignocellulose). At the canopy or stand level, the relevant signals also include leaf area index, leaf and branch clumping, leaf orientation, canopy height, foliage volume and plant density, which are presumably linked to the expression of leaf-level constituents. The distribution of these features at increasingly larger scales (e.g. landscapes) and across time (phenology) provides a consistent basis for scaling properties from local to global scales.
Based on multiple traits evident in vegetation optical properties, there have been a number of recent successful demonstrations of biodiversity assessment through optical diversity (Zutta, 2003; Carlson et al., 2007; Lucas & Carter, 2008; Asner et al., 2009a; Asner & Martin, 2009). Some of the studies argue that it is possible to characterize spectral types based on the underlying unique biochemical signatures, a concept termed ‘spectranomics’ (Asner & Martin, 2009). It should be noted that these demonstrations of consistent linkages between biochemistry and optics have so far been limited to the wet tropics, where seasonal contrasts are relatively small. Other studies in the seasonally dry tropics (Sánchez-Azofeifa et al., 2009) and in Mediterreanean climates (Zutta, 2003) have demonstrated that the ability to distinguish plant types based on underlying spectral features is strongly dependent on the phenological stages or environmental conditions. In addition, recent studies have demonstrated that biochemical features in reflectance signatures are strongly influenced by the three-dimensional structure of vegetation stands (Barton & North, 2001; Roberts et al., 2004; Sims et al., 2006; Hilker et al., 2009). Thus, we argue that to have global application, any successful classification scheme based on optical types must consider all three elements: biochemistry and physiology, structure, and phenology (Fig. 11). To be complete, this framework should consider both ecological theory (e.g. functional convergence and related economic concepts, Bloom et al., 1985; Field, 1991) and physical theory (e.g. radiative transfer theory, Jacquemoud et al., 2009).
As ecologists and remote sensing scientists continue to develop the concept of PFTs and understand how these plant properties relate to remotely sensed variables, improved informatics and cyberinfrastructural capabilities will be needed. To assist in understanding the links between remote sensing and functional types, metadata standards could be established for recording species’ composition, structure, phenology and physiological characteristics, together with GPS locations archived in web-based ecological databases. Existing efforts by FLUXNET, SpecNet, remote sensing and ecological communities represent an important beginning (Michener & Brunt, 2000; Gamon et al., 2006; Agarwal et al., 2008). However, most existing archival efforts lack the ability to easily integrate ecological data with time series of remote sensing data, or to manage the sheer volumes or dimensions of data involved. In this context, it is worth noting that NASA’s Earth Observing System Data and Information System (EOS-DIS), which delivers and manages remote sensing data from earth observing satellites, is the largest and most complex database effort ever attempted (Marshall, 1993), and further integration of these satellite data with field data or ecological data will continue to require innovative computing science approaches. Flexible ontological approaches are needed to link spectral and ecological data and to search metadata in meaningful ways that will advance the integration of ecology and remote sensing. Consequently, a full realization of the potential for monitoring functional types from remote sensing requires a new level of effort in informatics and cyberinfrastructure, such as described in NSF (2010) and reviewed in Gamon et al. (2010).
Despite the long history of remote sensing measurements to map vegetation, challenges remain to effectively link the observations made in particular bands of the electromagnetic spectrum with plant traits and environmental conditions as ecologists and botanists understand them (Schaepman et al., 2009). This seems a formidable task, particularly given the lack of universal agreement on how best to categorize groups of organisms. However, because remote sensing can detect fundamental vegetation properties that link physical properties to ecological theory, and provides spatial and temporal databases with consistent and complete coverage, it offers the potential to create a universal solution. Our ability to apply remote sensing for truly global tests of the hypothesis of optical types will depend, in part, on the capture of the essential plant optical characteristics which vary with the particular remote sensing methodologies. Hence the need to fundamentally understand the utility of BRDF and combined LiDAR–imaging spectrometer approaches. The datasets produced by these instruments provide rich data sources, allowing us to expand the resource hypervolume and properly assess the structure, physiology and phenology in the context of resource space (Figs 1, 11). A full understanding of the links between remote sensing and functional types requires a more complete integration of physical remote sensing concepts with a deep understanding of ecological theory. Ultimately, this demands greater multidisciplinary training, requiring remote sensing scientists to study biology, ecologists to become similarly well versed in the power of newer remote sensing technologies, and everyone to achieve greater facility with analyses of large, complex datasets.
S.U. wishes to thank the Department of Biological Science, Victoria University, Wellington, New Zealand for providing space and support during her sabbatical leave. We thank M. Andrew, D. Riaño, M. J. Santos and M. J. Whiting for comments on an earlier draft of the manuscript. We especially wish to thank the New Phytologist editors, Ms Jayne Young, for her exceptional patience with the many delays in completing the manuscript, and Peter Curtis and David Ackerly for helping us with additional revisions and edits.