NAME

r.nntool - Neural Network classification tool for raster maps.
(GRASS Raster Program)

SYNOPSIS

r.nntool

DESCRIPTION

r.nntool is a fully interactive, menu-driven GRASS routine that uses a neural network to train landuse classifications and build a classified image. r.nntool facilitates the use of neural networks and linear classifiers in supervised classification of raster cell files.
Neural networks are composed of simple non-linear computational units called neurons that are linked together and work cooperatively to solve complex mapping problems. Neural network output is validated by the use of a controlled training dataset. For application in the GRASS framework, each input unit (equal to one satellite image pixel) to the neural network is assigned a raster map layer, and training data for the network is collected on a cell by cell basis. Typically, a single map layer is used for selecting training sites, although this requirement can be relaxed so that output units can be assigned to more than one map layer, and the map layer used in selecting training sites need not be used as output to the neural network. Since a maximum likelihood classifier (i.maxlik) already exists in GRASS, many of the utilities of i.maxlik in selecting and analyzing training data were used in the neural network tool. Among these utilities is the ability to visualize and, if necessary, change histograms from each training site.

The program for the neural network tool is structured in such as a way that training classes selected in the neural network tool could also be used in the maximum likelihood classifier. This facilitates the implementation of i.maxlik within GRASS for validation of neural network outut.
In GRASS, the maximum likelihood classifier assumes a Gaussian distribution for the training data, which is a widely used method for satellie imagery classification. In the use of the tool, the user is asked to enter the name of the output map layer, the number of output classes, and the names of the input map layers.
Using the lump option of the menu, the tool selects the "dominant" category within a specified window and generates a new map layer. The user can reset the resolution to the specified window, or retain the old resolution in which he entered the tool. In existing GRASS routines, when resolution (window) of a region is enlarged, the middle pixel of the window in the lower resolution is selected.
Training areas are selected using the define areas option.
Using the zoom option the user can zoom out parts of the output map in which he wishes to delineate training areas. Training areas can be delineated by clicking on points, drawing circles, or by drawing polygons.
Using the delete function, the users can interactively select and delete polygons, with the number of samples after deletion shown in the window.
Histograms of training sites can be examined and signatures saved so that the user can use i.maxlik. Once the user is satisfied with the training sites selected, all input map layers are sampled for their data. At intersections of training areas with input map layers, training data for the neural network are gathered.
Training data are stored as an ASCII file so that the user may examine and change it, if necessary. Input data to the network is obtained cell-wise from all areas of the input maps.
The classes option of the neural network tool lets a user examine the distribution of data when two input map layers are used. For higher input dimensions, it is necessary to link the tool to a more sophisticated program such as xgobi. The user may eliminate outliers, and data conflicts by drawing rectangular boxes around data points. If necessary a whitening and diagonalization operation can be done on the data so that better class separability is achieved. Unlike in using traditional classifiers, careful preprocessing of the training data should be performed since neural networks give equal consideration to all data.
Once the user is satisfied with the class distributions, the configure option is selected. Here the user selects a quick propagation network, or the traditional back propagation The quick propagation network uses gradient descent to adjust weights and assumes a parabolic shape for global minimum. Iterations of the network are performed by the number of training cycles set by the user. Back propagation uses gradient descent and converges to a root mean square error value set by the user.
In r.nntool, performance of the network as training progresses is shown on the left half of the GRASS screen. Once training of the neural network is complete, the user propagates cell values of the input map layers through the network. The new map layer generated by the neural network can then be queried. Upon completion of network training, the user may save the neural network structure such as the number of input, hidden, and output units, and the network weights.

Notes

The training site I/O data is stored in a file called o_train (o_ for "old" training file). This is the default file used for training the neural network. If the random option is used to rearrange the training data, the network training data is stored in a file called TRAIN. The classes option can only be used right now for visualization of two input vectors. The color scheme isn't all that exciting and future upgrades are working on using a better color scheme to represent the training data in each class. [Users may wish to also ftp xgobi.tar.gz, and use that tool to visualize training samples in the o_train/TRAIN file]. The Bayes' classifier right now can only be used if you have access to IMSL (otherwise, you'll have to wait until routines are written for diagonalizing a covariant matrix etc in C). Users will have to do a priori determine the attribute value for each class in the output map:

Ex., Say there are 5 classes. Extend the attribute values for the classes from 0 to 100. So that,

       0 - class 1 
      25 - class 2 
      50 - class 3 
      75 - class 4 
     100 - class 5. 
This is a limitation of GRASS since the color intensities of a map are determined by the attribute values. The input values to the network are scaled by the highest attribute of each input. Users may wish to try other schemese such as sgn(x)(1 + ln|x|), or transform the data using a squashing function such as tahn(x). Users will have to look at the source code to do this (see nntool.c).

SEE ALSO

imagery
i.maxlik
r.reclass

AUTHORS

Ranjan Muttiah, TAES, Blackland Research Center
Bruce Byars, GRASS Research Group, Baylor University