THE PIT OPERATIONAL MANUAL
Supervised Classifications
The climax of our learning experience with PIT is now upon us - producing a
supervised classification of the Israel scene. In this, you will assume some
interpretive knowledge, based on your experience and common sense in identifying
various categories to establish the classes to be mapped onto the image. In
the Israel scene, several are obvious: the Mediterranean Sea; the sand dunes;
the towns, active growing vegetation, and fallow fields. This is well-displayed
in the standard false color composite, which we will adopt as the image to use
in specifying training sites. We will select the cell block method of picking
samples within classes that is the easiest to use in the PIT program.
We will start by specifying only 8 classes; later you may elect to rerun the
classification using a larger number that depends on your confidence in visually
picking out the sites where new classes seem best displayed. Our first attempt
will work with the full scene to choose the cells. That has problems which will
become obvious as you proceed. After you get your first classified image, you
will be encouraged to redo the process using movable enlarged portions of the
scene, which in effect makes the sites for cell blockage larger. Also, we will
start with the Maximum Likelihood Classifier; the PNN and PDM classifiers will
be explained later. So, onward.
- As you did before, bring up a band 4, 3, 2 (R,G,B) color image of the Israel
scene, going through the View, Display Image Control Window, RGB button routine.
After the scene is up, drag the RGB window off the image and place to the
right. Make any C and B adjustments you think make class distinctions easier.
- Inspection of the Israel image yields these obviously different classes,
whose shapes and colors lead to separability: (Sea)water; Town; Sand Dunes;
Active Crops (strong red tone). Four others show well enough to warrant designation
as classes but their identities are more nebulous: Other Crops (dark red);
Dark Fields (darkish grayish); Fallow Fields (grayish brown); Natural Surfaces
(yellow brown). Other classes are seemingly present but their areas are too
small to be sampled by the cell size we will choose. This is true, for example,
for linear features like roads and the airport runways. So, we will stick
with only 8 classes at this juncture.
- At this point, go to the PIT Window and click on Scheme. One option is in
gray; (not activated), the others in black (active). Click on "Add Class".
A window will drop down, labeled "Enter a Class Name and Color" In the first
Name Box, type "Water". Go to the Browse Color button and click. A long list
(requiring a scrolling button) will appear. Click on "Blue" when it appears:
part of the window will take on this color. Press Enter and the color name
will appear in the Color Box. Press Add and both boxes will be cleared. Next,
type in Town, go to the Color Menu, scroll down to Brown, and repeat the rest
of the procedure. For the rest of the classes we set up in this first try,
the names/colors are: Sand Dunes = Yellow; Mature Crops = Dark Green; Other
Crops = Pale Green; Dark Fields = Gray; Fallow Fields = Light Pink, and Natural
Surface = Medium Purple. The selection being done, press Done. (Note: if you
already know a color name [after familiarity with the list], you can elect
to type it directly into the Color box rather than scroll the list itself.)
- Raise (de-iconify) the Interpretation window (if it had been minimized).
The classes created will now appear as a list of color outlined rectanlges
at the left (or top). The scene may be the gray image or the color view, depending
on what was saved or minimized earlier. If the gray image, then convert it
to the false color version in the usual way (Display Image Control; RGB).
You will observe that a black grid, with widely spaced squares, is superimposed.
Each square encloses 64 x 64 pixels. For this scene, the square box contains
too many different classes - many in their visual expression are much smaller
than the box (hearafter called a Cell). It is necessary to either reduce the
Cell size (increase the total number) or, as we will do later, greatly enlarge
part of the scene (Zoom). To reduce the cell dimensions, go to View, then
down to Cell, and click. The small window to the right contains two options.
Clicking on Size produces a window with a series of n x n sizes, with the
64 x 64 size checked. Change to 16 x 16 by moving the cursor to that position
and clicking, which will check it. If you wish to change the color of the
grid, then click on Grid, the other option, to a new color. Here, we will
stay with black. Also, if you wish to remove the entire grid (perhaps temporarily),
click on View, then Show, and click off the checkmark in the right window
next to Grid; to restore, re-enter and click it on.
- You are now ready to start filling in the grid cells with local samples
of the class you interpret to be in each. Try to find the most "pure" examples
but a fraction of a cell containing one or more other classes (visual differences)
can be tolerated. Lets start with water in the upper left. Go to the legend
box labeled Water and click on the small circle at its left. A black dot (bullet)
will fill it. The mouse cursor will change shape to a larger white dot. In
the image find the sea (upper left), place that cursor in a cell enclosing
water and click. The cell fills with blue. Do this for a few more cells adjacent
or near by. Note that in the legend box, each time you add a cell, the total
(16 by 16 or 256) increases the score shown. You should try to have at least
1500 and perhaps 2500 or more cells thus picked for each class.
