in Figure 3(b), the increased posterior probability pertaining to
the right type is still not predominant to exceed the decreased
probability pertaining to the confusing types. However, the
effect that the SK method helps to improve the predicted
probability of the right land cover type is still traceable.
In Figure 3(d) the SK method fails to further improve the
originally prevailing posterior probabilities pertaining to the
right type, but instead meets the exact reverse. It may come
down to two reasons: (1) the input vectors are prone to
confusion which results in the slight advantage in the posterior
probability of the right type; (2) the training samples in a local
domain are sparse and unevenly distributed which results in an
unfaithful modelling of residual variation and a naive spatial
interpolation. That is, the less accurate residual variogram
models would easily reverse the slight advantage. However,
samples of this kind in Figure 3(b) only occupy 1.1 percent of
the total samples. On the contrary, samples of the kind in Figure
3(a), which are poorly classified but correctly revised, occupy
58 percent.
Moreover, compared to the producer's and user's accuracy
acquired by the SVM classification, as are listed in Table 2, the
corresponding accuracy fluctuations after the application of the
SK method may not equivalent to mean that the kriging method
is particularly suitable to some certain land cover types.
In order to further testify the efficiency of the kriging paradigm,
another TM image including 17 land cover types is adopted. A
total of ten variables were available: Landsat TM channels 1-5,
7, modified normalized difference vegetation index (MNDVI),
scaled elevation, slope in degrees, and a combined slope-aspect
variable. Further detail of this data set is documented in Zhang
and Goodchild (2007). The estimated Arif index manifests a
corresponding highest classification accuracy of 72.2296. The
overall accuracy and kappa coefficient achieved by generalized
linear model (abbreviated to GLM) are 65.55% and 0.62,
respectively. After residual corrections, the former is improved
to 75.45% and the latter is increased to 0.73. It is interesting to
notice that the revised accuracy of 75.45% is larger than the
estimated potential highest accuracy of 72.22%. The is that the
potential highest is just estimated by the input vectors without
considering the introduced spatial information during the post-
classification corrections.
4. CONCLUSION
The proposed two kriging methods are independent of the
specific classifiers initially adopted for image classification.
Hence, these methods may be treated as post-classification
approaches. They aim at compensating part of the information
consumed by classifiers due to the insufficient learning process.
Moreover, these methods are independent of the land cover
types, although the improvements vary in the producer’s and
user’s accuracies of different land covers. The factors, including
the sampling methods, sample sizes, and sample distributions,
need to be further investigated, for they are close related to the
spatial information of the training samples.
REFERENCE
Atkinson P. M., Lewis P., 2000. Geostatistical classification for
remote sensing: an introduction. Computers and Geosciences,
26, pp. 361-371.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
Arif M., Afsar F. A., Akram M. U., et al., 2009. Arif index for
predicting the classification accuracy of features and its
application in heart beat classification problem. In: Advances in
Knowledge Discovery and Data Mining: proceedings of the
13th Pacific-Asia Conference, Thailand.
Battiti R., 1995. Using mutual information for selecting features
in supervised neural net learning. /EEE Transactions on Neural
Networks, 5(4), pp. 537-551.
Food and Agriculture Organization of the United Nations, 2000.
Land Cover Classification System (LCCS): Classification
Concepts and User Manual.
http://www.fao.org/docrep/003/x0596e/x0596e00.htm(2011)
Foody G. M., 2002. Status of land cover classification accuracy
assessment. Remote Sensing of Environment, 80, pp.185-201.
Goovaerts P, 1997. Geostatistics for Natural Resources
Evaluation. Oxford University Press, New York, pp. 125-233.
Goovaerts P., 2002. Geostatistical incorporation of spatial
coordinates into supervised classification of hyperspectral data.
Geographical systems, 4, pp. 9-111.
Homer C., Dewitz J., Fry J., et al, 2007. Completion of the
2001 National Land Cover Database for the Conterminous
United States. Photogrammetric Engineering and Remote
Sensing, 73(4), pp. 337-341.
Homer C., Huang C. Q., Yang L. M., et al., 2004. Development
of a 2001 National Land-Cover Database for the United States.
Photogrammetric Engineering and Remote Sensing, 70(7), pp.
829—840.
Meyer D., 2009. Support Vector Machines-The Interface to
Libsvm in Package e1071.
http://cran.r-project.org/web/packages/e107 1/vignettes/
svmdoc.pdf (2011)
Zhang J. X., 2009. Scale, Uncertainty and Fusion of Spatial
Information. Wuhan University Press, Wuhan, pp. 221-234.
Zhang J. X., Goodchild, M. F., Steele, B. M. and et al., 2007. A
discriminant space-based framework for scalable area-class
mapping. In: Proceedings of SPIE-The International Society for
Optical Engineering, Nanjing, China, Vol. 6751.
ACKNOWLEDGEMENT
The research is partially supported by “973 Program” grants
(No. 41071286) and (No. 41171346 ).
