Normalized Eigenvalues in PCA

Microwave Toolbox
#1

Dear all,
I have some questions about using PCA in SNAP software.
I have some images in one stack (3 coherence images, 3 phase images, and 3 intensity images- overall 9 images-). I used PCA in SNAP and I got results like: PC0, PC1,PC2,PC3,PC4,PC5, PC6, PC7,PC8

I chose ‘show eigenvalues’ option and I got below result.

User Selected Bands:
coh_HH_30Mar2012_30Mar2012_Contrast_mst_30Mar2012
coh_HH_30Mar2012_30Mar2012_ASM_mst_30Mar2012
coh_HH_30Mar2012_30Mar2012_GLCMMean_mst_30Mar2012
Intensity_HH_mst_30Mar2012_Contrast_slv1_30Mar2012
Intensity_HH_mst_30Mar2012_Energy_slv5_30Mar2012
Intensity_HH_mst_30Mar2012_GLCMMean_slv8_30Mar2012
Phase_ifg_HH_30Mar2012_30Mar2012_Contrast_slv11_30Mar2012
Phase_ifg_HH_30Mar2012_30Mar2012_MAX_slv16_30Mar2012
Phase_ifg_HH_30Mar2012_30Mar2012_Entropy_slv17_30Mar2012

User Input Eigenvalue Threshold: 70.0 %

Number of PCA Images Output: 9

Normalized Eigenvalues:
4088.5073684301024
1197.4688017647886
791.4877969954761
412.24858028364145
27.62046656611915
2.4561881789628788
0.11794915384786647
0.030887788150756892
1.599021595534588E-11

My questions:

  1. Is it possible by these Normalized Eigenvalues know which one of bands or images (9 images) are more important in our process?
    I mean is there any option that organize selected input images in order of usefulness? For example like:
    A. Intensity_HH_mst_30Mar2012_GLCMMean_slv8_30Mar2012
    B. Phase_ifg_HH_30Mar2012_30Mar2012_MAX_slv16_30Mar2012
    and so on…
  2. How we can know how many of PCs (PC0, PC1,PC2,PC3,PC4,PC5, PC6, PC7,PC8) are good enough for our work?

Cheers,

#2

the eigenvalues are a measure of how much of information is stored in the first/second/third component. That means how much these differ in terms of information based on the original set of input products.
By breaking the inputs into different components you lose information on the single images’ content. The aim is to reduce the number of inputs by aggregating the contents of images into clearer images which are more disinct.

  1. How we can know how many of PCs (PC0, PC1,PC2,PC3,PC4,PC5, PC6, PC7,PC8) are good enough for our work?

You are the only person who can answer this. Look at the components. Do they contain information that might be of interest in later steps? Then keep them. If most of the image is described by the first 3 components, they might be enough. Totally depends on what your later steps are.

#3

Thanks. I got it now. Another question;
I used 33 layers (images) in PCA in SNAP and put it for running on university computer for one night but I do not know why it is still computing!!!
I do not think so it should take so much time!!!
Cheers,

1.PNG1289×1009 174 KB

SNAP hardware requirements and Tutorials Links toi
#4

Perhaps that is how long it takes? You are working in 33-dimensional space with probably millions of pixels. How about trying 1st with a tiny subset?

1 Like