Overview of Image Processing
71
heuristics. We have explained just one clustering approach. There are other
approaches to be found in the literature (Gersho and Gray 1992).
An illustration of forming classes through clustering is shown in Fig. 2.10.
A color image has been divided into five classes through unsupervised clustering. A pseudo color image is then created where each class is assigned a color.
The resulting classes reflect to a large degree actual ground objects such as
water, buildings and fields.
2.6.3
Crisp Classification Algorithms
Clearly, the objective of classification is to allocate each pixel of an image to
one of the classes. This is known as per-pixel, crisp or hard classification.
Maximum likelihood decision rule is most commonly used for per pixel based
supervised classification when the data satisfy Gaussian distribution assumption. However, often the spectral properties of the classes are far from the
assumed distribution. For example, images acquired in complex and heterogeneous environments actually record the reflectance from a number of different
classes. Many pixels are, therefore, mixed because the boundaries of mutually
exclusive classes meet within a pixel (boundary pixels) or a small proportion
of the classes exist within the major classes (sub-pixel phenomenon). The occurrence of mixed pixels is a major problem and has been discussed under
fuzzy classification algorithms.
Another issue associated with the statistical algorithm is the selection of
training data of appropriate size for accurate estimation of parameters governing that algorithm. In the absence of sufficient number of pure pixels to define
training data (generally 10 to 30 times the number of bands (Swain and Davis
1978», the performance of parametric or statistical classification algorithms
may deteriorate. Further, these classification algorithms have limitations in
integrating remote sensing data with ancillary data. Digital classification performed solely on the basis of the statistical analysis of spectral reflectance
values may create problems in classifying certain areas. For example, remote
sensing data are an ideal source for land cover mapping in hilly areas as it
minimizes accessibility problems. Hilly areas, on the other hand, are covered
with shadows due to high peaks and ridges thereby resulting in the change of
the reflectance characteristics of the objects underneath. To reduce the effect of
shadows in remote sensing image, the inclusion of information from the digital
elevation model (DEM) can be advantageous (Arora and Mathur 2001). Therefore, the ancillary (non-spectral) information from other sources such as DEM
(Bruzzone et al. 1997; Saha et al. 2004), and/or geological and soil maps (Gong
et al. 1993) may be more powerful in characterizing the classes of interest.
Moreover, due to the widespread availability of GIS, digital spatial data have
become more accessible than before. Therefore, greater attention is now being
paid to the use of ancillary data in classification using remote sensing imagery,
where parametric classifiers may result into inappropriate classification.
Non-parametric classification may be useful when the data distribution
assumptions are not satisfied. A number of nonparametric classification al-
71
heuristics. We have explained just one clustering approach. There are other
approaches to be found in the literature (Gersho and Gray 1992).
An illustration of forming classes through clustering is shown in Fig. 2.10.
A color image has been divided into five classes through unsupervised clustering. A pseudo color image is then created where each class is assigned a color.
The resulting classes reflect to a large degree actual ground objects such as
water, buildings and fields.
2.6.3
Crisp Classification Algorithms
Clearly, the objective of classification is to allocate each pixel of an image to
one of the classes. This is known as per-pixel, crisp or hard classification.
Maximum likelihood decision rule is most commonly used for per pixel based
supervised classification when the data satisfy Gaussian distribution assumption. However, often the spectral properties of the classes are far from the
assumed distribution. For example, images acquired in complex and heterogeneous environments actually record the reflectance from a number of different
classes. Many pixels are, therefore, mixed because the boundaries of mutually
exclusive classes meet within a pixel (boundary pixels) or a small proportion
of the classes exist within the major classes (sub-pixel phenomenon). The occurrence of mixed pixels is a major problem and has been discussed under
fuzzy classification algorithms.
Another issue associated with the statistical algorithm is the selection of
training data of appropriate size for accurate estimation of parameters governing that algorithm. In the absence of sufficient number of pure pixels to define
training data (generally 10 to 30 times the number of bands (Swain and Davis
1978», the performance of parametric or statistical classification algorithms
may deteriorate. Further, these classification algorithms have limitations in
integrating remote sensing data with ancillary data. Digital classification performed solely on the basis of the statistical analysis of spectral reflectance
values may create problems in classifying certain areas. For example, remote
sensing data are an ideal source for land cover mapping in hilly areas as it
minimizes accessibility problems. Hilly areas, on the other hand, are covered
with shadows due to high peaks and ridges thereby resulting in the change of
the reflectance characteristics of the objects underneath. To reduce the effect of
shadows in remote sensing image, the inclusion of information from the digital
elevation model (DEM) can be advantageous (Arora and Mathur 2001). Therefore, the ancillary (non-spectral) information from other sources such as DEM
(Bruzzone et al. 1997; Saha et al. 2004), and/or geological and soil maps (Gong
et al. 1993) may be more powerful in characterizing the classes of interest.
Moreover, due to the widespread availability of GIS, digital spatial data have
become more accessible than before. Therefore, greater attention is now being
paid to the use of ancillary data in classification using remote sensing imagery,
where parametric classifiers may result into inappropriate classification.
Non-parametric classification may be useful when the data distribution
assumptions are not satisfied. A number of nonparametric classification al-