106
was particularly difficult. As SVMs are traditionally binary classifiers, kernels are
necessary to combat the multivariate problem, and the chosen kernels for this study
were linear and used a radial basis function. The multivariate problem also includes
a choice between ‘one against the rest’ and ‘one against one’ methods, where the
latter was chosen based on its preference within the literature (Pal and Mather 2005;
Mountrakis et al. 2011).
Input Feature Selection
Feature selection to provide inputs for predicting classification algorithms is also a
key consideration that could affect the efficacy and accuracy of mapping outputs.
Another reason to be selective about input features is the curse of high dimensionality, or the Hughes phenomenon, where classification performance will reach a peak
without proportional increase in the training sample size, beyond which performance degrades (Landgrebe 2003). Therefore, high dimensions in the data need to
be reduced to ensure the predictive power of algorithms. Techniques such as
Principal Components Analysis (PCA) or Independent Component Analysis (ICA)
are often employed for this purpose where only the first few components are used as
feature vector input to the classifiers. However, in this process, some useful
Fig. 6 Example of an n-dimensional space between the most separable indices where classes
(bracken (orange), other vegetation (green)) are deemed inseparable. Indices are: CVI Chlorophyll
Vegetation Index (Hunt et al. 2011); SRGNIR1 Simple ratio of green and Near-Infrared 1 band;
IPVI Infrared Percentage Vegetation Index (Crippen 1990); Datt1 (Datt 1999); SRYB Simple ratio
of yellow and blue bands
G. Jones et al.
was particularly difficult. As SVMs are traditionally binary classifiers, kernels are
necessary to combat the multivariate problem, and the chosen kernels for this study
were linear and used a radial basis function. The multivariate problem also includes
a choice between ‘one against the rest’ and ‘one against one’ methods, where the
latter was chosen based on its preference within the literature (Pal and Mather 2005;
Mountrakis et al. 2011).
Input Feature Selection
Feature selection to provide inputs for predicting classification algorithms is also a
key consideration that could affect the efficacy and accuracy of mapping outputs.
Another reason to be selective about input features is the curse of high dimensionality, or the Hughes phenomenon, where classification performance will reach a peak
without proportional increase in the training sample size, beyond which performance degrades (Landgrebe 2003). Therefore, high dimensions in the data need to
be reduced to ensure the predictive power of algorithms. Techniques such as
Principal Components Analysis (PCA) or Independent Component Analysis (ICA)
are often employed for this purpose where only the first few components are used as
feature vector input to the classifiers. However, in this process, some useful
Fig. 6 Example of an n-dimensional space between the most separable indices where classes
(bracken (orange), other vegetation (green)) are deemed inseparable. Indices are: CVI Chlorophyll
Vegetation Index (Hunt et al. 2011); SRGNIR1 Simple ratio of green and Near-Infrared 1 band;
IPVI Infrared Percentage Vegetation Index (Crippen 1990); Datt1 (Datt 1999); SRYB Simple ratio
of yellow and blue bands
G. Jones et al.