studies used a different optimum scale for each type of land cover or land use
objects (Huang and Qi, 2010; Mallinis et al., 2008; Wang et al., 2009). Ke et al.
(2010) argued that for a particular study there may be no single optimum scale and
that a range of optimum scales may exist.
There have been efforts to develop a feasible framework for the selection of
optimum scale that can be applied to a wide range of OBIA applications. Lowe and
Guo (2011) investigated an optimum scale parameter in OBIA using semivariogram
and spatial autocorrelation concepts. Multiple scales were tested and the optimal
scale was identified when the average distance between neighboring image object
centroids was near the semivariogram lag distance. Estimation of scale parameter
by Drǎ gut et al. (2010) provides useful information in selecting a suitable range of
scales for image segmentation using local variance of object heterogeneity
within a scene. Crevier (2008) proposed a natural extension to the concept of
precision-recall curves and computationally efficient match measures to compare
multiple segmentations and identify optimum scale. In particular, the proposed
statistical tool can be effectively used to discriminate between segmentations of
different images.
In general, there appears to be a consensus on the need of a universally
applicable method to identify optimum scale in OBIA and to compare scales
among segmentations of different remote sensing data. Image segmentation
algorithms have different approaches (i.e., scales) that control the size of image
objects. Depending on the segmentation algorithm used, many factors can affect
object size, including pixel size, number of input layers, weighting scheme, or
radiometric resolution. In order to facilitate the selection of optimum scale or
compare segmentation results from different algorithms or data (Figure 10.4), it
would be good to provide additional metrics along with the arbitrary scale values
from the software tool used. Such metrics include the number of objects, the
number of objects divided by total area, average area of objects (e.g., in square
meters), and the average area of objects divided by pixel size. Metrics that
integrate the information could be used as surrogate variables to arbitrary scale
values that directly came from the software tools used. That way, it would be more
straightforward and intuitive to deal with the optimum scale concept in OBIA in
that the metrics can be compared even when different algorithms or data are used.
In fact, some studies provided additional information related to scale such as
the average area and number of image objects. For example, Ke et al. (2010)
provided the number of segmented objects of interest by scale and AguirreGutiérrez et al. (2012) reported the number of objects, area of biggest object, and
average object area.
According to Blaschke (2010), there is no recognizable relationship between the
scale parameter used in an image segmentation process and the spatial measures
specific to the image objects of interest. In addition, a single scale would not
successfully characterize the multitude of different image components; thus there
is a strong need for a universally applicable strategy to determine optimum scales for
different image components such as relatively homogeneous forests and heterogeneous urban features.
OPTIMUM SCALES IN OBIA
205
objects (Huang and Qi, 2010; Mallinis et al., 2008; Wang et al., 2009). Ke et al.
(2010) argued that for a particular study there may be no single optimum scale and
that a range of optimum scales may exist.
There have been efforts to develop a feasible framework for the selection of
optimum scale that can be applied to a wide range of OBIA applications. Lowe and
Guo (2011) investigated an optimum scale parameter in OBIA using semivariogram
and spatial autocorrelation concepts. Multiple scales were tested and the optimal
scale was identified when the average distance between neighboring image object
centroids was near the semivariogram lag distance. Estimation of scale parameter
by Drǎ gut et al. (2010) provides useful information in selecting a suitable range of
scales for image segmentation using local variance of object heterogeneity
within a scene. Crevier (2008) proposed a natural extension to the concept of
precision-recall curves and computationally efficient match measures to compare
multiple segmentations and identify optimum scale. In particular, the proposed
statistical tool can be effectively used to discriminate between segmentations of
different images.
In general, there appears to be a consensus on the need of a universally
applicable method to identify optimum scale in OBIA and to compare scales
among segmentations of different remote sensing data. Image segmentation
algorithms have different approaches (i.e., scales) that control the size of image
objects. Depending on the segmentation algorithm used, many factors can affect
object size, including pixel size, number of input layers, weighting scheme, or
radiometric resolution. In order to facilitate the selection of optimum scale or
compare segmentation results from different algorithms or data (Figure 10.4), it
would be good to provide additional metrics along with the arbitrary scale values
from the software tool used. Such metrics include the number of objects, the
number of objects divided by total area, average area of objects (e.g., in square
meters), and the average area of objects divided by pixel size. Metrics that
integrate the information could be used as surrogate variables to arbitrary scale
values that directly came from the software tools used. That way, it would be more
straightforward and intuitive to deal with the optimum scale concept in OBIA in
that the metrics can be compared even when different algorithms or data are used.
In fact, some studies provided additional information related to scale such as
the average area and number of image objects. For example, Ke et al. (2010)
provided the number of segmented objects of interest by scale and AguirreGutiérrez et al. (2012) reported the number of objects, area of biggest object, and
average object area.
According to Blaschke (2010), there is no recognizable relationship between the
scale parameter used in an image segmentation process and the spatial measures
specific to the image objects of interest. In addition, a single scale would not
successfully characterize the multitude of different image components; thus there
is a strong need for a universally applicable strategy to determine optimum scales for
different image components such as relatively homogeneous forests and heterogeneous urban features.
OPTIMUM SCALES IN OBIA
205