196
1 2 3 4 5 6
Fig.7. The object-oriented classification map based on MRF
and SYM.
1 2 3 4 5 6
Fig.8. The pixel-based classification map based on SVM.
3.3 Accuracy Assessment
Comparing Fig.7 and Fig.8, we notice that Fig.8 appears “salt
and pepper” phenomenon. Table 1 shows the confusion matrix
of classification based on pixel-based SVM, and table 2 shows
the confusion matrix of object-oriented classification based on
MRF and SVM.
The results show that the overall accuracy of pixel-based SVM
is 89.02%, whereas the overall accuracy of the object-oriented
classification based on MRF and SVM is 96.06%. The overall
accuracy is improved by 7.04%. In addition, the user’s
accuracy of each class using object-oriented classification is
higher than the pixel-based SVM method. In particular, the land
cover class, water and road, are obviously distinguished by the
object-oriented classification method.
4. CONCLUSION AND DISCUSSION
The study proposed a new object-oriented land cover
classification method based on MRF and SVM using HR
QuickBird data, which built connection between the domains of
raster analysis and vector analysis. The MRF segmentation can
generate image regions being homogeneous in spectral
distribution and textural properties. The SVM classification
method could solve sparse sampling, non-linear, high
dimensional data, and global optimum problems.
Compared with the pixel-based SVM classification method, the
results indicated that the proposed object-oriented classification
method can improve the efficiency of training and classification,
get accurate classification results, and update GIS database in a
quick and convenient way. Because the proposed classification
method could detect and reclassify the small misclassified
objects in an iteration way, we may get better classification
result. Additionally, this method has the potential to apply in
land cover monitoring at regional and global scale, object
detection, feature extraction, etc.
For the purpose of accurately classifying different ground
objects, it is important to consider scale effects and additional
effective features in the near future.
REFERENCE
Baatz, M., et al., 2004. eCognition User Guide 4,
www.definiens-imaging.com(Last visited March 10, 2007).
Bruzzone, L., Carlin, L. and Melgani, F., 2004. A multilevel
hierarchical approach to classification of high spatial resolution
images with support vector machines. IGARSS '04. pp. 20-24.
Burges, C. J. C., 1998. A tutorial on support vector machines
for pattern recognition. Knowledge Discovery and Data Mining,
2(2), pp. 121-167.
Chang, C. C. and Lin, C. J., 2001. LIBSVM: a library for
support vector machines, http://www.csie.ntu.edu.tw/~cjlin/,
(Last visited March 28, 2007).
Cuozzo, G., Elia, C. D. and Puzzolo,V., 2004. A method based
on Tree-Structured Markov Random Field for forest area
classification. IGARSS’04. pp. 2352-2354.
Li, H.T., Gu, H.Y., Han, Y.S. and Yang, J.H., 2007. Object-
oriented classification of very high-resolution remote sensing
imagery based on improved CSC and SVM. Geoinformatics’07.
Li, S.Z., 2001. Markov random field modeling in image
analysis. Springer.
