Endmembers Extraction Toolbox

Note: PAVIA dataset is used as an example.

Features:

• Identifies the number of endmembers from the hyperspectral data.

• Extracts individual endmember spectra and saves them in Excel and the ECOSTRESS format.

• Quantifies the area given by each endmember.

Select PPI function to estimate the endmembers by using the PPI approach. The PPI approach projects the pixel spectra to an orthogonal space and identifies extreme pixels in the projected space as endmembers. This is a non-iterative approach, and the results depend on the random unit vectors generated for orthogonal projection. To improve results, you must increase the random unit vectors for projection, which can be computationally expensive.

Use the FIPPI function to estimate the endmembers by using the FIPPI approach. The FIPPI approach is an iterative approach, which uses an automatic target generation process to estimate the initial set of unit vectors for orthogonal projection. The algorithm converges faster than the PPI approach and identifies endmembers that are distinct from one another.

Use the N-FINDR function to estimate the endmembers by using the N-FINDR method. N-FINDR is an iterative approach that constructs a simplex by using the pixel spectra. The approach assumes that the volume of a simplex formed by the endmembers is larger than the volume defined by any other combination of pixels. The set of pixel signatures for which the volume of the simplex is high are the endmembers.

Steps:

1.      Select the option in Endmember Calculation Method. Available options are:

a.      NWHFC

b.      PFA

c.      Noise Whiten -False

2.      Specify the method that calculates the number of endmembers under Endmember Estimation Method. Available options are:

a.      N-FINDR

b.      Specify the number of iterations. If the field is empty, the default is the Number of Iterations = 3 x number of endmembers. Note: The computation time of the algorithm increases with the increase in the number of iterations.

c.      FIPPI

d.      PPI 

e.      Alternatively, type the number of endmembers you expect to find (shown 5, default is 0). The number must be less or equal to the number of channels. If this box is filled with a number, the method selection is irrelevant.

3.      Select the reduction method. If this argument is specified, the function first reduces the spectral dimension of the input data by using the specified method. Then, it computes the endmember signatures from the reduced data. Available options are:

a.      Principal Component Analysis

b.      Maximum Noise Fraction (default)

4.      After the endmembers spectra are calculated (can be seen in the ENDMEMBER SPECTRA panel), single endmember images are produced in the ABUNDANCE MAP IMAGES panel. At the same time, the best three images are combined in the ESTIMATED RGB IMAGE panel.

5.      Select Monochromatic/PseudoRGB options to visualize the image on the IMAGE DISPLAY panel. Available options are:

c.      Monochromatic

d.      Pseudo RGB

Use a slider to move between the images (in a monochromatic mode) or combine the images in the pseudo-RGB mode. Up to three images can be combined.

Click the Quantify button to measure the area covered by each of the selected endmember. A new pop-up menu will be opened followed by a new panel (See the Quantify manual).

Tips: You can click on each image to see it in a larger separate window. Each image can be saved.

Each produced endmember spectrum can be saved either in Excel or the ECOSTRESS format. To save the spectra, right-click on the ENDMEMBERS SPECTRA panel. In the case of ECOSTRESS, the spectra have to be saved one by one. When Save Spectra in ECOSTRESS format is selected, a pop-up dialogue window ECOSTRESS Format Parameters will request additional information, like the Endmember number (mandatory field), name, type, etc. 

Additional Information:

The toolbox includes several types of algorithms: a) to find the number of endmembers, b) to identify their spectra, and c) to calculate endmembers’ abundance maps.

Noise-whitened Harsanyi–Farrand–Chang (NWHFC) finds the number of endmembers present in a hyperspectral datacube by using the NWHFC method. When 'NoiseWhiten – false' is checked, the algorithm does not perform noise-whitening of the data before extracting the endmembers. In that case, the method is also known as the Harsanyi–Farrand–Chang (HFC) method.

The method also requires the input for the Probability of False Alarm (PFA). The default value for PFA is 10-3 (default). A smaller value will provide more endmembers.

N-FINDR is an iterative approach for finding the endmembers of a hyperspectral dataset. The method assumes that the volume of a simplex formed by the endmembers (purest pixels) is larger than any other volume defined by any other combination of pixels. Computes principal component bands and reduces the spectral dimensionality of the input data by using MNF or PCA. The number of principal component bands to be extracted is set equal to the number of endmembers to be extracted. The endmembers are extracted from the principal component bands.

Pixel Purity Index (PPI) extracts endmember signatures from hyperspectral datacube by using the pixel purity index algorithm. The method computes the orthogonal projections of hyperspectral data values on a set of randomly generated unit vectors known as the skewers. Then, the method computes the PPI count for each data value. PPI count is the number of times a data value results in an extreme point when projected onto these skewers. Those data values with more than the expected number of PPI count comprise the endmembers of the hyperspectral data. PPI is a non-iterative method and the steps involved use MNF or PCA to reduce the dimensionality of the input data. The number of principal component bands to be extracted is set equal to the number of endmembers to be extracted.

Fast Iterative Pixel Purity Index (FIPPI) is an approach that iteratively selects the better candidates for endmembers after each iteration. Unlike the pixel purity index (PPI) technique, the FIPPI method selects the initial set of skewers by using the automatic target generation process (ATGP). As a result, the algorithm converges faster and generates a unique pixel for each endmember. The steps involved in FIPPI approach use MNF or PCA to reduce the dimensionality of the input data.

References:

NWHFC: Chang, C.-I., and Q. Du. “Estimation of Number of Spectrally Distinct Signal Sources in Hyperspectral Imagery.” IEEE Transactions on Geoscience and Remote Sensing 42, no. 3 (March 2004): 608–19. https://doi.org/10.1109/TGRS.2003.819189

N-FINDR: Winter, Michael E. “N-FINDR: An Algorithm for Fast Autonomous Spectral End-Member Determination in Hyperspectral Data.” Proc. SPIE Imaging Spectrometry V 3753, (October 1999): 266–75. https://doi.org/10.1117/12.366289.

PPI: J.W Boardman, F.A. Kruse and R.O. Green, "Mapping target signatures via partial unmixing of AVIRIS data.", Technical Report, California, USA, 1995.https://ntrs.nasa.gov/citations/19950027316

FIPPI: Chang, C.-I., and A. Plaza. “A Fast Iterative Algorithm for Implementation of Pixel Purity Index.” IEEE Geoscience and Remote Sensing Letters 3, no. 1 (January 2006): 63–67. https://doi.org/10.1109/LGRS.2005.856701.