Unsupervised Classification

Overview

You will copy to your home directory a subsetted image of Mt. Desert Island for this exercise, in this case /scratch/maine02/mdi.051900/mdi.L7.051900.img.

Read about how an unsupervised classification of an image identifies statistical patterns in the data without using any groundtruth information.

You will choose 15 classes for this exercise. The output file will have a gray scale color scheme. You'll use the Raster Attribute Editor to change the color scheme so it is easier to interpret the image.

You will be asked to hypothesize what each class might acually be on the ground.

The Nitty Gritty

Copy image /scratch/maine02/mdi.051900/mdi.L7.051900.img to a subdirectory on your U: drive to analyze. Open it in a viewer. Next choose the Classifier button in the icon panel. A Classification menu appears. Choose Unsupervised Classification. The Unsupervised Classification dialog box appears.

unsuper.class.dialog.gif (9593 bytes)

The Input Raster File will be the image displayed in your viewer. The Output Cluster Layer will be the name of the classified image you are creating. I suggest naming this file "mdi.L7.051900.unsuper15.img". The file name includes the date the image was taken, that it an unsupervised classification that has 15 classes. Make sure you click on the folder icon next to the blank line where you type in the file name to set the folder destination.

The Output Signiture Set should have the same prefix as the Output Cluster Layer with a suffix of .sig. For example, mdi.L7.051900.unsuper15.sig.

Reset the Number of Classes in the dialog box to 15. Leave all other parameters as default.

Click on OK and an isodata status box appears with the ok button blanked out..

isodata.dialog.box.jpg (9906 bytes)

When the isodata calculations are complete, the ok button will appear, Click on it to dismiss the program.

Display the classified image in a new viewer. It will display as a Pseudo Color image. A gray scale image will display with the 15 classes shown as different gray values. Not very informative.

In the active viewer left-click on Raster, then Attributes. The Raster Attribute Editor appears. Expand the window so all 15 classes display and is wide-enough to display all columns.

raster.attribute.editor.expanded.gif (16837 bytes)

Next you want to assign colors to each class to allow for easier interpretation. You could assign a color to each class, one at a time. To do that, right-click on a class in the color column and a color menu will appear. You can choose one of the colors or choose "Other" and create your own color palette. BORING!!!!!!!! as well as tedious. Has its uses though.

You are going to take a short-cut. In the Raster Attribute Editor dialog box, left-click on Edit. A menu appears. Left-click on Colors and a Colors dialog box appears. Choose all the defaults...... so just left-click on Apply. Wow!!! Look at those colors appear?  Play with the Color dialog box a little. Check out different color combinations..

The 15 classes the ISODATA model created may represent real land use/land cover (LULC) on the ground. Maybe a land use/land cover is represented by more than one class. In that case, combine the classes. Maybe a single class includes two or more land use/land cover classes. In that case more classes are needed to differentiate. In some cases the land use/land cover is too complicated or too similar to other LULC's and so can not be differentiated statistically. The only way to know for sure is getting ground truth. How can we get ground truth of the image? Travel there!!

In your write-up discuss the 15 classes and hypothesize as to their land use or land cover. Use maps, aerial photos and satellite images in your analysis. Discuss how we would prove/disprove your analysis when we go to MDI.

Publish the MDI subset image and your classified image to your web page.

Writeup to be typed and handed in.