edge-based features at various scales, and (c) to evaluate the developed methodology
in various remote sensing data sets.
In Figure 9.1 a general overview of the developed method is presented. Starting
from the initial image (Figure 9.1a), Gaussian filtering at different scales demonstrates the loss of edge information due its isotropic character (Figures 9.1f,k,p).
These results are directly compared with the AML scale-space representations at
equivalent scales. One can observe in Figures 9.1b,g,l,q that edge information is
preserved while the initial image is simplified. For example, in the tile roof of the
building the single tiles are more difficult to distinguish as scale increases. The
results from the application of the standard MSEG algorithm on the simplified
images using the same scale parameter value (100) are shown as well (Figures 9.1c,
h,m,r). The standard MSEG algorithm performs well across object boundaries but
produces over segmented results and the mean object size is increasing along with
scale. The edge and line feature extraction at various scales is demonstrated in
Figure 9.1 as well. The result from the application of the LSD in the original image
is shown in Figure 9.1d. This is an impressive result, demonstrating that LSD is
robust and works well for man-made objects, even if not all of the building sides
have been detected correctly. The results of the Canny edge detector are also
presented at different scales (Figures 9.1i,n,s). It can be observed that due to the
simplified data through the AML scale-space computation, more clear edge features
are detected which accurately describe object boundaries. Less false detections
have also been detected inside homogeneous regions as, for example, in the tile
roof region. Furthermore, results from the application of the improved MSEG
algorithm are presented in Figures 9.1e,j,o,t. The first result (Figure 9.1e) shows
how the developed algorithm is constrained by the detected LSD line features. On
the homogeneous regions there is not much difference, which is normal since the
same AML scale is used for both Figures 9.1c and e. The second result (Figure 9.1j)
shows how the Canny edges are preserved inside the roof segments and how the
improved segmentation method has been robustly constrained by edge information.
This result is better than Figure 9.1h, where image objects are over segmented and
arbitrarily set inside a homogeneous region of the image. The third result in Figure
9.1o shows that the combination of edge information with region merging in higher
scales is outperforming the standard MSEG algorithm (Figure 9.1m) at the same
AML scale and segmentation scale parameter. Moreover, the result in Figure 9.1t
shows how the scale space in combination with the edge-constrained segmentation
tackles the over-segmentation issue shown in Figure 9.1r.
These aforementioned results demonstrate that the developed method outperforms earlier efforts (Tzotsos and Argialas, 2008; Baatz and Schape, 2000; Tzotsos
et al., 2011). In addition, in order to further validate the developed algorithm’s
experimental results and demonstrate its performance under several type of data
sets and settings, a variety of remote sensing data have been selected with different
spatial and spectral characteristics. In the following sections, the developed method
was compared against previous research efforts (Tzotsos et al., 2011) and other
standard OBIA implementations implemented in Orfeo Toolbox (Inglada and
Christophe, 2009) and eCognition (Benz et al., 2004).
EVALUATION AND DISCUSSION
183
in various remote sensing data sets.
In Figure 9.1 a general overview of the developed method is presented. Starting
from the initial image (Figure 9.1a), Gaussian filtering at different scales demonstrates the loss of edge information due its isotropic character (Figures 9.1f,k,p).
These results are directly compared with the AML scale-space representations at
equivalent scales. One can observe in Figures 9.1b,g,l,q that edge information is
preserved while the initial image is simplified. For example, in the tile roof of the
building the single tiles are more difficult to distinguish as scale increases. The
results from the application of the standard MSEG algorithm on the simplified
images using the same scale parameter value (100) are shown as well (Figures 9.1c,
h,m,r). The standard MSEG algorithm performs well across object boundaries but
produces over segmented results and the mean object size is increasing along with
scale. The edge and line feature extraction at various scales is demonstrated in
Figure 9.1 as well. The result from the application of the LSD in the original image
is shown in Figure 9.1d. This is an impressive result, demonstrating that LSD is
robust and works well for man-made objects, even if not all of the building sides
have been detected correctly. The results of the Canny edge detector are also
presented at different scales (Figures 9.1i,n,s). It can be observed that due to the
simplified data through the AML scale-space computation, more clear edge features
are detected which accurately describe object boundaries. Less false detections
have also been detected inside homogeneous regions as, for example, in the tile
roof region. Furthermore, results from the application of the improved MSEG
algorithm are presented in Figures 9.1e,j,o,t. The first result (Figure 9.1e) shows
how the developed algorithm is constrained by the detected LSD line features. On
the homogeneous regions there is not much difference, which is normal since the
same AML scale is used for both Figures 9.1c and e. The second result (Figure 9.1j)
shows how the Canny edges are preserved inside the roof segments and how the
improved segmentation method has been robustly constrained by edge information.
This result is better than Figure 9.1h, where image objects are over segmented and
arbitrarily set inside a homogeneous region of the image. The third result in Figure
9.1o shows that the combination of edge information with region merging in higher
scales is outperforming the standard MSEG algorithm (Figure 9.1m) at the same
AML scale and segmentation scale parameter. Moreover, the result in Figure 9.1t
shows how the scale space in combination with the edge-constrained segmentation
tackles the over-segmentation issue shown in Figure 9.1r.
These aforementioned results demonstrate that the developed method outperforms earlier efforts (Tzotsos and Argialas, 2008; Baatz and Schape, 2000; Tzotsos
et al., 2011). In addition, in order to further validate the developed algorithm’s
experimental results and demonstrate its performance under several type of data
sets and settings, a variety of remote sensing data have been selected with different
spatial and spectral characteristics. In the following sections, the developed method
was compared against previous research efforts (Tzotsos et al., 2011) and other
standard OBIA implementations implemented in Orfeo Toolbox (Inglada and
Christophe, 2009) and eCognition (Benz et al., 2004).
EVALUATION AND DISCUSSION
183