the probability distribution of the training data sets (e.g., in ML), which might not
always coincide with reality. Other classifiers such as ANNs may require a significant
amount of effort in terms of calibration and fine tuning of neural nets before obtaining
a satisfactory level of classification accuracy. It is also generally argued that such
classification approaches do not make use of the spatial concept of the imagery, such
as textural or contextual information present in a remotely sensed imagery (Yan et al.,
2006). Finally, an important advantage of both classification techniques implemented
herein is that for both their implementation is not confined by the so-called Hughes
phenomenon known also as the “curse of dimensionality” (Hughes, 1968). This
occurs when the number of training samples is limited compared with the number of
input features (and thus of classifier parameters) and results in an accuracy loss as the
function of the data dimensionality increases (Dalponte et al., 2009; Zhang and Ma,
2009). Although this challenge does not exist in the classification approaches
examined in our study, this phenomenon can be important when parametric classifiers
(such as ML) are implemented with hyperspectral imagery.
15.3.2.2 Object-Based Classification Object-based methods were introduced in
the early 1970s (de Kok et al., 1999). They are based on the concept that information
necessary to interpret an image is represented not in single pixels but in meaningful
image objects. The first step in object-based classification is image segmentation
based on which remote sensing imagery is divided into regions where each is
homogeneous and no two adjustment regions are homogeneous (Pal and Pal,
1993). In the next step, the segmented image is used along with textural and
contextual information as well as the spectral information to produce a thematic
map of land use/cover. A hierarchy of levels of segmentation, ranging from a few
large objects to many smaller objects, can be obtained with each object belonging to a
superobject at a higher level of segmentation and smaller objects at a lower level of
segmentation. The spectral characteristics of each object include not only an average
value for each band from the pixels participating in the object but also statistical
values such as minimum, maximum, and standard deviation regarding each band. In
addition, the objects are described by shape, size, tone, texture, compactness, and
other characteristics describing the spatial features of the object (Bock et al., 2005).
All of those variables can be used in the classification process to assist in the
discrimination of the objects and their correct assignment to the land use/cover
classes. The characteristic of inclusion of spectral as well as spatial information of
image objects constitutes the main advantage of the object-based classifiers. Most
importantly, consideration of object attributes (e.g., shape, heterogeneity) results in
the reduction of the “salt-and-pepper” effect and “edge” usually seen in pixel-based
classification. Also, it should be noted that in comparison to pixel-based classifiers,
object-based classification generally requires higher user expertise to be implemented.
Also, generally, the accuracy of segmentation directly affects the performance of the
object-based classification, and several studies have shown that only good segmentation results can lead to object-oriented image classification outperforming pixelbased classification (e.g., Yan et al., 2006).
306
HYPERSPECTRAL REMOTE SENSING WITH EMPHASIS ON LAND COVER MAPPING
always coincide with reality. Other classifiers such as ANNs may require a significant
amount of effort in terms of calibration and fine tuning of neural nets before obtaining
a satisfactory level of classification accuracy. It is also generally argued that such
classification approaches do not make use of the spatial concept of the imagery, such
as textural or contextual information present in a remotely sensed imagery (Yan et al.,
2006). Finally, an important advantage of both classification techniques implemented
herein is that for both their implementation is not confined by the so-called Hughes
phenomenon known also as the “curse of dimensionality” (Hughes, 1968). This
occurs when the number of training samples is limited compared with the number of
input features (and thus of classifier parameters) and results in an accuracy loss as the
function of the data dimensionality increases (Dalponte et al., 2009; Zhang and Ma,
2009). Although this challenge does not exist in the classification approaches
examined in our study, this phenomenon can be important when parametric classifiers
(such as ML) are implemented with hyperspectral imagery.
15.3.2.2 Object-Based Classification Object-based methods were introduced in
the early 1970s (de Kok et al., 1999). They are based on the concept that information
necessary to interpret an image is represented not in single pixels but in meaningful
image objects. The first step in object-based classification is image segmentation
based on which remote sensing imagery is divided into regions where each is
homogeneous and no two adjustment regions are homogeneous (Pal and Pal,
1993). In the next step, the segmented image is used along with textural and
contextual information as well as the spectral information to produce a thematic
map of land use/cover. A hierarchy of levels of segmentation, ranging from a few
large objects to many smaller objects, can be obtained with each object belonging to a
superobject at a higher level of segmentation and smaller objects at a lower level of
segmentation. The spectral characteristics of each object include not only an average
value for each band from the pixels participating in the object but also statistical
values such as minimum, maximum, and standard deviation regarding each band. In
addition, the objects are described by shape, size, tone, texture, compactness, and
other characteristics describing the spatial features of the object (Bock et al., 2005).
All of those variables can be used in the classification process to assist in the
discrimination of the objects and their correct assignment to the land use/cover
classes. The characteristic of inclusion of spectral as well as spatial information of
image objects constitutes the main advantage of the object-based classifiers. Most
importantly, consideration of object attributes (e.g., shape, heterogeneity) results in
the reduction of the “salt-and-pepper” effect and “edge” usually seen in pixel-based
classification. Also, it should be noted that in comparison to pixel-based classifiers,
object-based classification generally requires higher user expertise to be implemented.
Also, generally, the accuracy of segmentation directly affects the performance of the
object-based classification, and several studies have shown that only good segmentation results can lead to object-oriented image classification outperforming pixelbased classification (e.g., Yan et al., 2006).
306
HYPERSPECTRAL REMOTE SENSING WITH EMPHASIS ON LAND COVER MAPPING