- Next, activate the black dot in the next box, Town, outlined in brown. Find
what you interpret to be examples in the image and click on enough cells (brown)
to meet quota. On to obvious sand dunes below the town. Activating its circle,
pick at least 8 boxes (filled with yellow). Do likewise to the remaining five
classes. The Mature Crops are those in bright red. You will be able to select
perhaps 2-3 cells at any one area of the image,so you must go to several areas
to activate at least 8 cells. For Other Crops (darker red), these being smaller
in size, you probably have to go to 6 to 9 different locations. Training sites
for Dark Fields are apparently sparse, at least in areas large enough to fill
a cell. Look at the right center margin for one such site; also in the lower
left. You may not find even 6 "good" cells; accept a smaller number. Fallow
Fields, present in the color composite as shades of gray-brown, are even smaller,
so you have to hunt for cells one at a time that meet the conditions. The
last class, Natural Surface, may be "artificial". But the terrain north of
Ashdod looks a bit different from sand dunes and may be a barren surface with
sand and soil. It has a more subdued yellow-brown color and some texture.
Enter several cells there and again at a small similar area in the lower left
corner.
- You have now selected samples of all classes, associated with certain cells
in the grid. There is no Save or Close Button, but the class selection information
is saved as long as you are actively working in PIT. However, you need to
remove this display while you are engaged in the next classification step.
There are two options: 1) you can just close the window by hitting the minimize
button [ - ] at the upper right; this will place a PIT - Interpretation button
at the bottom of your MS-Windows screen; or 2) you can seemingly close the
classes window by pressing [X] at the upper right; to recover this window,
just go to the main PIT window (which may be minimized also; click to activate),
then to Windows, and then Open - Interpretation - Left Scheme, and the image
with the class cells on it will come up; but, if you had saved through [X]
as a Zoom enlargement, this will not appear but instead you will see the full
scene with all colored cells located (if you wish to revert to an enlargement,
use the Zoom routine).
- Also, you may have made a mistake or two in coloring a cell with what you
decide is the wrong class; to correct that, select the proper class (from
the class scheme) and click with the mouse cursor in the cell being corrected.
And, you may decide to omit a class. If so, go to Scheme and select Delete
a Class; a menu will appear to the right with all classes listed; delete by
clicking on the desired class. There is also a Delete All Classes option,
if you wish to start over. Or, you can retain the classes and desire only
to choose a new set of cells; this involves Scheme - Clear Interpretations,
with a list of classes in the right window that appears; click on the class
whose cells you want to remove, and then on the Clear Button that appears;
or you can remove the entire group with the Clear All Interpretations button.
- At this stage, it should prove informative to look at the spectral signatures
of each class to judge how much real separability there is between any pair
of classes or all the classes together. You can do this now because you have
taken samples of each class so that appropriate statistics can be calculated.
Go back to the PIT window: click and drag on Windows - Open - Signature -
Small View (click). A new window labeled PIT - Signature shows up at the upper
left. Click/drag on Plot - Source - Interpretation - Spectral - Image (click).
Wait about 10 seconds. Then in the black window you should see spectral curves
(as straight line segments) for all 8 classes you set up, each with the color
assigned to it. The abscissa is simply the number of spectral bands, plotted
at equal intervals; the ordinate is the DN range from 0 - 255. There is a
scroll button: by dragging it down you will see each class spectral curve
by itself, with its name labeled. Some comments about what you can conclude
from the plots: all classes seem separable, i.e., even if several are close
in DN value for one band, there always are one or more bands that show significant
differences; Town and Fallow Field are most closely alike; there are peaks
at bands 3 and 5 (count from the origin to the third and fifth dots on the
curves) suggesting that overall, those bands are brighter; all band 2 values
are lower, implying that this band may be darker overall owing to either sensor
or calibration conditions.There are strong peaks for band 6 and all the curves
converge to a narrow range; this indicates that the DN values are similar
and radiances were not much different for the classes involved; this is borne
out by the image when displayed - it is tonally flat with little variation.
In general, if that is the case, it is wise to omit band 6 from the classification;
use band 6 only if there tends to be bright and dark patterns that indicate
hot spots and cool areas. For the classification we will now do, include band
6, but if you redo this at some other time, you can elect to drop band 6.
- We are now at the climax - ready to do the final classification. Make sure
the PIT - Interpretation window is minimized. Bring the PIT window up and
click on Windows. Then, follow the usual click/drag sequence: Windows - Open
- Classification - Supervised - Image - Left Scheme. This will bring forth
the large window with the image (Band 1, in this case) and the 8 class legend
(color outlined) to its left. Click on the Classifier button (on the menu
bar) and select ML... from the drop-down menu that appears. A dialog box will
be displayed in which the parameters for using the ML classifier may be selected.
