74
2: Raghuveer M. Rao, Manoj K. Arora
The successful implementation of the neural network classifier, however,
depends on a range of parameters related to the design of the neural network
architecture (Arora and Foody 1997). A recent review on setting of different
parameters in neutral network classification can be found in Kavzoglu and
Mather (2003). Moreover, neural networks are very slow in the learning phase
of the classification, which is a serious drawback when dealing with images of
very large sizes such as hyperspectral data. Further, apart from their complexities, neural network and other non-parametric approaches are restricted to
the analysis of data that is inherently in numerical form. Nevertheless, a GIS
database may consist of many spatial datasets that are non-numeric, which
may enhance the classification performance if they can be readily incorporated
(Richards and Jia 1999). Thus, a method involving qualitative reasoning would
be of particular value. The knowledge based approach offers great promise in
this regard. A number of decision tree (Hansen et al. 2000) and expert system
classifiers (Huang and Jensen 1997) have been developed to implement the
knowledge based approach. The applicability of these classifiers, however, depends upon the formation of suitable production rules, which are meaningful
and easy to understand. The generation of crisp, clear and objective rules is,
thus, the crux of the problem in any knowledge based approach.
2.6.4
Fuzzy Classification Algorithms
For a large country, mapping from remote sensing is often carried out at the
regional level. This necessitates the use of coarse spatial resolution images
from a number of satellite based sensors such as MODIS, MERIS and AVHRR
that provide data ranging from 250 m to 1.1 km spatial resolution. The images
obtained from these sensors are frequently contaminated by mixed pixels.
These are the pixels that do not represent a single homogeneous class but
instead two or more classes are present in a single pixel area. The result is large
variations in the spectral reflectance value of the pixel. Thus, occurrence of
mixed pixels in an image is a major problem.
The crisp classification algorithms force the mixed pixels to be allocated
to one and only one class thereby resulting in erroneous classification. The
problem may be resolved either by ignoring or removing the mixed pixels
from the classification process. This may, however, be undesirable since it may
result in loss of pertinent information hidden in these pixels. Moreover, since
the mixed pixel displays a composite spectral response, which may be dissimilar
to each of its component classes, the pixel may not be allocated to any of its
component classes. Therefore, error is likely to occur in the classification of
images that contain a large proportion of mixed pixels (Foody and Arora 1996).
Alternative approaches for classifying images dominated by mixed pixels are,
therefore, highly desirable.
The output of the maximum likelihood classification may be fuzzified to
represent multiple class memberships for each pixel. Thus, for example, the
a posteriori probabilities from MLC may reflect, to some extent, the class
composition of a mixed pixel (Foody 1992). Another widely used method to
2: Raghuveer M. Rao, Manoj K. Arora
The successful implementation of the neural network classifier, however,
depends on a range of parameters related to the design of the neural network
architecture (Arora and Foody 1997). A recent review on setting of different
parameters in neutral network classification can be found in Kavzoglu and
Mather (2003). Moreover, neural networks are very slow in the learning phase
of the classification, which is a serious drawback when dealing with images of
very large sizes such as hyperspectral data. Further, apart from their complexities, neural network and other non-parametric approaches are restricted to
the analysis of data that is inherently in numerical form. Nevertheless, a GIS
database may consist of many spatial datasets that are non-numeric, which
may enhance the classification performance if they can be readily incorporated
(Richards and Jia 1999). Thus, a method involving qualitative reasoning would
be of particular value. The knowledge based approach offers great promise in
this regard. A number of decision tree (Hansen et al. 2000) and expert system
classifiers (Huang and Jensen 1997) have been developed to implement the
knowledge based approach. The applicability of these classifiers, however, depends upon the formation of suitable production rules, which are meaningful
and easy to understand. The generation of crisp, clear and objective rules is,
thus, the crux of the problem in any knowledge based approach.
2.6.4
Fuzzy Classification Algorithms
For a large country, mapping from remote sensing is often carried out at the
regional level. This necessitates the use of coarse spatial resolution images
from a number of satellite based sensors such as MODIS, MERIS and AVHRR
that provide data ranging from 250 m to 1.1 km spatial resolution. The images
obtained from these sensors are frequently contaminated by mixed pixels.
These are the pixels that do not represent a single homogeneous class but
instead two or more classes are present in a single pixel area. The result is large
variations in the spectral reflectance value of the pixel. Thus, occurrence of
mixed pixels in an image is a major problem.
The crisp classification algorithms force the mixed pixels to be allocated
to one and only one class thereby resulting in erroneous classification. The
problem may be resolved either by ignoring or removing the mixed pixels
from the classification process. This may, however, be undesirable since it may
result in loss of pertinent information hidden in these pixels. Moreover, since
the mixed pixel displays a composite spectral response, which may be dissimilar
to each of its component classes, the pixel may not be allocated to any of its
component classes. Therefore, error is likely to occur in the classification of
images that contain a large proportion of mixed pixels (Foody and Arora 1996).
Alternative approaches for classifying images dominated by mixed pixels are,
therefore, highly desirable.
The output of the maximum likelihood classification may be fuzzified to
represent multiple class memberships for each pixel. Thus, for example, the
a posteriori probabilities from MLC may reflect, to some extent, the class
composition of a mixed pixel (Foody 1992). Another widely used method to