within a digital scene by many individual pixels; whereas low resolution (L-res)
implies that a single pixel represents the integrated signal of many smaller real-world
objects (e.g., leaves and branches within a single 30-m pixel) (Woodcock and
Strahler, 1987). Because a H-res scene typically models complex land cover
composed of varying sized, shaped, and spatially distributed image objects, there
is often no single “optimal” scale for analysis; rather there are many optimal scales
specific to the image objects that exist/emerge within a scene (Hay et al., 1994, 1997,
2002a,b; Marceau et al., 1994; Chen et al., 2011; Powers et al., 2012). Therefore, we
support the idea that multiscale analysis (including the generation of scale hierarchies)
should be guided by the innate spatial resolution of the salient landscape objects
composing a scene. To achieve this, we describe the concept of linear scale-space and
blob feature detection applied to a high-resolution remote sensing image and integrate
these methods with ideas from hierarchical theories.
8.2 METHODS
The following methods were developed to explore the following questions: What
does a scale domain look like, where is it located, and how can we visualize it?
Multiscale analysis requires two primary components: (i) the generation of a
multiscale representation and (ii) a (multiscale) object delineation tool. To satisfy
these requirements, we describe the use of linear scale-space and blob-feature detection
(SS) [as defined by Lindeberg (1994) and adapted by Hay et al. (2002a,b)] for
generating a multiscale representation of a complex agroforest landscape (Hall
et al., 2004) and for automatically delineating dominant multiscale components
from this representation.
8.2.1 Study Site and Data Set
Building on the work of Hay et al. (2002a), the data used in this study are a
panchromatic (PAN) 500 ´ 500-pixel subimage of an IKONOS-2 (Geo) scene
acquired in September 2001 (Figure 8.3a). IKONOS-2 provides 11-bit multispectral
data in the red, green, blue, and near-infrared channels at 4.0 m spatial resolution and
an 11-bit panchromatic (PAN) channel at 1.0 m resolution. Since the PAN channel
covers a significant portion of the wavelengths represented by the four multispectral
channels, a well-known 4.0-km
2 subarea of the 1.0-m PAN image was selected and
resampled to 4.0 m. This resampled spatial resolution represents a trade-off between a
fine enough grain and a reasonable spatial extent with which to evaluate object
evolution over multiple scales. Resampling was conducted using object-specific
upscaling, which is considered a robust (object-based) upscaling technique (Hay
et al., 1997, 2001, 2003, 2004, 2005). Based on the computational demands required
by SS processing, all data were linearly resampled to 8 bits (prior to analysis).
Geographically, this study area (Figure 8.3b) represents a portion of the highly
fragmented agroforested landscape typical of the Haut Saint-Laurent region of
southwestern Québec, Canada (Bouchard and Domon, 1997). The vegetation in
148
VISUALIZING SCALE-DOMAIN MANIFOLDS
implies that a single pixel represents the integrated signal of many smaller real-world
objects (e.g., leaves and branches within a single 30-m pixel) (Woodcock and
Strahler, 1987). Because a H-res scene typically models complex land cover
composed of varying sized, shaped, and spatially distributed image objects, there
is often no single “optimal” scale for analysis; rather there are many optimal scales
specific to the image objects that exist/emerge within a scene (Hay et al., 1994, 1997,
2002a,b; Marceau et al., 1994; Chen et al., 2011; Powers et al., 2012). Therefore, we
support the idea that multiscale analysis (including the generation of scale hierarchies)
should be guided by the innate spatial resolution of the salient landscape objects
composing a scene. To achieve this, we describe the concept of linear scale-space and
blob feature detection applied to a high-resolution remote sensing image and integrate
these methods with ideas from hierarchical theories.
8.2 METHODS
The following methods were developed to explore the following questions: What
does a scale domain look like, where is it located, and how can we visualize it?
Multiscale analysis requires two primary components: (i) the generation of a
multiscale representation and (ii) a (multiscale) object delineation tool. To satisfy
these requirements, we describe the use of linear scale-space and blob-feature detection
(SS) [as defined by Lindeberg (1994) and adapted by Hay et al. (2002a,b)] for
generating a multiscale representation of a complex agroforest landscape (Hall
et al., 2004) and for automatically delineating dominant multiscale components
from this representation.
8.2.1 Study Site and Data Set
Building on the work of Hay et al. (2002a), the data used in this study are a
panchromatic (PAN) 500 ´ 500-pixel subimage of an IKONOS-2 (Geo) scene
acquired in September 2001 (Figure 8.3a). IKONOS-2 provides 11-bit multispectral
data in the red, green, blue, and near-infrared channels at 4.0 m spatial resolution and
an 11-bit panchromatic (PAN) channel at 1.0 m resolution. Since the PAN channel
covers a significant portion of the wavelengths represented by the four multispectral
channels, a well-known 4.0-km
2 subarea of the 1.0-m PAN image was selected and
resampled to 4.0 m. This resampled spatial resolution represents a trade-off between a
fine enough grain and a reasonable spatial extent with which to evaluate object
evolution over multiple scales. Resampling was conducted using object-specific
upscaling, which is considered a robust (object-based) upscaling technique (Hay
et al., 1997, 2001, 2003, 2004, 2005). Based on the computational demands required
by SS processing, all data were linearly resampled to 8 bits (prior to analysis).
Geographically, this study area (Figure 8.3b) represents a portion of the highly
fragmented agroforested landscape typical of the Haut Saint-Laurent region of
southwestern Québec, Canada (Bouchard and Domon, 1997). The vegetation in
148
VISUALIZING SCALE-DOMAIN MANIFOLDS