- Top of page
- Supporting Information
In current vegetation science, much effort has been put into the definition of plant functional traits. These traits help us to understand how plants act on and react to their environment (Chapin et al. 1997; Lavorel & Garnier 2002; Diaz et al. 2004; Kleyer et al. 2008). The TRY database (Fast Track Initiative on Refining Plant Functional Classifications for Earth System Modelling; Kattge et al. 2011), for example, currently lists 1480 traits with relevance to ecosystem functioning. These traits have been identified to facilitate new interpretations of existing species distribution data or to feed new models of trait occurrences or quantities through scales (e.g. Kleyer et al. 2008; Kattge et al. 2011). A common goal of these efforts is to create spatially continuous information on ecosystem functioning and biogeochemical cycles.
If present patterns are of concern, remote sensing is an obvious option for mapping plant functional traits (Ustin & Gamon 2010). Various morphological traits show persistent and stable relations to canopy reflectance (Ollinger 2011). Even though at first glance certain traits such as dispersal strategy or life-history characteristics seem unrelated to canopy reflectance, they are often linked to optically relevant attributes and can therefore be mapped. Dispersal strategy, for example, is linked to flowering traits that are a frequent target in remote-sensing approaches (e.g. Hunt et al. 2004a,b; Chen et al. 2009). Life-history characteristics are related to leaf traits such as leaf architecture, which have a considerable influence on canopy reflectance (Kumar et al. 2001). The reason for these links is the covariation of traits (Grime et al. 1997) that can be particularly accentuated in communities formed by strong environmental filtering. Recurring trait combinations can be defined as plant functional types (Gitay & Noble 1997; Lavorel et al. 1997; Lavorel & Garnier 2002). Depending on the functions of interest, there are manifold options for defining such functional types and not all of them are equally detectable by remote sensing (Ustin & Gamon 2010). Apart from a search for ‘optical types’, as proposed by Ustin & Gamon (2010), existing systems of plant functional types should be checked for their potential to be mapped by remote sensing.
One of the most established systems for plant functional types is the strategy types proposed by Grime (1974, 1977). According to this approach, the flora consists of competitive plants, stress-tolerant plants and ruderal plants. These three types represent the extremes in a virtual triangle where every species has its particular place. Plants following a ruderal strategy can deal with mechanical disturbance, e.g. trampling or mowing, but are less capable of coping with other site conditions that restrict production, such as shortage of nutrients and water. This is the domain of stress-tolerant species. Both ruderals (R strategists) and stress tolerators (S strategists) avoid those environments that could be considered the ‘bright side of plant life,’ with a low degree of disturbance and low abiotic stress. These sites are occupied by plant species with pronounced competitive abilities (C strategists).
Even though there has been some criticism, especially with regard to the high level of generalization (Wilson & Lee 2000), the CSR system has been frequently used to describe variation in vegetation in a function-oriented way (Körner & Jeltsch 2008; Cerabolini et al. 2010; Bornhofen et al. 2011). Wilson & Lee (2000) considered the theory to be the most comprehensive and coherent in community ecology. The position of a species in the CSR trait space can be estimated using a few measured traits, including canopy height, dry matter content, flowering period, flowering start, lateral spread, leaf dry weight and specific leaf area (Hodgson et al. 1999). Most of these traits have a causal link to reflectance. Leaf traits, for example, are a crucial parameter in mechanistic models of canopy reflectance (e.g. Verhoef 1984). Other traits, such as lateral spread, are related to the orientation of leaves and branches and have a considerable influence on reflectance characteristics (e.g. Sandmeier et al. 1998). We can therefore assume that remote sensing could be a feasible method to detect CSR types. However, to our knowledge, there has been no such attempt until now. We therefore test whether remote sensing is able to reproduce spatial patterns of these primary strategies.
- Top of page
- Supporting Information
Between four and 21 species with C, S and R scores were recorded in a sample plot. The average number of species with scores per plot was ten and the overall number was 47. The distribution of strategies in the area of investigation (Fig. 2) mirrored the differences in plant species composition (Fig. 1). Competitive strategies (C) were more successful in the minerotrophic fen areas, and less successful in the raised bog plots; instead, stress strategists (S) were widespread. The most pronounced shift towards S strategies was found in the raised bog area; lower values were found in the unmanaged fen and along the brook side. Ruderal strategies (R) played a minor role, with negative scores throughout the area; the highest values were recorded near the rivulet. The weighted means of C and R scores were positively correlated (R² = 0.48 and 0.24 with and without log-transformation of the original species cover, respectively). Negative correlations were found for C and S (R² = 0.97 and 0.99) as well as for R and S (R² = 0.59 and 0.28). All mentioned coefficients of determination were significant at P < 0.01 or lower and were based on the Box–Cox-transformed values.
