392
increase in ecological niches with increasing energy or resources in ecosystems
(Brown 1981; Wright 1983; Bonn et al. 2004). In addition, advances in sensor technology allow capturing aboveground productivity in ecologically meaningful units,
such as the annual amount, variation, and minimum of photosynthetically active
radiation (fPAR), which has been found to explain global patterns of mammalian,
amphibian, and avian diversity to a substantial degree (Coops et al. 2018). The idea
behind spectral indices is that they are generally applicable and transferable; however, ground reference data are still needed to assess their accuracy. In addition,
site-specific data are also needed to recalibrate spectral indices, for example, by
selecting the optimal wavelengths for the sensor used and by estimating site-specific
model coefficients, because index responses often vary with the particular context.
Latent variable methods, such as partial least squares regression (PLSR) and
partial least squares discriminant analysis (PLSDA), were developed for chemometrics and specifically deal with the high degree of autocorrelation inherent in data
with high spectral resolution (Wold et al. 1983; Martens 2001). PLSR is a standard
method for modeling and predicting continuous vegetation characteristics and
PLSDA for determining class membership from spectral data. Clustering methods,
including principal component analysis (PCA), principal coordinate analysis
(PCoA), and linear discriminant analysis (LDA), are frequently used to explore patterns, such as the degree to which species or plant functional types cluster separately
from each other in spectral space. In addition, machine learning algorithms [e.g.,
random forest (RF) or support vector machines (SVNs)] and deep learning methods
[e.g., convolutional neural networks (CNNs)] can be used for classification problems, including the identification of vegetation types based on physiognomic attributes (e.g., forest, shrubland, grassland, cropland), for the detection of plant
pathogens and other stresses (Pontius et al. 2005; Herrmann et al. 2018), and for
species detection (Clark et al. 2005; Fassnacht et al. 2016; Kattenborn et al. 2019).
Physical models are based on causal physical relationships between electromagnetic radiation and vegetation properties; in spectroscopy, these models are called
radiative transfer models (RTMs). For leaf optical properties, the RTM PROSPECT
(Jacquemoud and Baret 1990; Jacquemoud et al. 2009) models leaf reflectance and
transmittance based on leaf chlorophyll a and b content, “brown pigment” content,
equivalent water thickness, leaf dry matter content, and a leaf structure parameter
indicating mesophyll thickness and density. For modeling optical properties of canopies, leaf-level RTMs can be combined with canopy RTMs (Jacquemoud et al.
2009), which incorporate structural canopy parameters, including leaf area index
(LAI, the ratio of leaf area to ground surface area), leaf inclination, a hot spot
parameter (a function of the ratio of leaf size to canopy height), as well as soil
reflectance and measurement characteristics, including sun and viewing angle.
Frequently used canopy RTMs include SAIL (Verhoef 1984), GeoSAIL (Huemmrich
2001), and DART (Gastellu-Etchegorry et al. 2004). Although RTMs do not model
all interactions between plants and light, because they cannot incorporate all characteristics of leaves and vegetation canopies that influence the spectral response,
they are useful for simulating spectra and retrieving estimates about plant
A. K. Schweiger
increase in ecological niches with increasing energy or resources in ecosystems
(Brown 1981; Wright 1983; Bonn et al. 2004). In addition, advances in sensor technology allow capturing aboveground productivity in ecologically meaningful units,
such as the annual amount, variation, and minimum of photosynthetically active
radiation (fPAR), which has been found to explain global patterns of mammalian,
amphibian, and avian diversity to a substantial degree (Coops et al. 2018). The idea
behind spectral indices is that they are generally applicable and transferable; however, ground reference data are still needed to assess their accuracy. In addition,
site-specific data are also needed to recalibrate spectral indices, for example, by
selecting the optimal wavelengths for the sensor used and by estimating site-specific
model coefficients, because index responses often vary with the particular context.
Latent variable methods, such as partial least squares regression (PLSR) and
partial least squares discriminant analysis (PLSDA), were developed for chemometrics and specifically deal with the high degree of autocorrelation inherent in data
with high spectral resolution (Wold et al. 1983; Martens 2001). PLSR is a standard
method for modeling and predicting continuous vegetation characteristics and
PLSDA for determining class membership from spectral data. Clustering methods,
including principal component analysis (PCA), principal coordinate analysis
(PCoA), and linear discriminant analysis (LDA), are frequently used to explore patterns, such as the degree to which species or plant functional types cluster separately
from each other in spectral space. In addition, machine learning algorithms [e.g.,
random forest (RF) or support vector machines (SVNs)] and deep learning methods
[e.g., convolutional neural networks (CNNs)] can be used for classification problems, including the identification of vegetation types based on physiognomic attributes (e.g., forest, shrubland, grassland, cropland), for the detection of plant
pathogens and other stresses (Pontius et al. 2005; Herrmann et al. 2018), and for
species detection (Clark et al. 2005; Fassnacht et al. 2016; Kattenborn et al. 2019).
Physical models are based on causal physical relationships between electromagnetic radiation and vegetation properties; in spectroscopy, these models are called
radiative transfer models (RTMs). For leaf optical properties, the RTM PROSPECT
(Jacquemoud and Baret 1990; Jacquemoud et al. 2009) models leaf reflectance and
transmittance based on leaf chlorophyll a and b content, “brown pigment” content,
equivalent water thickness, leaf dry matter content, and a leaf structure parameter
indicating mesophyll thickness and density. For modeling optical properties of canopies, leaf-level RTMs can be combined with canopy RTMs (Jacquemoud et al.
2009), which incorporate structural canopy parameters, including leaf area index
(LAI, the ratio of leaf area to ground surface area), leaf inclination, a hot spot
parameter (a function of the ratio of leaf size to canopy height), as well as soil
reflectance and measurement characteristics, including sun and viewing angle.
Frequently used canopy RTMs include SAIL (Verhoef 1984), GeoSAIL (Huemmrich
2001), and DART (Gastellu-Etchegorry et al. 2004). Although RTMs do not model
all interactions between plants and light, because they cannot incorporate all characteristics of leaves and vegetation canopies that influence the spectral response,
they are useful for simulating spectra and retrieving estimates about plant
A. K. Schweiger