ne.
%
mean
2.9
1.2
8.7
2.6
0
0
1.4
1.6
0.2
0.8
1.2
3.1
57.6
3.1
0.9
2
27.1
1.9
100
irvey methods
Ificient and
s reflecting
It has been
ibson (1980)
substantial
round based
mes as long,
lo not locate
as
nstrated the
sing aerial
of specific
nain problems
ncils obtain
s only. in
lard (1983)
l this, ’The
LI percentage
but like any
appreciated
hanges taking
iitions'. It
ty of local
nd therefore
a needs to be
is readily
al is suited
es. The use
umented for
ses, however
high spatial
such as SPOT
•an studies.
with the new
is a brief
ability for
:ive urban
used in the
with limited
it studies in
ate on those
on, such as
have realised
rban land use
ages over MSS
at may figure
lysis. Anuta
me clustering
exhibits 42
from MSS.
- increased
to MSS is
nponent (PC)
hi (1982 and
PC images in
facilitates
detailed examination of urban structures.
Additional dimensionality over MSS and the
ability of the third PC of 4 band data for
detection/discrimination of built up
areas/bare soil has been noted by Sadowski
(1983) .
Increased spatial and spectral resolution
do not necessarily mean that this data
provides an uncomplicated answer to urban
monitoring problems. The nature of urban
cover must be considered, Owe et al (1984)
state that heterogenous urban features often
cause classification error. The reflectance
from mature trees, which are often higher
than residential units, has been found by
Baumann (1979) to influence classification
accuracy. Other investigators, Bryant (1971)
and Forster (1981 and 1982) have highlighted
the problems presented by abrupt changes of
urban land use over short distances. These
take the form of considerable inter and intra
pixel differences, which seriously alter the
reflection from one cover class that is
surrounded by classes giving dissimilar
readings. The difficulties associated with
such conditions are summarised by Clark
(1979) who states 'Physical or spectral
conditions of a land use do not always divide
as sharply as the cultural definitions of the
land use' .
Boundary definition is very important in
the urban area, however sharp contrasts are
infrequently seen. Instead pixels lie along
boundary lines and introduce cover mixing
effects. Merickel et al (1984) consider
that up to 60% of the pixels in some Landsat
scenes are mixed, whilst Owe et al (1984) say
that with the TM, more pixels per unit area
do not lessen the chance of boundary features
being crossed. A programme, 'CASCADE',
introduced by Merickel et al (1984),
specifically to combat the mixed pixel
effect, assigns pixels to homogenous regions
in a neighbourhood after judging that region
responsible for the mixing effect.
A further influence on category
discrimination, particularly in heterogenous
regions, is the sensor Point Spread Function
(PSF) . Acting over a 3x3 pixel area, this
seriously affects the signature from cover
classes with dissimilar neighbours.
According to Forshaw (1983) a better
representation of resolution would be a
deconvolved PSF rather than a pixel only
estimate. Forshaw considers that as a term,
'spatial resolution' is 'poorly defined and
improperly used' and is artificially high in
order to compensate for the rapid data
sampling rates of modern satellite systems.
The views on the increased spatial resolution
of sensors are as mixed as those on the
improved spectral range. Irons (1984)
suggests that a 'stalemate' is evident, where
increasing resolution does not affect
accuracy because; a) category spectral
variability increase hinders classification;
and b) a decrease in mixed pixels (by up to
24%) enhances classification. A 'Point of
diminishing return is reached and 30m IFOV
should be the best for multispectral
classification' is the stated view of Clark
(1979) in an analysis of multi resolution TM
Simulation data in an urban environment.
Clark also recognised that as resolution
decreased, classification accuracy actually
increased, probably due to the heterogenous
nature of urban sites being 'smoothed' out.
Finally, to summarize Forster (1982),
higher resolutions will, in urban areas;
1. Reduce mixed pixels
2. Aid contextual identification
3. Aid in registration
4. Reduce the PSF effect
5. Higher data redundancy will allow more
accurate judgement of surface percentages
6. Higher pixel homogenity will aid in
clustering procedures
7. Texture studies may be implemented
The increase in information content of TM
data has proved problematic, not least in
terms of data handling, but also regarding
classification techniques. PC and canonical
analysis have been used as data reduction and
feature extraction techniques by a number of
workers, Brumfield (1981), Jackson (1984) and
Sadowski (1983), non traditional methods such
as canonical analysis can result in upwards
of a 20% improvement in classification
accuracy. The need for new classification
algorithms for TM imagery has been recognised
by Irons (1984), and classification schemes
currently being developed at the Natural
Environment Research Council by Jackson et al
(1984) exploit per-pixel, textural and
contextual algorithms. Wang (1984) considers
that such techniques need careful development
because the 'Averaging process smooths out a
certain amount of the data's unique
qualities', while Forshaw (1983) implies that
resampling may limit resolution to 2 times
the pixel size. It is evident that although
significant classification accuracy can be
obtained with TM data, better results can be
expected and a hiatus exists with current
methods unable to realise the full potential
of TM data.
Much of the work being done on TM imagery
and urban environments originated from the
USA, which features a different urban make-up
to the UK. Jackson et al (1984) define the
problem with a statement on the quality of
the first TM images, 'in rural areas, there
is a very significant improvement over the
MSS, whereas in urban areas the improvement
is much less marked, probably as a result of
the high density of English urban
development’. For this reason the
possibility of incorporating TM data into a
GIS is currently being investigated with
regard to derelict land. Although
cartographic fidelity does not appear to be a
problem with TM data (Welch et al 1984), data
transformation does alter pixel values,
damaging their essential qualities.
The difficulties in combining digital data
with different sorts of ancillary information
are well known, and to quote Brooner (1982),
'A new generation of information systems
whose design bridges both cell and polygon
inputs, characteristics of 'conventional'
systems is needed'. The problem is
compounded by the quality of present and
proposed satellite data, which should not be
excluded from a comprehensive GIS.
Clark (1979) suggests 30m resolution as
the optimum for urban studies, Forshaw (1983)
considers that 'resolutions rather better
than 10m will be necessary for consistently
high recognition accuracies', and Jackson et
al (1984), in recognition of the dense nature
of English development states 20m as a
minimum resolution requirement. Forster's
(1982) theory is that TM should be used to
determine surface types, while SPOT
panchromatic data, 10m resolution, could
provide high resolution cartographic and
contextual information.
It is the authors opinion that while these
suggestions are valid, it is essential to
include data from aerial photography in a GIS
to facilitate the provision of qualitative