Geoscience Reference
In-Depth Information
Table 16.1 The Average Relative Error for Eight
Different Landscape Metrics, Four
Based on Identifying Landscape
Patches and Four Based on Description
of Boundaries in a Continuous Image
Metric
Average Relative Error
Percentage forest
0.23
Edge density
0.35
No. of patches
0.75
Mean patch size
1.52
No. of boundary elements
0.02
No. of subgraphs
0.11
No. of singletons
0.24
Max subgraph length
0.40
Comparing the patch- vs. boundary-based metrics indicated that the majority of boundary
metrics had greater precision than the patch-based statistics (Table 16.1). This can best be explained
by the way in which changes in
precision were affected by the procedures used to calculate the
metric values. All of the patch-based metrics involved an image classification step, and two of them
added a patch identification step. Both of these steps are sensitive to spatial variations in image
quality and to the specific procedures used. Because the boundary-based metrics were calculated
directly from the NDVI images, there was less opportunity for propagation of the spatial pattern
of error. Further, the boundary-based metrics used only local information to characterize pattern,
but the patch-based metrics used global information (i.e., spectral signatures from throughout the
image). This use of global information introduced more opportunities for error in metric calculation.
Additionally, we evaluated the effects of various processing choices on the precision of metrics
(Brown et al., 2000a). The results of this work suggest that haze in the images and differences in
seasonal timing were important determinants of metric variability. Specifically, less precision
resulted from hazier images and image pairs that were separated by more Julian days, irrespective
of the year. Also, summarizing landscape metrics over larger areas (i.e., using larger landscape
partitions) increased the precision of the estimates, although it reduced the spatial resolution.
Further, postclassification processing, such as sieving and filtering, did not consistently increase
the precision, and can actually reduce the precision.
The obvious cost associated with obtaining precise estimates through the empirical approach
of redundant mapping is that the areas need to be mapped twice. However, the costs may be lower
than the costs of obtaining reference data for accuracy assessment, and redundant mapping can
provide reasonable estimates of precision in a pattern analysis context, where comparison with a
reference data set is much more problematic. Guindon et al. (2003) used a similar approach to
dealing with the precision of LC maps.
16.4.2
Comparing Class Definitions
16.4.2.1
Comparing TM Classifications
Across all landscape metrics tested, our forest cover classification of the Huron River watershed
suggested that the landscape was much less fragmented than did the NLCD forest class (i.e., that
there was more forest, in fewer but larger patches, with less forested/nonforested edge and more
core area) (Table 16.2). Comparisons of forested cells indicated that forest cover occurred in several
of the nonforest NLCD classes. The definitions of NLCD classes allowed for substantial amounts
of forest cover in nonforest classes. For example, in the low-density residential class “vegetation”
could account for 20 to 70% of the cover (USGS, 2001). Also, the NLCD forest classes were not
100% forested. Although 65% of the forested cover in the region (by our definition) was contained