We'll use the defaults so just click the Run button. The progress will be
shown at the lower-left: Creating temporary PIT file - Creating temporary
training set - Running classifier - Determining classifier boxes - Drawing
classifier boxes. When that last statement arrives, you will begin to see
colors at the top of the gray image that will progress downward until the
entire image is filled. This is the Maximum Likelihood classification you
sought.
- In general, the result should look believable (yours will vary from mine
because you almost certainly chose different cell sites for the classes; but
there should be strong similarities). The ocean water should just be in blue
in the upper left; there is a small lake near the scene center which may or
may not be in blue. Probably, the color assigned to Town shows up more widespread
than you expect. It should be nearly solid and continuous near the wharves
and dominant but mixed with other classes just inland; this color shows up
also in local concentrations within the vast agricultural part of the image,
and denotes small settlements, the airport, and the village of Lob. Both green
patterns look realistic as indicators of active crop growth. Fallow fields,
in pink, probably are the most prevalent class in the scene - you may judge
that there is perhaps too much of this class, depending on how you selected
your cells. The natural surface (purple) seems meaningful.Look at the percentage
of each class in the Legend.
- One class, Dark Fields, will likely be around 5%. It does not stand out
- the gray color given to it doesn't contrast enough. It is easy to change.
Click on Scheme, look in the window that drops for Modify Class, and a new
window with the eight classes list will appear to the right. Click on Dark
Fields and a small version of the window you used to select Class Name and
Color appears. Keep the name and browse through the color list. Check any
new one you wish, but we suggest Grey 30. After a few moments, that dark color
will replace the light gray. Consider this to be the final version - in effect
a land use map of the Ashdod coast and inland agriculture. You can now elect
to do one of three things: 1) keep it active, but minimized; 2) save it, or
3) delete it. For now, minimize it. Note that both the label at the top of
the classified scene and the minimize button call this Classification/1.
- So, now we ask you to pause a moment and think through this question: What
could you have done to have made this classification easier to do? PAUSE.
Well, the biggest problem you no doubt had was to find single cells that were
relatively "pure".Other Crops, and particularly, Fallow Fields and Dark Fields,
were usually hard to find as the predominant class in many/most cells. They
are often just too small and must share the cell with at least a second -
sometimes third - class. How can you get around this problem. Make the cells
smaller - not practical under the circumstance - OR make the image larger,
that is, zoom it up. Lets try the latter. To do this, restore the PIT interpretation
window from its minimized button. Click on Scheme and then Clear All Interpretations.
This eliminates your previous cell selection. Note that PIT Interpretation
has a Mode Button. Click on it and note that its window has a Zoom In option.
Click on this. A square cursor appears that can be placed anywhere in the
full scene image. Put it somewhere near the middle of the upper left quadrant.
Click, and an enlarged (by 2 x 2, the default) scene replaces the full one.
But, the cell size in the grid remains the same. However, each individual
cell (which encompasses 8 x 8 or 64 cells) in the grid "straddles smaller
parts of the scene, so that there is a higher likelihood of finding "pure"
classes within some given cell. Note, too, that there are both horizontal
and vertical scroll bars. When you move right or down or both, you will traverse
through the entire scene (fully down and right moves the image to its lower
right corner in the full scene). Thus, you can still choose cell classes over
the entire image - each cell filled in just samples smaller areas.
- Do this, that is, choose new class cells, minimize this Interpretation,
and proceed thru the PIT- Classification as you did before. We suspect you
will find it easier to select good examples of dominantly Dark or Fallow Fields,
and you can block out the town better than before. After your classification
is displayed (as Classification/2), see if you have obtained a reduction in
the percent of Fallow Fields and greater percentage of Dark Fields. The Town
may also be more "compact" and realistic. As a general rule of thumb, we have
concluded that, if you know a fair amount about the categories present in
any scene you plan to classify and if there is a high proportion of small
areas that nevertheless appear to be valid classes, you will achieve better
results if you select your class training cells from at least one level up
in Zoom (zooming in too much tends to present a scene with a patchwork of
blocky pixels). Close this classification by minimizing it.
- What might you do next. We suggest that you repeat the classification, in
the steps outlined above, but with one change. Choose PNN instead of ML. Try
this now. But, heed these warnings first: The PNN (neural network) classifier
is much slower than ML (maximum likelihood) - it took about 15 minutes on
a 200 MHz machine to complete "Running Classifier". Part way into an ultimately
successful run, the ScreenSaver pattern came on. The processing continued
but the image and legend disappeared only to restore towards the end, after
repeated pushing on random keys in the usual way to restore the screen. Suggest
you hit a key periodically in a time interval less than drives the Screen
Saver. In any event, the final PNN classification appeared and subjectively
was judged the better of the two (PNN vs ML). The Town was sharply delineated.
The amount of Other Crops seemed a better representation in the PNN version.
But, reserve judgment for yourself after you succeed with this PNN classification.
