Matt Spencer
New Member
Posts:14  
| 21 Aug 2017 06:52 AM |
|
|
Hi, I'm using MNF as a dimensionality reduction technique for hyperspectral image classification. I have estimated the noise statistics from the data using a spectrally uniform area, and have output 9 MNF components for use in classifications. I am using endmembers derived from the image (ROIs) for the classifications. I have read in the literature some using the Inverse MNF to transform back to the original data space and number of channels. Am I correct in thinking there is no need to use an Inverse MNF in this instance, as I am using the Forward MNF as a data dimensionality reduction method?
Thanks.
|
|
|
|
MariM
Veteran Member
Posts:2396 
| 21 Aug 2017 09:54 AM |
|
|
It is a valid approach to classify MNF data. One thing to consider - If your endmembers (spectra) derived from ROIs were taken from the original image, and not the MNF image, then they will need to be transformed into the same MNF space as the data. You can use the Apply MNF to Spectra tool and select the MNF statistics file used to transform the original data to MNF to transform the spectral library into the same space as the MNF data.
If you are going to use the ROIs rather than a spectral library, they can be used directly with the MNF data.
|
|
|
|
Matt Spencer
New Member
Posts:14
| 24 Aug 2017 08:53 AM |
|
|
Thanks, I'm using ROIs from original image as endmembers, so no need to transform? Also, is it recommended to estimate the noise from a spatial subset that is spectrally uniform, i.e. an area of the ocean?
Thanks.
|
|
|
|