9.2 RELATED WORK
9.2.1 Multiscale Object-Based Image Analysis
Along with the gradual availability of Earth observation data with higher spatial and
spectral resolution, research efforts in classifying remote sensing data have been
shifting in the last decade from pixel-based approaches to object-based ones
(Blaschke, 2010; Myint et al., 2011; Tzotsos et al., 2011). Assigning land cover
classes to individual pixels can be intuitively proper and functional for low-resolution
data. However, this is not the case for the emerging applications which arise from the
continuously improving remote sensing sensors (Aplin and Smith, 2008; Blaschke
et al., 2008). This is mostly because at higher resolutions it is a connected group of
pixels that is likely to be associated with a land cover class and not just an individual
pixel (Tzotsos et al., 2011).
In addition, the Earth surface exhibits various regular and irregular structures
which are represented with a certain spatial heterogeneity in images composing their
intensity, scale, and texture. Several important aspects of the Earth surface cannot be
analyzed based on pixel information but can only be exploited based on contextual
information and the topological relations of the objects of interest (Liu et al., 2008)
through a multiscale image analysis (Blaschke and Hay, 2001; Hay et al., 2002; Hall
and Hay, 2003; Benz et al., 2004; Stewart et al., 2004; Jimenez et al., 2005; DuarteCarvajalino et al., 2008; Ouma et al., 2008; Dragut et al., 2010; Tzotsos et al., 2011).
Starting with the observed spatial heterogeneity and variability, meaningful spatial
aggregations (primitive objects) can be formed at certain image scales configuring a
relationship between ground objects and image objects. Ground objects refer to realworld objects or areas of specific land cover class connected by complex spatial and
contextual relations. Image objects, on the other hand, refer to knowledge-free areas
(primitive objects) of an image that a segmentation algorithm provides based on
various criteria. In order to configure a relationship between ground and image
objects, a multiscale knowledge representation is needed, which is usually provided
by the object-oriented paradigm. With such an object-based multiscale representation
and analysis, which is based on certain hierarchically structured rules, the relationship
between the different scales of the spatial entities is described.
During the last decade, a number of object-based image analysis software were
developed in the form of proprietary software (eCognition, ENVI FX, ERDAS
Objective) (Baatz and Schape, 2000) or in the form of free software (Orfeo Toolbox,
EDISON, MSEG) (Tzotsos and Argialas, 2006; Inglada and Christophe, 2009;
Christophe and Inglada, 2009), enabling the broad application on various engineering
and environmental remote sensing studies (Benz et al., 2004; Zhou et al., 2009;
Dragut et al., 2009; Blaschke, 2010; Mladinich et al., 2010). In all cases, the challenge
was to construct an efficient scale-space object representation through certain multiscale (region merging or other) segmentation techniques (Blaschke et al., 2004;
Carleer et al., 2005; Jimenez et al., 2005; Neubert et al., 2006; Tzotsos and Argialas,
2006), which partition the image on several regions/objects, based on the spectral
homogeneity in a local neighborhood. In addition to the spectral homogeneity
172
MULTISCALE SEGMENTATION AND CLASSIFICATION
9.2.1 Multiscale Object-Based Image Analysis
Along with the gradual availability of Earth observation data with higher spatial and
spectral resolution, research efforts in classifying remote sensing data have been
shifting in the last decade from pixel-based approaches to object-based ones
(Blaschke, 2010; Myint et al., 2011; Tzotsos et al., 2011). Assigning land cover
classes to individual pixels can be intuitively proper and functional for low-resolution
data. However, this is not the case for the emerging applications which arise from the
continuously improving remote sensing sensors (Aplin and Smith, 2008; Blaschke
et al., 2008). This is mostly because at higher resolutions it is a connected group of
pixels that is likely to be associated with a land cover class and not just an individual
pixel (Tzotsos et al., 2011).
In addition, the Earth surface exhibits various regular and irregular structures
which are represented with a certain spatial heterogeneity in images composing their
intensity, scale, and texture. Several important aspects of the Earth surface cannot be
analyzed based on pixel information but can only be exploited based on contextual
information and the topological relations of the objects of interest (Liu et al., 2008)
through a multiscale image analysis (Blaschke and Hay, 2001; Hay et al., 2002; Hall
and Hay, 2003; Benz et al., 2004; Stewart et al., 2004; Jimenez et al., 2005; DuarteCarvajalino et al., 2008; Ouma et al., 2008; Dragut et al., 2010; Tzotsos et al., 2011).
Starting with the observed spatial heterogeneity and variability, meaningful spatial
aggregations (primitive objects) can be formed at certain image scales configuring a
relationship between ground objects and image objects. Ground objects refer to realworld objects or areas of specific land cover class connected by complex spatial and
contextual relations. Image objects, on the other hand, refer to knowledge-free areas
(primitive objects) of an image that a segmentation algorithm provides based on
various criteria. In order to configure a relationship between ground and image
objects, a multiscale knowledge representation is needed, which is usually provided
by the object-oriented paradigm. With such an object-based multiscale representation
and analysis, which is based on certain hierarchically structured rules, the relationship
between the different scales of the spatial entities is described.
During the last decade, a number of object-based image analysis software were
developed in the form of proprietary software (eCognition, ENVI FX, ERDAS
Objective) (Baatz and Schape, 2000) or in the form of free software (Orfeo Toolbox,
EDISON, MSEG) (Tzotsos and Argialas, 2006; Inglada and Christophe, 2009;
Christophe and Inglada, 2009), enabling the broad application on various engineering
and environmental remote sensing studies (Benz et al., 2004; Zhou et al., 2009;
Dragut et al., 2009; Blaschke, 2010; Mladinich et al., 2010). In all cases, the challenge
was to construct an efficient scale-space object representation through certain multiscale (region merging or other) segmentation techniques (Blaschke et al., 2004;
Carleer et al., 2005; Jimenez et al., 2005; Neubert et al., 2006; Tzotsos and Argialas,
2006), which partition the image on several regions/objects, based on the spectral
homogeneity in a local neighborhood. In addition to the spectral homogeneity
172
MULTISCALE SEGMENTATION AND CLASSIFICATION