An Introduction to Remotely Sensed Data Analysis - Sampling Spatial Units for Agricultural Surveys - page 79

1, and a p

subject to Cov (y p , y l ) 1. For greater technical details

on PCA and on the methods for choosing components see Jolliffe ( 2002 ).
For other methods that are specific to the analysis of certain phenomena, the
reader can see Richards and Jia ( 2006 ).

¼ 0, for p > l a p ¼

4.6 The Thematic Extraction of Information

After the restoration phase, an RS image is processed to extract information. The
information extraction procedures work simultaneously on multiple bands. This is
different from restoration techniques, which typically operate on the data in each
spectral band separately.
There are two different classes of operations for the thematic extraction of
information: segmentation and classification. There is little agreement on defini-
tions of segmentation and classification. These terms are often confused, or even
used as synonyms (Ripley 1988 ). Segmentation is generally defined as the problem
of dividing the image to identify discontinuities that emphasize physical boundaries
(Geman et al. 1990 ). The main goal of image segmentation is to divide a digital
image into multiple segments by identifying boundaries between objects. Here, the
aim is to simplify the representation of an image into something that is more
meaningful and easier to analyze. The labels are simply binary, and represent the
presence (1) or absence (0) of boundary elements.
Image classification uses the spectral information represented by the digital
numbers in one or more spectral bands to classify each pixel. The aim is to assign
all pixels in the image to particular classes or themes (e.g., water, coniferous forest,
deciduous forest, corn, wheat, etc.). The resulting classified image consists of a
mosaic of pixels, each of which belongs to a particular theme. It essentially
represents a thematic map of the original image (Richards and Jia 2006 ).
There are two broad classes of classification procedure: unsupervised and
supervised. In unsupervised classification, an image is divided into unknown
classes. Then the researcher labels the classes in the image. Unsupervised classifi-
cation groups pixels with similar spectral reflective characteristics into distinct
clusters. These spectral clusters are then labeled with a certain class name. Super-
vised classification uses a set of user-defined spectral signatures to classify an
image. The spectral signatures are derived from training areas (or sets) that are
created according to the features of interest. The important difference between the
two approaches is that unsupervised classification does not require an a priori
definition of the classes.
An example can clarify these definitions. Figure 4.3a represents a segmented
image. The labels are binary (i.e., they represent the absence or presence of a
border), and detect only boundaries or limits between different zones. Figure 4.3b
contains a classified image. If the labels are a priori defined, the image has been
classified using supervised methods; otherwise an unsupervised algorithm has been