Classification Tutorial - MultiSpec 32 PC


Modified from the GLOBE Toolkit Unsupervised Clustering Tutorial

To use this tutorial, you will need one image file; corpus_christi_600X600.lan. These should be included in your MultiSpec \images directory.

Select the folder "MultiSpec." The files, bevsub.lan and bevsub.cls will work with either the Macintosh or PC versions of MultiSpec.

Objectives:

The principal objective of this exercise is to provide expertise in classifying image data using MultiSpec.

In this tutorial, you will:

1) utilize the software to automatically create an unsupervised classification based solely on pixel characteristics.
2) manually train the software to recognize regions (AOI's) that contain known attributes on the ground, therby producing a supervised classification.

Overview:

Each pixel in your LandSat TM image contains a wealth of information about the surface materials that reflected light from that pixel to the satellite sensors. Each pixel contains a value which can range from 0 to 255, for each TM band supplied with your image. If, for instance, your image contains data for five bands, then each pixel contains five pieces of data, each potentially ranging from 0 to 255, as shown in the sample pixel diagram below.

This means that your image could contain 2565 (that's approximately 1.1 billion) different possible spectral combinations. Each of these combinations does not represent a different type of land cover; most of these variations represent very small and, to us, "unseeable" differences in surface reflectance.

In most instances, your computer monitor will be displaying only 256 different colors, hence only 256 different pixels. Even set to "thousands" of colors, only a small part of the many different pixels can be displayed. Even if a monitor could display all the different possible pixels, your eyes could recognize only a small number of differences in their appearance.

Because there is a limited number of different land cover types (the Modified UNESCO Classifications scheme, MUC, contains about 157 different types), and no GLOBE study site will have all of those different land cover types, it is necessary to group pixels together into a smaller number of closely related "classes." This process, whereby pixels with similar spectral characteristics are grouped, is called "Classification," and is done in two different ways.

In a supervised classification, you "train" the software to recognize that certain types of pixels represent specific land cover types. This is done on the basis of your knowledge of your own area, and field work you may do. The software then classifies the pixels of your image into the groups you have specified.

In an unsupervised classification, or "Clustering", we enter the number of groups, or "clusters," we wish to have, and certain other specifications. The software then examines the pixels in the image and groups them according to similar spectral characteristics. These groupings are not made on the basis of land cover, but on the similarity of the spectral characteristics of the pixels.

Part 1:  Unsupervised Classification or "Clustering"

To demonstrate clustering, you will use subscene of Path 26, Row 41 image centered on Corpus Christi, Texas. This 600 x 600 pixel subscene will allow the demonstration process to proceed more quickly than the clustering of a larger image, and will allow you to follow exactly the steps outlined in this tutorial. A new window, the Set ISODATA Cluster Specifications window will open.  It is in this window that you tell MultiSpec how you want the clustering to proceed. The information you need to provide is:
In the lower left-hand corner of the box is the Classification threshold: entry box. Change the value in this box to "100".
Setting this "threshold" value to 100 forces the system to assign every pixel in the image to one of the clusters. A value of less than 100 specifies the tolerance for assignment of pixels. A value of less than 100 will result in some pixels not being assigned to clusters. In this clustering, you are interested in large, fairly homogeneous areas, so individual pixels of slightly different spectral characteristics dotting the map are unnecessary.

The Results of Clustering

There are two results of clustering:
  1. A description of clustering activity and a "text map" in the TEXT OUTPUT window,
  2. A clustered Thematic image.

Also produced is a text map of the clustered area. The system assigns a number or letter to each of the clusters, and then displays a map of the clustered area using this code. A sample code is shown below.

A sample portion of the Text Map from the Clustering process. Each number/letter represents a pixel and the clustered group to which it belongs. You can see, even from this representation, that the system has identified several large, homogeneous areas, identified by the appearance of the same letter/number in an area.

Examining the Clustered Image

At this point, you will need to close all projects and images that are currently open in MultiSpec.
Select Project, Close Project, answer yes to "Save UntitledProject.Prj, and save as cc_unsupervised.Prj.  Remember, you have already saved your classification image as cc_unsuper.clu.
Next, clear the text output window by selecting Edit, Select All Text.., hit the delete key.

Part 2:  Supervised Classification

To demonstrate supervised classification, you will use the same subscene of Path 26, Row 41 image centered on Corpus Christi, Texas. The object here is to select large areas of homogeneous pixel "groups" that you will assign as training fields.  For example, the image contains large bodies of water that could be classed as "water" in a training field.  To do this, ***It is important that you try to identify as many KNOWN classes as possible, and use multiple training fields for each.  If you are unsure of what the pixel represents, it's probably not a good idea to use it as a training class.  Keep it simple.

Once you are satisfied with your selection of training fields, you can classify the image.  At this point it is advisable to save your project.
Select Save Project As... and name your project cc_supervised.Prj.

How Valid are these Classification Processes?

It is necessary for you to be confident that the process of "unsupervised classification" actually yields clusters that are related to land cover types, just as the supervised classification does with manual training of known regions. To compare the results of each classification process: You should see that the unsupervised clustering provides, at least in this case, a good indication of the locations of large areas of uniform land cover that could be investigated for verification studies.

Desk Verification

It is important to use as much data as you can obtain to validate your classification schemes.  The desk verification process could involve the use of local maps (topographic, land cover, soil, political, etc.), other local references (aerial photos, people, agencies, etc.) and the combined experiences of both you and your students to identify some of the clusters produced by MultiSpec. Use whatever resources you can to identify your classifications

Field Verification

If there are clusters that you cannot identify "from the desk," you will have to go out into the field to determine what they are. Ground truthing is an integral part of remote sensing and should be done to verify and validate your classification processes, supervised or unsupervised.

Renaming the Clusters

In  your unsupervised classification, the software produced clusters identified only by a number, and arranged in order of decreasing brightness. Once you have identified the land cover for each of these clusters, your Thematic Map display may be customized to show these clusters either by name or by MUC identification code. You can, in effect, produce two different Thematic Maps on the same image; one in which each cluster is identified by a name (e.g. Ocean, Transportation) and the other by MUC disignations (e.g. 72, 93.)

The secret to this process is that your Thematic Map can display both "Groups" and "Classes." When it is produced, both "Groups" and "Classes" have the same set of colors and labels. To see this:

You might decide that the "Groups" will contain descriptive names, while "Classes" contains MUC labes.

To change the name of a cluster in either view, at any time:

Once you have entered this data, you should save your work.  Saving your data can be done in a number of ways. You can return to either of your classification images at any time to change descriptive information.