Research Methodology - Historical Land Use/Land Cover Classification Using Remote Sensing - page 122

two multi-case-methods, other approaches directly identify the SVM as one
multiclass problem (Sebald Hsu and Lin 2002 ). A simultaneous separation of more
than two classes presents a more complex optimization problem (Sebald and
Bucklew 2001 ). Thus, such approaches may be less professional in comparison to
conventional multiclass approaches. In Melgani and Bruzzone ( 2004 ) a compu-
tationally promising hierarchical tree-based SVM was presented as an alternative
concept.
SVMs work very well with high dimensional data. Their computational cost
does not depend on data dimensionality and they need no feature selection. So,
classification results for multisource data classification from a non-parametric
classifier in particular, is maybe better than that received from a parametric
classifier, since a non-parametric classifier can solve some of the problems of a
stacked vector approach (Watanachaturaporn et al. 2008 ). SVM learning generally
requires large memory, a great deal of computation time and small training sets
(Su 2009 ). Some of the issues that influence the classification accuracy of SVM-
classifiers (Huang et al. 2002 ) are: choice of kernel used (linear, polynomial, radial
basis function, and sigmoid); and choice of the parameters related to a particular
kernel (degree of kernel polynomial, bias in kernel function, gamma in kernel
function,

penalty parameter, pyramid levels, and classification probability

threshold).
Recent studies have shown that the use of SVMs in remotely sensed data
classification might present results with higher accuracy than other classifiers (Tso
and Mather 2009 ). SVMs have been used for classification of RADAR-data
(Shimoni et al. 2009 ), ASTER-data (Zhu and Blumberg 2002 ; Marcal et al. 2005 ),
LANDSAT-TM-data (Keuchel et al. 2003 ) and hyper-spectral-data (Melgani and
Bruzzone 2004 ). Only a few studies are known which have used SVMs for clas-
sifying multisource or multi-temporal data (Camps-Valls et al. 2006 ). Foody and
Mathur ( 2004a , b , 2006 ) have examined both the characteristics and the size of
training samples in SVMs. The paper from Hernandez et al. (2007) confirmed that
applying a classification approach based on SVMs such as the SVDD could be
used to provide more accuracy (97.5 %) than a MLC (90.0 %). Other significant
papers on this topic include: Bruzzone and Marconcini ( 2009 ), and Su ( 2009 ).
The SVM-options that were used in the study were: Kernel type (polynomial);
degree of kernel polynomial (2); bias in kernel function (1,000); gamma in kernel
function (0.111); penalty parameter (100,000); pyramid levels (0); and classifi-
cation probability threshold (0).

5.7.2 Results and Evaluation

This section present the results (thematic maps) of the comparison study which
used the remotely sensed data obtained from LANDSAT: MSS-June-1975 with
60 m spatial resolution and four spectral bands (Fig. 5.41 ); TM-May-2007/30 m
and six bands (Fig. 5.42 ); and TERRA: ASTER-May-2005/15 m and three bands