Image Processing Reference
In-Depth Information
launches of satellites with sensors capable of super fine spatial resolutions
of 4 m and less (for example, IKONOS and QuickBird) may have improved
visual clarity of urban features but at the same time have also compounded
classification ambiguity (Welch
1982
; Corbley
1996
). The reason is that
because of the severe spatial heterogeneity in the composition of urban
surfaces, pixels at fine spatial resolution will represent variable mixtures of
similar urban objects. Any divergence in the proportion of reflected energy
representing urban surfaces (no matter how small) will result in dissimilar
pixel values, which in turn, will increase the range of thematic classes. Even
adjacent pixels may represent slightly different spectral signatures. This is
known as spectral “noise” but is less of a problem for sensors of coarser
spatial resolution as variations in urban reflectance are averaged to more similar
pixel values and hence a narrower range of possible thematic classes. The
use of finer spatial resolution data is further complicated by the degradation
of the radiometric resolution. Increasing the spatial resolution capabilities of
a sensor essentially narrows the instantaneous field of view (IFOV), which
means less energy is collected by the sensor in a shorter period of time (Fisher
1997
). In both fine and coarse spatial resolutions the standard classification
assumes that only one thematic class is to be allocated to each individual pixel.
This perpetuates the aggregated nature of the pixel; where it is conceivable
that two pixels of identical multispectral values may represent quite variable
combinations of urban land cover properties. In response, rather than
deterministically allocating each pixel one exclusive class ('hard' or 'crisp'
classification) some techniques employ stochastic mechanisms capable of
estimating multiple class memberships within individual pixels ('soft' or
'fuzzy' classification). Hard, or crisp, classification assumes each pixel is
assigned one, and only one, thematic class. Soft classification, on the other
hand, lifts the restriction and allows each output pixel to be labeled with
multiple classes. In a sense, the principles of soft classification are more in
tune with the continuous nature of the earth's surface, especially land defined
as urban. However, from a pragmatic standpoint, single classes produce simpler
maps and are still deemed to be more aesthetically communicable, especially
to users with little or no knowledge of remote sensing.
8.2.2
Spectral Classifications
As noted already accurate classifications of remotely sensed data representing
urban areas are difficult, and are, on the whole, dependent on the scale of the image
data and the scope of the application. Even a coarse classification of built-up land
cover will frequently fail the 85% minimum level of accuracy set out by the
Anderson criteria (Anderson et al.
1976
). The spatial configuration of urban areas
