This
research has quantitatively examined the ability of DIRSIG to replicate real world
texture characteristics, specifically for a grass target. Research discovered a
large gap between the amount of information that is contained in images of real
world data and the synthetic images that were generated by DIRSIG. Expanding
the spectral database did produce a slight increase in results, increasing the
qualitative appearance of the image as well. Still, the amount of data
contained in the DIRSIG image remained significantly less than that contained
in the real image. Increasing the number of bandpasses employed by DIRSIG in
determining which spectra to map onto a particular pixel, resulted in two
different concepts.
Adding one bandpass
increased the ability of DIRSIG to map correctly the most appropriate spectra.
Additional bandpasses may increase the quality of an image, however, choosing
too many bandpass regions for an image can cause the resulting quality of the
image to decrease. The process is very complex, the output quality is dependent
on the particular imaging system that is being modeled, the bandpasses that are
chosen, and the amount of bandpasses that are used as references.
The conclusion of my
research is that while the quality of the image is increased, more research
must be done into texture characterization techniques, and the current methods
of texture characterization may have to be greatly changed in order to obtain
any significant increase in the quality of the DIRSIG images. Also a much
closer examination in to the particular band regions examined may provide
a greater understanding on how multiple bands may be used together more
effectively.
MISI was simulated in DIRSIG through the incorporation of MISI response files into the DIRSIG's configuration file. The response files were generated by inputting data from MISI's response curves. The Digital Image and Remote Sensing group at the Rochester Institute of Technology provided information for the response curves. Spectra for DIRSIG's database were taken from grass spectra that were recorded as ground truth during the MISI acquisition flight we were attempting to simulate. The MISI sensor was successfully incorporated into DIRSIG, however the simulation of MISI data was determined to have low quality. This was due to the lack in variance of the curves used in the spectral database. The curve set used was taken while held over one particular area of grass. While thirty curves were generated, these curves varied slightly from each other when compared. A second set of field data was input into the spectral database, and the same error was found with this acquired data. Techniques for ground truth acquisition do not incorporate the acquisition of grass spectra over a large area of grass, and were therefore lacking in needed variance.
Spectra for DIRSIG's spectral database were successfully extracted from the HYDICE scene data. A previously implemented HYDICE model was used for HYDICE runs. The HYDICE data provided for a more accurate analysis of DIRSIG's abilities. The fact that the spectrum was derived directly from the scene ensured that there was an appropriate curve for each pixel. This produced a more accurate test for DIRSIG's ability to match the correct curve within it's database to a given pixel since we knew that there was a perfect match for each pixel in the image provided in the database. When a principle components analysis was completed on both the HYDICE image, and the DIRSIG simulation, even with the extracted scene spectra there was a definite difference in the amount of information contained in the images. The DIRSIG image was found to contain a far lesser amount of information than the HYDICE image.
Once the
quality of the DIRSIG image was assessed, the distinct methods that were
employed to change DIRSIG's texture characterization routines were evaluated.
While all of these routines did create an improvement in the quality of the
DIRSIG image, the principal components analysis’s that were done, revealed
a significant difference between the synthesized and real images in each case.
However, more important than the quantized values for the images, several of
the techniques produced an increase in image quality that while not described
by the numbers of the image, are quite apparent visually. This result alludes
to the complexity of texture itself, the reason why it was so closely examined
in this research. While the numbers may show no significant change, a
great change is apparent within the images.
Expanding
the spectral library was ineffective in improving the quality of DIRSIG's MISI
simulation due to the fact that the reference spectra did not contain enough variance.
A second acquisition of ground truth was completed in the hopes to provide a
more variant database. Once again the ground truth was found to lack the
variance that would be required to successfully measure the difference between
the two databases. While the curve set was expanded from thirty to two hundred,
the base set of curves that was used for the statistical comparison, was to
similar to produce a curve set with a high enough variance to significantly
effect the results.
For the HYDICE simulation,
the original extracted set of reflectance curves consisted of two hundred and
fifty six curves. This set of curves was than used with
the expand_emissivity_file2
utility to generate a DIRSIG database of ten spectra. The extracted HYDICE
spectra was than also used in order to construct a database with the same
number of curves as the set that was originally extracted, in this way to test
DIRIG's simulation of the statistics and it's effects on the image quality.
Figure 2a shows the resulting images for expanding the database, and the
DIRSIG image generated with the original extracted spectra.



