CEE 615: Digital Image Processing Lab 12: Continuum removal p. 3

Hyperspectral Processing: continuum removal

A major advantage of hyperspectral data is the possibility of identifying specific materials. This is generally only possible to the extent that the material in question has unique absorption features or, more likely, a set of absorption features that together allow a high level of specificity. A problem with this type of identification is that the absorption features tend to be relatively subtle, and the variation in the absorption features tends to be even more subtle. One obvious consequence is that one cannot equate variance in an image with the quality of the information in that image. This makes it more difficult to sort out the useful information from the background noise.

In this lab the task is to isolate a pair of subtle water absorption features in order to look for the presence of water on or very near the surface in a very arid environment. The processing will include steps to remove noise as well as to characterize variations in the depth of the absorption bands.

The image, collected by the Hyperion instrument flown on the EO-1 satellite, is of a portion of desert in Qatar. The image was collected on 3 March 2010 at 9:45 AM local time. The imaging system collects 220 bands with a 10-nanometer spacing from 0.4-2.6 microns. It is a proof-of-concept instrument and samples at a 30m pixel spacing with a FOV of 7.5km.

The processing sequence is as follows:

  1. Load the original image data: Qatar2010.img
  2. Apply an FFT filter to remove banding noise
  3. Perform an MNF transform on a limited spectral range of the data containing the target water absorption feature(s)..
  4. Analyze the results by displaying the spectrum (z-profile) for both the unfiltered and filtered image
  5. Apply the continuum removal algorithm on the resulting data to optimize for water absorption.

Now, step by step:

  1. Load the original image data: Qatar2010.img. The image should automatically display as a CIR image. Red will generally correspond to vegetation.
  1. Apply an FFT filter to remove banding noise

The banding noise in this imagery is rather severe and will make it difficult to observe subtle band-to-band variations. Since the banding noise is very identifiable in the frequency domain, an appropriate FFT filter can be very effective at eliminating a majority of the banding noise without an appreciable effect on the image data. With this in mind:

a.  Apply the Fast Fourier Transform (Filter > FFT Filtering > Forward FFT) to the Qatar2010.img

b.  Apply the inverse FFT using the filter, banding_rectangular-filter.img, naming the result Qatar2010_invFFT.

c.  Display the filtered image and inspect.

  1. In order to optimize the information content of the image set, perform a Minimum-Noise-Fraction (MNF) filtering on the target spectral range (671-884nm; bands 32-53) containing the target absorption features.

a.  Select: TransformMNF RotationForward MNFEstimate Noise Statistics from Data

b.  Select the image: Qatar2010_invFFT. OK.

c.  Select the spectral band subset 32-53. (Highlighted bands are the bands that are selected)

d.  Create a file for the noise statistics: Qatar2010_invFFT_MNF32-53-noise.sta

e.  Create a file for the MNF statistics: Qatar2010_ invFFT_MNF32-53.sta

f.  Name the output file: Qatar2010_ invFFT_MNF32-53.img

g.  Select OK

h.  Examine the MNF image data.

i.  The first MNF image is dominated by viewing angle effects. Since there is a strong relationship between the water vapor content and the atmospheric path length, this is probably indicative of water vapor content. MNF bands 2, 3, 4 and 5 all appear to have significant information about the land surface (non-noise) and should certainly be kept. MNF bands 6 and 7 contain little information about the land surface, but do appear to contain non-random features that are probably related to atmospheric variability. Bands 8 and 9 are primarily noise but do have minor features related to the dunes and some of the vegetation patches.

j.  Select the useful MNF images. Certainly keep bands 2-5. Keep band 1 if you want to preserve the effects of the atmospheric path. I would opt to keep bands 6 and 7 as well since they are not illustrating random noise. Bands 8 and 9 are more questionable, but beyond 9, none of the MNF images have discernable features. At most, retaining 9-bands (1-9) is sufficient to retain anything but random noise.

k.  Apply inverse transform.

i.  Select TransformInverse MNF Transform.

l.  Select the image: Qatar2010_ invFFT_invMNF33-53_bi-bf.img (where bi is the first MNF band used and bf is the final MNF band used.

ii.  Select OK.

iii.  Select the MNF stats file: Qatar2010_ invFFT_MNF33-53.sta

iv.  Name the output file Qatar2010_ invFFT_invMNF33-53_b2-x.img (where x represents the highest MNF band being retained).

v.  Select OK

  1. Analyze the results by displaying the spectrum (z-profile) for both the unfiltered and filtered image. Link the images and move the cursor around the image. If this worked the spectra should be relatively stable, but the depth of the water absorption band will change.
  1. Apply the continuum removal algorithm on the resulting image to optimize for water absorption.

Finally, the data are ready for continuum removal.

a.  Select SpectralMapping MethodsContinuum Removal

b.  Select Qatar2010_ invFFT_invMNF33-53_b2-x.img as the input file. OK.

c.  Select Memory and OK.

Note that in the final image set, the bands forming the tie points for the normalization are all the same gray value and appear black. Other images in the data sets display a variety of features.

Pay particular attention to the images corresponding to wavelengths 721nm and 823nm, corresponding to the center of two of the water absorption bands. These are the most likely to show the effects of atmospheric water absorption.