Classification

Objectives

• To define and describe digital image classification

• To distinguish between supervised and unsupervised classification and
describe how each is produced
• To discuss the application of digital image classification data to remote sensing

Introduction

An important part of image analysis is identifying groups of pixels that have similar
spectral characteristics and to determine the various features or land cover classes
represented by these groups. This form of analysis is known as classification. Visual
classification relies on the analyst's ability to use visual elements (tone, contrast, shape,
etc) to classify an image. Digital image classification is based on the spectral
information used to create the image and classifies each individual pixel based on its
spectral characteristics. The result of a classification is that all pixels in an image are
assigned to particular classes or themes (e.g. water, coniferous forest, deciduous
forest, corn, wheat, etc.), resulting in a classified image that is essentially a thematic
map of the original image. The theme of the classification is selectable, thus a
classification can be performed to observe land use patterns, geology, vegetation
types, or rainfall.

The analyst classifying an image must distinguish between spectral classes and
information classes. Spectral classes are groups of pixels that have nearly uniform
spectral characteristics. Information classes are the various themes or groups the
analyst is attempting to identify in an image. Information classes may include such
classes as deciduous and coniferous forests, various agricultural crop types, or inland
bodies of water. The objective of image classification is to match the spectral classes in
the data to the information classes of interest.

Though any image can be classified, multispectral imagery tends to be used most
often. One band classification is usually very difficult to classify since more than one
surface type will exhibit the same digital number. Thus, any spectral classes in a single
band classification will likely contain several information classes, and distinguishing
between them would be difficult. Normally two or more bands are used for
classification, and their combined digital numbers are used to identify the spectral
signatures of the spectral classes present in the image. The more bands used to create
a classification, the more likely the analyst will get a set of unique land cover classes.

Supervised Classification

A supervised classification is performed when some prior or acquired knowledge of the

classes in a scene is used to identify representative samples of different surface cover
types. These samples, known as training sites, are set up to identify the spectral
characteristics of each class of interest. The determination of training sites is based on
the analyst's knowledge of the geographical region and the surface cover types present
in the image. Once the training sites have been established, the numerical information
in all of the image's spectral bands are used to define the spectral "signature" of each
class. Once the computer has determined the signatures for each class, it will compare
every pixel to the signatures and label it as the class that it is mathematically closest to.
Thus, in a supervised classification, the analyst starts with information classes and uses
these to define spectral classes. Each pixel in the image is then assigned to the class
which it most closely resembles.

Unsupervised Classification

An unsupervised classification is essentially the opposite of a supervised classification.

The pixels in an image are examined by the computer and classified into spectral
classes. The grouping is based solely on the numerical information in the data and the
spectral classes are later matched by the analyst to information classes. In order to
create an unsupervised classification the analyst typically determines the number of
spectral classes to identify and a computer algorithm will find pixels with similar spectral
properties and group them accordingly. The following image is an example of a 40-level
unsupervised classification of a portion of Howard County, Maryland. This means that
there are 40 distinct spectral classes in this image, each of which is assigned a gray
tone value ranging from black to white, with intermediate shades of gray.

[40-level unsupervised classification]

Programs, called clustering algorithms, are used to determine the statistical groupings
in the data. Usually, the analyst specifies how the initial classification should proceed.
In addition to specifying the desired number of classes, the analyst may specify
parameters to determine how close pixels' digital numbers must be to be considered in
the same class. Once the clustering process has run, the analyst may want to combine
or further break down some clusters. Thus, unlike its name suggests, an unsupervised
classification often requires interaction with an analyst.
© CGIS
at
Towson
Universit
y