400
planning phase. Often, it is easier to discuss issues and concerns in person, and seeing the conditions on site facilitates decision-making. Joining another research
group for some time in the field or accompanying a person who knows the area well
provides a great way to get to know an area. Covering the heterogeneity of a research
area is important for all aspects of remote sensing of plant biodiversity, including
collecting spectral references for image processing and sampling ground data (such
as vegetation spectra and vegetation samples) for model calibration, validation, and
interpretation. For example, empirical line correction (ELC, Kruse et al. 1990,
Broge and Leblanc 2001), a common method for correcting atmospheric and instrument influences on remotely sensed data, utilizes ground measurements of invariant
surfaces (e.g., pavements, rocky outcrops, snow, water, calibration tarps) that can be
readily identified in the image and/or for which accurate location data has been
acquired. As for any empirical method, the performance of ELC depends on the
representativeness and accuracy of the input data, and model transferability is limited. Thus, target surfaces for ELC should ideally be distributed across the area of
interest (e.g., located across all flight lines) and cover differences in altitude, slope,
and aspect (see Sect. 15.3.3.2).
Sampling biodiversity in a way that fits remotely sensed data means incorporating the heterogeneity of a research area but also requires thinking about the size and
shape of sampling units. Generally, sampling units should be delineated to encompass areas with similar environmental conditions. Remotely sensed imagery are
usually raster data, so ground measurements need to represent areas rather than
points. The optimal size of the sampling units on the ground depends on the spatial
heterogeneity and spatial resolution of the imagery. Sampling units that are smaller
or the same size as the pixels in the imagery are usually unrepresentative (Justice
and Townshend 1981). Pixel shifts are a common consequence of image processing,
and averaging RS data across several pixels is common practice for noise reduction.
As a general guideline, the minimal dimensions of a representative sampling unit
can be calculated as A = P ∗ (1 + 2 L) (Justice and Townshend 1981), with P being
the pixel dimensions of the image and L the accuracy of image alignment in number
of pixels. For example, if the spatial resolution of an image is 3 m and a one-pixel
shift is expected to occur during image processing, the minimal size of an internally
homogeneous sampling area would be 3∗(1 + 2 ∗ 1) = 9 m × 9 m. This makes it
possible to capture similar environmental conditions even when the image pixel that
should align with the sampling unit’s center pixel has shifted one pixel in either
direction or when averaging the center pixel and its neighboring pixels (Fig. 15.1).
In this particular context, internal homogeneity does not mean that the sampling
unit can only consist of one particular feature but rather that all features should be
evenly distributed throughout the unit. In other words, to be considered internally
homogeneous, a sampling unit does not have to consist of a single plant species of
one particular age or size class; it can consist of different species and individuals as
long as their spatial distribution is comparable among the pixels within that unit. For
example, the center pixel in Fig. 15.1a is not representative for the sampling unit
because species abundance varies among the nine pixels; the sampling area is internally heterogeneous. In contrast, the center pixel in Fig. 15.1b is representative for
A. K. Schweiger
planning phase. Often, it is easier to discuss issues and concerns in person, and seeing the conditions on site facilitates decision-making. Joining another research
group for some time in the field or accompanying a person who knows the area well
provides a great way to get to know an area. Covering the heterogeneity of a research
area is important for all aspects of remote sensing of plant biodiversity, including
collecting spectral references for image processing and sampling ground data (such
as vegetation spectra and vegetation samples) for model calibration, validation, and
interpretation. For example, empirical line correction (ELC, Kruse et al. 1990,
Broge and Leblanc 2001), a common method for correcting atmospheric and instrument influences on remotely sensed data, utilizes ground measurements of invariant
surfaces (e.g., pavements, rocky outcrops, snow, water, calibration tarps) that can be
readily identified in the image and/or for which accurate location data has been
acquired. As for any empirical method, the performance of ELC depends on the
representativeness and accuracy of the input data, and model transferability is limited. Thus, target surfaces for ELC should ideally be distributed across the area of
interest (e.g., located across all flight lines) and cover differences in altitude, slope,
and aspect (see Sect. 15.3.3.2).
Sampling biodiversity in a way that fits remotely sensed data means incorporating the heterogeneity of a research area but also requires thinking about the size and
shape of sampling units. Generally, sampling units should be delineated to encompass areas with similar environmental conditions. Remotely sensed imagery are
usually raster data, so ground measurements need to represent areas rather than
points. The optimal size of the sampling units on the ground depends on the spatial
heterogeneity and spatial resolution of the imagery. Sampling units that are smaller
or the same size as the pixels in the imagery are usually unrepresentative (Justice
and Townshend 1981). Pixel shifts are a common consequence of image processing,
and averaging RS data across several pixels is common practice for noise reduction.
As a general guideline, the minimal dimensions of a representative sampling unit
can be calculated as A = P ∗ (1 + 2 L) (Justice and Townshend 1981), with P being
the pixel dimensions of the image and L the accuracy of image alignment in number
of pixels. For example, if the spatial resolution of an image is 3 m and a one-pixel
shift is expected to occur during image processing, the minimal size of an internally
homogeneous sampling area would be 3∗(1 + 2 ∗ 1) = 9 m × 9 m. This makes it
possible to capture similar environmental conditions even when the image pixel that
should align with the sampling unit’s center pixel has shifted one pixel in either
direction or when averaging the center pixel and its neighboring pixels (Fig. 15.1).
In this particular context, internal homogeneity does not mean that the sampling
unit can only consist of one particular feature but rather that all features should be
evenly distributed throughout the unit. In other words, to be considered internally
homogeneous, a sampling unit does not have to consist of a single plant species of
one particular age or size class; it can consist of different species and individuals as
long as their spatial distribution is comparable among the pixels within that unit. For
example, the center pixel in Fig. 15.1a is not representative for the sampling unit
because species abundance varies among the nine pixels; the sampling area is internally heterogeneous. In contrast, the center pixel in Fig. 15.1b is representative for
A. K. Schweiger