Support Vector Machines for Classificationof Multi- and Hyperspectral Data
241
Finally, an important issue in the implementation ofSVMs is the selection of
the type of optimizer to be used to find the support vectors. Different methods
for this purpose have been proposed (see Sect. 5.5). An optimizer based on
QP or LP methods is a natural choice. However, these optimizers may not be
able to efficiently handle large data volume such as the one found in remote
sensing images. In such cases, more efficient optimizers must be employed.
A more appropriate choice for an optimizer, for instance, can be a subset
selection or an iterative method. In the subset selection method, such as the
chunking-decomposition method, a large problem is solved by breaking it into
sub-problems. Each sub-problem is solved separately by either a QP or an
LP optimizer. Examples of optimizers based on the subset selection method
may be found in Chang and Lin (2002) and Joachims (2002). Alternatively, an
iterative method, such as the LSVM (Mangasarian and Musicant 2000), which
is a fast and simple algorithm, may also be used.
Many of these issues will be investigated in this chapter through a number
of experiments on classification of multi and hyperspectral remote sensing
images, which are described in the next section.
10.3
Remote Sensing Images
10.3.1
Multispectral Image
A UTM rectified Landsat 7 ETM + multispectral image acquired in 8 bands
has been used. The image was acquired in 1999 and covers an urban region of
Syracuse, NY. Only 7 bands excluding the panchromatic band, due to its finer
spatial resolution than the remaining bands were considered.
Approximately 10% of the area in the image selected is covered by water
whereas the remaining 90% covers a combination of built up area, small vegetation, and trees. The size of the image is 445 x 595 pixels, which has been
resampled to 25 m at the source (see Fig. 10.1).
A large number of pure pixels were extracted from the Landsat ETM + image
for the six classes of interest, namely water, highways/runways, grassland,
Table 10.1. Total pure pixels, training, and testing pixels from the Landsat ETM+ image
Class Name
Number of pure pixels identified
Water
2100
Highway/Runway
529
Grass
951
Commercial
1626
Tree
1378
Residential
3004
Total
9588
241
Finally, an important issue in the implementation ofSVMs is the selection of
the type of optimizer to be used to find the support vectors. Different methods
for this purpose have been proposed (see Sect. 5.5). An optimizer based on
QP or LP methods is a natural choice. However, these optimizers may not be
able to efficiently handle large data volume such as the one found in remote
sensing images. In such cases, more efficient optimizers must be employed.
A more appropriate choice for an optimizer, for instance, can be a subset
selection or an iterative method. In the subset selection method, such as the
chunking-decomposition method, a large problem is solved by breaking it into
sub-problems. Each sub-problem is solved separately by either a QP or an
LP optimizer. Examples of optimizers based on the subset selection method
may be found in Chang and Lin (2002) and Joachims (2002). Alternatively, an
iterative method, such as the LSVM (Mangasarian and Musicant 2000), which
is a fast and simple algorithm, may also be used.
Many of these issues will be investigated in this chapter through a number
of experiments on classification of multi and hyperspectral remote sensing
images, which are described in the next section.
10.3
Remote Sensing Images
10.3.1
Multispectral Image
A UTM rectified Landsat 7 ETM + multispectral image acquired in 8 bands
has been used. The image was acquired in 1999 and covers an urban region of
Syracuse, NY. Only 7 bands excluding the panchromatic band, due to its finer
spatial resolution than the remaining bands were considered.
Approximately 10% of the area in the image selected is covered by water
whereas the remaining 90% covers a combination of built up area, small vegetation, and trees. The size of the image is 445 x 595 pixels, which has been
resampled to 25 m at the source (see Fig. 10.1).
A large number of pure pixels were extracted from the Landsat ETM + image
for the six classes of interest, namely water, highways/runways, grassland,
Table 10.1. Total pure pixels, training, and testing pixels from the Landsat ETM+ image
Class Name
Number of pure pixels identified
Water
2100
Highway/Runway
529
Grass
951
Commercial
1626
Tree
1378
Residential
3004
Total
9588