Image Processing Reference
In-Depth Information
Summaries of residential and commercial point patterns using the nearest-neighbor
techniques can now be used to identify similar spatial patterns in classified imagery.
In addition, postal points can also be used to identify two types of misclassifications;
confusion between built and non-built land covers (especially bare rock or bare
soil), and confusion between residential and commercial built land uses. Postal
points, on the other hand, represent the entire distribution of individual dwelling
units within a city, and as such are the ultimate in disaggregated surfaces. They
convey valuable information on local spatial association - density and arrangement -
information that is surprisingly overlooked in research on urban image classifi-
cation, especially given the spatial nature of class distributions and the inherent
limitations of spectral data. However, and despite the area attribute, postal points
are still models of reality and as such do not precisely delineate the spatial bound-
aries of buildings. In extreme examples, especially in high density residential areas,
the area attribute frequently causes points to overlap. Nonetheless, postal points
represent the location of all individual buildings and as such are an invaluable
source of information for identifying misclassified pixels that represent not only
omitted buildings (errors of omission) but also have also erroneously included the
location of buildings (errors of commission). The identification of both types of
errors should inevitably lead to a better understanding of the reasons behind
misclassified urban pixels.
In an example taken from Belfast (Mesev
2005
), the use of postal points has
identified a number of misclassified urban pixels from IKONOS imagery. Most of
these are at the urban fringe and include the spectral similarities between pixels
representing built land cover and those representing bare rock or bare soil. In addition,
some misclassifications have been identified between pixels representing residen-
tial and commercial land use. In both instances, the identification of misclassified
pixels follows the categorization of IKONOS imagery using the ISODATA algorithm
available from the ERDAS Imagine
TM
8.6 proprietary software. Spatial masks and
iterative spectral clustering using panchromatic-sharpening of all four available
multispectral IKONOS bands (blue, green, red, and near infrared) then yielded a
reasonably accurate classified image of 88% for the built land cover and 72% and
67% for the residential and commercial land uses respectively. Ground data for
these accuracy assessments was provided by the interpretation of digital aerial
photographs collected in September 2001 by GeoInformation International in
Cambridge as part of its Cities Revealed
TM
series. The aerial photographs were
recorded at spatial resolutions of 12.5 and 25 cm, with a 15.25 cm camera focal
length, and at a height of 3,200 m. Figure
8.5
illustrates the close correspondence
between COMPAS
TM
points and the 15 cm aerial photograph of Belfast. In addition
to identifying misclassified pixels, the spatial distribution of postal points, charac-
terized in terms of density and arrangement, can be used to infer types of residential
and commercial developments identifiable in imagery classified as built.
Figure
8.6
illustrates the spatial correspondence between a subset of the classi-
fied IKONOS image of Belfast and the location of postal points from the
COMPAS™ dataset. Note that the postal points (open squares for residential and
shaded squares for commercial) have been scaled in proportion to the area of the
