various segmentation methods is performed. Initially, the standard MSEG segmentation algorithm is tested in Figure 9.2a at a scale of 100. The MSEG algorithm is
applied on a scale-space AML representation and the result is achieved without any
parameter tuning.
In Figure 9.2d results from the application of the developed edge-constrained
MSEG algorithm are demonstrated. The edge objects are not merged to the rest of the
image objects, and they remain unmerged until a final step concludes the segmentation procedure and produces the result in Figure 9.2e. A comparison of the developed
algorithm with mean-shift algorithm (Comaniciu and Meer, 2002) (Figure 9.2b)
shows that while in some image regions the mean shift can merge large parts of the
image into one object, it fails to do so in other areas of the same texture. This behavior
is not optimal and can lead to problems for classification steps since the mean object
size varies. On the other hand, multiresolution segmentation as provided by eCognition (Figure 9.2c) manages to segment the image with a homogeneous object size,
but it suffers from oversegmentation problems, especially on the roof objects. For
both algorithms the default values were used to avoid parameter tuning. The
developed method (Figures 9.2e,f) manages to obtain similar objects in size, which
can be very applicable in a multiscale OBIA approach to classification. The main
difference between Figure 9.2e and Figure 9.2f is that image edges are larger in
number as derived from the Canny algorithm, and this leads to better results in areas
that LSD (Figure 9.2f) has not detected any straight lines in the image. So the LSD
method is not very suitable for curved or small edges.
In order to evaluate the developed method and to show the advantages of edgeconstrained segmentation algorithm, a test similar to the one performed in Tzotsos
et al. (2011) was deployed. For this test a very high resolution aerial scanner image
was used with four spectral bands in order to perform full-scale object-based image
analysis tests. The initial image was segmented using a simple MSEG algorithm
without parameter tuning (default values of scale parameter 100, color 0.8, and shape
0.2 were selected). After primitive objects were obtained, a training set was given to a
kernel-based classifier (SVM) to perform learning based on the feature space
introduced by object spectral and shape properties. For this test four generic land
cover classes were used: vegetation, tile roofs, bright roofs, and asphaltlike materials.
A set of training samples/objects was introduced to the SVM and a classification was
performed. Using ground truth data, a quantitative evaluation was performed and a
confusion matrix is presented in Table 9.1. The accuracy of the object-based
classification was 88.07% similar to the results reported for this approach in Tzotsos
and Argialas (2008).
A similar approach was then followed for the same image, with the same
segmentation parameters and the same training and testing samples. This time, a
scale-space AML representation was used to provide anisotropic diffusion and
simplification of the initial data set. After SVM classification and evaluation of
results (Table 9.1) an overall accuracy of 89.29% was achieved, similar to the
accuracy reported in Tzotsos et al. (2011).
Finally the developed method was applied in a similar manner to the same data. A
scale-space AML method was used to simplify the initial data set. A segmentation
EVALUATION AND DISCUSSION
185
applied on a scale-space AML representation and the result is achieved without any
parameter tuning.
In Figure 9.2d results from the application of the developed edge-constrained
MSEG algorithm are demonstrated. The edge objects are not merged to the rest of the
image objects, and they remain unmerged until a final step concludes the segmentation procedure and produces the result in Figure 9.2e. A comparison of the developed
algorithm with mean-shift algorithm (Comaniciu and Meer, 2002) (Figure 9.2b)
shows that while in some image regions the mean shift can merge large parts of the
image into one object, it fails to do so in other areas of the same texture. This behavior
is not optimal and can lead to problems for classification steps since the mean object
size varies. On the other hand, multiresolution segmentation as provided by eCognition (Figure 9.2c) manages to segment the image with a homogeneous object size,
but it suffers from oversegmentation problems, especially on the roof objects. For
both algorithms the default values were used to avoid parameter tuning. The
developed method (Figures 9.2e,f) manages to obtain similar objects in size, which
can be very applicable in a multiscale OBIA approach to classification. The main
difference between Figure 9.2e and Figure 9.2f is that image edges are larger in
number as derived from the Canny algorithm, and this leads to better results in areas
that LSD (Figure 9.2f) has not detected any straight lines in the image. So the LSD
method is not very suitable for curved or small edges.
In order to evaluate the developed method and to show the advantages of edgeconstrained segmentation algorithm, a test similar to the one performed in Tzotsos
et al. (2011) was deployed. For this test a very high resolution aerial scanner image
was used with four spectral bands in order to perform full-scale object-based image
analysis tests. The initial image was segmented using a simple MSEG algorithm
without parameter tuning (default values of scale parameter 100, color 0.8, and shape
0.2 were selected). After primitive objects were obtained, a training set was given to a
kernel-based classifier (SVM) to perform learning based on the feature space
introduced by object spectral and shape properties. For this test four generic land
cover classes were used: vegetation, tile roofs, bright roofs, and asphaltlike materials.
A set of training samples/objects was introduced to the SVM and a classification was
performed. Using ground truth data, a quantitative evaluation was performed and a
confusion matrix is presented in Table 9.1. The accuracy of the object-based
classification was 88.07% similar to the results reported for this approach in Tzotsos
and Argialas (2008).
A similar approach was then followed for the same image, with the same
segmentation parameters and the same training and testing samples. This time, a
scale-space AML representation was used to provide anisotropic diffusion and
simplification of the initial data set. After SVM classification and evaluation of
results (Table 9.1) an overall accuracy of 89.29% was achieved, similar to the
accuracy reported in Tzotsos et al. (2011).
Finally the developed method was applied in a similar manner to the same data. A
scale-space AML method was used to simplify the initial data set. A segmentation
EVALUATION AND DISCUSSION
185