av-
ing
s of
evi-
the
e of
lard
lues
the
rent
le 1
Table 2: The normalised standard deviation of
the element of difference matrix
0.075 0.055 0.069 0.066 0.077
0.057 0.040 0.064 0.054 0.066
0.056 0.040 0.073 0.046 0.060
0.071 0.050 0.062 0.046 0.062
0.094 0.067 0.063 0.055 0.075
is the averaged difference matrix, it can be con-
sidered as the difference matrix for any segments
with little loss of accuracy.
Based on the above analysis, in the case of single-
channel images, only the mean values of seg-
ments are significant for classification, as either
the RSTM or the difference matrix are very sim-
ilar for all segments, as long as natural vegeta-
tion cover is being considered. However, this is a
great advantage to the classification implemen-
tation as the classification rules can be greatly
simplified. Given a mean difference as small
as 0.3 dB, eight classes can be discriminated
from the segmented image shown in Fig. 3. The
classified image is shown in Fig.6, in which all
known areas such as bare soil, wet Melaleuca,
dry Melaleuca and the mixed woodland are clas-
sified. The number of segments in the classified
image shown in Fig. 6 may differ from the num-
ber of segments in the segmented image shown
in Fig 3, as the spatially adjacent segments may
be classified as the same class and the boundary
between them is therefore removed.
4. CONCLUSIONS
Highly speckled SAR image data require area-
based analysis in order to improve the accu-
racy of classification. Images are segmented us-
ing the Gaussian markov random field model
with techniques of wavelet filtering, edge detec-
tion and watershed process. The classification is
then implemented based on segment analysis. In
the case of single-channel images, the most im-
portant statistical parameters of a segment are
mean, standard deviation and the difference ma-
trix (the description of textures). For areas of
natural vegetation, it has been found that the
ratio of the standard deviation to the mean is al-
most constant, and the difference matrix is also
very similar for all segments. Therefore, the clas-
sification rules for segments depend only on the
mean values of segments, making the implemen-
tation of classification relatively simple. Areas
where the difference of backscattering coefficient
is as small as 0.3 dB can be discriminated. The
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
classified image was found to be accurate when
compared to the ground truth.
5. REFERENCES
Baraldi, A., and Parmiggiani, F. (1996): Single
linkage region growing algorithms based on the
vector degree of match. IEEE Trans on Geo-
science and Remote Sensing, 34(1), pp. 137-148.
Derin, H., Kelly, P. A., Vezina, G., and Labitt, S.
G. (1990): Modelling and segmentation of speck-
led images using complex data. IEEE Trans on
Geoscience and Remote Sensing, 28(1), pp. 76-
87. *
Dobson, M. C., Ulaby, F. T., Pierce, L. E.,
Sharik, T. L., Bergen, K. M., Kellndorfer, J.,
Kendra, J. R., Li, E., Lin, Y. C., Nashashibi, A.,
Sarabandi, K., and Siqueira, P. (1995): Estima-
tion of forest biophysical characteristics in north-
ern Michigan with SIR-C/X-SAR. IEEE Trans
on Geoscience and Remote Sensing, 33(4), pp.
871-895.
Dobson, M. C., Pierce, L. E., and Ulaby, F.
T. (1996): Knowledge-based land-cover classi-
fication using ERS-1/JERS-1 SAR composites.
IEEE Trans on Geoscience and Remote Sensing,
34(1), pp. 83-99.
Dong, Y., Forster, B. C., and Milne, A. K.
(1998a): Evaluation of radar image segmenta-
tion by Markov random field model with Gaus-
sian distribution and Gamma distribution, Pro-
. ceedings of IGARSS '98, IEEE Inc, New Jersey.
Dong, Y., Forster, B. C., and Milne, A. K.
(1998b): Segmentation of radar imagery using
the Gaussian Markov random field model. In-
ternational Journal of Remote Sensing, in Press.
Dong, Y., Forster, B. C., and Ticehurst, C.
(1996): Decomposition and interpretation of
radar polarisation signatures. Proceedings of
IGARSS '96, pp. 1556-1558. IEEE Inc. New
Jersey.
Freeman, A., and Durden, S. (1992): A three-
component scattering model to describe po-
larimetric SAR data. Proceedings of SPIE
Radar Polarimetry Conference, pp. 213-224, San
Diego, July 23-24, 1992.
Geman, S., and Geman, D. (1984): Stochas-
tic relaxation, Gibbs distribution, and Bayesian
restoration of images. IEEE Trans on Pattern
Analysis and Machine Intelligence, 6(6), 721-
521
