Theoretical Background and State of the Art - Historical Land Use/Land Cover Classification Using Remote Sensing - page 18

training samples position on the image (Mather 2004 ). These training samples
have to be homogeneous spectrally to represent specific LULC-classes. A super-
vised algorithm, after the training samples stage, uses the distribution of the
training samples for each class to assess density functions in the feature space
statistically and to divide the space into class regions (Fukunaga 1990 ). In other
words, the used image's processing software recognizes the spectral signature of
each training site based on its statistics, and then classifies the images in different
LULC-classes according to the applied classification algorithm (Jensen 2005 ).
Here, the information required from the training data differs from one algorithm to
another. The most general and used supervised approaches are: The Maximum
Likelihood Classifier (MLC) and the Minimum Distance Classifier (MDC). The
advanced supervised classification algorithms are: The Artificial Neural Network
(ANN), the Decision Tree Classifier (DTC), the Nearest Neighbor Classifier
(NNC) and the Support Vector Machines classifier (SVM).
The supervised approach is more popular but requires more detailed a priori
knowledge of the study area and analyst expertise, to identify suitable training sites
and the resultant spectra for classification (ERDAS 1999 ). The characteristics of
the training sites selected by the analyst have a great impact on the dependability
and the functioning of a supervised classification process. This approach has a
more subjective impact on the analyst during the defining of the LULC-categories
characteristics and its representative training samples. Supervised classification
approaches need more user-data-software interaction, especially in the collection
of training data.
A general introduction to pattern recognition and classification is given in the
textbooks by Duda et al. ( 2000 ) and Bishop ( 1995 , 2006 ), and in the review paper
by Jain et al. ( 2000 ). A detailed introduction in the context of remote sensing is
given by Richards and Jia ( 2003 ).
2.3.3 Remote Sensing Applications in Land Use/Land Cover
Mapping
The broad utilization of remote sensing is to extract and represent LULC-infor-
mation from multispectral imagery as thematic maps, data and GIS-layers (Donnay
et al. 2001 ). Research proves that remote sensing can be considered as a useful tool
for studying arid and semi-arid ecosystems (Tucker et al. 1983 ; Justice and
Hiernaux 1986 ; Townshend and Justice 1986 ; Maselli et al. 1993 ; Bastin et al.
1995 ; Hobbs 1995 ; Schmidt and Karnieli 2000 ; Kheiry 2003 ; Suliman 2003 ).
In comparison to the more classical classification methodologies such as basic
aerial photo interpretation, LULC-mapping using satellite imagery has four dis-
tinct advantages: (1) LULC-classes can be mapped faster and often with lower
costs; (2) fast and inexpensive updating of LULC-map products is possible, where
the satellite imagery are captured for the same geographic area at a high repeat