Figure 2. Distribution of weighted means of C, S and R scores in sampling plots (original species cover log-transformed and untransformed). In theory, values between −2 and 2 could have been reached. The point sizes are rescaled from minimum to maximum. Note that values for R are negative throughout the area.
Download figure to PowerPoint
The pls model parameters and validation results are given in Table 2. Irrespective of using original or log-transformed species cover values, C scores and S scores could be modelled with good accuracy (R² in validation 0.80–0.82). Models of R scores were weaker, with R² in validation of 0.54 when species cover was log-transformed before entering the weighted averaging procedure, and with R² of 0.35 without that transformation. The poor model fit for R scores was in accordance with the absence of R strategists in the area. None of the models showed significant spatial autocorrelation in the residuals.
Table 2. Basic statistics of the pls models and validation results. spc = transformation of species cover data used in weighted averaging of C, S and R values; # bd = number of spectral bands used; # lv = number of latent vectors used; R²cal = coefficient of determination in calibration; R²val = coefficient of determination in ten-fold cross-validation; rmsecal = root mean squared error in calibration; rmseval = root mean squared error in ten-fold cross-validation. nrmsecal = normalized root mean squared error in calibration; nrmseval = normalized root mean squared error in ten-fold cross-validation. The errors are in original units.
| ||spc||# bd||# lv||R²cal||R²val||rmsecal||rmseval||nrmsecal||nrmseval|
Figure 3 shows how C, S and R scores were related to reflectance and how these specifics translated into the regression equations. The relationship with reflectance (Fig. 3b) was similar for C and R and complementary for S. The C and R scores were higher in areas with a brighter green reflectance (480–560 nm) and near-infrared reflectance (700–1400 nm). This effect was more pronounced for C than for R. Instead, S scores were more related to reddish tones (635–700 nm), with low reflectance in the near-infrared. This led to differences in regression coefficients (see Fig. 3a and Discussion). All models relied on bands from the entire range of wavelengths, including the short-infrared beyond 1400 nm.
Figure 3. (a) Regression coefficients (Bws) of individual spectral bands that entered the pls equations for the prediction of weighted averages of C, S and R (original species cover log-transformed). All values were weighted to exclude the influence of absolute data ranges of bands and rescaled to a range of −1 to 1. (b) Correlogram showing the linear relation between reflectance and weighted averages of C, S and R scores (original species cover log-transformed). The missing spectral regions are water absorption bands and were skipped in the analyses.
Download figure to PowerPoint
The resulting models can be presented as single maps (Fig. 4a–c) and as a colour composite map (Fig. 5), with colours assigned to pixels according to their respective modelled C, S and R scores. The maps show patterns of different strategies that are in accordance with field observations. Competitive strategies prevailed in the unmanaged fen sites and near the rivulet, where P. australis was dominant. Several of the accompanying species, such as Cirsium palustre, had high ruderal scores. Stress tolerators were typical of the rain-fed raised bog area and also for the managed fen plots, where competitors such as reed are reduced by mowing. The species in these two habitats were different, with calcifuge plants such as Vaccinium oxycoccos in the raised bog and calcicole plants such as Parnassia palustris in the fen. In drier spots of the raised bog area, especially along a drainage trench and the more inclined margins, some species with ruderal characteristics were present. The corresponding higher values for R were related to a relatively high cover of Rhynchospora alba.
Figure 4. Map representations of the distributions of averaged C, S and R scores (a–c) in the area of investigation (original species cover log-transformed). Note the different scaling of strategy scores. The typical validation error (root mean squared error, rmse) of these maps amounts to 0.14, 0.21 and 0.14, respectively, measured in original units. The colours saturate at the 2 and 98% quantiles. Areas covered by forest or open water were not sampled and are hence masked.
Download figure to PowerPoint
Figure 5. Colour composite map showing the relative distributions of C, S and R scores in the area of investigation (original species cover log-transformed). The meanings of colours are given in the ternary plot. Colours saturate at the 2% and 98% quantiles of each parameter. Areas covered by forest or open water were not sampled and are hence masked.
