Geography Reference
In-Depth Information
2.3.2 The General Classification Techniques
Researchers have presented various approaches for image classification, which can
be divided into three general groups (Fig. 2.5 ) (Pal and Pal 1993 ).
Statistical classifiers: these are ideally suitable for data that have information
with an assumed theoretical model based distribution within each of the classes.
The representative algorithms for this group are: MLC; PPC; k-NNC and MDC.
Corresponding literature for these algorithms can be found in Swain and Davis
( 1978 ) and Hastie et al. ( 2001 ). Fuzzy mathematical approaches: Zadeh ( 1965 )
presented the concept of fuzzy sets in which unclear knowledge can be used to
delineate a result. Artificial intelligence (AI): here, supervised classification
approaches were developed from the starting of the 1970s, with the well-known
''Arch Concept Learning'' problem presented by Winston ( 1975 ). These methods
based on the learning from descriptions of a constructive pattern, and therefore
gave up the value-attribute based model that was used in other methods. AI-type
models were constructed based on semantic networks and on predicate logic.
Liu and Mason ( 2009 ) summarized the classification approaches in seven cat-
egories: unsupervised classification; supervised classification; hybrid classifica-
tion; single pass classification; iterative classification; image scanning
classification; and feature space partition. In most cases, image classification
approaches included: supervised and unsupervised; parametric and nonparametric;
hard and soft (fuzzy) classification; per-pixel, sub-pixel, object-oriented and per-
field; spectral classifiers, contextual classifiers and spectral-contextual classifiers;
or combinative approaches of multiple classifiers (Lu and Weng 2007 ). This article
presents: present practices; remotely sensed data classification troubles and sce-
narios. It highlighted the main advanced classification approaches, in addition to
those techniques that can improve the at-end classification accuracy.
Unsupervised classification: when insufficient ground reference information is
available (e.g., field work measurements) about the characteristics of specific
classes for classification processes, an unsupervised classification is used to
identify natural homogeneous groups (clusters) within the remotely sensed data.
Unsupervised classification approaches are based on non-parametric statistical
approaches, such as Iterative Self-Organizing Data Analysis Technique (ISO-
DATA) (Tou and Gonzalez 1974 ), K-means-clustering (Johnson and Wichern
1988 ) algorithms, and the advanced unsupervised neural classification method
Self-Organizing feature Mapping (SOM) (Kohonen 1989 ). In this approach, the
Fig. 2.5 Major approaches
for image segmentation/
classification (Source
modified after Pal and Pal
1993 )
Image segmentation/ classification approaches
Statistical
classifier
Fuzzy mathematical
approaches
Artificial
intelligence
Search WWH ::




Custom Search