64
While the PLSR approach produces algorithms that “weight” wavelengths by
their importance in the prediction (Wold et al. 2001) of the functional traits of interest (e.g., Serbin et al. 2014), some researchers have also explored modifications to
the standard PLSR approach that provide additional reductions in data dimensionality. For example, Li et al. (2008) coupled PLSR with a genetic algorithm (GA)
approach to select a smaller subset of wavelengths to use in the final PLSR model
for predicting leaf water content, measured as equivalent water thickness (EWT).
DuBois et al. (2018) combined the SVI and PLSR approach by using all two-band
AVIRIS wavelength combinations to model the relationship between spectral reflectance and ecosystem carbon fluxes across a water-limited environment. To date, the
spectra-trait PLSR modeling approach has shown the capacity to characterize the
widest array of leaf functional traits using the optical properties of plants across a
broad range of species and ecosystems (e.g., Dahlin et al. 2013; Asner et al. 2014;
Asner et al. 2015; Serbin et al. 2015; Singh et al. 2015; Couture et al. 2016).
Similar to the PLSR approach, researchers have leveraged various machinelearning approaches to connect RS observations to functional traits (e.g., Féret et al.
2018). Schweiger (Chap. 15) describes two commonly used machine-learning
approaches in RS; several other approaches have also been used to model trait variation as a function of spectral measurements. More recently, Gaussian processes
regression (GPR) has been recommended as superior to other machine-learning
approaches for trait mapping from imaging spectroscopy data (Verrelst et al. 2012;
Verrelst et al. 2016). GPR is a nonlinear nonparametric probabilistic approach similar to kernel ridge regression that directly generates uncertainty (or confidence) levels for the prediction (Wang et al. 2019). This is in contrast to PLSR uncertainties,
generally assessed through permutation (Singh et al. 2015; Serbin et al. 2015). PLSR
and GPR yield very similar results, both in terms of absolute trait predictions and
relative scaling of uncertainties (Wang et al. 2019). PLSR is much more computationally efficient, and results are readily interpretable in terms of wavelength quantitative contribution to prediction (see Fig. 3.1 in Schimel et al., Chap. 19), whereas
GPR only identifies relatively important wavelengths.
The challenge with most machine-learning approaches is that some level of data
reduction is required for optimal performance. Standard approaches, such as principle
component analysis (PCA) or minimum noise fraction (MNF) transformations, may
reduce data dimensionality. However, features important to trait estimation may be
buried in lower principle components, as high contrast variation (albedo, greenness,
water content) dominate scene properties. In contrast, PLSR rotates the data into
latent vectors optimized to the empirical dependent variables, which generally yields
strong models for calibration data but can lead to poor model performance when confronted with new data that differ considerably from the model-building data sets.
3.2.2.2 Radiative Transfer Models and Scaling Functional Traits
An alternative to statistical, field-based, and empirical approaches for connecting
leaf and canopy optical properties with plant functional traits, RTMs can be used
either at the leaf and canopy scales to directly retrieve leaf traits (e.g., Colombo
S. P. Serbin and P. A. Townsend
While the PLSR approach produces algorithms that “weight” wavelengths by
their importance in the prediction (Wold et al. 2001) of the functional traits of interest (e.g., Serbin et al. 2014), some researchers have also explored modifications to
the standard PLSR approach that provide additional reductions in data dimensionality. For example, Li et al. (2008) coupled PLSR with a genetic algorithm (GA)
approach to select a smaller subset of wavelengths to use in the final PLSR model
for predicting leaf water content, measured as equivalent water thickness (EWT).
DuBois et al. (2018) combined the SVI and PLSR approach by using all two-band
AVIRIS wavelength combinations to model the relationship between spectral reflectance and ecosystem carbon fluxes across a water-limited environment. To date, the
spectra-trait PLSR modeling approach has shown the capacity to characterize the
widest array of leaf functional traits using the optical properties of plants across a
broad range of species and ecosystems (e.g., Dahlin et al. 2013; Asner et al. 2014;
Asner et al. 2015; Serbin et al. 2015; Singh et al. 2015; Couture et al. 2016).
Similar to the PLSR approach, researchers have leveraged various machinelearning approaches to connect RS observations to functional traits (e.g., Féret et al.
2018). Schweiger (Chap. 15) describes two commonly used machine-learning
approaches in RS; several other approaches have also been used to model trait variation as a function of spectral measurements. More recently, Gaussian processes
regression (GPR) has been recommended as superior to other machine-learning
approaches for trait mapping from imaging spectroscopy data (Verrelst et al. 2012;
Verrelst et al. 2016). GPR is a nonlinear nonparametric probabilistic approach similar to kernel ridge regression that directly generates uncertainty (or confidence) levels for the prediction (Wang et al. 2019). This is in contrast to PLSR uncertainties,
generally assessed through permutation (Singh et al. 2015; Serbin et al. 2015). PLSR
and GPR yield very similar results, both in terms of absolute trait predictions and
relative scaling of uncertainties (Wang et al. 2019). PLSR is much more computationally efficient, and results are readily interpretable in terms of wavelength quantitative contribution to prediction (see Fig. 3.1 in Schimel et al., Chap. 19), whereas
GPR only identifies relatively important wavelengths.
The challenge with most machine-learning approaches is that some level of data
reduction is required for optimal performance. Standard approaches, such as principle
component analysis (PCA) or minimum noise fraction (MNF) transformations, may
reduce data dimensionality. However, features important to trait estimation may be
buried in lower principle components, as high contrast variation (albedo, greenness,
water content) dominate scene properties. In contrast, PLSR rotates the data into
latent vectors optimized to the empirical dependent variables, which generally yields
strong models for calibration data but can lead to poor model performance when confronted with new data that differ considerably from the model-building data sets.
3.2.2.2 Radiative Transfer Models and Scaling Functional Traits
An alternative to statistical, field-based, and empirical approaches for connecting
leaf and canopy optical properties with plant functional traits, RTMs can be used
either at the leaf and canopy scales to directly retrieve leaf traits (e.g., Colombo
S. P. Serbin and P. A. Townsend