Lehmann (2008) but also was developed independently in MSEG (Tzotsos and
Argialas, 2006). The computed properties are bound to each object by a unique
identifier within the object hierarchy of the image. Some of the objects are selected
as samples and their properties formed a training set for the SVM (Tzotsos and
Argialas, 2008).
The SVM classifier seeks to find the optimal separating hyperplane between
classes by focusing on the training data (support vectors) that are placed at the edge of
the class descriptors. Training data other than support vectors are discarded. Thus, not
only an optimal hyperplane is fitted but less training samples are effectively used as
well (Tzotsos and Argialas, 2008). This method works very well for classes that are
linearly separable. In the case that image classes are not linearly separable, the SVM
maps the feature space into a higher dimensionality using kernels (Vapnik, 1998;
Theodoridis and Koutroumbas, 2003) and then separates classes in that new feature
space forming the support vectors.
Since the SVM classification method was initially designed for binary classification problems, a heuristic one-against-one strategy was employed for multiclass
classification (Hsu and Lin, 2002). Many binary classifiers were applied for each pair
of classes and for every object of the image and then a max-win operator determined
the final classification of the object. An n-fold cross-validation scheme was also used
in order to define the parameters needed for the training procedure. After specification
of parameters, the training set was used to train the classifier. Primitive image objects
were classified using this trained SVM algorithm using the one-against-one strategy.
Finally a quality assessment took place using ground truth data that were not used
during the training procedure.
The above classification procedure was repeated for all scales (values starting from
1 to a maximum value defined by the user) in the scale-space representation in order to
determine the best classification accuracy, as proposed in Tzotsos et al. (2011). The
best scale was then selected based on the best classification accuracy. After determining the best scale to perform classification, a final classification step took place to
produce the optimal results.
To sum up, the initial data set was simplified and a successive series of simplified
images were constructed forming a nonlinear scale space. The simplified imagery that
was derived was then used to extract edge and line features using advanced methods.
An edge-enhanced multiscale image segmentation algorithm was employed to
provide primitive image objects from the scale-space images without the tuning
of any standard parameter. Finally, a classification step was performed to complete the
OBIA tasks and to evaluate the developed method.
9.4 EVALUATION AND DISCUSSION
As already stated, the overall objective of the present research was (a) to introduce a
generic and robust framework able to process any kind of remote sensing data without
tuning any parameters (scale, texture, color, etc.) during the computation, (b) to
introduce a multiscale segmentation algorithm which is constrained by advanced
182
MULTISCALE SEGMENTATION AND CLASSIFICATION
Argialas, 2006). The computed properties are bound to each object by a unique
identifier within the object hierarchy of the image. Some of the objects are selected
as samples and their properties formed a training set for the SVM (Tzotsos and
Argialas, 2008).
The SVM classifier seeks to find the optimal separating hyperplane between
classes by focusing on the training data (support vectors) that are placed at the edge of
the class descriptors. Training data other than support vectors are discarded. Thus, not
only an optimal hyperplane is fitted but less training samples are effectively used as
well (Tzotsos and Argialas, 2008). This method works very well for classes that are
linearly separable. In the case that image classes are not linearly separable, the SVM
maps the feature space into a higher dimensionality using kernels (Vapnik, 1998;
Theodoridis and Koutroumbas, 2003) and then separates classes in that new feature
space forming the support vectors.
Since the SVM classification method was initially designed for binary classification problems, a heuristic one-against-one strategy was employed for multiclass
classification (Hsu and Lin, 2002). Many binary classifiers were applied for each pair
of classes and for every object of the image and then a max-win operator determined
the final classification of the object. An n-fold cross-validation scheme was also used
in order to define the parameters needed for the training procedure. After specification
of parameters, the training set was used to train the classifier. Primitive image objects
were classified using this trained SVM algorithm using the one-against-one strategy.
Finally a quality assessment took place using ground truth data that were not used
during the training procedure.
The above classification procedure was repeated for all scales (values starting from
1 to a maximum value defined by the user) in the scale-space representation in order to
determine the best classification accuracy, as proposed in Tzotsos et al. (2011). The
best scale was then selected based on the best classification accuracy. After determining the best scale to perform classification, a final classification step took place to
produce the optimal results.
To sum up, the initial data set was simplified and a successive series of simplified
images were constructed forming a nonlinear scale space. The simplified imagery that
was derived was then used to extract edge and line features using advanced methods.
An edge-enhanced multiscale image segmentation algorithm was employed to
provide primitive image objects from the scale-space images without the tuning
of any standard parameter. Finally, a classification step was performed to complete the
OBIA tasks and to evaluate the developed method.
9.4 EVALUATION AND DISCUSSION
As already stated, the overall objective of the present research was (a) to introduce a
generic and robust framework able to process any kind of remote sensing data without
tuning any parameters (scale, texture, color, etc.) during the computation, (b) to
introduce a multiscale segmentation algorithm which is constrained by advanced
182
MULTISCALE SEGMENTATION AND CLASSIFICATION