Stavroula Giannakopoulou
New Member
Posts:49  
| 14 Aug 2019 07:13 AM |
|
|
Hello everybody,
I have a Hyperion image, in which I implemented an atmospheric correction, georeference and two masks, in order to avoid the vegetation and the water in the image. I have already removed the "Bad bands" and my image includes 112 bands. The problem is that I want to apply the n-D Visualizer, but at the diagram of the eigenvalues, at the step of MNF, the values that I received for the eigenvalue is better than the value of 1 (specifically the lowest value is 1.02). So I would like to ask if there is somebody who knows the reason for these values and if there is a way to find out a mistake that I made during the process. Thank you very much in advance for your help!
|
|
|
|
MariM
Veteran Member
Posts:2396 
| 14 Aug 2019 08:40 AM |
|
|
Are you saying that all 112 MNF bands have a noise floor above 1.0? If so, all bands are 'relevant' and include some unique information in this situation and you should not reduce dimensionality and continue on the workflow with all 112 bands. It is not a requirement that you reduce dimensionality, but was one method of improving processing speed when computers were less performant.
|
|
|
|
Stavroula Giannakopoulou
New Member
Posts:49
| 14 Aug 2019 09:03 AM |
|
|
Yes, I think that we are talking about the same... The eigenvalues are greater than the value of 1, which indicates that there isn't a noisy band at my image..Right? So, I should keep all the bands at the data dimensionality step of the spectral hourglass wizard, because of that every band includes information?
|
|
|
|
MariM
Veteran Member
Posts:2396
| 14 Aug 2019 09:28 AM |
|
|
Yes, that is the theory. I would double check my atmospheric correction result is as expected by looking at the min/max values of the bands and some spectra of known materials to see if they look correct. If your data cube is accurate, then I would keep all bands above the noise threshold of 1.0 - so all bands in this case.
|
|
|
|
Stavroula Giannakopoulou
New Member
Posts:49
| 14 Aug 2019 09:40 AM |
|
|
Ok!!Thank you very much! One last question: the theory refers that if the total number of your bands has eigenvalues greater than 1.0, then our data are not technically hyperspectral...So if I use all the MNF bands, then my data are not characterized as hyperspectral?
|
|
|
|
MariM
Veteran Member
Posts:2396
| 14 Aug 2019 10:58 AM |
|
|
I would not say your data are not 'hyperspectral' and I am not sure there is a clear definition of what makes data 'hyper spectral'.
The workflow is intended to reduce redundancy in spectral information and if all MNF bands contain more information than noise (by using the noise floor of 1.0), you can say that your data cube is not 'over-sampled' (where 'samples' are the narrow bands intended to characterize the true spectral response of the pixel/material on the surface).
In cases where some MNF transformed bands fall below the noise threshold, you can say that the data cube was 'over-sampled' enough to remove some spectral information that is redundant or indistinguishable from noise.
|
|
|
|