DIGITAL IMAGE CLASSIFICATION

N.S. Maphosa
Department of Geomatics and Survey
 Introduction

 Image classification is the extraction of differentiated classes or themes, from

remotely sensed digital satellite data using classification algorithms
 The objective of image classification is to automatically categorize all pixels in
an image into thematic classes such land cover
 Image classification is based on the spectral characteristics of different
materials on the earth's surface
Spectral reflectance curves
 Image Space
 To understand the image classification concept we
have
to comprehend the concept of image
and feature space
 Digital image is a 2D array of pixels
 Each pixel stores a DN or Brightness value e.g.
an 8-bit record in the range 0-255
 Each DN corresponds to the EM radiation
reflected or emitted from a ground
resolution cell
 The spatial distribution of the DNs defines
the image space
The image space
 Feature Space

Feature space
 In one pixel two bands can be regarded as
components of a 2D vector, the feature vector
 An example of a feature vector is (13,15) which
tells that 13 DN and 15 DN are stored in
band 1 and band 2 respectively
 This vector can be plotted on a 2D.
 Such graphs are called scatter plots
Feature space
Chivero Landsat TM band 5:4:3
Chivero Landsat TM band
4 and 3 Scatter

Soil Line
Band 4 and 3 Scatter: Water
Band 4 and 3 Scatter: Woodland

Band 4 and 3 Scatter: Bare

 Ban 4 an 3 Scatter
d d

NIR

RED

Feature space

Distances and clusters in feature space

 Distance in feature space expressed as Euclidian

Distance (straight line between 2 points & units

are DN)
 In the feature space the distance can be

calculated using Pythagoras theorem

 For example the distance between (10,10) and

(40,30) is equal to the square root of (40-10)2

+ (30-10)2
 For 3 or more dimensions the distance is

calculated in a similar way

 Distance and clusters in feature
space

Feature space
 Scatter plot gives information about the
distribution of corresponding pixel values in
two bands of an image
 Each cluster of a feature vector occupies its own
area in the feature space
 Basic assumption of image classification is that one
cluster can be separated from the other
 A specific part of a feature space corresponds
to a specific class
 The principle of image classification is that the pixel is
assigned a class according to its feature vector by
comparing it to predefined clusters in the feature
space
 Doing so for all image pixels results in a classified
image
 The principle of image
classification
 Classes to be distinguished in an image need to
have different spectral characteristics e.g.,
comparing reflectance curves
 Once classes have been defined in the feature space,
each image pixel can be compared to these classes &
assigned to the corresponding class
 Definition of clusters is an interactive process and is
carried out during a training process
 Comparison of individual pixels with the clusters takes
place using classifier algorithms

Principle of image classification:

Example of a feature space showing respective clusters of 6
classes
 Methods of image classification
 The traditional methods of classification
mainly follow two approaches: unsupervised
 and supervised
Unsupervised: statistical "clustering“
algorithms used to select classes inherent
 to the data, more computer- automated
Supervised: image analyst "supervises" the
selection of regions/clusters that represent
patterns/features that the analyst can
recognize

Image classification process

There are five steps in the image classification

1. Preparing image data: Depending on the thematic
map to be produced , the most appropriate sensor
and date of image acquisition, selection of
wavelength bands appropriate for the purpose
2. Definition of clusters in feature space: In supervised
classification the operator defines the clusters during
the training process and in unsupervised classification a
clustering algorithm automatically finds and defines the
clusters in the feature space
3. Selection of a classification algorithm: once
spectral classes have been defined the use will
decide how pixels are allocated to classes
 Image classification process

4. Running the classification. Once the

training data have been established
and the classifier algorithm selected
image classification can be performed
i.e. pixels assigned to classes
5. Validation of the result using the
error matrix. Assess the quality of
the classified image by comparing it
to reference data(ground truth).
Requires selection of a sampling
technique

Supervised Classification

 Prior knowledge about the area is needed. It

is done in steps:
 Sampling training pixels: The user selects
samples representative of each class. These
are called training samples
 Analysis of training samples: This is done to
find out whether clusters are well
defined or see if there is overlap. The
user may decide to combine some
clusters
 Select classification algorithm
* Box classifier/ parallelpiped classifier
* Minimum distance to mean
* Maximum likelihood classifier
 Supervised classification process

Supervised classification
algorithms

Trainin es plot
e using
 Supervised classification
algorithms
 Box classifier or parallelepiped requires only the lower
and upper limit pixel values
 It is a quick and cheap method to classify
 Problems when parallelepipeds overlap (due
to high covariance)
 To compensate, rectangles with stepped
borders might be used

