61
spectrum and thus selects specific wavelengths, and therefore SVI, for their analysis.
An alternative approach is to explore the spectra and trait space to identify new or
previously unknown SVIs that maximize the correspondence between optical properties and traits of interest (e.g., Inoue et al. 2008), akin to a data mining exercise.
A challenge of this approach can be interpretation of the selected SVIs, where the
resulting vegetation indices may not contain wavelengths with known absorption
features relating to the trait of interest. The same general approach can also leverage
multiple SVIs, provided the research avoids highly correlated portions of the spectrum (Grossman et al., 1996), to attempt to capture how variation in the trait of
interest is reflected in various portions of the EM spectrum to other sites and plant
species. However, a limitation to the use of SVIs has been the ability to generalize
across broad canopy architectures, species, and environments due to the often sitespecific modeling results or potential signal saturation issues with some SVIs
(Shabanov et al. 2005; Glenn et al. 2008).
Continuous spectral wavelet transforms have been used to reduce the dimensionality of spectral data prior to developing simple statistical models (e.g., Blackburn
and Ferwerda 2008). Wavelets are functions that are used to decompose a full, complex signal into simpler component sub-signals. When used with spectral data, the
full reflectance signature can be decomposed in a way that allows the resulting
wavelet coefficients assigned to each sub-signal to be related to concentrations of
chemical constituents or other traits of interest, through standard statistical modeling approaches (e.g., linear regression). Previous studies have explored the use of
wavelet methods to retrieve a host of functional traits, including pigments, water,
and nitrogen content (e.g., Blackburn and Ferwerda 2008; Cheng et al. 2011; Li
et al. 2018; Wang et al. 2018). Continuum removal together with band-depth analysis (Kokaly and Clark 1999) has also been utilized as a means to retrieve the chemical composition of leaves. In this approach, continuum removal lines are fit through
the absorption features of interest based on those regions not in the areas of interest,
then the original spectra are divided by corresponding values of the continuum
removal line. The band centers can then be found by finding the minimum of the
continuum-removed spectra. Normalization of the band centers is often used to
standardize the values across samples. These data are then used to develop models
to predict functional traits at the leaf and canopy scales, including foliar nitrogen
and recalcitrant properties, such as the amount of lignin and cellulose (Kokaly
et al. 2009).
In addition to the empirical SVI approach, as discussed in Schweiger (Chap. 15),
partial least-squares regression (PLSR) modeling has been used extensively in the
development of spectra-trait models for measuring, scaling, and mapping plant
functional traits (e.g., Ollinger et al. 2002; Townsend et al. 2003; Asner and Martin
2008; Martin et al. 2008; Dahlin et al. 2013; Singh et al. 2015; Ely et al. 2019). A
key attribute of PLSR is the capacity to utilize the entire measured portion of the
EM spectrum as predictors (i.e., X matrix) without requiring a priori selection of
wavelengths or SVIs (Wold et al. 1984; Geladi and Kowalski 1986; Wold et al.
2001). PLSR avoids collinearity (i.e., spectral autocorrelation across wavelengths)
in the predictor variables (i.e., reflectance wavelengths), even if predictors exceed
3 Scaling Functional Traits from Leaves to Canopies
spectrum and thus selects specific wavelengths, and therefore SVI, for their analysis.
An alternative approach is to explore the spectra and trait space to identify new or
previously unknown SVIs that maximize the correspondence between optical properties and traits of interest (e.g., Inoue et al. 2008), akin to a data mining exercise.
A challenge of this approach can be interpretation of the selected SVIs, where the
resulting vegetation indices may not contain wavelengths with known absorption
features relating to the trait of interest. The same general approach can also leverage
multiple SVIs, provided the research avoids highly correlated portions of the spectrum (Grossman et al., 1996), to attempt to capture how variation in the trait of
interest is reflected in various portions of the EM spectrum to other sites and plant
species. However, a limitation to the use of SVIs has been the ability to generalize
across broad canopy architectures, species, and environments due to the often sitespecific modeling results or potential signal saturation issues with some SVIs
(Shabanov et al. 2005; Glenn et al. 2008).
Continuous spectral wavelet transforms have been used to reduce the dimensionality of spectral data prior to developing simple statistical models (e.g., Blackburn
and Ferwerda 2008). Wavelets are functions that are used to decompose a full, complex signal into simpler component sub-signals. When used with spectral data, the
full reflectance signature can be decomposed in a way that allows the resulting
wavelet coefficients assigned to each sub-signal to be related to concentrations of
chemical constituents or other traits of interest, through standard statistical modeling approaches (e.g., linear regression). Previous studies have explored the use of
wavelet methods to retrieve a host of functional traits, including pigments, water,
and nitrogen content (e.g., Blackburn and Ferwerda 2008; Cheng et al. 2011; Li
et al. 2018; Wang et al. 2018). Continuum removal together with band-depth analysis (Kokaly and Clark 1999) has also been utilized as a means to retrieve the chemical composition of leaves. In this approach, continuum removal lines are fit through
the absorption features of interest based on those regions not in the areas of interest,
then the original spectra are divided by corresponding values of the continuum
removal line. The band centers can then be found by finding the minimum of the
continuum-removed spectra. Normalization of the band centers is often used to
standardize the values across samples. These data are then used to develop models
to predict functional traits at the leaf and canopy scales, including foliar nitrogen
and recalcitrant properties, such as the amount of lignin and cellulose (Kokaly
et al. 2009).
In addition to the empirical SVI approach, as discussed in Schweiger (Chap. 15),
partial least-squares regression (PLSR) modeling has been used extensively in the
development of spectra-trait models for measuring, scaling, and mapping plant
functional traits (e.g., Ollinger et al. 2002; Townsend et al. 2003; Asner and Martin
2008; Martin et al. 2008; Dahlin et al. 2013; Singh et al. 2015; Ely et al. 2019). A
key attribute of PLSR is the capacity to utilize the entire measured portion of the
EM spectrum as predictors (i.e., X matrix) without requiring a priori selection of
wavelengths or SVIs (Wold et al. 1984; Geladi and Kowalski 1986; Wold et al.
2001). PLSR avoids collinearity (i.e., spectral autocorrelation across wavelengths)
in the predictor variables (i.e., reflectance wavelengths), even if predictors exceed
3 Scaling Functional Traits from Leaves to Canopies