ENS 243 Introduction to

Remote Sensing
Image Interpretation – Image Classification
 Image Classification
What is Image Classification?
• Image classification is a procedure of categorizing all pixels in an
image into land cover classes or themes based on the spectral data
embodied in the digital numbers (DNs).
• The spectral pattern present within the data for each pixel is used as the
numerical basis for categorization.
• If a pixel satisfies a certain criteria, the pixel is assigned to the class that
corresponds to those criteria.
• The goal of image classification is to assign all pixels in the image to
classes/themes.
• The classified image is a thematic map of the original image and is
made up of a mosaic of pixels, each of which belongs to a specific
theme.
 Image Classification
Information Classes Vs Spectral Classes
• Information Classes: classes that the analyst is actual trying to identify
from analysis of image. E.g. forest types, crops, water, etc.
• Spectral Classes: groups of pixels which are similar or near similar
with respect to brightness values in the different spectral bands of
the data.
• The aim is to match the spectral classes in the data to the information classes
of interest.
 Approaches to Classification
There are 2 Approaches to image classification:
o Supervised classification
o Unsupervised classification
• Supervised classification: in this type of classification the image analyst
“supervises” the pixel categorization process by specifying, to the
computer algorithm, numerical descriptors of the various land cover types
present in the scene.

• Representative sample sites of known cover type, called training areas are
used to compile numerical “interpretation key” that describes spectral
attributes for each feature of interest
• interpretation keys: shape, size, pattern, etc.
• classification algorithms used in supervised classification: includes
maximum likelihood and minimum-distance classification.
 Approaches to Classification
• Unsupervised: the process of automatic identification of classes.
• The aim of this method is to split data into groups as similar as
possible, without knowing the nature of such groups
• It is objective and completely data driven
• This method mainly suited to images of objects/areas where there is
no ground knowledge
• Clustering algorithms are used to identify most frequent land cover
types

• Most algorithms used: K-means and ISODATA

 Unsupervised Classification
• Unsupervised classifiers do not utilize training data as the basis for
classification

• Rather, the classifiers involve algorithms that examine the unknown

pixels in an image and aggregate them into a number of classes based
on the natural groupings/clusters in the image values

• Main assumption is that

• values within a given cover type should be close together in the
measurement space

• Whereas data in different classes should be comparatively well separated

 Unsupervised Classification
• The classes that result from unsupervised classification are spectral
classes.

• After classification, the analyst must compare the classified data with
some form of reference data to determine the identity and
informational value of the spectral classes.
 Supervise vs. Unsupervised
• The fundamental difference between supervised and unsupervised
classification is that

• Supervised classification involves a training step followed by a classification step

• In the unsupervised classification the image data are first classified by aggregating
them into the natural spectral groupings, or clusters present in the scene.
• No of clusters in unsupervised classification is subjective and depends on the following
factors:
• Size of area you are trying to classify
• How diverse (heterogeneous) the landscape is
• Resolution of the data you will be using
• Spatial
• Spectral
• The number of classes you will be mapping
 Supervise vs. Unsupervised

• In the supervised approach we define useful information categories

and then examine their spectral separability.

• In the unsupervised approach we determine spectrally separable

classes and then define their informational utility
 Supervised Classification
Advantages and Limitations:
Advantages Disadvantages
Analyst has control over the Time-consuming and costly
classification
Errors can be detected and often Analyst imposes a structure on data, which
rectified may not match reality
Processing is tied to specific areas of Training classes are generally based on field
known identity identification and not on spectral properties
hence spectral signatures are forced
Training data selected by the analyst may
not be representative of conditions present
throughout the image
Unable to recognise and represent special
or unique categories not represented in the
training data
 Unsupervised Classification
Advantages and Limitations:
Advantages Disadvantages
Human error is minimised Analyst has little control over classes
No extensive prior knowledge of the Not all spectral classes are
study area is required information classes
Unique classes are recognized as Spectral properties change over
distinct units time hence detailed spectral
knowledge of different features
may be necessary
 Approaches to Classification
• Combination of supervised and unsupervised classification processes
to develop classified maps.
• This is called a ‘hybrid’ approach
• Perform unsupervised classification interpret results using
ground reference data perform supervised classification on
the original image.
• This approach provides more objective and reliable results.
 Steps in Supervised Classification
Four (4) Basic Steps:

