391
evaluating model performance and transferability. External validation involves
either leaving out samples from the internal calibration-validation process or collecting additional independent data, followed by the evaluation of agreement
between model output and observed data without any attempt to modify model
parameters to improve fit.
Spectral models can be categorized into empirical-statistical and physical models and combinations between the two (see Verrelst et al. 2015). Empirical-statistical
models are based on the relationship between the spectral behavior of certain spectral bands or the entire spectrum (the predictors or independent variables) and the
vegetation characteristic(s) of interest (the dependent variable(s)). Empiricalstatistical models generally use regression or clustering algorithms and aim to predict vegetation characteristics or class membership from a population of spectral
data that were not included in the modeling process. For calibrating empiricalstatistical models, it is essential to collect representative field data (see Sect. 15.2.3).
Generally, this means that for the ecosystem, time of year, and area of interest, data
should cover the range of values for which predictions are intended to be made or
the number of vegetation classes with suitable replication, as well as the range of
environmental conditions present in that area. Sample size should be large enough
and samples distributed evenly across the expected range of values, classes, and
environmental gradients to allow samples to be left out from model development
and enable external validation.
Regression techniques are generally used for modeling and predicting continuous vegetation characteristics, such as biomass and chemical or structural composition (Ustin et al. 2009; Serbin et al. 2014; Schweiger et al. 2015a, b; Couture et al.
2016), or relative proportions of vegetation properties, such as the abundances of
species, plant functional types, or vegetation types (Schmidtlein et al. 2012; Lopatin
et al. 2017; Fassnacht et al. 2016; Féret and Asner 2014; Schweiger et al. 2017).
Univariate, multivariate, linear, and nonlinear regressions are common for modeling
and predicting vegetation characteristics from few spectral bands or from spectral
indices. Spectral indices are used to infer vegetation status, including plant stress,
and ecosystem parameters, inducing productivity, from empirical or physical relationships between spectra and plant traits. Many spectral indices have been published (see, e.g., https://cubert-gmbh.com/applications/vegetation-indices/). The
most widely used indices include the normalized difference vegetation index
(NDVI; Rouse Jr et al. 1974; Tucker 1979), an indicator of vegetation greenness,
and modified versions such as the soil-adjusted vegetation index (SAVI; Huete
1988) and the photochemical reflectance index (PRI; Gamon et al. 1992). The NDVI
and its variants have been shown to correlate well with biomass, LAI, and the photosynthetic capacity of canopies. The PRI estimates light-use efficiency and can be
used to estimate gross primary productivity (GPP) and assess environmental stress
(Sims and Gamon 2002). Spectral indices can also be used directly to estimate certain environmental characteristics (Anderson et al. 2010; Pettorelli et al. 2011;
Wang et al. 2016). For example, the NDVI has been used for predicting and mapping taxonomic diversity (e.g., Gould 2000), the rationale being the expected
15 Spectral Field Campaigns: Planning and Data Collection
evaluating model performance and transferability. External validation involves
either leaving out samples from the internal calibration-validation process or collecting additional independent data, followed by the evaluation of agreement
between model output and observed data without any attempt to modify model
parameters to improve fit.
Spectral models can be categorized into empirical-statistical and physical models and combinations between the two (see Verrelst et al. 2015). Empirical-statistical
models are based on the relationship between the spectral behavior of certain spectral bands or the entire spectrum (the predictors or independent variables) and the
vegetation characteristic(s) of interest (the dependent variable(s)). Empiricalstatistical models generally use regression or clustering algorithms and aim to predict vegetation characteristics or class membership from a population of spectral
data that were not included in the modeling process. For calibrating empiricalstatistical models, it is essential to collect representative field data (see Sect. 15.2.3).
Generally, this means that for the ecosystem, time of year, and area of interest, data
should cover the range of values for which predictions are intended to be made or
the number of vegetation classes with suitable replication, as well as the range of
environmental conditions present in that area. Sample size should be large enough
and samples distributed evenly across the expected range of values, classes, and
environmental gradients to allow samples to be left out from model development
and enable external validation.
Regression techniques are generally used for modeling and predicting continuous vegetation characteristics, such as biomass and chemical or structural composition (Ustin et al. 2009; Serbin et al. 2014; Schweiger et al. 2015a, b; Couture et al.
2016), or relative proportions of vegetation properties, such as the abundances of
species, plant functional types, or vegetation types (Schmidtlein et al. 2012; Lopatin
et al. 2017; Fassnacht et al. 2016; Féret and Asner 2014; Schweiger et al. 2017).
Univariate, multivariate, linear, and nonlinear regressions are common for modeling
and predicting vegetation characteristics from few spectral bands or from spectral
indices. Spectral indices are used to infer vegetation status, including plant stress,
and ecosystem parameters, inducing productivity, from empirical or physical relationships between spectra and plant traits. Many spectral indices have been published (see, e.g., https://cubert-gmbh.com/applications/vegetation-indices/). The
most widely used indices include the normalized difference vegetation index
(NDVI; Rouse Jr et al. 1974; Tucker 1979), an indicator of vegetation greenness,
and modified versions such as the soil-adjusted vegetation index (SAVI; Huete
1988) and the photochemical reflectance index (PRI; Gamon et al. 1992). The NDVI
and its variants have been shown to correlate well with biomass, LAI, and the photosynthetic capacity of canopies. The PRI estimates light-use efficiency and can be
used to estimate gross primary productivity (GPP) and assess environmental stress
(Sims and Gamon 2002). Spectral indices can also be used directly to estimate certain environmental characteristics (Anderson et al. 2010; Pettorelli et al. 2011;
Wang et al. 2016). For example, the NDVI has been used for predicting and mapping taxonomic diversity (e.g., Gould 2000), the rationale being the expected
15 Spectral Field Campaigns: Planning and Data Collection