UNMIX

Description

UNMIX creates fraction images for a set of class signatures. The input signatures are assumed to represent spectrally pure classes or "endmembers". The output fraction images store values indicating the percentage of each pixel that is composed of each class.

Parameters

Name	Type	Caption	Length	Value range
FILE *	String	Input file name	1 - 192
SIGFILE	String	Input signature file	0 - 192
DBEM *	Integer	Input endmember signature segments	2 - 16	-1024 -
DBOC *	Integer	Output endmember fraction channels	2 - 16	-1024 -
RMSCHAN	Integer	Output RMS-error channel	0 - 1
MASK	Integer	Area mask	0 - 4	Xoffset, Yoffset, Xsize, Ysize
RANGE	Integer	Scaling range (min, max)	0 - 2	Default: 0,255
NORM	String	Normalize results	0 - 3	YES | NO Default: NO
REPORT	String	Report mode	0 - 192	Quick links

* Required parameter

Parameter descriptions

FILE

The name of the PCIDSK file that contains the image channels and endmember signatures to unmix.

SIGFILE

The name of the PCIDSK file that contains the signature segments. If no value is specified for this parameter, the value specified for FILE will be used.

This parameter is optional.

DBEM

The segments to use as endmember signatures. Signatures (type 121) must be created by running the CSG algorithm.

You can specify from two to 16 signature segments; however, you must not specify duplicate signatures.

Ranges of channels or segments can be specified with negative values. For example, {1,-4,10} is internally expanded to {1,2,3,4,10}. When you are not specifying a range in this way, only 48 numbers can be specified explicitly.

DBOC

The output channels to which to write the endmember fraction images.

You must specify one output channel for each input signature segment.

You can specify from two to 16 signature segments; however, you must not specify duplicate signatures.

Ranges of channels or segments can be specified with negative values. For example, {1,-4,10} is internally expanded to {1,2,3,4,10}. When you are not specifying a range in this way, only 48 numbers can be specified explicitly.

RMSCHAN

The output channel to which to write the residual-error image.

If no value is specified for this parameter, the residual-error image is not saved.

MASK

The window or bitmap that defines the area to process in the input raster.

If you specify a single value, it represents the channel number of the bitmap segment in the input file. Only the pixels under the bitmap are processed; the rest of the image remains unchanged.

If you specify four values, they define the x and y offsets and the x and y dimensions of a rectangular window identifying the area to process. Xoffset, Yoffset define the upper-left starting pixel coordinates of the window. Xsize is the number of pixels that define the window width. Ysize is the number of lines that define the window height.

If no value is specified, the entire channel is processed.

RANGE

Specifies the minimum and maximum output values to use for scaling the fraction images (unscaled values are typically between 0 and 1).

NORM

Specifies whether to normalize output channels.

REPORT

Specifies where to direct the generated report.

Available options are:

• TERM: generates a report on the terminal (default)
• DISK: generates a report on file "IMPRPT.LST"
• OFF: turns off report generation
• <filename>: appends a report to the specified file

Details

UNMIX performs linear spectral unmixing given set of class signature segments (DBEM) created by CSG. Each signature must have been created using a training site bitmap that is "spectrally pure", meaning that every pixel in the training site is 100% in a given class. The output fraction images (DBOC) contain values representing the percentage that each pixel covers in each class. Optionally, the RMS (root mean squared) error can be saved (RMSCHAN).

Before scaling, the output fraction images contain values that nominally represent fractional percentages between 0.0 and 1.0. Ideally, the sum of all unscaled fraction images should add up to approximately 1.0 at each pixel location. Typically, however, output fraction images contain values outside this range. The RANGE parameter (Scaling Minimum, Scaling Maximum) is used to specify the desired scaled output values that correspond to the unscaled values of 0.0 and 1.0 respectively.

The Scaling Range (Scaling Minimum and Scaling Maximum) defaults to 0 and 255, but we recommend setting the minimum and maximum values to greater than 0 and less than 255, respectively.

By default, the output images are not normalized. If the last class input endmember signature (DBEM) represents a shadow area to be removed from the data, set normalization to "YES" to remove the effect of the shadow from the previous class signatures before saving the output fractional images.

Example

EASI>file	=	'irvine.pix'
EASI>pciop	=	'ADD'
EASI>pcival	=	8

EASI>run PCIMOD

EASI>sigfile	=	'irvine.pix'
EASI>dbem	=	17,-24
EASI>dboc	=	8,-15
EASI>rmschan	=	16
EASI>mask	=	
EASI>range	=	100,200
EASI>norm	=	

EASI>run UNMIX

Algorithm

The algorithm used to perform linear spectral unmixing is described in the paper cited in References.

Land-cover maps are important tools for environment quality assessment, climatological modeling, and other studies in Earth sciences. The traditional approach to land-cover mapping from remotely sensed data is through classification of each pixel to one land-cover type. Classification methods are not appropriate, however, when quantitative land-cover information is required at the sub-pixel level. This applies to mapping urban land cover from Landsat Thematic Mapper (TM) data, in which a large number of pixels exist containing spectral contributions of more than one land cover.

An alternative approach to land-cover mapping is the linear spectral mixing analysis method. It has been used in many geological applications. Unlike laboratory spectral reflectance, which is usually measured from pure materials, a large proportion of remotely sensed data is spectrally mixed. In spectral mixing analysis, it is assumed that signatures of a subset number of surface elements can reproduce the observed spectra when mixed together in various proportions. This subset may be referred to as endmembers, components, or factors; they may be mixtures themselves.

Suppose there are m spectral bands in a remotely sensed image, and there are n endmembers. Let rij represent the spectral reflectance, radiance or digital number of jth endmember at ith band. All the reflectances, radiances, or digital numbers can be arranged in an m X n matrix, R, in the following form:

           /                    \
          | r11   r12  ...   r1n | 
          | r21   r22  ...   r2n |
      R = | ....                 | 
          | rm1   rm2  ...   rmn |
          \                     /

For each individual pixel, there are m responses observed by the sensor, corresponding to m spectral bands. In a vector form, they are represented as D = (d1, d2,...,dm)'. All area fractions for the n endmembers are denoted as f = (f1, f2, ..., fn)' and they should sum to 1. The linear mixing model can be described in the form:

        d = R f

with the following constraints on f:

        fj > or = 0 and   SUM fj = 1, j=1,2,...,n.

Fractions of each endmember, i.e., f, can be obtained by using the singular value decomposition method with least squares. To obtain a deterministic solution, the number of endmembers should not exceed the number of spectral bands; that is, n < or = m. After the image endmembers are identified, the entire image can be unmixed, pixel by pixel. When f is obtained for each pixel, the appropriateness of the least squares estimation of f can also be obtained.

References

Gong, Miller, Freemantle and Chen. "Spectral Decomposition of Landsat Thematic Mapper Data For Urbal Land-Cover Mapping", Proceedings of the 14th Canadian Symposium on Remote Sensing, Calgary, Alberta, Canada, May 1991.