(ANDERSON & WITMER, 1976). Level I is urban (or built up areas) and Level II 
is a further breakdown into residential, commercial, industrial etc. Only 
limited success has been achieved classifying current Landsat data into Level 
II. Some researchers have further broken the residential general class into 
subclasses, but typically these are older and new housing classes, residential 
and mixed residential or some other dichotomous division (CHRISTENSEN & 
LACHOWSKI, 1976). Other studies have suggested the use of processed Landsat 
data as a surrogate for residential quality and population density and have 
applied the concept with some encouraging results (FORSTER, 1981, LANDINI & 
McLEOD 1979, MURAI, 1974). 
One of the major difficulties encountered using current Landsat data is the 
complexity of the urban landscape which exhibits an extremely heterogeneous 
surface cover, with considerable inter-pixel and intra-pixel change occurring. 
This variation severely limits the standard clustering approach to classifi- 
cation where it is assumed that the area of study is comprised of a number of 
unique internally homogeneous classes. 
The spatial resolution of the current Landsat data is considered the primary 
limitation to more detailed classification,and this problem,and the advantages 
afforded by the new generation of satellite sensors,will be discussed in this 
paper. Some consideration will also be given to spectral resolution. Parti- 
cular reference will be made to Sydney (Australia) where the writer has 
previously undertaken an extensive study (FORSTER, 1981). 
CURRENT PROBLEMS 
There are a number of limitations to the current methods being used to analyse 
Landsat data over urban areas. These limitations are due to a number of inter- 
related causes (FORSTER, 1981, 22-23). 
(a) As the sensor becomes further removed from the scene, and resolution 
decreases, the interpreter becomes further removed from the contextural 
clues of site and association so essential in manual interpretation of 
urban areas from medium to large scale imagery, with the result that only 
cover classes and not use classes can be inferred. 
(b) Urban areas are typically heterogeneous containing many cover types, for 
example, an average residential Landsat pixel over the Sydney metropolitan 
area contained approximately 28% roof cover, 14% road cover, 5% concrete 
cover, 16% tree cover, 36% grass cover and 1% of water and soil cover. In 
such areas the radiation received from a single ground element will com- 
prise radiation from each of the surfaces which individually may have dis- 
tinct spectral signatures, giving an additive response that is not 
representative of any one class. 
(c) The point spread function (psf) of a sensor integrates the response from 
an observed (or target pixel) and its surrounding pixels. For the Landsat 
sensor the psf essentially affects an area covered by a 3 x 3 pixel array. 
The response from the central pixel accounts for only 50% (approximately) 
of the response recorded at the sensor (neglecting additive path radiance 
due to the atmosphere) with the eight surrounding neighbours accounting 
for the balance (FORSTER, 1981, 49-52). Over homogeneous agricultural 
areas this is not a problem, except at the interface of two separate 
classes, because the target pixel has the same response as its neighbours, 
however, the psf effect can significantly affect the signature from a 
single cover class if the surrounding cover is dissimilar, which occurs 
quite frequently in urban areas. 
690 
BE EE 
P2 UO] 
I-— an € 0 0À 
IT 
The 
con 
the 
tur 
all 
spr 
The 
pri 
and 
pre 
the 
der 
the 
198í 
tha 
sum 
can 
In i 
and 
the: 
COV( 
cor