Download figure to PowerPoint
Strategy scores were more closely allied with ancillary site information (cover of moss, cover of red and green peat mosses, cover of dead plant material) when the underlying species data were log-transformed prior to the weighted averaging procedure (Table 3). However, it was not possible to derive good quality models for these ancillary site properties using reflectance as a predictor. The best model was obtained for total moss cover, with an R² in validation of 0.4. Red and green peat mosses could be modelled, with R² values of 0.15 and 0.32, respectively. Dead plant material could not be predicted from reflectance.
Table 3. Spearman rank correlations (ρ) between averaged C, S and R scores and auxiliary site attributes. The first column of ρ refers to strategy scores derived from untransformed species cover data and the second column refers to strategy scores derived from log-transformed species data. Significance level P < 0.05 (*), n.s., not significant.
| ||ρ (raw)||ρ (log)|
|C vs. cover of Mosses||−0.39*||−0.46*|
|C vs. cover of Litter||−0.65*||−0.70*|
|C vs. cover of Red Peat Mosses||−0.42*||−0.47*|
|C vs. cover of Green Peat Mosses||−0.66*||−0.70*|
|C vs. cover of Other Mosses||−0.14 n.s.||−0.21 n.s.|
|S vs. cover of Mosses||0.39*||0.46*|
|S vs. cover of Litter||0.63*||0.67*|
|S vs. cover of Red Peat Mosses||0.42*||0.48*|
|S vs. cover of green peat mosses||0.66*||0.71*|
|S vs. cover of Other Mosses||0.13 n.s.||0.20 n.s.|
|R vs. cover of Mosses||−0.21 n.s.||−0.27 n.s.|
|R vs. cover of Litter||−0.17 n.s.||−0.33 *|
|R vs. cover of red peat mosses||0.01 n.s.||−0.15 n.s.|
|R vs. cover of Green Peat Mosses||−0.23 n.s.||−0.38 *|
|R vs. cover of Other Mosses||−0.29 n.s.||−0.27 n.s.|
In most plots only a few species, typically two or three, exceeded a total cover of 5% area. Although these dominant plants covered a total of 92% area (SD = 5%) on average, their cover was statistically less related to reflectance than the cover of less frequent plants (Mantel statistics of 0.47 vs 0.61, P < 0.001). The same principles apply to the derived CSR scores. It was not possible to establish meaningful models without the contribution of low-cover species, as can be seen from the results of the permutation tests: for C and S strategies, the average coefficients of determination (R2) were negative if strategy scores for low-cover species were assigned randomly (more details in App. S2). In 100 permutations with randomization, these models never approached the original model quality. Randomizing dominant species instead of low-cover species was still detrimental for model quality but usually less so than randomizing low-cover species. In the case of R strategists with their low overall variance, dominant species were more relevant.
The Mantel statistic of the relation between overall species composition and modelled CSR scores gave values of 0.65 (log-transformed species data) and 0.54 (untransformed species data; in all cases P < 0.001). This suggests that species composition was partly related to modelled CSR scores, but that CSR scores were not merely a representation of species composition. Where raw cover values were used as a basis, the statistical link to CSR scores used in model calibration was weaker than the link to scores resulting from model validation (Mantel statistic of 0.32 instead of 0.54). This effect was not observed if log-transformed cover values were used as a starting point.
Shannon's entropy of strategies (Fig. 6) was often higher than 0.75. A value of H = 0 would indicate the presence of only one strategy; a value of H = 1.1 indicated equal shares of all three strategies. This usually occurred where species with intermediate strategies prevailed.
Figure 6. Shannon's entropy of plant strategies. Low values indicate places where strategic specialization leads to success; high values correspond to less exclusive conditions. Areas covered by forest or open water were not sampled and are hence masked.
Download figure to PowerPoint
- Top of page
- Supporting Information
A glimpse of how the models work is provided by Fig. 3a: negative regression coefficients in the red and red edge wavelengths (680–730 nm) for C are an expression of the strong contrast between this section of the spectrum and adjacent wavelength regions (see also Fig. 3b). This high contrast can be considered a typical feature of vital and dense vegetation (Kumar et al. 2001). Indeed, competitive ability ‘depends upon plant characteristics that maximize vegetative growth in productive, relatively undisturbed conditions’ (Grime 1977: 1183). This finding fits well with the observed spectra. Instead, S selection brings about ‘reductions in both vegetative and reproductive vigor’ (ibid, 1183). Accordingly, coefficients for S were complementary to C and reflectance in the near-infrared was low, corresponding to scarce vegetation (Kumar et al. 2001). The coefficients in the R model can be attributed to a reflectance that is moderately bright in the entire visible and near-infrared area, without the strong deflections observed for C. The ruderal strategy is typically more successful in productive, disturbed habitats, which are largely poorly represented in the area due to the absence of disturbance agents. Nevertheless, this spectral response confirms that ruderal strategies were less excluded from productive habitats. Shannon's entropy of strategies was relatively low where strategic specialization led to success. This was particularly the case in the raised bog area where others than stress strategists are unable to survive. High values were reached where conditions were less extreme, giving rise to plants with intermediate strategy scores. This was often the case in ecotone areas.
