criterion, shape parameters are used to define geometric properties that the segmentation algorithm must take into account when computing the overall homogeneity
(scale parameter) of each image object during search for optimal merges.
In 2008, a texture optimization procedure was introduced for the MSEG algorithm
(Tzotsos et al., 2008) integrating gray-level co-occurrence matrices and introducing
an object-based cost measure for texture homogeneity as an additional parameter to
the segmentation procedure. Such an integration of spatial and spectral information
can produce a multiscale object representation but only through an iterative and
exhaustive tuning (based on trial-and-error investigation) of certain parameters, for
example, shape, scale, and texture (Baatz and Schape, 2000; Benz et al., 2004;
Blaschke et al., 2004; Carleer et al., 2005; Hay et al., 2005; Tzotsos and Argialas,
2006; Ouma et al., 2008; Dragut et al., 2009; Zhou et al., 2009).
Other research efforts were based on the construction of linear scale spaces for the
multiscale analysis of several landscape structures (Blaschke and Hay, 2001; Hay
et al., 2002, 2003; Stewart et al., 2004) or on the construction of multiscale representations through object-specific analysis and upscaling through the computation of a
number of coarse and fine scales by sampling the initial image (Hall and Hay, 2003).
Furthermore, other studies employed unsupervised classification algorithms for both
optical and radar data (Derrode and Mercier, 2007; Jung, 2007) or multiple hierarchical segmentation (Akcay and Aksoy, 2008).
More recent research efforts are focusing on optimizing the segmentation procedure through a data-driven thresholding approach (Martha et al., 2011) and on
constructing advanced nonlinear scale-space representations for efficient supervised
classification (Tzotsos et al., 2011) and change detection over urban areas (Doxani
et al., 2012).
9.2.2 Scale-Space Remote Sensing Data Representations
Earth surface objects cannot be represented to a single scale but rather to many. The
use of scale-space image representations is thus of fundamental importance for a
number of image analysis and computer vision tasks. It dates back to the 1960s and
was first introduced by Iijima (Weickert et al., 1999). In western literature many linear
scale-space methods were introduced (Witkin, 1983; Koenderink, 1984; Lindeberg,
1994), and respectively many isotropic multiscale operators were developed. Either
through Gaussian filtering or through isotropic multiresolution analysis (by downsampling the initial data), all linear scale-space approaches present the same important
drawback: image edges are blurred and new nonsemantic objects may appear at coarse
scales (Witkin, 1983; Paragios et al., 2005; Ouma et al., 2008). Under a hierarchical
multiscale segmentation or an object-based classification framework, the thematic
information to be extracted is directly related with the primitive image objects
computed at every scale. The better these primitive objects represent real-world
entities, the better they can describe the semantics of the image (Hay and Castilla,
2006; Blaschke et al., 2008; Hofmann et al., 2008; Tzotsos et al., 2011). Therefore, the
selection of the appropriate approach for constructing the multiscale image and
hierarchical object representation is of great importance.
RELATED WORK
173
(scale parameter) of each image object during search for optimal merges.
In 2008, a texture optimization procedure was introduced for the MSEG algorithm
(Tzotsos et al., 2008) integrating gray-level co-occurrence matrices and introducing
an object-based cost measure for texture homogeneity as an additional parameter to
the segmentation procedure. Such an integration of spatial and spectral information
can produce a multiscale object representation but only through an iterative and
exhaustive tuning (based on trial-and-error investigation) of certain parameters, for
example, shape, scale, and texture (Baatz and Schape, 2000; Benz et al., 2004;
Blaschke et al., 2004; Carleer et al., 2005; Hay et al., 2005; Tzotsos and Argialas,
2006; Ouma et al., 2008; Dragut et al., 2009; Zhou et al., 2009).
Other research efforts were based on the construction of linear scale spaces for the
multiscale analysis of several landscape structures (Blaschke and Hay, 2001; Hay
et al., 2002, 2003; Stewart et al., 2004) or on the construction of multiscale representations through object-specific analysis and upscaling through the computation of a
number of coarse and fine scales by sampling the initial image (Hall and Hay, 2003).
Furthermore, other studies employed unsupervised classification algorithms for both
optical and radar data (Derrode and Mercier, 2007; Jung, 2007) or multiple hierarchical segmentation (Akcay and Aksoy, 2008).
More recent research efforts are focusing on optimizing the segmentation procedure through a data-driven thresholding approach (Martha et al., 2011) and on
constructing advanced nonlinear scale-space representations for efficient supervised
classification (Tzotsos et al., 2011) and change detection over urban areas (Doxani
et al., 2012).
9.2.2 Scale-Space Remote Sensing Data Representations
Earth surface objects cannot be represented to a single scale but rather to many. The
use of scale-space image representations is thus of fundamental importance for a
number of image analysis and computer vision tasks. It dates back to the 1960s and
was first introduced by Iijima (Weickert et al., 1999). In western literature many linear
scale-space methods were introduced (Witkin, 1983; Koenderink, 1984; Lindeberg,
1994), and respectively many isotropic multiscale operators were developed. Either
through Gaussian filtering or through isotropic multiresolution analysis (by downsampling the initial data), all linear scale-space approaches present the same important
drawback: image edges are blurred and new nonsemantic objects may appear at coarse
scales (Witkin, 1983; Paragios et al., 2005; Ouma et al., 2008). Under a hierarchical
multiscale segmentation or an object-based classification framework, the thematic
information to be extracted is directly related with the primitive image objects
computed at every scale. The better these primitive objects represent real-world
entities, the better they can describe the semantics of the image (Hay and Castilla,
2006; Blaschke et al., 2008; Hofmann et al., 2008; Tzotsos et al., 2011). Therefore, the
selection of the appropriate approach for constructing the multiscale image and
hierarchical object representation is of great importance.
RELATED WORK
173