68
sampling should be as close to the RS collection date as possible, as an optimal
approach, but at least be selected to match the phenological stage of the vegetation
during the imagery collection, if leveraging sample campaigns in following year(s).
A number of different methods have been used to collect plant functional traits
to link with RS imagery (e.g., Wang et al. 2019). Common approaches for the collection of canopy leaf samples include the use of slingshot, pruning pole, and shotgun (Lausch et al., Chap. 13), but also include line-launcher and air cannon (e.g.,
Serbin et al. 2014); simpler tools and hand shears are often used for accessible,
shorter canopies. Regardless of the sample collection approach, harvested leaves
should be reasonably intact and minimally damaged in order to avoid any issues
with changes in leaf chemistry from physical damage or stress. In addition, leaves
should be immediately measured for leaf optical properties and fresh mass, if these
are of interest, then stored in humidified and sealed bags and placed in a cool, dark
place prior to transport for further processing. Processing should then be completed
within 2–4 hours of sampling—though a much shorter time between sample and
measurement or different sample storage and handling (e.g., flash freezing in liquid
nitrogen) may be needed for specific biochemical traits. Typically top-of-canopy,
sunlit samples have been the main focus; however, more recent work has also begun
to focus on collection of canopy and subcanopy samples (e.g., Serbin et al. 2014;
Singh et al. 2015). This provides the ability to evaluate the depth in the canopy
needed to link traits with image, which may vary by vegetation type or LAI.
3.3.2 Evaluating Functional Trait Maps and the Need
to Quantify Uncertainties
Maps of plant functional traits are useful for a wide variety of applications. From an
ecological perspective, maps of plant traits across broad biotic and abiotic gradients
can be used to explore the drivers of plant trait variation in relation to climate, soils,
and vegetation types (e.g., McNeil et al. 2008). Modeling activities can leverage
these trait maps as either inputs for model parameterization across space and time
(Ollinger and Smith 2005) or to evaluate prognostic plant trait predictions. However,
to maximize the utility of functional trait maps a detailed understanding of the their
uncertainties across space and time is required.
In the earliest functional trait mapping work, predictive model uncertainties were
limited to the “goodness of fit” and overall model root mean square error (RMSE)
statistics provided by the modeling approach (e.g., Wessman et al. 1988; Martin and
Aber 1997; Townsend et al. 2003). While this information is helpful to understand
the accuracy of the model fit, that level of accuracy assessment is insufficient for
characterizing the uncertainty of the trait maps themselves. Mapping efforts should
instead provide an accounting of the trait measurement, scaling, and algorithm
uncertainties and provide this information in the resulting trait map data products.
However, detailed error propagation is not trivial, particularly with respect to empirical modeling approaches, and is an ongoing and active area of research in the RS
S. P. Serbin and P. A. Townsend
sampling should be as close to the RS collection date as possible, as an optimal
approach, but at least be selected to match the phenological stage of the vegetation
during the imagery collection, if leveraging sample campaigns in following year(s).
A number of different methods have been used to collect plant functional traits
to link with RS imagery (e.g., Wang et al. 2019). Common approaches for the collection of canopy leaf samples include the use of slingshot, pruning pole, and shotgun (Lausch et al., Chap. 13), but also include line-launcher and air cannon (e.g.,
Serbin et al. 2014); simpler tools and hand shears are often used for accessible,
shorter canopies. Regardless of the sample collection approach, harvested leaves
should be reasonably intact and minimally damaged in order to avoid any issues
with changes in leaf chemistry from physical damage or stress. In addition, leaves
should be immediately measured for leaf optical properties and fresh mass, if these
are of interest, then stored in humidified and sealed bags and placed in a cool, dark
place prior to transport for further processing. Processing should then be completed
within 2–4 hours of sampling—though a much shorter time between sample and
measurement or different sample storage and handling (e.g., flash freezing in liquid
nitrogen) may be needed for specific biochemical traits. Typically top-of-canopy,
sunlit samples have been the main focus; however, more recent work has also begun
to focus on collection of canopy and subcanopy samples (e.g., Serbin et al. 2014;
Singh et al. 2015). This provides the ability to evaluate the depth in the canopy
needed to link traits with image, which may vary by vegetation type or LAI.
3.3.2 Evaluating Functional Trait Maps and the Need
to Quantify Uncertainties
Maps of plant functional traits are useful for a wide variety of applications. From an
ecological perspective, maps of plant traits across broad biotic and abiotic gradients
can be used to explore the drivers of plant trait variation in relation to climate, soils,
and vegetation types (e.g., McNeil et al. 2008). Modeling activities can leverage
these trait maps as either inputs for model parameterization across space and time
(Ollinger and Smith 2005) or to evaluate prognostic plant trait predictions. However,
to maximize the utility of functional trait maps a detailed understanding of the their
uncertainties across space and time is required.
In the earliest functional trait mapping work, predictive model uncertainties were
limited to the “goodness of fit” and overall model root mean square error (RMSE)
statistics provided by the modeling approach (e.g., Wessman et al. 1988; Martin and
Aber 1997; Townsend et al. 2003). While this information is helpful to understand
the accuracy of the model fit, that level of accuracy assessment is insufficient for
characterizing the uncertainty of the trait maps themselves. Mapping efforts should
instead provide an accounting of the trait measurement, scaling, and algorithm
uncertainties and provide this information in the resulting trait map data products.
However, detailed error propagation is not trivial, particularly with respect to empirical modeling approaches, and is an ongoing and active area of research in the RS
S. P. Serbin and P. A. Townsend