The pls models were derived empirically and there is to date no physically based alternative if we deal with complex mixed stands of vegetation. Mixtures of species, differing in growth rates, levels of stress or health and ancillary site attributes, produce a variation in reflectance that is difficult to reconstruct, even given progress in research in this direction (Jacquemoud et al. 2009; Féret & Asner 2011). This means that even in the longer term, fieldwork will not be replaced by remote sensing of species-related vegetation attributes. On the other hand, observed statistical relations are rarely transferable in time and to new sites (Feilhauer & Schmidtlein 2011). However, this should not be criticized too much as long as such maps cannot be derived in non-empirical ways. The current study was a rather simple exercise due to the small extent of the area and due to the clear and steep gradients in environment and species composition. In order to show that the fundamental principles can be transferred, more data are needed. Problems must be expected where multiple canopy layers are involved, especially in forested landscapes where variation in the field layer is covered. However, due to the interrelatedness and covariation of phenomena in vegetation, visible attributes of the canopy may often act as proxies for invisible attributes. Our study provides examples for this phenomenon as will be discussed below.
One could argue that different colours on the maps of strategies represent differences in species composition rather than differences in strategy per se. However, this is not the case: the Mantel statistics describing statistical links between modelled scores and observed species composition did not exceed a value of 0.65. Managed minerotrophic fen sites and raised bog areas feature different species (Fig. 1; see also Schmidtlein et al. 2007 for a gradient map of species composition of this area) but they are both colonized by stress strategists (Fig. 2). This deviant pattern is accurately reproduced in the models (Figs 4, 5). When deriving the original CSR scores from raw species cover data instead of log-transformed data, the link between CSR scores and overall species composition was weak. In this case, models tended to re-introduce species-related reflectance information into the CSR maps by filtering out noise. The effect of log-transformation of species coverage on model fits underlines the complexity of causal links between vegetation attributes and reflectance. While this transformation had little effect on model qualities for C and S strategies, it was very useful to enhance the model for the R strategy. The R strategy had almost no importance in the area; scores were low and showed little variation. It is therefore unlikely that plant traits related to the R strategy were picked up by canopy reflectance. Nevertheless, a model could be fitted because strategies are interrelated and because strategies follow environmental gradients. Environmental gradients affect, beside vascular plant species, ancillary surface properties relevant for reflectance. Examples in the study area are the cover of mosses in general and the cover of red and green peat mosses in particular. Vascular plant traits that are correlated with these surface attributes have a good chance of being linked to overall reflectance, even if they contribute little to reflectance themselves. Spearman rank correlation analyses show that strategy scores are often more related to these surface attributes if the former are based on log-transformed species cover. In the case of R strategies, this could have been crucial for the modelling performance. This does not mean that all models are exclusively based on ancillary site attributes instead of vascular plant traits. None of the mentioned site attributes could be well modelled using reflectance alone as a predictor. Nevertheless, we can hypothesize that ancillary site properties contribute to the strategy models and are sometimes (as with R in our case) decisive for the success. This is supported by the observation that the small group of dominant species was usually not sufficient to establish statistical links between CSR scores and reflectance. Dominant species are often ubiquitous and therefore less suited for a detailed differentiation.
The translation of CSR space into a colour space has the unique potential of giving an impression of the relative contributions of plant strategies across sites. The ternary plot in Fig. 5 does not represent the entire realized trait space. The reason is that the true CSR trait space and the corresponding RGB colour space are three-dimensional. The projection onto a two-dimensional triangle brings a loss of represented combinations of strategies. However, we consider the ternary plot to be more readable than three-dimensional solutions.
It should be noted that airborne hyperspectral data are expensive and difficult to obtain, even though their availability has greatly improved in recent years. We can expect a much more favourable situation when suitable hyperspectral satellites such as EnMap become operational (Buckingham & Staenz 2008). First simulations suggest that even satellites such as the upcoming Sentinel-2, with 13 wavelength bands, will allow for similar analyses with relatively high accuracy.