dqi is the nearest-neighbour distance, min is the minimum distance,
and dist is the Euclidian distance between the query pixel q and a
pixel xi belonging to a specific spectral class Si, which is a subset
of the entire image.
In step 4, the contextual feature vectors are classified using
clustering techniques that group the pixels with similar contextual
feature vectors to create new classes that are based on contextual
information. This classification can either be supervised or unsupervised.
Supervised classification entails identifying a number of
cluster centres using training areas. A basic example is an area of
forest that has been marked as a training area for the class forest
and an urban area that has been marked as urban. Based on the
contextual feature vectors covered by the training areas, a mean
cluster centre is calculated. All of the contextual feature vectors
are then compared with the mean cluster centres and are assigned
the closest centre. An example of unsupervised classification is the
common method of a moving K-mean cluster analysis in which the
user decides the number of clusters (K). For each cluster, a mean
vector is located within the multidimensional space created by
the contextual feature vectors. The dimensions are the distance
to each individual spectral class. Therefore, if there are ten classes,
then there are ten dimensions, and each pixel has a position in the
ten-dimensional space. The moving K-mean clustering algorithm
applied to the contextual feature can be defined as: