Image Processing Reference
In-Depth Information
evidence to suggest that high density housing is underestimated (less pixels classi-
fied), and conversely that low density housing is over-estimated (more pixels clas-
sified). Further accuracy assessments of other settlements, including one with a
detailed accuracy evaluation, can be found in Mesev (
1998
), with all indicating
slight to moderate improvements.
Accuracy from ML per-pixel classifiers is generally increased if the assumption
of multi-normality is sustained. However, accuracy is dramatically impaired when
spectral distributions representing cover types are far from normal, as in the case of
complex heterogeneous urban environments. Moreover, parametric classifiers fre-
quent fail to preserve areal estimates; the ability to classify classes with relation to
known areal properties. This is particularly noticeable when the distribution of
cover types in a scene is highly variable. In attempts to overcome some of the prob-
lems, research has led to the development of entirely non-parametric methods using
class probabilities from the gray-level frequency histograms to alleviate areal mis-
estimation as well as hybrid methods which preserve the benefits of both the para-
metric and non-parametric approaches.
8.4
Nearest Neighbor Spatial Pattern Recognition
In contrast to the spectral-based ML decision rule the nearest neighbor spatial pat-
tern recognition system is a much more recent development (Mesev
2003, 2005
).
The system is designed to utilize new disaggregated point-based GIS data for repre-
senting individual buildings. Although point based data can be considered the ulti-
mate in disaggregation, capable of representing precise locations, in Euclidean
geometry terms they are dimensionless, and as such are incapable of measuring the
size and shape of objects they represent. Nonetheless, the potential exists for analyz-
ing localized two-dimensional spatial patterns of groups of point data for character-
izing various morphologies of both residential and commercial developments.
Examples of point-based databases include two in the United Kingdom that were
designed to represent the location and type of every postal delivery address. One
dataset is known as ADDRESS-POINT™ and is created by the Ordnance Survey of
Great Britain, the other as COMPAS™ (COMputerised Point Address Service) in
Northern Ireland. Planimetric coordinates of the point data representing postal delivery
buildings are claimed to be precise to within 0.1 m (50 m in some rural areas) of the
actual location of the building. To achieve this standard of precision both databases
were created primarily using the Royal Mail's Postcode Address File (PAF) along with
routine ground survey measurements, which are updated every 3 months. Information
in both is identical other than an additional “area” attribute in COMPAS
TM
, which
indicates the area of the point representing the location of a building. This areal
feature is a regular approximation of the actual two-dimensional size but not shape
of the building. It is worth noting that postal points are unconstrained, which means
that if the area attribute is inappropriately activated it is not inconceivable for build-
ings to be represented as overlapping. Nevertheless, the distribution of postal points,
including area attributes, are an invaluable source for calculating measures such as
