Eti MEE ga ts E 
  
  
  
  
  
(d) The standard approach to classification of remotely sensed data is to 
assume that the area of study is comprised of a number of unique, inter- 
nally homogeneous classes and to apply a cluster based classification 
procedure. Within broad general urban land use classes, particularly 
residential, a continuum of cover classes exist which cannot be easily 
broken into discrete nominal classes. For example, in one area residential 
density may be low with few established trees, another may have a similar 
density but mature vegetation, while a third may be of high density with 
Tittle or no vegetation; with intermediate examples occuring between 
each of these. Urban residential response is therefore not particularly 
amenable to cluster or similar analysis, because either clusters will have 
considerable overlap or be extended linear clusters. This makes it ex- 
tremely difficult or impossible to determine the probability of a feature 
value being part of one or other classification. Similar conclusions were 
drawn by BILLINGSLEY (1972) when referring to geological applications. 
(e) The radiation reflected by a surface to a satellite sensor will be attenua- 
ted by the atmosphere as it passes through it, and an extraneous component 
of scattered radiation will be added to the transmitted component. This 
additive component is partially dependent on background reflectance which 
is relatively constant from homogeneous areas, but is not so from 
spatially heterogeneous areas. In urban areas, therefore, the degradation 
of the recorded response due to atmospheric effects will be spatially 
variable increasing the difficulties of classification. 
(f) An essential component of urban analysis is the monitoring of temporal 
change. There is thus a requirement for accurate registration of the same 
area imaged at different times. For either a single scene or a temporal 
sequence there is also a need to spatially relate ground truth areas to 
the Landsat digital image at the sub-pixel level, if the correlation 
between ground and image is to be fruitfully examined. Using 100 ground 
control points over an urban area FORSTER (1980) achieved circular stan- 
dard areas of only 30 m, (which is larger than many change elements), 
this was essentially due to limited accuracy in control point selection in 
the image. 
The low spatial resolution of the current Landsat sensors is the primary or a 
contributing factor in most of the above. If the sensor resolution was greater 
the amount of mixing of surface cover in each pixel would be reduced, contex- 
tural clues would be enhanced, control point selection would be more accurate, 
allowing for better data registration, and the relative effect of the point 
spread function would be reduced. 
There are also a number of spectral limitations to urban Landsat data, the 
primary one being number of bands. Because urban surfaces are heterogeneous 
and represent a continuum of cover classes, clustering techniques for cover 
prediction are not always appropriate. It is necessary, therefore, to predict 
the percentage of particular urban surfaces in a picture element using equations 
derived from multiple linear regression techniques, using percentage cover as 
the dependent variable and band response as the independent variable (FORSTER, 
1980, 1981). Theoretically the number of dependent percentage cover variables 
that can be predicted is one plus the number of bands (one percentage being the 
sum of the others subtracted from 100%), with four Landsat bands, five variables 
can be determined. 
In urban areas there are five main cover types, grass, trees, asphalt, concrete 
and roofs (predominantly red tile in the Sydney region) and while it is 
theoretically possible to predict the proportions per pixel of each of the five 
cover classes from the four bands, the lack of redundant data and the high 
correlation of bands 4 and 5 and 6 and 7, limits the confidence that can be 
$91 
a mw 
BES 
  
A 
- NOM x cocks d 
Ere RS SEG