This part of Appendix B is essentially a step-by-step training manual, or
"Cookbook", intended to aid the user (you) in learning how to operate
PIT, the Photo Interpretation Tool. The most recommended approach for those displaying this page from a CD-ROM or from the Internet is to print out pages B-8 through B-13. There is, however, a separate manual that was prepared by Dr. Jon Robinson; it's style may be confusing but it sometimes answers uncertainties that the Cookbook may fail to address. To access and print it out for consultation, follow this pathway. C --> July1_2002 --> PIT --> PITnew --> RAC Documentation --> RAC-1 --> User Manual --> PIT. From the list (or icon) then displayed, highlight and click on: PhotoInterpretationTool.htm . This brings up the manual; inspect it and decide whether you want to save or print its contents.
At the outset, be advised that PIT will not work properly on any screen resolution with less pixels than 1024 x 768. If you plan to work with PIT, you must configure to that size. This will, of course, cause images from any part of the Tutorial to appear smaller (relative to across-page text length) on your screen. If that is unacceptable, but you want to try PIT, simply restore your favored resolution afterwards. PIT was originally written in Unix but the program has been modified to run
under Windows. The version you will download from the CD-ROM or from the Internet
is designed to operate only in Microsoft Windows (NT, XP, 98[?}), i.e., is oriented
towards PC systems. PIT currently can handle standard Band Sequential format
which, for Landsat, expresses each image point in a DN range from 0-255 (single
byte; 8 bits). Both Landsat MSS and TM are accepted, as is properly formatted
AVHRR and GOES data; radar and hyperspectral data sets are not supported.
The data sets you will create presently cannot be saved under certain circumstances
(exceptions will be noted). Thus, PIT is essentially a training tool for learning
by doing the basics of image processing. It is not yet designed to yield permanent
images that can be ported from the program to external directories for other
uses. PIT can display a raw image,
or a specially stretched single image representing some given band. This image enhancement is not through a program such as linear stretch or histogram equalization (described in Section 1), but by simply repositioning movable bars associated with Brightness (B) and Contrast (C) that appear when individual bands are displayed. PIT does
not routinely allow for simultaneous display of multiple images, so that visual
comparison must be carried out by creating or calling for one image at a time.
PIT also permits color composites, using three input images assigned arbitrarily
to the red, green, and blue monitor guns. These can be natural color (TM band
1,2,3), standard false color (2,3,4), or other color-band combinations (e.g,
7R, 5G, 1B). PIT is also capable of producing single ratio images in black and
white, or colored ratio composites. It is likewise able to conduct Principal
Components Analysis (PCA) generating as many Principal Components Images (PCI)
as input bands. These also can be combined in groups of three to generate color
composites - both true and false.
PIT's primary use is in image classification. Clustering routines are a starting point. Both unsupervised and supervised
programs are included. The three supervised methods are Maximum Likelihood (ML),
Probablistic Neural Network (PNN), and Polynomial Discriminate Method (PDM) (the Minimum Distance Classifier is not included).
Outputs are color-coded images with a legend of classes (best confined to 20 or less) exhibiting color assignments shown
either on the top or the left of the classified image. Training sites for each
class are selected by filling in some number of cells that appear as a grid
over the image being interpreted; the cell size can be varied. Statistics on
the class distribution are available.
Once chosen, the class DN values can be displayed as spectral signatures. Histograms for each input band can also be shown. A portion of a
band histogram can be selected and then all pixels in the image having the DN
values within that segment may be highlighted in color for those pixels within
the image falling within this DN range. Scatter diagrams plotting the distribution
of DN values for any two bands are producible.
