human visual system, several multiscale low-level processes (i.e., filtering, segmentation, etc.) have been developed during which a series of representations of the same
image are computed (from fine to coarse) and used for the recognition of Earth surface
objects (Blaschke and Hay, 2001; Hay et al., 2002; Hall and Hay, 2003; Benz et al.,
2004; Stewart et al., 2004; Jimenez et al., 2005; Karantzalos and Argialas, 2006;
Duarte-Carvajalino et al., 2008; Ouma et al., 2008). The mathematical models and the
manner for constructing these scale space representations are of fundamental
importance.
In addition, during the last decade the way of classifying remotely sensed imagery
has been changing, and instead of classifying individual pixels into discrete land cover
classes, object-based classification approaches construct a hierarchical object representation of an image and the classifier is responsible for associating them with a land
cover class (Blaschke, 2010). Therefore, it is not just the spectral signature of each
pixel but the statistical, geometric, and topological characteristics of each object that
play a key role during classification. Object-based classification is considered optimal
for the analysis of very high resolution remote sensing imagery (with spatial
resolution of 5 m per pixel or less), since this kind of imagery is introducing
more complexity to the classification tasks, due to increased heterogeneity and the
increased number of land cover classes that can be observed. The statistical,
geometric, and contextual characteristics of the image primitives are considered
by the object-based methods (much like in photointerpretation), in contrast to pixelbased approaches (Blaschke, 2010). Recent studies are highlighting that the determination of one or more optimal filtering scales for image segmentation is still a
challenge and a multiscale object-based classification is a significantly better
approach than the classical per-pixel classification procedure (Myint et al., 2011;
Tzotsos et al., 2011).
In this chapter, an object-based image analysis framework was developed which
integrates advanced scale space representations, edge and line feature detection,
multiscale segmentation, and a kernel-based classification. The contributions of this
approach are twofold:
• A generic framework able to process any remotely acquired raster data (satellite/
airborne data, multispectral/hyperspectral data, and radar data of any spatial
resolution) without the need of tuning any parameter (scale, color, texture, etc.).
• A new robust multiscale segmentation procedure (replacing an earlier segmentation algorithm in the framework) which is constrained by advanced edgebased features.
The chapter is structured as follows. In Section 9.2, the related work on scale-space
representations, multiscale object-based analysis, and edge-based image segmentation is briefly presented. The developed object-based image analysis framework is
detailed in Section 9.3, along with a description and detailed analysis of its different
processing steps. Experimental results and the performed quantitative evaluation
are given in Section 9.4. Finally, conclusions and perspectives for future work are
in Section 9.5.
INTRODUCTION
171
image are computed (from fine to coarse) and used for the recognition of Earth surface
objects (Blaschke and Hay, 2001; Hay et al., 2002; Hall and Hay, 2003; Benz et al.,
2004; Stewart et al., 2004; Jimenez et al., 2005; Karantzalos and Argialas, 2006;
Duarte-Carvajalino et al., 2008; Ouma et al., 2008). The mathematical models and the
manner for constructing these scale space representations are of fundamental
importance.
In addition, during the last decade the way of classifying remotely sensed imagery
has been changing, and instead of classifying individual pixels into discrete land cover
classes, object-based classification approaches construct a hierarchical object representation of an image and the classifier is responsible for associating them with a land
cover class (Blaschke, 2010). Therefore, it is not just the spectral signature of each
pixel but the statistical, geometric, and topological characteristics of each object that
play a key role during classification. Object-based classification is considered optimal
for the analysis of very high resolution remote sensing imagery (with spatial
resolution of 5 m per pixel or less), since this kind of imagery is introducing
more complexity to the classification tasks, due to increased heterogeneity and the
increased number of land cover classes that can be observed. The statistical,
geometric, and contextual characteristics of the image primitives are considered
by the object-based methods (much like in photointerpretation), in contrast to pixelbased approaches (Blaschke, 2010). Recent studies are highlighting that the determination of one or more optimal filtering scales for image segmentation is still a
challenge and a multiscale object-based classification is a significantly better
approach than the classical per-pixel classification procedure (Myint et al., 2011;
Tzotsos et al., 2011).
In this chapter, an object-based image analysis framework was developed which
integrates advanced scale space representations, edge and line feature detection,
multiscale segmentation, and a kernel-based classification. The contributions of this
approach are twofold:
• A generic framework able to process any remotely acquired raster data (satellite/
airborne data, multispectral/hyperspectral data, and radar data of any spatial
resolution) without the need of tuning any parameter (scale, color, texture, etc.).
• A new robust multiscale segmentation procedure (replacing an earlier segmentation algorithm in the framework) which is constrained by advanced edgebased features.
The chapter is structured as follows. In Section 9.2, the related work on scale-space
representations, multiscale object-based analysis, and edge-based image segmentation is briefly presented. The developed object-based image analysis framework is
detailed in Section 9.3, along with a description and detailed analysis of its different
processing steps. Experimental results and the performed quantitative evaluation
are given in Section 9.4. Finally, conclusions and perspectives for future work are
in Section 9.5.
INTRODUCTION
171