Image A: DIRSIG Image
generated
Image B: DIRSIG Image generated Image C:
DIRSIG Image generated
with HYDICE extracted
spectra
using a database of 10
spectra
using a database of 256 spectra
Comparing the two images
generated with expanded emissivity files on the right with the image generated
using the HYDICE extracted spectra on the left, we can qualitatively
examine the effect that the size of the database has on the quality of the
DIRSIG output. Image B replicates some of the larger artifacts found in image
A, but fails to replicate a lot of the detail. Image C, which was generated
with the larger spectral database, demonstrates the increase in structure that
is found with a larger set of spectra to select from. A careful examination of
the right side of the images could be used as one example to illustrate the
increase in quality of the DIRISIG image due to a larger database. In image B,
many areas are dark that are not in image A. The decreased number of spectra in
the database in image B causes an apparent decrease in the detail of the image.
This decreases ability of DIRSIG to replicate the structure that is present in
the real image.
While the DIRSIG image with
the larger database does reconstruct a large amount of the structure of the
texture more accurately, it fails to completely reproduce the structure found
in the original image. Even the DIRSIG image produced using the extracted
spectra for it's database has a great difference from the actually HYDICE
image. Figure 2b, allows us to look closely at these two images side by side.
The high quality spatial and spectral variance found in the HYDICE image is not
contained in the DIRSIG simulation. Image E does demonstrate the capability
that DIRSIG has of incorporating high amounts of spatial and spectral
variation, while also demonstrating the fact that it still falls short of
replicating the complex spatial and spectral variations found in the real
image.