• Training stage (1): The analyst identifies representative training areas

and develops a numerical description of the spectral attribute of each
land cover type of interest .
• Training samples: are sets of pixels that represent what is
recognized as a potential class.

• Training can be carried in 2 forms: either supervised training or

unsupervised training.

• There are no specific rules regarding the number of training sites per
class
 Steps in Supervised Classification
• Collect Signatures (2): The analyst uses signature collection tools to
collect signatures of training samples for potential classes.

• Signature: a set of data that defines a training sample and statistical

information such as minimum, maximum, mean vector, pixel number, number
of bands, and covariance matrix of the potential class.

• Each signature corresponds to a class.

 Steps in Supervised Classification
• Signature Evaluation (3): Examine the spectral signatures of each class in the
image dataset if is categorized into the land cover class it most closely
resembles.
• To ensure a minimum of spectral overlap between signatures of different
classes.

• If the pixels is insufficiently similar to any training data, it is usually labeled

“unknown”
 Steps in Supervised Classification
• Decision Making/Image Classification (4): presentation of the results after the
entire data set has been categorized
• It is accomplished using algorithm known as decision rules.
• These rules set certain criteria
• If pixels meet the set criteria, it is assigned to a particular class
• Pixels failing to satisfy any criteria remain unclassified

Stages in the process of image classification

Training Collect Signatures Signature Evaluation Decision Making

 R R F F F R F F F F
Basic Steps in Supervised Classification F R R F F R F F F F
F F R R U U U U F F
F F R U U U U U F F

Classification F U R U U U U U F F
Blue Band
Stage F U U R U U U U F F
Green Band
F U U R U U U U F F
Red Band
F U U R U U U U F F
NIR
F F U U R U U U U F
67 (B) Forest F F F U U R U F F F
62 50 22 35
54 (G)
River
R R F F F R F F F F
21(R) 12 11 19 22
F R R F R F F F F F
37 Urban F F R R U U U U F F
(NIR) 121 99 78 09 F F R U U U U U F F
F U R U U U U U F F
F U U R U U U U F F
F U U R U U U U F F
Training
F U U R U U U U F F
Stage
F F U U R U U U U F
Output Stage F F F U U R U F F F
Image Classification – Supervised
 Supervised Algorithm-Minimum distance
classification
• The First Law of Geography, according to Waldo Tobler, is "everything
is related to everything else, but near things are more related than
distant things”.

• This first law is the foundation of the fundamental concepts of spatial

dependence and spatial autocorrelation

• Utilized specifically for the inverse distance weighting method

for spatial interpolation and to support the regionalized variable
theory for kriging
 Image Classification – Supervised (Minimum
distance classification)
254
+
Step 1
+
Compute mean (average) spectral value in each band
+
By considering the two-channel pixel values as
positional coordinates, a pixel of unknown identity
+ maybe classified by computing the distance between
the value of the unknown pixel and each of the
category means
+ +

0 254
 Maximum Likelihood classifier
• Most widely used and powerful classification algorithm
• A pixel is assigned to the class for which it has maximum likelihood of
association.
• It uses training data to estimate means and variances of the classes
• It considers the variability of brightness values in each class around
the mean.
Maximum Likelihood classifier
To classify the measurement
vector X of an unknown pixel into
a class, the maximum likelihood
decision rule computes the value
pc for each class.
Then it assigns the pixel to the
class that has the largest (or
maximum) value.
THE END