NAME

i.gensigset - generate statistics for i.smap from raster map layer.
(GRASS Imagery Program)

SYNOPSIS

i.gensigset
i.gensigset help
i.gensigset trainingmap=name group=name subgroup=name signaturefile=name [maxsig=value]

DESCRIPTION

i.gensigset is a non-interactive method for generating input into i.smap. It is used as the first pass in the a two-pass classification process. It reads a raster map layer, called the training map, which has some of the pixels or regions already classified. i.gensigset will then extract spectral signatures from an image based on the classification of the pixels in the training map and make these signatures available to i.smap.

The user would then execute the GRASS program i.smap to create the final classified map.

OPTIONS

Parameters

trainingmap=name
ground truth training map

This raster layer, supplied as input by the user, has some of its pixels already classified, and the rest (probably most) of the pixels unclassified. Classified means that the pixel has a non-zero value and unclassified means that the pixel has a zero value.

This map must be prepared by the user in advance. The user must use r.digit, a combination of v.digit and v.to.rast, or some other import/developement process (e.g., v.in.transects) to define the areas representative of the classes in the image.

At present, there is no fully-interactive tool specifically designed for producing this layer.

group=name
imagery group

This is the name of the group that contains the band files which comprise the image to be analyzed. The i.group command is used to construct groups of raster layers which comprise an image.

subgroup=name
subgroup containing image files

This names the subgroup within the group that selects a subset of the bands to be analyzed. The i.group command is also used to prepare this subgroup. The subgroup mechanism allows the user to select a subset of all the band files that form an image.

signaturefile=name
resultant signature file

This is the resultant signature file (containing the means and covariance matrices) for each class in the training map that is associated with the band files in the subgroup selected.

maxsig=value
maximum number of sub-signatures in any class
default: 10

The spectral signatures which are produced by this program are "mixed" signatures (see NOTES). Each signature contains one or more subsignatures (represeting subclasses). The algorithm in this program starts with a maximum number of subclasses and reduces this number to a minimal number of subclasses which are spectrally distinct. The user has the option to set this starting value with this option.

INTERACTIVE MODE

If none of the arguments are specified on the command line, i.gensigset will interactively prompt for the names of these maps and files.

It should be noted that interactive mode here only means interactive prompting for maps and files. It does not mean visualization of the signatures that result from the process.

NOTES

The algorithm in i.gensigset determines the parameters of a spectral class model known as a Gaussian mixture distribution. The parameters are estimated using multispectral image data and a training map which labels the class of a subset of the image pixels. The mixture class parameters are stored as a class signature which can be used for subsequent segmentation (i.e., classification) of the multispectral image.

The Gaussian mixture class is a useful model because it can be used to describe the behavior of an information class which contains pixels with a variety of distinct spectral characteristics. For example, forest, grasslands or urban areas are examples of information classes that a user may wish to separate in an image. However, each of these information classes may contain subclasses each with its own distinctive spectral characteristic. For example, a forest may contain a variety of different tree species each with its own spectral behavior.

The objective of mixture classes is to improve segmentation performance by modeling each information class as a probabilistic mixture with a variety of subclasses. The mixture class model also removes the need to perform an initial unsupervised segmentation for the purposes of identifying these subclasses. However, if misclassified samples are used in the training process, these erroneous samples may be grouped as a separate undesired subclass. Therefore, care should be taken to provided accurate training data.

This clustering algorithm estimates both the number of distinct subclasses in each class, and the spectral mean and covariance for each subclass. The number of subclasses is estimated using Rissanen's minimum description length (MDL) criteria [1]. This criteria attempts to determine the number of subclasses which "best" describe the data. The approximate maximum likelihood estimates of the mean and covariance of the subclasses are computed using the expectation maximization (EM) algorithm [2,3].

REFERENCES

  1. J. Rissanen, "A Universal Prior for Integers and Estimation by Minimum Description Length," Annals of Statistics, vol. 11, no. 2, pp. 417-431, 1983.
  2. A. Dempster, N. Laird and D. Rubin, "Maximum Likelihood from Incomplete Data via the EM Algorithm," J. Roy. Statist. Soc. B, vol. 39, no. 1, pp. 1-38, 1977.
  3. E. Redner and H. Walker, "Mixture Densities, Maximum Likelihood and the EM Algorithm," SIAM Review, vol. 26, no. 2, April 1984.

SEE ALSO

i.group for creating groups and subgroups

v.digit and r.digit for interactively creating the training map.

i.smap for creating a final classification layer from the signatures generated by i.gensigset.

AUTHORS

Charles Bouman, School of Electrical Engineering, Purdue University
Michael Shapiro, U.S.Army Construction Engineering Research Laboratory