extends histogram thresholding into multidimensions (i.e., when more than one
feature is used); (3) region-based approaches (Definiens Imaging, 2005) that
include region growing, region splitting, region merging, or their combinations;
and (4) edge-based methods (Shih and Cheng, 2005) that first detect edges within an
image. The most widely used image segmentation algorithm is perhaps the
multiresolution algorithm used in eCognition software. This algorithm is a bottom-up, region-merging algorithm based on the fractal net evolution approach
(FNEA) (Baatz and Schäpe, 2000; Baatz et al., 2004). An overview of different
segmentation techniques and related issues can be found in Baatz and Schäpe
(2000) and Schöpfer et al. (2008).
OBIA might be more appropriate than pixel-based analysis in many applications,
especially when high-spatial-resolution data were used. One reason is that OBIA can
utilize not only spectral information but also other information such as the shape,
texture, and contextual relationships of image objects, which are called object-based
metrics. Another feature of OBIA in the remote sensing classification literature is that
OBIA-based classification results are relatively free from salt-and-pepper noise,
which is a common problem in pixel-based analysis (Lillesand et al., 2008).
Numerous studies have compared OBIA with traditional pixel-based approaches
and many found that OBIA generally produced higher accuracy than pixel-based
methods (Im et al., 2008b; Jensen et al., 2006; Myint et al., 2011; Vieira et al., 2012;
Whiteside and Ahmad, 2005).
Although OBIA has already gained much popularity in remote sensing applications, there are several issues that need further exploration. Successful OBIA requires
that image segmentation produces image objects of high quality (Baatz and Schäpe,
2000). As new or improved image segmentation methods continue to be introduced,
the quantitative assessment of segmented results has gained great interest. Many
studies examined the quality of image objects simply based on visual inspection of the
image objects or accuracy of a subsequent analysis such as classification (Ouyang
et al., 2011). However, some studies argue that the quantitative assessment of image
objects in terms of real-world objects is crucial for successful classification or feature
extraction (Ke et al., 2010; Möller et al., 2007; Wang et al., 2010). The methods that
assess segmentation results typically compare segmented objects with the corresponding reference objects and measure similarity between them. However, there is
no single method that provides a standardized way to quantitatively assess image
segmentation results.
A second issue in OBIA that requires attention is object-based change detection.
Unlike OBIA based on single-date imagery, object-based change detection requires
additional considerations because the geometries of different dates of remote sensing
data can vary. Object-based change detection approaches are often broadly grouped
into (1) post-object-based classification comparison and (2) multitemporal image
object change classification (Stow, 2010). Chen et al. (2012) added another group—
image object change detection—that directly compares image objects between
multiple dates. While post-object-based classification comparison may be the most
widely used object-based change detection method, its performance is dependent on
the accuracy of the object-based classifications. Regardless of the approach to be used,
198
OPTIMUM SCALE IN OBJECT-BASED IMAGE ANALYSIS
feature is used); (3) region-based approaches (Definiens Imaging, 2005) that
include region growing, region splitting, region merging, or their combinations;
and (4) edge-based methods (Shih and Cheng, 2005) that first detect edges within an
image. The most widely used image segmentation algorithm is perhaps the
multiresolution algorithm used in eCognition software. This algorithm is a bottom-up, region-merging algorithm based on the fractal net evolution approach
(FNEA) (Baatz and Schäpe, 2000; Baatz et al., 2004). An overview of different
segmentation techniques and related issues can be found in Baatz and Schäpe
(2000) and Schöpfer et al. (2008).
OBIA might be more appropriate than pixel-based analysis in many applications,
especially when high-spatial-resolution data were used. One reason is that OBIA can
utilize not only spectral information but also other information such as the shape,
texture, and contextual relationships of image objects, which are called object-based
metrics. Another feature of OBIA in the remote sensing classification literature is that
OBIA-based classification results are relatively free from salt-and-pepper noise,
which is a common problem in pixel-based analysis (Lillesand et al., 2008).
Numerous studies have compared OBIA with traditional pixel-based approaches
and many found that OBIA generally produced higher accuracy than pixel-based
methods (Im et al., 2008b; Jensen et al., 2006; Myint et al., 2011; Vieira et al., 2012;
Whiteside and Ahmad, 2005).
Although OBIA has already gained much popularity in remote sensing applications, there are several issues that need further exploration. Successful OBIA requires
that image segmentation produces image objects of high quality (Baatz and Schäpe,
2000). As new or improved image segmentation methods continue to be introduced,
the quantitative assessment of segmented results has gained great interest. Many
studies examined the quality of image objects simply based on visual inspection of the
image objects or accuracy of a subsequent analysis such as classification (Ouyang
et al., 2011). However, some studies argue that the quantitative assessment of image
objects in terms of real-world objects is crucial for successful classification or feature
extraction (Ke et al., 2010; Möller et al., 2007; Wang et al., 2010). The methods that
assess segmentation results typically compare segmented objects with the corresponding reference objects and measure similarity between them. However, there is
no single method that provides a standardized way to quantitatively assess image
segmentation results.
A second issue in OBIA that requires attention is object-based change detection.
Unlike OBIA based on single-date imagery, object-based change detection requires
additional considerations because the geometries of different dates of remote sensing
data can vary. Object-based change detection approaches are often broadly grouped
into (1) post-object-based classification comparison and (2) multitemporal image
object change classification (Stow, 2010). Chen et al. (2012) added another group—
image object change detection—that directly compares image objects between
multiple dates. While post-object-based classification comparison may be the most
widely used object-based change detection method, its performance is dependent on
the accuracy of the object-based classifications. Regardless of the approach to be used,
198
OPTIMUM SCALE IN OBJECT-BASED IMAGE ANALYSIS