Theoretical Background and State of the Art - Historical Land Use/Land Cover Classification Using Remote Sensing - page 15

Sensor at aireal- or space- platform

Multi-spectral, multi-temporal and multi-source images; plus ancillary GIS or contextual information

Pre-processing: One or more of these steps (Geometric-, atmospheric- and radiometric- correction;
enhancement: Color, spatial, radiometric and spectral; plus mosaic and subset)

The pre-processed
multi-spectral image

Feature
extraction

Classifier

Classes to be
classified

Auxiliary data (Maps, field
work: GPS-Points, …etc.)

K-D feature
image

Discriminant Function based
on training statistics + Choise
the classification algorithm

K-D feature space

Select training pixels

Categorical lables

Post-processing: (Filtering … etc.)
to refine LULC-Classes

Accuracy assessment

Final thematic map

Fig. 2.4 The classification process (Source modified from Townshend and Justice 1986 , Tutz
2000 , Wilkinson 2005 and Schowengerdt 2007 )

of the data to shorten the computing time needed by the classifier, and thus to
raise the effectiveness of statistical estimators in a statistical classifier;
• Training: the term ''training'' is the choosing of the pixels to train the classifier
to identify the preferred themes,orclasses, and the selection of decision
boundaries. Here, the drawing of boundaries around geographically located
pixels has to be homogeneous, or suitably heterogeneous. This phase can be
carried out either supervised or unsupervised; and
• Labeling: it is the process of allocating diverse pixels to their most likely class
based on the use of the feature space decision boundaries. This process can be
supervised or unsupervised. If a pixel is not spectrally alike to any of the
available classes, then it can be assigned to an unknown class. There are two
kinds of relationships between the object and the class label: one-to-one (pro-
ducing a hard classification); or one-to-many (producing a fuzzy classification).
The object may be a single pixel or a group of neighboring pixels forming a
geographical unit. As a result, a thematic map is produced, presentating every
pixel with a class label. The end result is a transformation of the digital image
data into descriptive labels that classify unlike Earth surface objects or

conditions.