- As an option, run the PDM (Polynomial Discriminate Method) classifier. Do
exactly the same as before, except select the PDM option. It will take about
the same time as PNN. Results are similar to PNN but there is a real difference
from ML. In the dual run we made, ML distribution for classes Dark Fields
and Fallow Fields was 8.5% and 40% and with PDM was 22% and 23% respectively.
This could depend much on the choice of specific cells but in this comparison
the same cells were used. The distribution of percentages is thus sensitive
to the particular areas in which the training site cells are located and on
the classifier used. Which classification is best can only be determined by
comparing with actual ground truth, but intuition helps.
- While we're at it, lets run a revealing experiment. Let us peform a standard
ML classification but on only bands 2, 3, and 4. This will in effect simulate
a Landsat MSS image. Lets see how well this reduced number of bands can achieve
a suitable distribution of classes. Request a ML classification as before
but before clicking Run specify that only bands 2, 3, and 4 should be used.
Do this by clicking on the buttons labeled 1, 5, 6, and 7 in the "Spectral
Bands:" area of the dialog box. This will cause them to change to a light
gray (rather than black) indicating that they will not be used. Now click
Run to start the classification. After the classification is displayed, look
at it and, if you retained any other classification, compare the two. Your
conclusion will likely be: the "MSS" classification did almost as well as
those based on 6 or 7 TM bands. Why? Largely, because this Israel scene is
dominated by vegetation, so MSS 6 and 7 pretty much match TM 4. If the scene
had contained considerable rock materials, and certain other classes, these
differentiate better when TM bands 5 and 7 are available to distinguish their
special characteristics. When (if) you decide to classify the several other
scenes in this PIT Appendix, those that contain rock materials (e.g., the
Waterpocket Fold scene) should benefit from utilizing TM bands 5 and 7, and
probably 6 also.
- As an aside, the writer (NMS) experimented with a maximum likelihood classification
using TM bands 1 through 4 and selecting 6 of the 8 original classes, eliminating
Dark Fields and Natural Surfaces. The result was to make Towns appear more
realistic - Ashdod was more widespread - and to replace Dark Fields and Natural
Surfaces with Fallow Fields. The overall effect was a "cleaner" (sharper)
classification but at the cost of omitting two classes that are discrete and
probably real and worth mapping.
- PIT also facilitates classification using PCIs rather than spectral bands.
If you are curious about the nature and appearance of such a classification,
make some number of PCIs, as described before, and run a ML classification,
specifying PCIs instead of spectral bands in the dialog box. Interesting,
eh!
- You may have noticed a button labeled Training Set that appears on both
the PIT - Interpretation and the PIT - Classification title bars. Since we
won't be using this function, ignore it. But, its purpose is to establish
training sites for use with a classifier (e.g., Miminum Distance) not a part
of PIT but one that can be used in some other processing software package
or with a classifier that can be imported to PIT. Also, under the Image button
is an option called Palette. We will not use this either, but it refers to
the use of a color palette image control. There are several other functions
and procedures on PIT that, again, will not be integral to this training exercise.
You can learn something about most of these by scanning through the Help explanation.
- The last thing you will want to add to your PIT skills is the ability to
save your work. From the PIT window menu bar, click on the PIT button and
select "Save As..." from the drop-down menu. A dialog box will appear. Navigate
the the PITimages folder (most likely by clicking "Parent Directory" and then
double-clicking on PITimage in the new list of files displayed). In the "File:"
field type in the name of a file with a ".pit" extension and then click Save.
This will save your current PIT session to that file (called a PIT file).
A PIT file contains the image you were working on, the image controls selected,
classes created, and all of your interpretations. Now, to see it, let's suppose
you want to check out the PIT - Interpretation group of training cells you
diligently selected earlier. And let's assume you had exited the entire PIT
program (after saving the PIT file!). Get back into PIT from scratch. Hit
the PIT button on the left end of the title bar. Then on Open. A Select a
PIT File window appears. Enter the PITimages directory in the Dir. Box and
then insert the file name you chose (can access it through Parent Directory,
and it will show up on a list [click on the file desired and it enters automatically;
if you remember its name, just type it in) and press Enter. It is now in active
memory. To see the training site display image, from the PIT bar, click on
Window - Open - Interpretation - Left Scheme, as you've done before, and yesterday's
work reappears. You can also save the results of a classification as a GIF
file. To do so click the View button on the menu bar of the Classification
window and select "Save As - GIF..." from the drop-down menu. You will be
prompted for the name of the GIF file to save.
Primary Author: Nicholas M. Short, Sr.
email: nmshort@epix.net
Collaborators: Code 935
NASA GSFC, GST, USAF
Academy
Contributor Information
Last Updated: September '99
Webmaster: Bill Dickinson Jr.
Site Curator: Nannette Fekete
Please direct any comments to rstweb@gst.com.