Geography Reference
In-Depth Information
Image Data Set
(
4 spectral bands per pi
xel)
LULC- Signatures
(based on
training
samples )
Pixel Grey Value
Signature
Classified Image Resul
t
RR
R
F
F
FF F F
F
F
F
F
69
F
F
Urban
(234, 96, 75,
05
R
U UUUU
U
U
UUU
UU
R
U
FF
F
FF
F
56
R
F
F
F F
F
F
F
River
(23, 18, 38, 2
0)
F
R
U
U
F
F
F
F
F
FF
F
F
F
F
U
U
UU
U
U
U
R
U
U
U
U
U
Spectral
Bands:
22
U
U
U
U
UU
R
U
F
1
Forest
(66, 44, 18, 3
0)
U
R
F
U
U
U
2
38
F
F
U
F
R R
F
F
F
3
F
F
F
F
4
F
Compare
Classify
Fig. 2.6 The principal idea of the supervised classification approach for multi-spectral remote
sensing (Source modified from Eastman
2006
)
image processing software groups pixels that have similar properties (in feature
space and in adequate representative spectrally-separable clusters for the ground
surface features), based on the statistics of the radiometric value/digital number of
each pixel. Then the analyst evaluates the classified map with field survey data,
aerial photographs and other reference data, and labels these clusters (spectral
classes) with its equivalent in the real world to information classes, without having
a prior knowledge of the classes. Generally, some clusters must be subdivided or
combined to make this equivalence. Results of an unsupervised classification can
be used to define the training samples, which are a main input in the supervised
classification, or the labeled cluster map can be just accepted as the final map
(Schowengerdt
2007
).
Supervised classification: supervised approaches, as seen in Fig.
2.6
, are based
upon training sites, and can assure the former but not the latter; unsupervised
approaches can assure the latter but not the former (Tso and Mather
2009
). Each
image is characterized by n-observations (the values in n-data bands). Supervised
image classification is an approach in which the analyst delineates the training
samples (vectors in an n-dimensional feature space) on the image which are
representative of each interested LULC-class (Mather
2004
). A basic step in
supervised classification and mapping is the design of a realistic classification
scheme, which satisfies a clear definition of separable discrete informational
LULC-categories within the available data (Cingolani et al.
2004
). Training sites/
samples can be created from fieldwork, aerial photography and other existing maps
based on analyst knowledge (e.g., Google Earth), and are then used as reference
information (Lillesand et al.
2008
). Visual interpretation is used to locate the
Search WWH ::
Custom Search