with various pavement materials. For grassland, training samples were collected for
two spectral classes with one for golf course with a bright color and the other for
urban green spaces with low woody cover. Two spectral classes were defined
for evergreen forest with one for highland evergreen forest and the other for
wetland evergreen forest. For mixed forest, training samples were collected for
two spectral classes that vary due to soil types. We calculated the spectral separability for each pair of the spectral classes, and finally selected 20 classes for use in
the training phase of the SVM classification that will be discussed later.
13.3.3 SVM Configuration and Classification
As discussed before, SVM parameter settings can affect the classification performance (Huang et al. 2002; Kavzoglu and Colkesen 2009). Among them, the kernel
type, error penalty, and Gamma term are the three most critical parameters (Yang
2011). We configured a support vector machine with radial basis function as the
kernel type, a moderate error penalty value (C ¼ 100), and a Gamma term equaling
Fig. 13.3 The Landsat Thematic Mapper (TM) image used in this study. It was clipped to match
the geographic coverage of Gwinnett County, Georgia. Note that the image is displayed in false
color composite
272
D. Shi and X. Yang
two spectral classes with one for golf course with a bright color and the other for
urban green spaces with low woody cover. Two spectral classes were defined
for evergreen forest with one for highland evergreen forest and the other for
wetland evergreen forest. For mixed forest, training samples were collected for
two spectral classes that vary due to soil types. We calculated the spectral separability for each pair of the spectral classes, and finally selected 20 classes for use in
the training phase of the SVM classification that will be discussed later.
13.3.3 SVM Configuration and Classification
As discussed before, SVM parameter settings can affect the classification performance (Huang et al. 2002; Kavzoglu and Colkesen 2009). Among them, the kernel
type, error penalty, and Gamma term are the three most critical parameters (Yang
2011). We configured a support vector machine with radial basis function as the
kernel type, a moderate error penalty value (C ¼ 100), and a Gamma term equaling
Fig. 13.3 The Landsat Thematic Mapper (TM) image used in this study. It was clipped to match
the geographic coverage of Gwinnett County, Georgia. Note that the image is displayed in false
color composite
272
D. Shi and X. Yang