X
PrevPrev Go to previous topic
NextNext Go to next topic
Last Post 21 Mar 2016 02:14 PM by  anon
Apply forward MNF to Spectra
 1 Replies
Sort:
You are not authorized to post a reply.
Author Messages

anon



New Member


Posts:
New Member


--
21 Mar 2016 02:14 PM
    I am using MTMF to match image spectra to library end member spectra. Hyperspectral Image data is in emissivity, ranging generally from 0.9 to 1. Forward MNF transformation on the image data yields an MNF image with ~10-20 high signal bands, with MNF values ranging from -10 to 10, but most MNF values for the first 20 bands range from -2 to +2. The MNF transformation is very good, it gives very good image discrimination of different materials. Laboratory spectra are in emissivity, ranging from ~0.7 to 1.0. This is the typical difference in intensity between laboratory and field spectra. Laboratory spectra should be sufficient to approximate the majority of the subpixel end members in the image Laboratory spectra are resampled to match the band configuration of the file input to the MNF transformation, then I have applied the forward MNF to spectra tool using the MNF stats file from the image transformation. The resulting MNF transformed spectra have values ranging from -50,000 to 50,000, most in the range of -5000 to +5000 (in the first 20 bands that should correspond with the high signal bands from the image data) Is this an expected transformed spectra result, when the input spectra have moderately more amplitude variation than the image spectra? My MTMF results do not yield results that match up well with known target (Subpixel) compositions, and I'm trying to determine if the error lies within the spectra transformation, or one of the other possible sources of error (poor laboratory end member choices, sub-pixel mixing assumptions, etc) Thanks

    MariM



    Veteran Member


    Posts:2396
    Veteran Member


    --
    22 Mar 2016 12:38 PM
    It could be an issue with the lab spectra not really representing the image data very well. One test you might try is to collect spectra from your images, save them to a spectral library and transform them using the image MNF statistics. I would expect the transformed image spectra to be with the same range as the image spectra. You could then use the Spectral Analyst to compare the image spectra to the lab spectra to help identify the materials of interest. It might also give you an idea of how well the library spectra represent the materials found in the image. Also be sure to select the MNF stats and not the noise statistics which would give incorrect results (assuming ENVI would not error).
    You are not authorized to post a reply.