Parallelepiped
 Box classifier or
parallelepiped
requires only the
lower and upper
limit pixel values
 It is a quick and
cheap method to
classify
 DISADVANTAGE
🞑 overlapping
boxes
🞑 poorly
adapted to
cluster shape
 BOX CLASSIFICATION

Partitioned Feature Space

255

Band 2

B
a
n
d

2
5
5

Supervised classification algorithms

 Minimum distance to means classifier: the mean vector of
each training class is computed and the Euclidean
distance from each pixel to each class will be
calculated and labeling of the pixel will be based on
the shortest distance
 Compute mean of each desired class and then classify

unknown pixels into class with closest mean using

simple Euclidean distance
 This is done by assigning a threshold distance and

any pixel outside the threshold distance will be

labeled as unknown
 This is done to avoid over classification i.e. each
pixel going to more than one class
 Supervised Classification

 Minimum distance to
means 45

40

🞑 Find mean value of

35

30

pixels

Band 4
of training sets in n-
25

20

dimensional space
15

10
5

🞑 All pixels in image

0
0 2 4 6 8 10 12 14 16 18 20

Band 3

classified according to
the
class mean to which
they
are closest

MINIMUM DISTANCE TO MEANS

CHARACTERISTICS
 emphasis on the
location of cluster
centre
 class
labelling by
considering
minimum distance to
the cluster centres
DISADVANTAGE
 avoids the
presence of
variability within a
class
 shape and size of
the clusters are not
important
 Supervised classification algorithms

 Minimum distance to means classifier

 It is insensitive to different degrees of
variance in the spectral response data
 This method is not used in applications where
spectral classes are close to one another in the
measurement space and have high variance

Minimum Distance to Means

Classifier: Euclidian Distance
 Supervised classification algorithms

 Maximum likelihood classifier

 It uses the statistical parameters of mean vector,
variance and co-variance of training samples
 It calculates the probability that a pixel can belong
to a class and a pixel is assigned to a class to which
it has the highest probability
 Unidentified pixels can then be given a probability
of being in any one class
 Assign the new pixel to the class with the highest
probability – or unclassified if all probabilities low

Maximum likelihood example

 Unsupervised Classification

 The unsupervised approach is often referred

to as clustering and results in statistics that
are for spectral, statistical clusters
 This is carried out by spectral groupings
made by the computer taking into account
certain similarity measures
 This is only carried out where there
is no prior knowledge of the area
 Unsupervised classifiers do not utilize training
data as the basis for classification

Unsupervised Classification
 The classes that result from unsupervised
classification are spectral classes as they are
based solely on the natural groupings in
the image values
 The identity of the spectral classes will not
be initially known
 The analyst must compare the classified
data with some form of reference data
(such as larger scale imagery or maps) to
determine the identity and informational
value of the spectral classes
 Results of
Unsupervised
Classification

Problems in image
classification
 Spectral classes-different type of grasses
each with a different spectral signature
 Same type of grass may have different
spectral characteristics over large areas due
to different soils, climate conditions
 Mixel: one pixel assigned to one class,
what if that pixel has two or more
cover types?
 Assign to several classes?
 Validation: Accuracy Assessment

Accuracy Assessment
 Quality assessment is done by comparing the
classified image to true world class
 One way to do this is by creating an error

matrix from which different accuracy

measures may be calculated
 In the evaluation of
classification errors, a classification
error matrix is typically formed
🞑 This matrix is sometimes called confusion
matrix or contingency table
 In this table, classification is given as rows

and verification (ground truth) is given as

columns for each sample point

 Error Matrix

Error Matrix
 Overall accuracy/proportion correctly
classified: sum diagonal (correctly
classified)/total number of pixels checked.
 This is the Observed
 The expected value for each cell is calculated
as (row total/grand total * column
total/grand total)
Error Matrix
 From the confusion matrix, it can be seen
that at least two methods can be used to
determine individual category accuracies
 The ratio between the number of
correctly classified and the row total
🞑 the user's accuracy - because users are
concerned about what percentage of the classes
has been correctly classified
 The ratio between the number of
correctly classified and the column total
🞑 is called the producer's accuracy

Kappa Statistic

k OE
1E
Kappa Interpretation

Calculating Kappa for different

classes
 Class 1: (Row total/Grand
Total)*(column Total/Grand
total)
 Do this for all the classes and then ADD
the results you would have obtained for
individual classes
 From the calculation you get the Expected
 The observed is just the Sum of
correctly classified pixels/Grand total
 After this then use the Kappa statistic
formula to determine the accuracy of
your classification
Thank you for your attention