Image D: HYDICE
Image
Image E: HYDICE extracted spectra
The
spatial pattern (texture) of a material will vary as a function of wavelength.
Images of the same area of grass at different wavelengths are not the same. We
want to be able to introduce spatial and spectral variations (texture) to a
modeled material. This method for improvement considered the fact that
usually images of the material exist for a few spectral bands. Current
techniques utilized by DIRSIG made use of only one of these images as a texture
band. The use of only one band meant that the spectral database was matched by
using the data for that spectral region only. The spectrum that was chosen
based on that bandpass region was than used for all bandpass regions of the given
pixel. Considering the addition of other bandpass regions allowed us to compare
the spectra in different areas of the curve. The theory was that by adding more
spectral bandpass, DIRSIG would be better able to match the entire shape of the
curve for one pixel.
Two additional bandpass
region were incorporated into DIRSIG's current techniques for mapping texture.
For each bandpass to be examined, a texture image was created in DIRSIG. These
bands were than specified in DIRSIG using the current technique for the input
of one texture band. This resulted in a list of z-scores to be examined for
each spectrum in the database, DIRSIG than selects the spectra that fits all of
the bandpass regions most closely. This results in a list of parameters for
DIRSIG to fit for a spectrum instead of the one parameter that it was examining
before. In this way a number of regions could be examined to ensure that the
accuracy in one region was not disregarded due to the need for accuracy in
another.
In order
to analyze the information contained in each DIRSIG image, to quantify the
results, ENVI's Principle Components Analysis was utilized. The forward PC
rotation uses a linear transform to maximize the variance of the data through
statistical means. This transformation can be used in order to determine the
amount of information that is contained within each band of an image. Table one
is a chart that displays the normalized eigen values for DIRSIG images
with multiple bands.
|
DIRSIG |
DIRSIG |
HYDICE |
DIRSIG |
DIRSIG |
DIRSIG |
Truth: |
|
Band Number |
Wavelength (nm) |
Wavelength (nm) |
One Band |
Two Bands |
Three Bands |
HYDICE |
|
1 |
394.214 |
|
1 |
1 |
1 |
|
|
2 |
397.51 |
|
0.077012 |
0.0520512 |
0.0577245 |
|
|
3 |
400.806 |
400.60999 |
0.032065 |
0.0289281 |
0.0241239 |
1 |
|
4 |
404.104 |
403.90601 |
0.005005 |
0.0044415 |
0.00427459 |
0.059038 |
|
5 |
407.428 |
407.229 |
0.004497 |
0.0040892 |
0.00351993 |
0.028432 |
|
. |
. |
. |
. |
. |
. |
. |
|
13 |
435.348 |
435.12799 |
0.000277 |
0.0002466 |
0.00022195 |
0.000764 |
|
. |
. |
. |
. |
. |
. |
. |
|
18 |
454.433 |
454.19501 |
0.000124 |
0.0001274 |
0.0001059 |
0.000359 |
|
19 |
458.444 |
458.20099 |
0.000105 |
0.0001124 |
9.15E-05 |
0.000339 |
|
20 |
462.526 |
462.27899 |
8.43E-05 |
9.17E-05 |
7.53E-05 |
0.000328 |
|
21 |
466.683 |
466.431 |
6.66E-05 |
8.14E-05 |
7.48E-05 |
0.000302 |
|
22 |
470.919 |
470.66199 |
5.90E-05 |
7.73E-05 |
6.51E-05 |
0.000276 |
|
. |
. |
. |
. |
. |
. |
. |
|
37 |
546.582 |
546.22198 |
8.47E-06 |
1.46E-05 |
1.30E-05 |
9.20E-05 |
|
38 |
552.639 |
552.27002 |
7.97E-06 |
1.32E-05 |
1.26E-05 |
8.69E-05 |
|
. |
. |
. |
. |
. |
. |
. |
|
50 |
639.123 |
638.61603 |
1.29E-06 |
4.47E-06 |
4.53E-06 |
3.89E-05 |
|
51 |
647.651 |
647.13098 |
5.61E-09 |
4.22E-06 |
4.22E-06 |
3.50E-05 |
|
52 |
656.408 |
655.87402 |
0 |
4.11E-06 |
3.60E-06 |
3.44E-05 |
|
. |
. |
. |
. |
. |
. |
. |
|
109 |
1419.11 |
1418.28003 |
0 |
2.57E-08 |
2.90E-09 |
6.64E-07 |
|
110 |
1433.01 |
1432.18994 |
0 |
2.45E-08 |
0.00E+00 |
6.32E-07 |
|
. |
. |
. |
. |
. |
. |
. |
|
121 |
1580.55 |
1579.78003 |
0 |
3.54E-09 |
0 |
3.27E-07 |
|
122 |
1593.47 |
1592.70996 |
0 |
2.71E-09 |
0 |
3.20E-07 |
|
123 |
1606.32 |
1605.56006 |
0 |
0 |
0 |
3.00E-07 |
|
194 |
2352.83 |
2352.31006 |
0 |
0 |
0 |
6.04E-09 |
|
195 |
2361.57 |
2361.05005 |
0 |
0 |
0 |
0 |
|
. |
. |
. |
. |
. |
. |
. |
The normalized eigen
values in table are used to measure the amount of data contained in the DIRSIG
images for each of the runs at certain wavelengths. Eigen values for the HYDICE
image are also included in the table, in order to demonstrate the relationship
that existed between the numbers of wavelength ranges that contained in the
DIRSIG images, to the number that contained information in the truth image, the
HYDICE image. For the truth image there is information contained over every
bandpass, the DIRSIG images fall short of accomplishing this. The results of
the principle components analysis for the images indicate that there is a great
gap in the amount of information that can be found in the HYDICE image, and the
amount of information that contained in any of the DIRSIG images. The
maximum eigen value for the HYDICE image was 8,649,461. The maximum eigen
values for the DIRSIG images showed a significant difference in the amount of
information present within the images.
Adding a second bandpass
region did create a slight increase in the eigen values, but it fell far short
of bridging the gap of informational content that exists between the synthetic
data and real world data. With the use of two regions, the amount of
information contained in the image covers a wider range of wavelengths. After
the addition of the third bandpass region, the amount of information within the
image decreased, as did the range of bandpasses that contained useful data.
This system for selecting and attempting to reconstruct texture is very complex.
The number of different combinations of texture bandpasses that could be
utilized was very large. Many factors could be used in real world scenarios to
determine which bandpass regions to use, such as, wavelengths that would be
examined by the sensor or algorithms to be tested, the bandpasses that were
available for texture images, and the reliability of those bandpass regions
(some bandpass regions may contain distortions caused by, for
instance, water absorption features).
After a examining the HYDICE
image in order to determine which band numbers to use, three bandpass regions
were identified that appeared to be images that were good examples of
structured texture, and contained data that the other bandpass regions did not.
The three bands used in this research as texture bands were: Band #1 (0.431
microns), Band #2 (0.828 microns), and Band #3 (1.260 microns). The
texture band used in the single DIRSIG run was at 431 nm, the run with a second
band also included the 828 nm texture band, and finally, the DIRSIG run with
three bands incorporated the texture bands for all three of the wavelengths
(431 nm, 828 nm, 1260 nm).

