Image Processing Reference
In-Depth Information
and spatial properties that are indicative of real urban
patterns. Of course, as discussed in Chapters 4, 5 and 7, the
representation of reality is heavily simplified by the inherent
limitations of the sensing instrumentation, which is further
hampered by the elevated platform of satellite sensors high
above the earth. Nevertheless, remote sensing is rapidly
gaining a crucial role in many urban applications, from
monitoring urban growth and density attenuation to models
that predict quality of life and sustainable development (see
Mesev 2003 ). The interest in urban remote sensing is further
fuelled by the availability of satellite sensor imagery at a
scale that begins to replicate that of aerial photography. It is worth noting at this
stage that although aerial photographic interpretation has revolutionized the map-
ping of urban areas, all reference to remotely sensed data in this chapter will imply
digital representations from multispectral satellite sensors. The reason is that satel-
lite sensors provide a much more flexible data source, capable of comprehensive
coverage at multispectral wavelengths, captured across multitemporal intervals, and
globally available at relatively lower costs than aerial photography.
The vast majority of contemporary research into urban classification involves the
manipulation of image pixels to represent either single
(hard/crisp) or multiple (soft) thematic classes, and to be
treated as either individual entities (spectral based classifica-
tion) or as groups and objects (spatial based classification).
This chapter will outline the main technological and meth-
odological developments in each by demonstrating modifi-
cations to the popular per-pixel spectral maximum likelihood
decision rule (Strahler 1980 ; Mesev 1998, 2001 ) and a spa-
tial method based on nearest-neighbor calculations (Mesev
2005 ; Mesev and McKenzie 2005 ). In both, strong cases
will be made for the incorporation of ancillary data from
beyond the spectral domain for improving classification
accuracy - where ancillary data are most conveniently handled by geographical
information systems (GIS). In addition, coverage will be given to the conceptual
distinctions between urban land cover and urban land use (Dobson 1993 ); a sensitive
dichotomy that underpins all urban remote sensing applications (Mesev 2003 ).
algorithms seek
to reveal the
intrinsic spectral
and spatial
properties of
imagery that are
indicative of real
patterns and
classification can
be considered in
different ways,
whether it is
spectral based
or spatial based;
hard or soft;
per-pixel or
Urban Image Classification
In simple terms, the classification of remotely sensed data is the process of generat-
ing thematic interpretations from digital signals that represent the world. In even
stricter terms, thematic image classification is little more than a statistical data
reduction procedure, generalizing from continuous to categorical, or can be visual-
ized as a conversion of data, usually from interval to nominal levels of measurement.
For example, an image with an 8-bit radiometric resolution of 256 graylevels at
Search WWH ::

Custom Search