In many instances the most useful image processing output is a classified scene. This is because you are entering a partnership with the processing program to add information from the real world into the image you are viewing, in a systematic way, in which you try to associate names of real features or objects with the spectral/spatial patterns evident in inidividual bands, color composites, or PCI images. PIT is capable of producing both unsupervised and supervised classifications (review the relevant parts in Section 1, starting on page 1-16, if you need a refresher on principles). In the next pages we will walk you through the steps in developing each type of classification.

Unsupervised Classification

Here the strategy is to determine statistically separable distributions of DN values in multispectral space. Rules establish where a given pixel is located in this space. PIT utilizes two such classifiers, labeled UC1 and UC2. (We will examine only UC1 here because UC2, which can be a better discriminator, takes a very long [hour +] time to run). The default number of spectral classes is set at 27. Many of these contain less than 2% of the pixels (as distributed spatially over the image) and are commonly hard to see in the final result. There are two ways, both described below, to eliminate or adjust for these minor classes. The steps involved in PIT unsupervised classification are:

  1. With all spectral bands loaded, go through the sequence of clicking/dragging by starting at Windows,then Open-->Classification-->Unsupervised-->Partial-->Image--> Left Scheme, clicking on the last one. A new window appears with the Thumbnail image (default) or any other you have chosen using the View routine. At the left side is a series of rectangles in a column, the first labeled Class 1, the second Class 2, and so on, and each outlined by some color (with 27 total, many of these colors will prove hard to distinguish from one or more others). The first 15 Class boxes are visible; those to Class 27 can be accessed with the vertical scroll bar.
  2. On this window in the upper right is a Classifier button. Click on it and select UC1... from the drop-down menu that appears. A dialog box will be displayed in which the parameters for using the UC1 classifier may be selected. For now we'll simply use the defaults. Make sure there is a checkmark in each box (if not that box will be rendered in gray, meaning not classified). Press Run, and sit back while the classification proceeds through these steps, shown dynamically at bottom left: Running Classifier (%); Determining Classifier Boxes (quick, unmarked); and Drawing Classified Boxes. As this last step begins, look at the image - you will start to see various colors superimpose on the image, gradually working from top to bottom. After a few minutes, the entire image is fully colored, often with a truly esthetic (modern art-like) pattern. Note that each class box on the left has a small box with a checkmark and at its right, a percentage that indicates the percent of the total number of pixels that belongs to this class as well as the actual number of pixels within the grand total of the 262,144 pixels in the full scene. Of the 27 classes, 16 will be greater than about 1.5%. At the end, the gray lettering in this legend turns black (i.e., completed).
  3. For the Israel scene, about 9 colors associate with enough area to be distinguishable at first glance or after a brief, but careful look. In order of decreasing percentage, the most common color in the part of the image in the fields (east from the town) is green followed by yellow, blue, orange, peach, purple, red (red is so conspicuous, it seems to have a larger area of display than actual) and brown. The ocean (at the left) may be green or a wine purple or some other color and the sand dunes are a mix of blue-green and darker orange (but other colors may be possible depending on the number of classes you chose). What all this means is that there are probably about 10 classes common enough to be specified during the supervised classification we will next conduct. This is a prime use for unsupervised classification, to identify those classes that are spatially significant, and to display where they are located,so as to assist you in defining classes to include in the supervised classification. Most of the 8 colors found in the agricultural part of the scene are likely crops of different kinds or stages of growth, possible tree groves, and fallow fields.
  4. You can also determine the spatial distribution of any given class by clicking on the checkmark in its left square. After it disappears, the lettering in the legend class boxes goes gray, and a time elapses while the class removed disappears from the scene. This is often very hard to see for classes less than about 5 to 10%. Try it on the red. Eventually, the underlying gray tone of the image will appear in the areas of the scene where the class occurs. Then click in the square to restore the checkmark, and then the red will start filling in top to bottom until all the lettering again becomes black.
  5. Now, let's see what reducing the number of classes leads to. Go through the same procedure as before up into the window containing the Run button. Go through the same procedure as before, but change the number of classes to 12 before clicking Run. Wait until the full unsupervised classification is finished. There clearly are less colors, less of a hodgepodge. Note that the color assignments have changed: the red of the 27 classification has been replaced by blue in the same pattern; the red itself is now the largest single class and is widespread. You can try other specified numbers of class if you wish, to see if simplification results, but remember to restore the "12" to "27" when finished.
  6. We reiterate: An unsupervised classification may or may not actually identify a specific class and locate it accurately. The principal value of this classification is to develop color patterns in the image that correspond to spectrally distinct (separable) classes. Hopefully, many of these will approximate the spatial locations of the individual unsupervised classes in such a way that they could match some land cover features that have unique signatures (and multidimensional spectral values that are not overlapping with other features). The product is thus a guide to seeking out the identities of each class, gambling that it also is a real and separable feature. If you know these identities beforehand, then the supervised classification (next page) will almost always lead to a more accurate "map" of the feature-named class.
  7. It is suggested that, if you plan to now (or soon) do a supervised classification, that you retain the unsupervised one as a reference. Hit the Minimize button [ - ] in the upper right and in Windows fashion it will appear at the bottom button bar. Be warned that when you wish to put it back on screen, and click on that bottom button, its frame outline will appear but the image itself will take a few minutes to be restored to the full screen. Or, you may wish to save the classified image for use in a later session. To do this, go to View at the top, then scroll to Save at the bottom, click and a Window As-->GIF sequence follows. Click on GIF and a Save GIF as window appears. Go to the Directory slot and bring up PITimages in the usual manner. For file enter a name, such as "ISRAELunsup" and hit Save.


Nicholas M. Short, Sr. email:
Jeff Love, PIT Developer ( Next Previous Next Table of Contents Previous