MultiSpec Tutorial: Unsupervised Classification (Cluster Analysis)
Requirements: MultiSpec application and image titled “ag020522_DPAC_cd.lan”.
Two Clustering algorithms are available in MultiSpec. They are useful in grouping similar pixels
in the image into clusters or categories. One algorithm implemented is a simple one-pass type.
The second is an iterative type called ISODATA. We will use the ISODATA algorithm for this
tutorial.
To start this tutorial, be sure that the “ag020522_DPAC_cd.lan” image that was used in tutorial 1
is open. Also clear any selections in the image window by striking the “Delete Key”
A cluster analysis will be run using the image file represented by the active (top-most)
multispectral image window.
3.1 From the Processor menu, select Cluster… to bring up the cluster specifications dialog
box. Select “Do Not Save” in the Cluster Stats: popup menu. Select “Cluster mask file”
and select “Image window overlay” under the “Write Cluster Report/Map To” group.
This will cause a cluster map to be created as a thematic image disk file and displayed as an
overlay on the multispectral image window.
Revised 9/10/2015 Page 1 of 5
MultiSpec Tutorial 3 Unsupervised Classification (Cluster Analysis)
3.2 Then select “ISODATA…” This will cause the ISODATA Specifications dialog box to be
displayed. Enter 7 for the number of clusters, 100 for the convergence percentage and set
the line and column intervals to 1 (if needed) and then select OK.
3.3 You are now back to the Cluster Specifications dialog box. Select OK to close this dialog
box and start the clustering operation.
You will be prompted to enter a name for the cluster map disk file and where to save the
file. Just use the defaults by selecting OK in the Save Cluster Map dialog box.
A cluster map will now be created with around 7 classes in an unsupervised manner. You
will notice the colors change in the image window as the pixels are sorted into cluster
classes during each iteration. After the final iteration, a thematic image file with a map of
the cluster classes will be saved to disk.
The text output for the cluster operation will be at the end of text output window. The
information includes the mean values for each of the channels for each cluster for both the
intial condition and the ending condition. If the map information is available for the image,
the final area for each cluster is listed in the units specified in the coordinate view for the
image window.
Usually the convergence is set for a little less than 100 so that the process does not take too
long to complete. We used 100 in this example so that you have a chance to watch the
pixels change cluster classes which illustrates the nature of the ISODATA algorithm.
The cluster map overlay on the multispectral image window will look similar to the
following.
Page 2 of 5 Revised 9/10/2015
Unsupervised Classification (Cluster Analysis) MultiSpec Tutorial 3
You can turn the
overlay on and off by
using the “Red O”
popup menu button in
the toolbar to the right
of the “small mountain”
zoom button.
3.4 Now open the cluster map thematic image file. This will be the same image as is shown in
the overlay above but you will have more control over the cluster classes.
From the File menu, select Open Image… to bring up the open image dialog box. You
may have to change the “Files of type” popup to All Files or Thematic Files. Then select
“ag020522_DPAC_cd_clMask.gis” and then select OK. One may need to select
“Thematic” for the Files of Type
popup menu.
The Thematic Display
Specifications dialog box to the
right will be displayed.
The default settings are fine;
select “OK” in the Display
Thematic Image dialog box.
This opens a thematic image
window.
Page 3 of 5 Revised 9/10/2015
MultiSpec Tutorial 3 Unsupervised Classification (Cluster Analysis)
3.5 The cluster class legend is on the left in the thematic image window below.
Compare the Cluster map with the original image window. Note clusters that represent
trees, sparse vegetation, thicker vegetation, different soil colors and shadows.
There are several things one can do to evaluate the image. One can move the cursor over a
color chip, hold the shift key down (cursor will change to an open eye) and click the left
mouse button down and up to cause the colors for that class to blink off and on (alternate
between white and the color). If one holds down both the shift and ctrl keys and then click
the left mouse button down and up, then all of the other classes will blink off and on. These
procedures are helpful in understanding the extent of the classes in the image.
One can also change the class color by double clicking on the color chip.
One can change the cluster class names by double clicking on the name to the right of the
color chip. (The list of interpreted cluster names that I come up with are: Light soil,
Medium colored soil, Sparse vegetation (weeds), Dark soil, Dense vegetation (trees, wheat,
weeds), Shadowed trees, Image blank.
3.6 One can also group the cluster classes together in information groups by selecting
Classes/Groups in the popup menu above the legend. Then drag the cluster classes into
similar information group categories. Again one can double click on the group name to
change the name. For example one could change the appearance of the thematic image to
represent Bare Soil, Vegetation and Non-image informational classes. The popup in the
legend allows one to display the original cluster (spectral) classes.
Page 4 of 5 Revised 9/10/2015
Unsupervised Classification (Cluster Analysis) MultiSpec Tutorial 3
Concluding Remarks
There are many other operations that one can do with MultiSpec including several Reformatting
processes. See other tutorial or the MultiSpec Introduction at the MultiSpec web site for more
information. The MultiSpec web site is: https://engineering.purdue.edu/~biehl/MultiSpec/. Or
contact Larry Biehl at biehl@purdue.edu with questions.
Page 5 of 5 Revised 9/10/2015
