Image Processing Reference
In-Depth Information
8.5
Concluding Thoughts Regarding Classifications
The classification of remotely sensed data is a highly subjective process. Converting
radiometric values to user-specified thematic categories requires a level of interpre-
tation that forgoes objective multivariate measurements of reflected and emitted
energy for the sake of semantic expediency. The translation is even more tenuous
when urban areas are classified from satellite sensor imag-
ery, during both the initial land cover interpretation stage
and at the inference of land use from land cover. Spatial
variability and spectral heterogeneity, inherent in all images
of urban morphologies, severely hamper accurate intra-
urban classifications, tempered of course by the scale of the
image data and the scope of the application. It is therefore
unsurprising that a variety of classification and pattern rec-
ognition systems have been developed with urban areas
particularly in mind. This chapter has reviewed some of these techniques, focusing
on the contrast between hard and soft spectral classifications and those oriented on
the spatial configuration of pixel groups.
Per-pixel spectral techniques, like the Bayes' modification of the ML, allocate
one “likely” class to one pixel; whereas soft classifiers, such as mixture models and
fuzzy sets, assign proportions of one or more classes per pixel. As pixels represent-
ing urban areas are virtually all composed of mixed surfaces, the difference between
the two types of classifiers is a matter of scale, generalization, and classification
scheme. The development of soft classifiers was widely heralded as a breakthrough
in techniques that at last began to mirror the continuum of reality. However, given
the severe spatial heterogeneity in the arrangement of urban structures, there is a
case to move away from detail (from soft classifiers) in favor of more aggregated
or “averaged” pixels (from hard classifiers). In other words, the breakdown of
mixed pixels may produce too much information that cannot be easily categorized
within a scheme: a situation of “not seeing the wood for the trees”. This is particu-
larly evident from new high spatial resolution imagery. Moreover, hard classes,
such as the ones generated by the ML illustrated in this chapter, are semantically
more communicable where the three urban classes of residential density, high,
medium, and low, are directly representative of variations in proportions of build-
ings to vegetation ratios. For instance, pixels representing mostly, if not all, build-
ings are classified as high residential density, and pixels representing sizeable
proportions of vegetation in conjunction with buildings as low density residential.
The second type of urban classification examined the role of postal point data as
major conceptual advances in the de-construction of the geography of urban areas;
or at least as an alternative to traditional aggregated census tracts. In disaggregating
the geographical distributions of individual households these point data provide
unique opportunities to view the urban landscape as a surface of discrete entities
rather than the traditional and administratively convenient patchwork of aggregated
and uniform surfaces partitioned only by artificial zones. Disaggregated surfaces
classification of
remotely sensed
data requires a
level of
interpretation
that forgoes
objectivity
