402
inventory within 1 m
2
, ideally in the plot center. Likewise, it can be expected that the
chemical composition of the biomass clipped in a 1 m × 20 cm strip in the central
1 m
2
is representative for the chemical composition of the vegetation in the entire
9 m × 9 m plot, given that the clip strip captures the variation in species composition, height, and age-class distribution within the central 1 m
2
.
The accuracy and precision of the surveying equipment used for measuring plot
coordinates is another aspect to consider when determining the minimal size of
homogeneous sampling units; ideally, measurement errors should be estimated
under field conditions and added to the minimal size of the sampling unit.
Additionally, edge effects influence remotely sensed data. Sampling units should be
placed sufficiently far from landscape features, such as open soil, gravel, snow,
water, large rocks, roads, footpaths, bridges, and trees (when working in grasslands), that influence the spectral properties of adjacent areas. Depending on the
time of day of image acquisition, shadow effects from tall objects such as trees,
mountains, or buildings need to be taken into account as well.
15.2.3.2 Range of Values
For the remote sensing of biodiversity, the range of conditions (time of year, value
range, species, and environmental context) used to calibrate a model should cover
the range of conditions for which inferences or predictions should be made. For
example, extrapolating beyond the range of values relies on the assumption that the
estimated relationship holds beyond the investigated range. This cannot be assumed
without additional information, because nonlinearities (e.g., saturating curves) are
common in ecological data, especially when covering large areas and multiple environmental gradients and ecosystems.
For modeling continuous data with regression-style empirical-statistical
approaches, the sampling design should cover the expected range of values in the
area of interest with a sufficient number of evenly distributed samples. Predictions
outside the calibrated range are not reliable because deviations from the 1:1 line
between measured and predicted values (Fig. 15.2a) increase at the lower and upper
ends of the distribution (Fig. 15.2b). During model validation the entire range of
values should be covered as well, specifically paying attention to the value range
most important for the research question(s). When the tails of the distribution are of
interest for predictions, it is good practice to include a number of extreme values in
the validation. These values can be used for updating the model, extending model
validity beyond the previously calibrated range or beyond previously covered environmental contexts (Fig. 15.3c).
However, a larger range of values and environmental contexts is not automatically better. Empirical-statistical models are context specific. Transferability beyond
the time of year, value range, species, and environmental context for which they are
calibrated cannot be assumed without a test. The power of empirical-statistical
methods lies in their ability to fit the data. Thus, empirical-statistical models should
be calibrated for the range of values and the environmental conditions that are most
A. K. Schweiger
inventory within 1 m
2
, ideally in the plot center. Likewise, it can be expected that the
chemical composition of the biomass clipped in a 1 m × 20 cm strip in the central
1 m
2
is representative for the chemical composition of the vegetation in the entire
9 m × 9 m plot, given that the clip strip captures the variation in species composition, height, and age-class distribution within the central 1 m
2
.
The accuracy and precision of the surveying equipment used for measuring plot
coordinates is another aspect to consider when determining the minimal size of
homogeneous sampling units; ideally, measurement errors should be estimated
under field conditions and added to the minimal size of the sampling unit.
Additionally, edge effects influence remotely sensed data. Sampling units should be
placed sufficiently far from landscape features, such as open soil, gravel, snow,
water, large rocks, roads, footpaths, bridges, and trees (when working in grasslands), that influence the spectral properties of adjacent areas. Depending on the
time of day of image acquisition, shadow effects from tall objects such as trees,
mountains, or buildings need to be taken into account as well.
15.2.3.2 Range of Values
For the remote sensing of biodiversity, the range of conditions (time of year, value
range, species, and environmental context) used to calibrate a model should cover
the range of conditions for which inferences or predictions should be made. For
example, extrapolating beyond the range of values relies on the assumption that the
estimated relationship holds beyond the investigated range. This cannot be assumed
without additional information, because nonlinearities (e.g., saturating curves) are
common in ecological data, especially when covering large areas and multiple environmental gradients and ecosystems.
For modeling continuous data with regression-style empirical-statistical
approaches, the sampling design should cover the expected range of values in the
area of interest with a sufficient number of evenly distributed samples. Predictions
outside the calibrated range are not reliable because deviations from the 1:1 line
between measured and predicted values (Fig. 15.2a) increase at the lower and upper
ends of the distribution (Fig. 15.2b). During model validation the entire range of
values should be covered as well, specifically paying attention to the value range
most important for the research question(s). When the tails of the distribution are of
interest for predictions, it is good practice to include a number of extreme values in
the validation. These values can be used for updating the model, extending model
validity beyond the previously calibrated range or beyond previously covered environmental contexts (Fig. 15.3c).
However, a larger range of values and environmental contexts is not automatically better. Empirical-statistical models are context specific. Transferability beyond
the time of year, value range, species, and environmental context for which they are
calibrated cannot be assumed without a test. The power of empirical-statistical
methods lies in their ability to fit the data. Thus, empirical-statistical models should
be calibrated for the range of values and the environmental conditions that are most
A. K. Schweiger