10.6 Random ICA-Based EEA (RICA-EEA)

The idea of RICA-EEA is derived from a recent work on an independent component analysis (ICA)-based approach to DR (Wang and Chang, 2006a) where FastICA developed by Hyvarinen and Oja (1997) is used to generate independent components (ICs). Since FastICA also uses random projection vectors as its initial condition to initialize its algorithm, it also encounters the same problem as both PPI and N-FINDR do. Because the initial projection vector for each IC is randomly generated by FastICA for each run, the ICs generated by each run are also different. Nevertheless, if the information contained in an IC is significant, such an IC will always appear in each run. With this assumption, if FastICA is repeatedly implemented, the ICs that are common in all runs should be the desired ICs for endmember extraction. The detailed implementation of RICA-EEA is summarized as follows.

RICA-EEA Algorithm

1. Initialization
Assume that the number of initial endmembers is p. Set img.
2. At each n, run FastICA to find p independent components, img, where each independent component, img can be formed as a vector denoted by img. It should be noted that FastICA randomly generates a unit vector as an initial projection vector.
3. If img, img and go to step 2. Otherwise, continue.
4. Find common ICs for all runs up to the nth run. Two independent components for different runs, img and img, are considered to be distinct if the SAM between their corresponding vectors,img and img, is greater than a prescribed threshold ε. Let img denote the common ICs obtained for all runs, img.
5. If img, go to step 2. Otherwise, the algorithm is terminated and img is the desired set of ICs for endmembers.
6. For each of FastICA-generated IC images in img, find a pixel with maximum absolute value, which is referred to as endmember pixel. The spectral signature of such found pixel is then selected as an endmember.
7. The spectral signatures of the endmember pixels produced in step 6 are our desirable endmembers.

Like RPPI and RN-FINDR, VD is only used to provide a good reasonable upper bound on the number of random initial endmembers to be used for RICA-EEA.

..................Content has been hidden....................

You can't read the all page of ebook, please click here login for view all page.
Reset
18.222.239.77