Image Processing Reference
In-Depth Information
is intrinsically complex and typically unpredictable, regardless
of the spatial resolution of the data. Logically, any spectral
classification, either hard or soft, will generally suffer lower
accuracy as a result of mixed pixels (mixels). Inherent spatial
variability of urban land cover and acute spectral heterogeneity
between pixel values has typically led to lower interpretation
accuracy (Baraldi and Parmiggiani
1990
). To increase accu-
racy, methodologies have attempted to incorporate ancillary
information (Mesev
1997
), usually during the classification
process (for example, fuzzy sets, neural nets (Dreyer
1993
), and Bayesian modi-
fications (Mesev
1998, 2001
)). Other approaches have focused on developing the
three-dimensional perspective of urban remote sensing by using, for example,
LIDAR data (Laser-induced Detection and Ranging) (Barnsley et al.
2003
), inter-
ferometric SAR (Synthetic Aperture Radar) data (Grey et al.
2003
) and hyper-
spectral image interpretation (see Chapter 9). The creation of soft classification
algorithms, where pixels are allowed to represent more than one class, was a direct
response to the inflexibility of hard classifications to adequately represent the contin-
uum of reality (Ji and Jensen
1999
). Fuzzy set theory, multiple endmember spectral
mixture models (Rashed et al.
2003
) and decision rules based on artificial neural
networks were employed to fit spectral values using complex recursive mathemati-
cal relationships. Output is multivariate, and as such is closer to a more representa-
tive interpretation of the continuum of the natural landscape than output from hard
classifications. However, soft classified images are also more difficult to understand
and their multiple class schemes are less comparable with GIS data.
the “targeted”
level of image
classification
depends on the
scale of the
image data and
the scope of the
application
8.2.3
Spatial Classifications
Given the limitations of spectral-based classifications, there
has been a recent re-focus on techniques that are more
responsive to the inherent complex spatial configuration of
urban areas. Spatial classifications can be loosely separated
into those that focus on traditional tonal contrasts by mea-
suring grayscale variations of adjacent pixels, and those that
examine the broader inter-relationships between pre-defined
groups of pixels representing complete objects. Research
on the spatial characterization of urban land cover has
gained momentum in recent years bringing about the devel-
opment of a number of quantitative indices. Amongst the
most widely applied have been the scale invariant properties
of fractal geometry (see relevant discussion in Chapter 12),
which many proponents have argued are capable of measur-
ing the structural complexity and fragmentation of highly
heterogeneous urban land cover (Batty and Longley
1994
;
spatial
classifications are
separated into
those that focus
on traditional
tonal contrasts
among pixels,
those that examine
the broader
inter-relationships
between
pre-defined
groups of pixels
representing
distinct objects
