144
B.G.H. Gorte
Various clustering algorithms exist. Usually, they are not completely automatic:
The user must specify some parameters, such as the number of clusters that he approximately wants, the maximum cluster size (in the feature space!), the minimum
distance (also in the feature space) that is allowed between different clusters etc.
The software "builds" clusters as it is scanning through the image. Typically, when
a cluster becomes larger than the maximum size, it is split into two clusters; on the
other hand, when two clusters get nearer to each other than the minimum distance,
they are merged into one.
Supervised Classification. In order to make a classifier work with thematic (instead of spectral) classes, some knowledge is needed about the relationship between
classes and feature vectors.
Theoretically, this knowledge could come from a data base in which the relationships between (thematic) classes and feature vectors is stored. It is tempting to
assume that in the past enough images of each kind of sensor have been analyzed as
to know the spectral characteristics of all relevant classes.
Unfortunately, the observed feature vectors in a particular image are influenced
by a large amount of other factors than land cover, such as: atmospheric conditions,
sun angle (as function of latitude/time of day/date and as function of terrain relief),
soil type, soil humidity, vegetation growing stage, wind, etc.
Trying to take all these influences into account is practically impossible, even if
vast amounts of additionally required data (OEMs, soil maps etc.) were available.
Much more widely used are, therefore, classification methods where the relationship between reflections and cover classes is established for each separate image
under consideration. Supervised classification is divided into two phases: a training
phase, where the user trains the computer by telling for a limited number of pixels to
what classes they belong in this particular image, followed by the decision phase,
where the computer assigns a class label to all (other) image pixels, by looking for
each pixel to which of the trained classes this pixel is most similar.
Training. During the training phase the user decides what classes to use. About
each class he needs some ground truth, a number of places in the image area that are
known to belong to that class. This knowledge must have been acquired beforehand,
for instance as a result of fieldwork, or from an existing map, assuming that in some
areas the class membership has not changed since the map was produced.
Suitable training software allows to display the image on the computer screen, to
indicate the places where the ground truth is known and to enter the corresponding
class names. Meanwhile the feature space is visible, where the training samples are
plotted using distinct colors for the different classes, such that the user can judge
whether the classes are spectrally distinguishable, whether each class corresponds
to only one spectral cluster etc.
Decision Making. It is the task of the decision making algorithm to make a partitioning of the feature space, according to the training data. For every possible feature
vector in the feature space (at least for those that actually occur in the image), the
program decides to which of the sets of training pixels this feature vector is most
B.G.H. Gorte
Various clustering algorithms exist. Usually, they are not completely automatic:
The user must specify some parameters, such as the number of clusters that he approximately wants, the maximum cluster size (in the feature space!), the minimum
distance (also in the feature space) that is allowed between different clusters etc.
The software "builds" clusters as it is scanning through the image. Typically, when
a cluster becomes larger than the maximum size, it is split into two clusters; on the
other hand, when two clusters get nearer to each other than the minimum distance,
they are merged into one.
Supervised Classification. In order to make a classifier work with thematic (instead of spectral) classes, some knowledge is needed about the relationship between
classes and feature vectors.
Theoretically, this knowledge could come from a data base in which the relationships between (thematic) classes and feature vectors is stored. It is tempting to
assume that in the past enough images of each kind of sensor have been analyzed as
to know the spectral characteristics of all relevant classes.
Unfortunately, the observed feature vectors in a particular image are influenced
by a large amount of other factors than land cover, such as: atmospheric conditions,
sun angle (as function of latitude/time of day/date and as function of terrain relief),
soil type, soil humidity, vegetation growing stage, wind, etc.
Trying to take all these influences into account is practically impossible, even if
vast amounts of additionally required data (OEMs, soil maps etc.) were available.
Much more widely used are, therefore, classification methods where the relationship between reflections and cover classes is established for each separate image
under consideration. Supervised classification is divided into two phases: a training
phase, where the user trains the computer by telling for a limited number of pixels to
what classes they belong in this particular image, followed by the decision phase,
where the computer assigns a class label to all (other) image pixels, by looking for
each pixel to which of the trained classes this pixel is most similar.
Training. During the training phase the user decides what classes to use. About
each class he needs some ground truth, a number of places in the image area that are
known to belong to that class. This knowledge must have been acquired beforehand,
for instance as a result of fieldwork, or from an existing map, assuming that in some
areas the class membership has not changed since the map was produced.
Suitable training software allows to display the image on the computer screen, to
indicate the places where the ground truth is known and to enter the corresponding
class names. Meanwhile the feature space is visible, where the training samples are
plotted using distinct colors for the different classes, such that the user can judge
whether the classes are spectrally distinguishable, whether each class corresponds
to only one spectral cluster etc.
Decision Making. It is the task of the decision making algorithm to make a partitioning of the feature space, according to the training data. For every possible feature
vector in the feature space (at least for those that actually occur in the image), the
program decides to which of the sets of training pixels this feature vector is most