specific test, it was impossible for those algorithms to be applied to the full spectral
resolution of the hyperspectral data set, since both algorithms are not designed to work
on a large number of image bands. For this test, a subset of bands were used to derive
the mean-shift and watershed results (Figures 9.6b,c). Both those algorithms produced similar results for this test. A mixture of small- and large-scaled objects were
obtained at the same time, with the watershed being more accurate in this case,
providing better results in edge areas (Figure 9.6c).
On the other hand, MSEG is designed to be applied to images of any spectral
resolution, up to 65,535 bands. It is demonstrated in Figure 9.6d that the simple MSEG
algorithm is performing very well given that the default scale parameter is easily reached
(since it is a heterogeneity threshold) with a big number of bands contributing to object
heterogeneity. Therefore, this is causing oversegmentation (Figure 9.6d) that cannot be
considered a major problem; rather it is an effect caused by the nature of this data set.
After application of a strong simplification of the AML filtering, the results were
improved (Figure 9.6e). In all cases of MSEG application it is shown that MSEG respects
the scale of the image objects in a better way that the other algorithms tested. This is very
crucial for OBIA applications, especially when multiscale approaches are necessary.
Finally, the enhanced MSEG algorithm was tested in Figure 9.6f, and the results were
obviously better that other approaches shown here. The edge information was preserved
(Figure 9.6f) and the size of the image objects still is similar for all objects on this scale.
9.5 CONCLUSIONS AND FUTURE PERSPECTIVES
A new object-based image analysis framework was proposed and developed in this
research based on advanced edge features incorporated in a multiscale region merging
algorithm. Advanced scale-space representations were used in order to avoid tuning
of segmentation and feature extraction parameters, and a kernel-based classification
was implemented to complete the OBIA framework. The developed image segmentation algorithm was shown to work on any type of remote sensing data, outperforming some widely used segmentation algorithms in some cases. The
improvement of the MSEG segmantation results was demonstrated, and the edge
enhancements were shown to make the algorithm robust and generic for multiscale
OBIA applications. The performed qualitative and quantitative evaluation reported
that the developed algorithm outperformed previous efforts, regarding both the
construction of the object representations and the classification results.
In terms of performance, the proposed improvement of the MSEG algorithm had
a major impact on execution times. For the very high resolution airborne image
(Figure 9.1a), the execution time for the original MSEG algorithm was 7–8 sec on a
3-GHz PC with 4 GB of RAM running a GNU/Linux operating system. With the
edge-based optimization, the number of possible object merges dropped significantly
per segmentation pass, and the execution time was around 4–5 sec. Still, there is room
for speed optimization improvements since for the same image the execution times
were 2–3 sec for Multiresolution segmentation (eCognition) and 3–4 sec for the mean
shift (Orfeo Toolbox).
CONCLUSIONS AND FUTURE PERSPECTIVES
191
resolution of the hyperspectral data set, since both algorithms are not designed to work
on a large number of image bands. For this test, a subset of bands were used to derive
the mean-shift and watershed results (Figures 9.6b,c). Both those algorithms produced similar results for this test. A mixture of small- and large-scaled objects were
obtained at the same time, with the watershed being more accurate in this case,
providing better results in edge areas (Figure 9.6c).
On the other hand, MSEG is designed to be applied to images of any spectral
resolution, up to 65,535 bands. It is demonstrated in Figure 9.6d that the simple MSEG
algorithm is performing very well given that the default scale parameter is easily reached
(since it is a heterogeneity threshold) with a big number of bands contributing to object
heterogeneity. Therefore, this is causing oversegmentation (Figure 9.6d) that cannot be
considered a major problem; rather it is an effect caused by the nature of this data set.
After application of a strong simplification of the AML filtering, the results were
improved (Figure 9.6e). In all cases of MSEG application it is shown that MSEG respects
the scale of the image objects in a better way that the other algorithms tested. This is very
crucial for OBIA applications, especially when multiscale approaches are necessary.
Finally, the enhanced MSEG algorithm was tested in Figure 9.6f, and the results were
obviously better that other approaches shown here. The edge information was preserved
(Figure 9.6f) and the size of the image objects still is similar for all objects on this scale.
9.5 CONCLUSIONS AND FUTURE PERSPECTIVES
A new object-based image analysis framework was proposed and developed in this
research based on advanced edge features incorporated in a multiscale region merging
algorithm. Advanced scale-space representations were used in order to avoid tuning
of segmentation and feature extraction parameters, and a kernel-based classification
was implemented to complete the OBIA framework. The developed image segmentation algorithm was shown to work on any type of remote sensing data, outperforming some widely used segmentation algorithms in some cases. The
improvement of the MSEG segmantation results was demonstrated, and the edge
enhancements were shown to make the algorithm robust and generic for multiscale
OBIA applications. The performed qualitative and quantitative evaluation reported
that the developed algorithm outperformed previous efforts, regarding both the
construction of the object representations and the classification results.
In terms of performance, the proposed improvement of the MSEG algorithm had
a major impact on execution times. For the very high resolution airborne image
(Figure 9.1a), the execution time for the original MSEG algorithm was 7–8 sec on a
3-GHz PC with 4 GB of RAM running a GNU/Linux operating system. With the
edge-based optimization, the number of possible object merges dropped significantly
per segmentation pass, and the execution time was around 4–5 sec. Still, there is room
for speed optimization improvements since for the same image the execution times
were 2–3 sec for Multiresolution segmentation (eCognition) and 3–4 sec for the mean
shift (Orfeo Toolbox).
CONCLUSIONS AND FUTURE PERSPECTIVES
191