×

You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

Title
Remote sensing for resources development and environmental management
Author
Damen, M. C. J.

518
differences between the Landsat MSS and SIR-B radar obviously
complicated their co-registration.
Once the Landsat and SIR-B were successfully co-registered,
maximum likelihood classification of the Landsat and SIR-B images
was undertaken. Landsat and SIR-B images were also separately
classified, as were the first three principal components of a
principal components transformation of the combined Landsat and
SIR-B images.
2. Methods
All processing of the SIR-B and Landsat data was performed on a
Dipix Aries II image analysis system at the Centre for Remote
Sensing, University of New South Wales.
A static average filter was chosen to reduce speckle. It
consists of a grid cell moving over the raw radar image, the size ol
which is user specified (in our case a 2 x 2 cell). The algorithm
calculates the average value for all pixels within the grid cell, and
-this value then becomes the new grid cell value, replacing all other
pixel values. This filter effectively reduced the 12.5 m pixel size tc
25 m, as well as lessening the speckle effect.
Co-registration of the two images was multistaged. The SIR-B
image was rotated clockwise through 300 degrees to align it with
the Landsat image. Next, the SIR-b image was registered to the
Landsat image by a first order polynomial using 11 well distributed
tie-points. Cubic convolution interpolation resampled the SIR-B
image to the same resolution as the Landsat. Corresponding ground
control points on the Landsat band 7 image and the basemap (a
composite 1 -.100,000 topographic mapsheet with an Australian Map
Grid (AMG) at 1 km intervals) were located and used to calculate a
second order polynomial with acceptable residual errors. Geometric
correction of both the Landsat and SIR-B images was performed with
this single polynomial function. Cubic convolution interpolation was
again used to resample the Landsat and Sir-B pixel size to 50 m
intervals.
Three separate maximum likelihood classification strategies
were attempted on the full five feature combined Landsat and SlR-B
data set; on the first three principal components of the full five
feature data set; on the four feature Landsat MSS data; and on the
SIR-B data alone. A representative subscene encompassing all the
major cover types (viz. Box forest, River Red Gum forests of site
qualities 1,2 and 3, swamp, water and agriculture) was selected to
/educe the classification time.
The classified maps produced were quantitatively assessed fot
accuracy using a modified stratified random sample technique
proposed by Kalensky and Sherk (1975). This strategy provides an
estimate of mapping accuracy or the image (i.e. it accounts for
positional accuracy) as well as the more frequently quoted
classification accuracy. The sampling intensity was estimated using
the procedure developed by Van Genderen et al. (1979), where taking
30 randomly located sample units per strata gives a classification
accuracy estimate within 95 per cent confidence limits.
3. Results and Discussion
The static average filter used to reduce radar speckle did not
degrade edges and boundaries to the same extent as a mean filter,
which was also run on the data. The other advantage of the static
average filter was that spatial resolution was reduced in one
operation without having to interpolate using cubic convolution or
nearest-neighbour and attract the respective disadvantages of these
techniques Tviz. an overly smoothed image or a reduction in spatial
accuracy). Figure 2 shows the radar image after speckle reduction
and qeometric correction.
Co-registration of SIR-B radar and Landsat MSS proved
acceptable, with residual errors of less than one third of a pixel.
Visual comparison of the Landsat and SIR-B images, and the
base-map indicated that the registration had been very successful!.
Figure 2 - Radar image after speckle reduction
The Landsat MSS and SIR-B images displayed many prominent
features which could be readily ground truthea to known surface
features. The Landsat colour composite showed differences in
vegetation vigour within the forest, with the Red Gum site quality 1
ana site quality 2 stands (which had recently been flooded) appearinc
much "redder. As the vegetation became drier (i.e. vegetation on
higher ground) the tone became less dense and the colour lighter.
Water and swamp areas on both the Landsat and radar appeared dark.
Inspection of the SIR-B image, and investigation of bispectrai
plots of cover class pixel dumps for the SIR-B and Landsat images,
the following qualitative assessment of the range of radar
brightness value responses for the cover classes was developed
(Table 3).
Table 3 - Range of radar brightness values for the cover classes
Bright ¡«Intermediate >Dark
Forest
or cover
type
Box
Red Gum SQ1
Red Gum SQ2
Red Gum SQ3
Swamp
Agriculture
Water
Water, swamp and agriculture appeared dark as they acted as
specular reflectors away from the satellite's receiving antennae. The
forest complex is essentially uneven-aged, and so the density of the
different forest types generally reflect the site quality (Forestry
Commision of N.S.W.). in general, the denser the forest type (i.e. the
higher the basal area or stocking), the brighter the radar response.
This observation agrees with hoekman (1985), who found that
backscatter from XT>and SLAR increased for increasing age (or
density) for spruce plantations in Holland, (during the needle
forming period tor the tree). Wu (1984) also found that for L-Band
SIR-A data, radar return strength is highly correlated with tree
height or age (and hence density) for three types of pine forest in
South Carolina, U.S..
Wu (1984) also found that X-band SAR had higher backscatter
for cypress forests over standing water. This phenomena may be
occuring over the Riverina forests, which were subject to an 80%
flood at the time of the SIR-B overflight. This aspect is to be
discussed in a forthcoming paper by the authors.
Supervised classification yielded the overall classification
results shown in Table 4. Note that as the SIR-B data classification
result was so poor, a detailed estimate of mapping accuracy was nol
attempted for that data source alone.
Table 4 - Overall classification results for the supervised
classification
Data type Overall classification accuracy
Landsat and SIR-B combined 65.1%
Landsat only 602%
First 3 principal components 56.4%
Landsat combined with SfR-B gave the best overall result.
SIR-B obviously provided additional information which increased
classification accuracy, but the amount of contributed information
is poor in comparison to that obtained from the MSS. The Landsat and
SIR-B in combination also gave the best individual class mapping
accuracies, and the confusion table (after Kalensky and Sherk, 1975)
is shown in Table 5.
~Nb other workers have attempted to map site quality classes
for uneven-aged forests using remotely sensed data, so these result!
are not comparable with previous work. However, Benning et al.
(1981) had limited success at classifying exotic forest types into
age classes using Landsat data alone in New Zealand., with
classification accuracies ranging from 16% to 58%; a poor result
compared with that stated here. Inkster et al. (1980) used a four
channel SAR over forested sites in Canada and estimated the
resolution required for accurate forest mapping was 6 m. Guidon et
al. (1980) used Landsat MSS data, an 11 channel airborne MSS and a
four channel airborne scanner over rugged forest terrain, and showed
higher classification accuracies were possible with the MSS imagery
than with SAR imagery, which agrees with our results. The airborne
MSS gave much better results than the Landsat MSS, due to the
superior resolution and spectral range of this scanner.
Supervised classification of the first three principal
components proved less accurate overall than the Landsat MSS data
alone. Inspection of the eigenvector-eigenvalue matrix produced by
the principal component analysis showed that 84.2% of the total
Variance in the first three principal components was due to the radar
The radar, which dominated the first three components, generally
'shows only minor discrimination of the forest.