Digital Image Processing-Information
Extraction
1
(ii) Unsupervised Classification
Remote Sensing: M5L2 D. Nagesh Kumar, IISc
Objectives
2
Introduction
Chain method of unsupervised classification
ISODATA algorithm
K means algorithm
Remote Sensing: M5L2 D. Nagesh Kumar, IISc
Introduction
3
Unsupervised classification is a technique that
groups the pixels into clusters based upon the
distribution of the digital numbers in the image
An unsupervised classification operates in an
iterative fashion.
Initially it assigns arbitrary means to the classes
and allocates each pixel in the image to the class
mean to which it is closest.
New class means are then calculated and each pixel
is then again compared to the new class means.
This procedure is repeated over number of
iterations
Remote Sensing: M5L2 D. Nagesh Kumar, IISc
Introduction
4
Pixels move between clusters following each
iteration until threshold is reached.
A threshold of 0.98 means that the program
terminates when less than 2% of the pixels move
between adjacent iterations.
The classes produced from unsupervised
classification are spectral classes and may not
correlate exactly with ‘information classes’ as
determined by supervised classification.
Remote Sensing: M5L2 D. Nagesh Kumar, IISc
Chain method of
Unsupervised Classification
5
This algorithm operates in two steps
a) Generation of clusters
To generate clusters, the algorithm requires the following
information from the user:
- Radius to estimate the formation of a new cluster
- Parameter to merge clusters
- Number of pixels to be evaluated
- Maximum number of clusters which needs to be identified
b) Merging of clusters
- Based on the user defined thresholds
Remote Sensing: M5L2 D. Nagesh Kumar, IISc
ISODATA Method
6
ISODATA is self organizing because it requires relatively
little human input
Classification using the ISODATA algorithm normally requires
the analyst to specify the following criteria:
a) Maximum number of clusters to be identified by the
algorithm.
b) The maximum percentage of pixels whose class values are
allowed to be unchanged between iterations.
c) The maximum number of times ISODATA has to classify pixels
so that cluster mean vectors can be recalculated.
d) Minimum members in a cluster
e) Maximum standard deviation
f) Minimum distance between cluster means
Remote Sensing: M5L2 D. Nagesh Kumar, IISc
ISODATA Method
7
The mean vectors of all clusters are arbitrarity assigned.
With the mean vectors, the algorithm searches through the
image data wherein each pixel gets compared to each cluster
mean using some distance measure and is assigned to that
cluster mean to which it lies closest in the spectral feature
space
ISODATA can either progress line by line or block by block
At the end of first iteration, a new mean vector is calculated
for each cluster based on the new pixel locations that have
been assigned to each cluster based on distance measure.
This process is iteratively repeated untill the algorithm
termination condition gets fulfilled.
Remote Sensing: M5L2 D. Nagesh Kumar, IISc
ISODATA Method
8
(a) ISODATA initial distribution of 5
hypothetical mean vectors in both bands
as beginning and ending points (b) In the
first iteration, each candidate pixel is
compared to each cluster mean and
assigned to the cluster whose mean is
closest in Euclidean distance (c) during
the second iteration, a new mean is
calculated for each cluster based on the
actual spectral locations of the pixels
assigned to each cluster, instead of the
intial arbitrary calculation. This involves
analysis of several parameters to merge or
split clusters. After the new cluster mean
vectors are selected, every pixel in the
scene is once again assigned to one of the
new clusters. (d) This split-merge-assign
process continues until there is little
change in class assignment between
iterations (the threshold is reached) or the
maximum number of iterations is reached.
Remote Sensing: M5L2 D. Nagesh Kumar, IISc
K-Means Classification
9
Most commonly used non-parametric unsupervised
clustering algorithms
Similar to ISODATA, k means algorithm also
assigns initial cluster vector. The difference is that
k means algorithm assumes that the number of
clusters is known a priori
Remote Sensing: M5L2 D. Nagesh Kumar, IISc
K-means Classification
10
Steps
a) Location of initial cluster centers
Ifx represents the sample space of data having elements
N as x,
be the total number of elements in sample
Bndspace and be the
total number of bands, then the mean for data points for i th cluster in
jth dimension is given by :
N
Bnd xkj
vij k 1 ,1 i c
j 1 N
The objective function to be minimized is given by:
c Bnd N
2
J xkj vij
i 1 j 1 k 1
Remote Sensing: M5L2 D. Nagesh Kumar, IISc
K-Means Classification
11
b) Merging of clusters
Several measures are available for cluster
merging like root mean square euclidean distance
for each cluster, matrix of euclidean distances
between cluster centers etc.
K mean clustering approach is iteratively
performed till the objective function is minimized.
This technique is simple, computationally fast that
results in tighter clusters.
Remote Sensing: M5L2 D. Nagesh Kumar, IISc
Thank You
Remote Sensing: M5L2 12 D. Nagesh Kumar, IISc
