CSG -- Signature Generator

Creates signatures for a particular window or a region under a bitmap mask using data from a set of image channels. Each signature consists of channel means, deviations, thresholds and relative 'a priori' probabilities. The user can select the channels, threshold, and 'a priori' probability for the class.

Signatures are most often generated as input to the classification program MLC but also have useful applications in Principal Component Analysis.

See Also: CSR, MLC, MINDIS, PCA

PARAMETERS

CSG is controlled by the following global parameters: 

Name     Prompt                                  Count     Type
FILE     Database File Name                      1-64      Char
DBIC     Database Input Channel List             1-16      Int
MASK     Area Mask (Window or Bitmap)            0-4       Int
VALU     Grey-level Value List                   1         Int
THRS     Gaussian Threshold                      0-1       Real
BIAS     Class Bias                              0-1       Real
REPORT   Report mode: TERM/OFF/filename          1-64      Char

The following parameter receives output:

LASC     Last Segment Created                    1         Int

FILE

Specifies the name of the PCIDSK image file used to derive signature segments. 

 EASI>FILE="filespec"

DBIC

Specifies the image channels to be sampled for signature creation. 

 EASI>DBIC=i,j,...,p

MASK

Specifies the window or bitmap used to define the area to sample, within the input channels, in order to generate the signature. 

 EASI>MASK=i            | use area under bitmap
 EASI>MASK=i,j,k,l      | use area under defined window
 EASI>MASK=             | use entire image

VALU

Specifies an integer value (1-254) that will be assigned during image classification, to the pixels considered as belonging to this spectral signature. 

 EASI>VALU=n
Values of 0 and 255 are invalid, as they will be assigned to the null class and multiply class, respectively.

THRS

Specifies the Gaussian feature space threshold. The greater this value, the larger the hyperellipsoid or parallelepiped in feature space. THRS is a REAL number in units of standard deviation. 

 EASI>THRS=n                    
 EASI>THRS=              | defaults to 3.0

BIAS

Specifies the BIAS or relative `a priori' probabilities for the signature. If all signatures have equal bias, their `a priori' probabilities are equal. BIAS `biases' the maximum likelihood classifier's assignment of a pixel when it lies within overlapping portions of class signatures in feature space. BIAS is a REAL number. 

 EASI>BIAS=nn.nn
 EASI>BIAS=              | defaults to 1.0

REPORT

Specifies the file to which the generated report is appended. 

 EASI>REPORT="filename"
 EASI>REPORT=            | defaults to terminal output
Note: The following names have special meaning: 

 EASI>REPORT="TERM"      | generates reports on your terminal
 EASI>REPORT="DISK"      | generates reports on file "IMPRPT.LST"
 EASI>REPORT="OFF"       | switch off report generation

LASC

The segment number of the created signature is stored in this parameter.

DETAILS

Spectral signature data is derived from image data on selected database channels (DBIC) as sampled under a selected window or bitmap (MASK). This data is principally used to define partitions in the feature space of the image, which will subsequently be used to classify the data. Signatures can also be used in the PCA program to define the new measurement axes.

CSG creates a new signature segment, and stores the following:

The Gaussian threshold (THRS) and relative 'a priori' probability (BIAS) of the signature segment may be specified by the user prior to signature generation. THRS defaults to 3.0 standard deviation units, and BIAS defaults to 1.0. Other values, such as channel means, standard deviations, and lower and/or upper parallelepiped limits may be modified using CSE (Note - initially, the lower and upper parallelepiped limits (LOLIM and UPLIM) are defaulted to the value of THRS).

The `segment-number' of this data segment will be stored in the parameter LASC. Many programs that will use/modify this data will expect the `segment-number' to be stored in the parameter DSIG. The EASI command DSIG=LASC, executed IMMEDIATELY after the execution of CSG, will move this value into DSIG.

Use CSR to generate a soft- or hard-copy signature report on the report device of your choice.

Use MLC for multi-class classification of this signature in conjunction with other class signatures.

Use PCA to optimize the enhancement for a particular image feature or to transform the image data according to the statistics of a different image dataset.

REPORT

CSG only reports alterations that it forces on the class CORRELATION matrix to ensure that the COVARIANCE matrix is well conditioned. If a report from CSG does NOT appear, even though you asked for one, this means that no `conditioning' of the CORRELATION (and concomitantly, the COVARIANCE) matrix has occurred.

CSG's matrix alteration report will appear on your terminal even if you have turned reporting off.

To get a full printout of the statistical vectors and matrices to be used in subsequent classification, see the CSR program documentation.

MESSAGES and WARNINGS

The mathematics used in creating a signature segment is based on statistics and involves matrix manipulations. A number of situations can occur which will result in statistical or mathematical problems. These include the following:

In general, CSG will attempt to reduce these problems by introducing minor modifications of values. The user is informed of these actions and re-conditionings with one of the following messages: 

 Error 81: Sample size too small.
For mathematical reasons, a training site must have at LEAST one pixel set for each channel used in DBIC. For example, if DBIC specifies 4 channels then the training bitmap specified by MASK must have at least 4 bits (i.e., pixels) set. 

 Sample size = xxx  (for n channels, yyy is recommended).
This is a warning only and can be ignored. For statistical reasons (involving confidence in estimates), training sites should be at least 5*(n*n+n) pixels large. Larger training sites usually yield better statistics and thus better classification results. 

 Standard deviation xx was forced from xxxx to 0.0001.
This is a warning only and can be ignored. Usually, this message indicates that all (or almost all) the pixels in a channel under the training site were of the same grey level. To prevent mathematical problems, a small variance (0.0001) was introduced. 

 Correlation matrix cell (n,m) was forced from xxx.xx to yyy.yy.
This is a warning only and can be ignored. Usually, this message indicates that two or more channels in the training site were highly correlated (very similar). A small change is introduced to prevent mathematical problems.

EXAMPLE

Example 1: This example creates a spectral signature (based upon TM bands 5, 6, 7, and 8) for the residential land use of the IRVINE.PIX database. The contents of this signature can later be listed (using CSR) or edited (using CSE) before it is used as input to the maximum likelihood classifier (MLC). 

 EASI>FILE="irvine.pix"   | database file
 EASI>DBIC=5,6,7,8        | image channels used to define sig.
 EASI>MASK=12             | bitmap of residential training areas
 EASI>VALU=64             | theme value assigned during classification
 EASI>THRS=               | default threshold
 EASI>BIAS=               | default bias
 EASI>RUN CSG
Example 2: This example generates a signature (based upon TM channels 5, 6, 7, and 8) that will serve as the basis for generating enhanced images (using principal component analysis - PCA). Once generated, this signature can be used repeatedly in PCA to apply the same transformation to datasets representing adjacent scenes (the same channels) or the same scene from different dates. The purpose of this is to maintain a consistent colour representation among several images of a particular scene. 

 EASI>FILE="irvine.pix"      | database file
 EASI>DBIC=5,6,7,8           | image channels used to define sig.
 EASI>MASK=12                | bitmap of residential training areas
 EASI>VALU=64                | value irrelevant in this context
 EASI>THRS=                  | value irrelevant in this context
 EASI>BIAS=                  | value irrelevant in this context
 EASI>RUN CSG
About PCI Help Gateway