Figure 5. Scattergram of radar (SIR-B) incident angle
43° versus 36°. Mean response values of Forest (F),
mature residential (R), soil {S), intense urban (I)
(commercial and industrial), and cardinally aligned
residential (C) are shown with an ellipse represent
ing two standard deviations.
Figure 6. Scattergram of Landsat Band 7 versus radar
(SIR-B) incident angle 36°. Mean response values of
Forest (F), mature residential (R), soil ,(S) , intense
urban (I) (commercial and industrial), and cardinally
aligned residential (C) are shown with an ellipse re
presenting two standard deviations.
the response from different incident angle radar
could discriminate between urban classes due to the
differing combinations of specular, diffuse and
corner reflector back scattering. Possible variations
in backscatter with incident angle, are illustrated
in figure 4 for various idealised urban surfaces.
The scattergram of incident angle 36° versus 43° is
shown in figure 5, and Landsat 7 versus 36° in figure
6, for cleared and heavily urbanised surfaces.
Similar difficulties with Landsat data have occurred
when attempting to discriminate older residential
areas, that typically are surrounded and overhung by
mature vegetation, from low density forest which has
a predominantly vegetation signature, modified by soil
and understorey, in the visible and the near infrared.
Because the SIR-B wavelength of 23.5 cm (L-Band) has
a reasonable penetrating capacity it was expected that
an increased response from the buildings underlying
the trees would result, allowing the separation of
these confused cover classes. Examples of their res
ponse from Landsat and SIR-B are also shown in the
scattergrams, figures 3, 5 and 6. Note that units
used in the scattergrams are count values on a 0 - 255
scale as recorded.
4. DISCUSSION OF RESULTS
From the scattergrams (figures 3, 5 and 6) it can be
seen that the radar response, for both incident angles,
has a much greater range of values than the equivalent
Landsat response, and the spread of values for indivi
dual cover types is again much greater for radar. The
greater overall range of radar values is due to the
increased number of surfaces that act in a specular
manner. Whilst the majority of surfaces in a Landsat
urban scene respond in a near diffuse manner, for
high sun angles, surfaces of roughness variation less
than approximately 3 cm, for L-Band radar, will cause
specular reflection. Such surfaces can include
grass, concrete, bitumen and buildings. Thus, depend
ing upon the relative angle between the incident
radiation and the surface, either no response or a
saturated response can result. This extreme dependence
on the alignment of the surfaces also holds true for
individual surface classes, where varying roof and
building facets, and orientation of leaves and branches
can cause considerable variation in radar backscatter
from a surface that appears essentially homogeneous
to Landsat.
Examining the plot of Landsat Bands 5 and 7 as shown
in figure 3, it can be seen that forest and mature
residential surface classes overlap and their mean
response values are only marginally separated by
approximately 10 count values in Band 5, making it
difficult to define a decision surface between them
for classification purposes. However in figures 5 and
6, while there is still some overlap, the mean values
have a greater separation. The introduction of radar
response has thus aided class discrimination. The
situation with soils and the intense urban class of
commercial and industrial land use, is not so clear.
Whilst they are relatively well separated in the two
dimensional space of Landsat Bands 5 and 7 (figure 3)
this may not be the case when darker soils are invol
ved because the Band 7 on Band 5 ratios of these sur
faces response are very similar, and thus a darker
soil response would move closer to, and possibly
overlap, the intense urban response cluster. This is
not the case in figure 6 where the lowered radar
response of soils, due to its specular reflection
away from the receiving antenna, and the high radar
response of the intense urban surfaces allows for
easier class separation with little possibility of
overlap due to their significantly different ratios.
A further surface 'cardinal residential' is also
shown in figures 5 and 6. This represents tree
covered residential areas where major street patterns
are aligned at right angles to the incident radar
radiation, resulting in the so called 'cardinal
effect' discussed early. Because of the significantly
higher backscatter from these areas, the separation
between forest and residential is even more pronoun
ced, even though the spread of values is considerably
greater than for 'non-cardinal' residential. A
similar effect would result from 'cardinally' aligned
intense urban (industrial and commercial) surfaces,
resulting in greater separation from soils, but
examples of these were not available in the study
area.
It is clear from the scattergrans that the great
est, difficulty in using radar backscatter in the
classification of urban surfaces is the wide range
of values displayed by each class. This internal
class 'noise' could be reduced by the use of the
mean value determined over say a three by th-atee
neighbourhood. The resultant reduction in variance
would significantly improve classification accuracy
and have the added advantage of being/more spatially
compatible with the Landsat data, £ut with the
resultant disadvantage of loss of spatial detail for/
interpretation.
5. CONCLUDING REMARKS
While the results from this study are limited at