Edward Kelly
New Member
Posts:6  
| 11 Jul 2017 12:06 PM |
|
|
Hi, I intend to apply the Spectral Angle Mapper (SAM), Maximum Likelihood (MLC) and Support Vector Machine (SVM) algorithms on a hyperspectral dataset but would just like to clarify the steps required. I have performed radiometric correction and a forward PCA rotation which resulted in 3 PCA bands. However, I am slightly confused regarding endmember selection procedures when using PCA.
For example, for the MLC algorithm I understand that if I wasnt using PCA I would simply create new ROIs to derive endmembers. However, when using PCA as a dimensionality reduction technique, I am not sure where to derive the endmembers from. Is it from the original radiance image or from the PCA components?
Thanks.
|
|
|
|
MariM
Veteran Member
Posts:2396 
| 12 Jul 2017 02:14 PM |
|
|
If you are going to use data dimensionality reduction, I would suggest you use the Spectral Hourglass wizard which walks you through the process with explanations:
https://www.harrisgeospatial.com/docs/SpectralHourglassWizard.html
You can get endmembers from the original data, PCA data or from pre-existing spectral libraries taken in the lab. What is key is that the endmembers are 'good' examples/representations of the materials of interest.
|
|
|
|
Edward Kelly
New Member
Posts:6
| 13 Jul 2017 01:54 PM |
|
|
Thanks, yes I have collected endmembers through ROIs which I will export to a spectral library.
|
|
|
|
Edward Kelly
New Member
Posts:6
| 13 Jul 2017 03:37 PM |
|
|
Is there a way to use PCA with Maximum Likelihood, as the Spectral Hourglass Wizard doesn't have an option for it?
Thanks
|
|
|
|
Edward Kelly
New Member
Posts:6
| 16 Jul 2017 01:54 PM |
|
|
Is it acceptable to use the output from the Forward PCA transform as the input for MLC in the Classification Workflow? Although my training data (which i derived from my original radiance image) don't seem to load from the .xml file when i try to use them in the Classification Workflow.
Thanks
|
|
|
|
Edward Kelly
New Member
Posts:6
| 16 Jul 2017 02:21 PM |
|
|
Sorry, I may have misunderstood, I think I need to use Inverse PCA rotation to recreate the original image data using the 3 PCA images, and then use this as input into the classification techniques.
|
|
|
|