132
Chapter 14
classification, examples of practical applications are so far almost nonexistent.
8.
CLASSIFICATIONS AND TERMINOLOGY
Given a set of objective procedures for classifying, or otherwise
segmenting remotely sensed data, there is no reason why a classification
scheme could not be applied to two or more radiometrically-corrected image
sets, with entirely reproducible results. The difficulty arises when the
remotely sensed data products come to be interpreted or compared with
external data sets. Remotely sensed data do not record land cover directly;
rather, land cover is inferred from observations of multi-spectral radiance or
SAR back-scatter. Objective spectral classes are then interpreted in terms of
land cover types that are often far from objective or reproducible. Wyatt et
al. (1994) clearly demonstrate the impacts of classification differences on
areal estimates of land cover in a variety of contemporary surveys, and hence
the need for greater objectivity and consistency in defining the land cover
classes that are to be mapped from remote sensing.
ALGORITHMS FOR QUANTIFYING CHANGE
Remarkably little attention has been given to the choice of algorithms for
quantifying changes observed from remotely sensed data. Whatever the
reason for this, the inference in much of the literature is that, once change
has been detected, the subsequent steps needed for quantitative analysis are
self-evident. The commonest approach is to measure differences between
data sets (Green, 1984; Rees and Williams, 1997). One approach is to
explore differences between the original multi-spectral data (which must first
be radiometrically corrected); alternatively, change analysis may be based on
differences between processed products, such as classified outputs.
There has been extensive use of Principal Components Analysis (e.g.,
Byrne, Crapper and Mayo, 1980; Fung and LeDrew, 1987) but it is not easy
to ensure that the results from this method are reproducible. Sader et al.
(1991) estimated change by use of ground reference data collected within a
statistical sampling framework designed for the purpose. Lambin and
Strahler, (1994a and b) successfully deployed a technique known as ‘change
vector analysis’.
9.
Chapter 14
classification, examples of practical applications are so far almost nonexistent.
8.
CLASSIFICATIONS AND TERMINOLOGY
Given a set of objective procedures for classifying, or otherwise
segmenting remotely sensed data, there is no reason why a classification
scheme could not be applied to two or more radiometrically-corrected image
sets, with entirely reproducible results. The difficulty arises when the
remotely sensed data products come to be interpreted or compared with
external data sets. Remotely sensed data do not record land cover directly;
rather, land cover is inferred from observations of multi-spectral radiance or
SAR back-scatter. Objective spectral classes are then interpreted in terms of
land cover types that are often far from objective or reproducible. Wyatt et
al. (1994) clearly demonstrate the impacts of classification differences on
areal estimates of land cover in a variety of contemporary surveys, and hence
the need for greater objectivity and consistency in defining the land cover
classes that are to be mapped from remote sensing.
ALGORITHMS FOR QUANTIFYING CHANGE
Remarkably little attention has been given to the choice of algorithms for
quantifying changes observed from remotely sensed data. Whatever the
reason for this, the inference in much of the literature is that, once change
has been detected, the subsequent steps needed for quantitative analysis are
self-evident. The commonest approach is to measure differences between
data sets (Green, 1984; Rees and Williams, 1997). One approach is to
explore differences between the original multi-spectral data (which must first
be radiometrically corrected); alternatively, change analysis may be based on
differences between processed products, such as classified outputs.
There has been extensive use of Principal Components Analysis (e.g.,
Byrne, Crapper and Mayo, 1980; Fung and LeDrew, 1987) but it is not easy
to ensure that the results from this method are reproducible. Sader et al.
(1991) estimated change by use of ground reference data collected within a
statistical sampling framework designed for the purpose. Lambin and
Strahler, (1994a and b) successfully deployed a technique known as ‘change
vector analysis’.
9.