"Phoenix Project" Neural Network Data

    The data used in the first series of experiments with neural
networks is drawn from work done at Johnson Space Center in 1974
on Spacelab photographs of the Phoenix metropolitan area.

    The photographs were re-photographed with a circular aperture
corresponding to a 2-km diameter circle on the ground, spaced so
that the area originally photographed was covered by 433 of these
circular "events".

    Each "event" photograph was then processed through equipment
which produced and photographed a Fraunhofer Diffraction Pattern.
Two distinct sampling geometries were used to record the light
intensity patterns of these FDP's: a radial wedge sampling and an
annular ring sampling.  The resulting 195 data points for each
event were the basis of further analysis.

    During the 1974 project, 57 "features" were empirically
extracted from the FDP data points.  A variety of methods were
used to identify classification groupings for the original events.
These were boiled down from 96 to 5 land use classes, and Fischer
discriminant tests were developed and refined to yield 90+%
classification accuracy, using as few as 9 of the features as
input.

    For the neural network experiments, 292 of the total events
were selected to provide coverage of the 5 land use classes.
These, and the frequency distribution with which they are
represented, are:

       Class                  # Events
       __________             ________
       R(esidential)               31
       F(arm)                      75
       M(ountain)                  29
       W(ater)                     31
       U(rban)                    126
                                _____
                                  292

    The data format for all files on these two volumes is that of
NeuralWorks input (.nna) files.  The first section of the record
contains all the floating point numbers which are the event data
points, plus an event ID number.  The second part of the record
contains the desired network output values, in this case a 5-
vector of binary digits, followed by the letter code for the land
use class (see table above), and an event number to match that in
the middle of the record. (The order on the end may vary from file
to file: event number + land use class, or vice versa.) Because the
records are more than one text line long, they consist of multiple
"continuation" lines.  NeuralWorks prefaces each text line after the
first within a single record with an "&" character. The records are all
ASCII.

NOTE ON THE FORMAT STRUCTURE FOR THE RECORDS: These data sets
have been converted from an earlier version of NeuralWorks which
separated the input and output data into distinct record types.
Each could carry comments after the data portion; the event ID and
the land use class are comments. The conversion routine does not
remove comments, or insert the current comment delimiter, "!".  The
data files you have thus have a comment (Event ID) in the middle of
each record, separating the input portion from the output portion.
When devleoping and specifying your network, you must take this into
account in the I/O Parameters section.  Be sure to specify that the
input begins in column 2, and the output in column 57, for the 55-
feature data (54 features plus and Event ID implies column 57 for
output data.)

    This diskette contains both the raw data for the 292 sample
events, that is the 195 light intensity data points from the FDPs,
and the feature data (features 1-54 only) for the same events.

    The Feature data sets contain the original data, without
normalization, and the feature data is presented in two data
sets: the 148-event subset used for network training; and the 
144-event subset used for recall and generalization testing.

      Feature Files   Content
      _____________   ___________________
      PhTrain55.nna   148 training events
      PhTest55.nna    144 recall events
      
    These have the following frequency distributions of the 5 land
use classes:

                        PhTrain   PhTest   

                        ___       ___  
      R(esidential)      14        17        
      F(arm)             40        35   
      M(ountain)         13        16   
      W(ater)            16        15    
      U(rban)            65        61   
                         ___       ___  
                        148       144   


    Also on the diskette are the raw FDP datapoint data sets.
These contain, for each event, the 100 radial wedge geometry
points, normalized to a scale of 0 -1000, and the 95 annular ring
geometry points, also normalized to a scale of 0 - 1000.

     Again, the files have the format described above.  The files
are:

     File Name        Contents
     _________        _____________________
     PhTrain195.NNA   100 training events
     PhTest195.NNA     77  recall events
                      115 events not used in this series

    Their land use class frequency distributions are:

                     PhTrain   PhTest   NotUsed   
                       ___      ___     ___   
      R(esidential)    18       13       0   
      F(arm)           21       22      32   
      M(ountain)       20        9       0   
      W(ater)          20       11       0    
      U(rban)          21       22      83   
                      ___       ___    ___     ___
                      100       77     115     292