Creating Denoised Model after PCA

[felixu...@gmail.com]

Dear all,

I have a question about creating a denoised EDX model after the PCA decomposition step. As I understand it, the number of components that should be used to construct a denoised model is determined by the 'elbow point' in the scree plot. However, I've found that in several of my datasets, the components above this elbow have factors and loadings that looks a lot like noise, while some components in the flat line below this elbow show legitimate peaks.

While constructing the model, would it be mathematically sound if I ignore these noise-like components and include those with actual peaks even though the former have higher variance values then the latter?

Thank you.

Kind regards,

Felix

[Thomas Aarholt]

Hi Felix,

The scree plot approach doesn't work terribly well for microscopy signals - I think partly due to how we often have a component present at low concentration in only a few pixels (so it's statistical significance is low) and since we often have noise that might be considered "real" (say, a poor dark reference).

The best approach is to get a rough idea of the number of components from the scree plot, then use s.plot_decomposition_results() and step through the pairs of loadings and factors. If the factors (ala EDS spectra) look like total garbage, they are probably safe to not include.

Sometimes I find (for instance) that components 0-3 are real, and 6 and 7 look a bit real as well. Maybe there's some contrast on the edge of my sample in the loadings.

Then I use s2 = s.get_decomposition_model([0,1,2,3,6,7]).

Often you can remove a significant component that just shows noise because it describes the dark reference I mentioned earlier. On stacks of TEM images PCA often shows you the edges of the quadrants of the detector, which can be nice to remove, as well.

Does that make sense?

Best,

Thomas

--
You received this message because you are subscribed to the Google Groups "hyperspy-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hyperspy-user...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

Thomas M. Aarholt

Post Doc, University of Oslo 

 

Mobile: +47 92855828

Email: thomas...@gmail.com

[felixu...@gmail.com]

Hi Thomas,

Great answer, very helpful. thank you!

Felix