International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
100 4
Local kappa (%)
3888882388
0+
Classes
Figure 7. Local classification accuracy
7. CONCLUSION
The purpose of this work is to design robust algorithm for
classification of remotely sensed images. Our experience
confirms that context information plays an important role in the
task of scene interpretation. At the pixel level, context information
provides neighbourhood information around a pixel, and helps to
increase the reliability of each detect object. Discrete random
fields, especially the Gibbs Random Fields (GRF) and Markov
Random Fields (MRF) provide a methodological framework
which allows the integration of context information in satellite data
classification. A powerful of these models is that the prior
probability density function modelled by the use of the contextual
information and the class conditional probability density function
modelled by the use of the observed data from one or more
sensors, can be easily combined through the use of suitable
energy function. Once the posterior energy model and the
associated parameters have been defined, pixel labelling is found
out by using the MAP estimate which is equivalent to a minimum
energy function in terms of GRF-MRF modelling. For a non-
convex energy function, the solution space may contain several
local minimum. To find a global minimum which is a truly MAP
estimate, the solution is to use an optimisation algorithm among
which ICM is the most know and used. The ICM algorithm is
sub-optimal and converges only to a local minimum of the energy
function. However, classification result of such algorithm is
acceptable and shows that the incorporation of contextual
information successfully improves classifier performances by
more than 10% in terms of global accuracy. However, algorithms
and methods to construct more complex modek and to efficiently
integrate context (context at object level which is useful for
obtaining a coherent interpretation of the whole scene) in order to
achieve higher classification accuracy, are still significant issues
worthy of further investigation.
References
Amadamn, M., and King, R. A., 1988. Low level segmentation of
multispectral images via agglomerative clustering of uniform
neighborhoods. Patterns recognition, 21, pp. 261-268.
Besag, J., 1986. On the Statistical Analysis of Dirty Pictures.
Journal Royal of Statistics: Soc. B, 48, 3, pp. 259-302.
Braathen, B., Pieczynski, W., and Masson, P., 1993. Global and
local methods of unsupervised Bayesian segmentation of images.
Machine Graphics and vision, vol. 2, no. 1, pp. 39-52.
Brogaard, S., Prieler, S., 1998. Land cover in the Horqin
Grasslands, North China. Detecting changes between 1975 and
16
1990 by means of remote sensing. /nterim report on work of
IIASA, 1R-98-044/]uly.
Congalton, R. G., 1991. A Review of Assessing the Accuracy of
Classifications of Remotely Sensed Data. Remote Sensing of
Environment, no. 37, pp. 35-46.
Desachy, J. 1991. Interprétation automatique d'images
satellitaires le système ICARE. Thèse de Doctorat en
informatique, Université Paul Sabatier, Toulouse, France.
Geman, S. and Geman, D. 1984. Stochastic Relaxation, Gibbs
Distribution, and the Bayesian Restoration of Images. IEEE
Trans. Pattern Analysis and Machine Intelligence, PAMI-6,
pp. 721-741.
Kartikeyan, B., Gopalakrishna, B., Kalubarme, M. H., and
Majumder, K. L, 1994. Contextual techniques for classification
of high and low resolution remote sensing data. /JRS, 1994, vol.
15, No. 5, pp. 1037-1051.
Kettig, R. L., d Landgrebe, D. A., 1987. Classification of
multispectral image data by extraction and classification of
homogenous objects. /EEE. Transactions on Geoscience and
Remote Sensing, GE-14, pp. 17-26.
Khedam, R., Belhadj-Aissa, A., 2001. General Multisource
Contextual Classification Model of Remotely Sensed Imagery
based on MRF. IEEE / ISPRS Workshop on Remote Sensing and
Data Fusion Over Urban Areas, Rome, Italy, November &9^
2001. /EEE Catalog Number 01 EX482C. ISBN 0-7803-7060-0.
Khedam, R., Belhadj-Aissa, A. and Ranchin, T., 2002. Study of
ICM parameters influence on imges satellite contextual
classification. In proceedings of the 22"^ symposium of the
European association
Prague, Czech, 4-6 june, pp. 79-85.
of remote sensing laboratories,
Khedam, R., Belhadj-Aissa, A., 2003. Contextual fusion by
genetic approach applied to the classification of satellite images.
Remote sensing in transition, Goossens (ed.), 2004
Millpresse, Rotterdam, ISBN 90 5966 007 2.
Lillesand, T. M., and Kiefer, R. W., 1987. Remote sensing and
image interpretation. John Wiley and sons: New York.
Marroquin, J., Mitter, S., Poggio, T., 1987. Probabilistic solution
of ill-posed problems in computational vision. Journal of the
American Statistical Association, no. 82, pp. 76-89.
Pieczynski, W., 2000. Segmentation Statistiques d'Images. Notes
de cours, Institut National des Télécommunications, France, 95 p.
Shabah, M. A., Dikshit, O., 2001. Improvement of classification
in urban areas by the use of textural features: the case study of
Lucknow city, Uttar Pradesh. /JRS, 22, 4, pp. 565-593.
Richards, J. A., Landgrebe, D. A., and Swain, P. H., 1981. Pixel
labelling by supervised probabilistic relaxation. /EEE.
Transactions on Pattern Analysis and Machine Intelligence,
PAMI-3, pp. 188-191.
Toussaint, G. C., 1978. The use of context in pattern recognition.
Pattern Recognition, 10, pp. 189-204.
«T
KEY
ABS!
This
conve
unbia
data-c
traini
intern
Rio F
concl
Rem‘
of i
Conv
(ML
used
geogi
frequ
distri
been
traini
inter]
data’
class
thous
categ
struc
netw
long
hour:
into
must
netw
need
class
prop
inde
class
categ
Func
prob
In re
class
cove
Thes
*- C