Hello,

 

I'm trying to locate mineral clusters in CRISM I/F data using a Self-Organizing Map, but ultimately, even though I normalize the data prior to running a SOM my output clusters correspond with bright and dark areas.

 

I am working with GRASS and R, so here are my processing steps:

 

- import CRISM data into R, each pixel contains a vector of 254 values corresponding to the I/F values in the first 254 image bands.

I am now working with a matrix, of n rows (number of pixels in image) by 254. This CRISM data has previously been though the CAT pre-preprocessing pipeline.

 

- Calculate the mean value of each pixel, so that each pixel is now a residual vector (difference of each band from the mean spectrum of the pixel) - This means that I can better compare pixels to one another, as they are not affected by differences in brightness (are not offset from one another, as they would be if I used the mean spectrum of the image).

 

- normalize each vector component [0,1]

 

- use this new residual matrix as input into a 50x50 SOM.

 

- The output codebook data from the SOM I simply classify with hierarchical clustering (for the moment) so I can see basic separable regions. (This basic number of clusters I estimate using a distance/similarity matrix.)

 

Unfortunately, despite the above steps these clusters still correspond to bright/shadowed areas, in the majority of cases.

 

I was hoping someone might be able to offer some suggestions as to why this might be?

 

 

Allendej,

I forwarded your question to Ray Arvidson, who is the PDS Geosciences Node director and a member of the CRISM team. He is on travel and will get back to you in a few days.

Susan Slavney

Allendej,

Ray would like to ask you some questions about the software you are using to process the CRISM data. Would you please contact him by email at arvidson @ wunder.wustl.edu?

Susan