Alberti 2008). Land cover patterns are observable and therefore can be mapped by
ground surveys or remote sensing. While ground surveys are largely limited by
logistical constraints, remote sensing makes direct observations across large areas
of the land surface, thus allowing land cover patterns to be mapped in a timely and
cost-effective mode. Both visual interpretation and computer-based digital classification can be used to extract information on land cover from a variety of remotely
sensed data varying in spatial, spectral, radiometric, and temporal resolutions.
Digital pattern classification is generally preferred over visual interpretation for
mapping land cover in large areas (Jensen 2005).
While conventional pattern classifiers (e.g., maximum likelihood) have been
widely used, they generally work well with medium-resolution images and in
relatively homogeneous areas rather than highly heterogeneous areas (Yang
2002). Over the years, substantial research efforts have been directed to improve
the performance of land cover mapping in heterogeneous areas (e.g. Hoffer 1978;
Richards et al. 1982; Skidmore et al. 1997; Duda et al. 2001; Yang and Lo 2002;
Schmidt et al. 2004; Del Frate et al. 2007; Foody 2008; Heikkinen et al. 2010; Zhou
and Yang 2011; Liu and Yang 2013).
This study targets support vector machines (SVM), a group of relatively novel
machine learning algorithms based on statistical learning theory that have not been
extensively exploited in the remote sensing community. They are found to
outperform most of the conventional classifiers (Huang et al. 2002; Keuchel
et al. 2003; Kavzoglu and Colkesen 2009; Su and Huang 2009). Moreover, SVM
were found to even outperform some novel pattern recognition methods, such as
neural networks (Huang et al. 2002; Foody and Mathur 2004a, b). Nevertheless,
there are some parametric and non-parametric factors that can affect the performance of SVM, and there is a need to investigate them so that SVM could be used
with improved performance (Yang 2011).
In this chapter, we examine the utilities of support vector machines (SVM) as a
pattern recognition technique for landscape mapping particular for heterogeneous
areas. It is organized into two major parts, beginning with a brief introduction of
some basic knowledge on SVM and a review on the research status and possible
challenges of using SVM for land cover mapping. The review focuses on some
comparative studies that demonstrated the effectiveness of SVM over other conventional classifiers. Based on the review, we further discuss several areas that need
additional research in order to improve SVM classification accuracies and reduce
computational burdens, which are mostly related to appropriate treatments of some
parametric and non-parametric factors. The second part of the paper discusses our
implementation of SVM to map various land cover types from a remote sensor
image covering an urban area, demonstrating the robustness of this type of pattern
recognition technique for mapping heterogeneous landscapes.
266
D. Shi and X. Yang
ground surveys or remote sensing. While ground surveys are largely limited by
logistical constraints, remote sensing makes direct observations across large areas
of the land surface, thus allowing land cover patterns to be mapped in a timely and
cost-effective mode. Both visual interpretation and computer-based digital classification can be used to extract information on land cover from a variety of remotely
sensed data varying in spatial, spectral, radiometric, and temporal resolutions.
Digital pattern classification is generally preferred over visual interpretation for
mapping land cover in large areas (Jensen 2005).
While conventional pattern classifiers (e.g., maximum likelihood) have been
widely used, they generally work well with medium-resolution images and in
relatively homogeneous areas rather than highly heterogeneous areas (Yang
2002). Over the years, substantial research efforts have been directed to improve
the performance of land cover mapping in heterogeneous areas (e.g. Hoffer 1978;
Richards et al. 1982; Skidmore et al. 1997; Duda et al. 2001; Yang and Lo 2002;
Schmidt et al. 2004; Del Frate et al. 2007; Foody 2008; Heikkinen et al. 2010; Zhou
and Yang 2011; Liu and Yang 2013).
This study targets support vector machines (SVM), a group of relatively novel
machine learning algorithms based on statistical learning theory that have not been
extensively exploited in the remote sensing community. They are found to
outperform most of the conventional classifiers (Huang et al. 2002; Keuchel
et al. 2003; Kavzoglu and Colkesen 2009; Su and Huang 2009). Moreover, SVM
were found to even outperform some novel pattern recognition methods, such as
neural networks (Huang et al. 2002; Foody and Mathur 2004a, b). Nevertheless,
there are some parametric and non-parametric factors that can affect the performance of SVM, and there is a need to investigate them so that SVM could be used
with improved performance (Yang 2011).
In this chapter, we examine the utilities of support vector machines (SVM) as a
pattern recognition technique for landscape mapping particular for heterogeneous
areas. It is organized into two major parts, beginning with a brief introduction of
some basic knowledge on SVM and a review on the research status and possible
challenges of using SVM for land cover mapping. The review focuses on some
comparative studies that demonstrated the effectiveness of SVM over other conventional classifiers. Based on the review, we further discuss several areas that need
additional research in order to improve SVM classification accuracies and reduce
computational burdens, which are mostly related to appropriate treatments of some
parametric and non-parametric factors. The second part of the paper discusses our
implementation of SVM to map various land cover types from a remote sensor
image covering an urban area, demonstrating the robustness of this type of pattern
recognition technique for mapping heterogeneous landscapes.
266
D. Shi and X. Yang