Figure three demonstrates
the difference that exists in the Eigen values found for DIRSIG runs with or
without multiple bands. When comparing the DIRSIG runs with the HYDICE
data, an obvious shift in the eigen values for all of the DIRSIG runs
toward lower wavelengths is clearly apparent.
The
greatest difference in the results found by incorporating the use of additional
bandpasses is demonstrated not quantitatively by a table of numbers, but
through a qualitative examination of the resulting images. It was very
difficult to gain an understanding of the texture contained within an image
using quantitative measurements. The greatest indication of the effect that
increasing the number of bandpass regions DIRSIG incorporates has on the
images, was to do a simple visual inspection and comparison of the
images. The resulting images for these three runs were examined over
several wavelengths for a comparison with the original HYDICE image. Figure
three shows the DIRSIG images for the several runs side by side next to the
HYDICE image.
Truth Image:
DIRSIG
DIRSIG
DIRSIG
HYDICE
One Band Run
Two Band
Run
Three Band Run




The image on the left in Figure 3a is the HYDICE image of the grass scene that was reproduced in DIRSIG. The DIRSIG images are to the right of it. When visually inspecting these images, there is an apparent structure to the grass image produced by HYDICE. The DIRSIG run using one texture image, or comparing one bandpass shows some of the same spectral structure, however, when a second bandpass is added, the texture in the image is noticeably closer to that of the HYDICE image. Adding a third bandpass at this wavelength seems to deteriorate the structure of the texture in the DIRSIG image, once again making it less similar to the HYDICE image.




Inspecting the image at different wavelengths will give us more information on how well the chosen spectra matches over a broad range of wavelengths. Once again, in the set of images in figure 3b, it is apparent that the DIRSIG image created with the use of two texture bands more closely replicates the visual texture properties of the HYDICE image to the left. A comparison of these images suggests that the texture in the two-band image has texture that is structured more similarly to that of the HYDICE image. In the image generated using one texture band, this texture is less apparent. The image to the far left, generated using three texture bands, once again seems to indicate that the third additional band is causing a break down of the textural structure.



It is apparent that at this wavelength, that the HYDICE image is best
reproduced when two bands are used. The lower left hand corner of the images
can be identified as one region in which details found in the HYDICE image are
replicated in the two-band image but are lost in the three-band image. The one
band image fails to reproduce the same amount of detail as the two band and
three band images.




The
second image, created using one texture band in DIRSIG shows an obvious lack of
textural structure, the third image which was completed using two texture
bands in DIRSIG and shows a definite increase in the structure of the texture
in the resulting DIRSIG image. When closely comparing this image, and the last
image to the original HYDICE image on the left, in this case, the image
generated using three texture bands matches the texture more closely. This
could be due to the wavelength which is being examined, this wavelength, which
is closer to that of the third band that was used, than the second or first.