407
distribution of values within classes, and the distance among the class centroids. As
before, the strength of the relationship between classes and their spectral characteristics and measurement accuracies should be taken into account when deciding on
sample sizes. As a rule of thumb, a minimum of 50 samples per class and 75–100
samples per class for more than 12 categories or areas larger than 4000 km
2
has been
suggested (Congalton and Green 1999), but fewer samples can provide sufficient
accuracy when classes are relatively dissimilar. Tracking error propagation is important for assessing the performance of spectral models (Singh et al. 2015, Wang et al.
2019, Serbin and Townsend, Chap. 3); ideally, the accuracy of laboratory analysis
should be included in error assessments, as well.
For empirical-statistical models that combine vegetation characteristic and spectral measurements, stratified random sampling is often a good choice. Stratifying a
research area based on ecologically relevant environmental variation helps cover the
heterogeneity of a research area and the range of values of the vegetation characteristics of interest. A range of methods for automating sampling designs are available
for R, including the packages spsurvey (Kincaid and Olsen 2016), spcosa (Walvoort
et al. 2010), spatstat (Baddeley and Turner 2005), and spatialEco (Evans 2017), and
ArcGIS, including the Geospatial Modeling Environment (Beyer 2010) and the
Reverse Randomized Quadrant-Recursive Raster algorithm (RRQRR, Theobald
et al. 2007). However, it can be difficult to automate sampling design completely,
especially in natural ecosystems with limited accessibility. Moreover, for studies
with a RS component, it can be difficult to select research plots that are internally
homogeneous and located far enough from objects that influence the spectral signal
of neighboring pixels (see Sect. 15.2.3.1) automatically. Under such circumstances,
a mix of automated sampling based on GIS data and informed decision-making
(convenience/purposive sampling) can be a good option. For example, information
about environmental factors and gradients influencing the vegetation characteristics
of interest and other relevant information about the study area, such as accessibility
and travel time, can be used as input into a GIS and used as strata. Random points
per stratum can be created automatically and used, for example, to define larger
polygons within which the exact location of research plots is determined in the field.
When vegetation characteristics are expected to vary along gradients, cluster sampling of plots at predefined intervals along these transects is a good choice—but
again, it might be necessary to adjust these distances to avoid objects influencing the
spectral signal of the plots or to find internally homogeneous areas. In this context,
areas can be considered “internally homogeneous” when their biotic and abiotic
characteristics are comparable, which means that they can actually show a high
degree of small-scale heterogeneity (e.g., situations changing every 5 cm) as long as
this small-scale heterogeneity creates a similar mosaic at the measurement scale
(e.g., 1 m
2
is comparable to the adjacent 1 m
2
; see Sect. 15.2.3.1). It is important to
report the reasons for deviating from common sampling schemes in the methods.
When working at the level of individual plants, sampling random points within
research plots makes it possible to capture interindividual variation, which can be
important, for example, when scaling functional traits of individual plants to plotlevel estimates (Wang et al. 2019; Serbin and Townsend, Chap. 3). Random
15 Spectral Field Campaigns: Planning and Data Collection
distribution of values within classes, and the distance among the class centroids. As
before, the strength of the relationship between classes and their spectral characteristics and measurement accuracies should be taken into account when deciding on
sample sizes. As a rule of thumb, a minimum of 50 samples per class and 75–100
samples per class for more than 12 categories or areas larger than 4000 km
2
has been
suggested (Congalton and Green 1999), but fewer samples can provide sufficient
accuracy when classes are relatively dissimilar. Tracking error propagation is important for assessing the performance of spectral models (Singh et al. 2015, Wang et al.
2019, Serbin and Townsend, Chap. 3); ideally, the accuracy of laboratory analysis
should be included in error assessments, as well.
For empirical-statistical models that combine vegetation characteristic and spectral measurements, stratified random sampling is often a good choice. Stratifying a
research area based on ecologically relevant environmental variation helps cover the
heterogeneity of a research area and the range of values of the vegetation characteristics of interest. A range of methods for automating sampling designs are available
for R, including the packages spsurvey (Kincaid and Olsen 2016), spcosa (Walvoort
et al. 2010), spatstat (Baddeley and Turner 2005), and spatialEco (Evans 2017), and
ArcGIS, including the Geospatial Modeling Environment (Beyer 2010) and the
Reverse Randomized Quadrant-Recursive Raster algorithm (RRQRR, Theobald
et al. 2007). However, it can be difficult to automate sampling design completely,
especially in natural ecosystems with limited accessibility. Moreover, for studies
with a RS component, it can be difficult to select research plots that are internally
homogeneous and located far enough from objects that influence the spectral signal
of neighboring pixels (see Sect. 15.2.3.1) automatically. Under such circumstances,
a mix of automated sampling based on GIS data and informed decision-making
(convenience/purposive sampling) can be a good option. For example, information
about environmental factors and gradients influencing the vegetation characteristics
of interest and other relevant information about the study area, such as accessibility
and travel time, can be used as input into a GIS and used as strata. Random points
per stratum can be created automatically and used, for example, to define larger
polygons within which the exact location of research plots is determined in the field.
When vegetation characteristics are expected to vary along gradients, cluster sampling of plots at predefined intervals along these transects is a good choice—but
again, it might be necessary to adjust these distances to avoid objects influencing the
spectral signal of the plots or to find internally homogeneous areas. In this context,
areas can be considered “internally homogeneous” when their biotic and abiotic
characteristics are comparable, which means that they can actually show a high
degree of small-scale heterogeneity (e.g., situations changing every 5 cm) as long as
this small-scale heterogeneity creates a similar mosaic at the measurement scale
(e.g., 1 m
2
is comparable to the adjacent 1 m
2
; see Sect. 15.2.3.1). It is important to
report the reasons for deviating from common sampling schemes in the methods.
When working at the level of individual plants, sampling random points within
research plots makes it possible to capture interindividual variation, which can be
important, for example, when scaling functional traits of individual plants to plotlevel estimates (Wang et al. 2019; Serbin and Townsend, Chap. 3). Random
15 Spectral Field Campaigns: Planning and Data Collection