Image Processing Reference
In-Depth Information
that the hybrid approach performed slightly better (compared
to the unsupervised clustering approach), and that the super-
vised fuzzy classification was efficient in low-density urban
use areas characterized by mixed pixels.
Recently though, researchers have used more advanced
sub-pixel methods such as spectral mixture models to quan-
tify and measure the composition of urban land cover (Rashed
et al.
2001
; Small
2001
; Wu and Murray
2003
). These
“soft” approaches to classification are becoming increasingly
more popular as researchers attempt to model the hetero-
geneity that exists within urban environments. Rashed et al. (
2001
) used spectral
mixture analysis (SMA) to derive physical measures of urban land cover that
describe the morphological characteristics of the Greater Cairo region. The
authors suggest that SMA may be superior to other standard classification meth-
ods, especially when used in an urban context. They applied their model to an
Indian Remote Sensing multispectral image (IRS-IC LISS-III) of Cairo, Egypt, in
an attempt to produce a replicable procedure to analyze the anatomy of cities. The
study of Cairo was followed with a multi-temporal study of change analysis
between the years 1987 and 1996 (Rashed et al.
2005
). The authors sought to
emphasize land cover change within classes (rather than between classes) in an
attempt to model urban morphology.
Since urban environments are particularly complex, given the diversity of land
cover and land use types, Herold et al. (
2004
) suggest the use of hyperspectral
data as a way to deal with the spectral complexities of urban environments. The
authors developed and analyzed a field spectral library in the 350-400 nm spec-
tral range, consisting of approximately 4,500 individual spectra. They also evalu-
ated the most suitable wavelength for the separation of urban land cover and
related them to the spectral bands of two sensor systems, IKONOS and Landsat
ETM+ (results of this evaluation are reviewed in Chapter 4). This interesting
comparison highlights the spectral limitation of most current sensor systems -
often the optimal bands lie outside or near the edge of the sensors' spectral range.
The results of their study indicate the future potential of hyperspectral data for
urban analysis (Chapter 9).
Image processing approaches such as image texture generation and spatial pat-
tern analysis have been applied to the study of urban structure. Application of
image texture to urban environments is particularly widespread. A variety of quan-
titative measures of image texture have been developed with the aim of capturing
the underlying spatial structure of the urban scene (Brivio and Zilioli
2001
).
Barnsley et al (
2001
) modeled image texture using
standard gray level co-occurrence matrices, and pro-
vided a good review of image texture methods. High
within class variability is a common characteristic of
urban regions, particularly in urban-rural transition
areas (Pesaresi and Bianchin
2001
). In an attempt to
increase classification accuracies, texture measures
“hybrid”
approaches
refer the
methods that
combine both
supervised and
unsupervised
classifications
spatial pattern analysis
is a method for
quantifying the spatial
arrangement of image
brightness components
in an image
