Environmental Engineering Reference
In-Depth Information
10
7
, a difference that was also significant at
3.19
0.0001. The conclusion is
that the loss of forest edge was greater than expected from increases in urban areas
and decreases in forest alone.
a ¼
15.5 Discussion
Measuring the rate and extent of land cover change, and determining the causes of
these changes continues to be a significant challenge for landscape ecology.
Although extensive data sets from satellite sources are now widely available
(Vogelmann et al. 1998, 2001), issues of accuracy and changing technologies
continues to complicate the detection of change through time (Homer et al. 2004;
Wickham et al. 2004). This chapter addresses a second issue associated with change
detection; that of making scientifically rigorous comparisons from these pattern-
rich data sets. The argument made here is that the appropriate approach should be
parsimonious, achieved by a careful match of specific questions with a minimum
set of relevant measures of landscape change. Analysis based on multiple metrics
results in multiple comparisons and consequently an uncontrolled statistical error
associated with inferences drawn from these comparisons. The familiar
level, or
the probability of accepting a false hypothesis is usually defined for a single
comparison as
a
0.05, meaning that there will be a 1 in 20 chance of accepting
a false hypothesis. With each additional comparison,
C
, the
a ¼
-level or more
a
e
(Kirk 1968) increases as [1.0 -
appropriately, the “experiment-wise error rate”,
a
)
C
]. This means that if
(1.0 -
for the comparison-wise error rate is fixed at 0.05
and one makes four comparisons, then
a
a
e
0.1855, a nearly 1 in 5 chance of
accepting a false hypothesis. A further complication is the lack of independence
among metrics (as illustrated in Fig.
15.4
) that complicates conclusions drawn from
the examination of a suite of landscape metrics.
These issues are not new and have been discussed within the landscape litera-
ture (e.g. Dale et al. 2002; Gardner and Urban 2007; Wagner and Fortin 2005).
The chief impediment for adopting a parsimonious approach that avoids this
problem seems to be the three-pronged issue of determining the specific question,
selecting the appropriate metric and performing the relevant statistical test. These
steps may all be performed using Qrule, but have been awkward in the past
because the user must be familiar with the data sets generated by Qrule. A series
of programmes written in R (Appendix) have been provided here to illustrate and
facilitate this process.
An example using data from the piedmont of Maryland has been presented to
illustrate the process of hypothesis testing. The question(s) of concern involved
the possible effects that increases in urban areas (over an 11-year period) had on
the simultaneous loss of forested habitat. A comparison of the two data sets had
shown that urban areas had increased and forest areas declined over this time
period. Three specific questions were tested: the first question was, “Could the
pattern of change be explained by two simple random processes?” Or more
a
¼
