UNIVERSITY OF CALIFORNIA, SAN DIEGO
3 1822 04429 0682
UC SAN DIEGO LIBRARY
UNIVERSITY OF CALIFOR
ATMOSPHERIC OPTICS GROUP
TECHNICAL NOTE NO. 257
3 1822 04429 0682
Version 1.1
Aug 2002
Offsite
(Annex-Jo
anals)
974.5
. T43
no. 257
\ Analytic Support of the Phillips Lab
Whole Sky Imager, 1997 - 2001
Published as Final Report for ONR Contract
N00014-97-C0385
UNIVERSITY
OF
CALIFORNIA
SAN DIEGO
J. E. Shields
M. E. Karr
A. R. Burden
R. W. Johnson
J. G. Baker
e.
....
IVES
eeee..
.
.
.....
desse
FORN
.
THE
RNIA
gedere
1868
ee
eeeee
The material contained in this note is to be considered
proprietary in nature and is not authorized for distribution
without the prior consent of the Marine Physical Laboratory.
SCRIPPS
INSTITUTION
OF
OCEANOGRAPHY
MARINE PHYSICAL LAB San Diego, CA 92152-6400
UC SAN DIEGO LIBRARY
UNIVERSITY OF CALIFORNIA, SAN DIEGO
II
1111
3 1822 04429 0682
Analytic Support of the Phillips Lab
Whole Sky Imager, 1997 - 2001
Published as Final Report for ONR Contract
N00014-97-C0385
J. E. Shields, M. E. Karr, A. R. Burden,
R. W. Johnson and J. G. Baker
Analytic Support of the Phillips Lab Whole Sky Imager, 1997 - 2001
Janet E. Shields, Monette E. Karr, Art R. Burden,
Richard W. Johnson, and Justin G. Baker
1. Introduction
The Atmospheric Optics Group at the Marine Physical Lab, Scripps Institution of
Oceanography, University of California San Diego, has been developing Whole Sky
Imagers (WSI) for many years (Johnson 1989, 1991). A Day/Night WSI which had 24-
hour a day capability was initially developed in the early 1990's, and further developed
since that time (Shields 1993, 1998). One of these instruments, Day/Night WSI Unit 5,
was delivered to the Air Force Phillips Lab in December 95. This work was funded
under Contract N00014-93-D-0141 DO #11 (Shields 1997a).
A follow-on contract, N00014-96-D-0065-D3 (Shields 1997b) was funded to support
additional documentation and to support enhancement of the software provided with the
instrument. Under that contract, a new Operations Manual and Theory of Operations
Manual were delivered, and the archival and communications software was upgraded to
enable several new control options.
This report documents the work that was performed under a second follow-on contract,
Contract N00014-97-C-0385, which was funded from July 97 through June 01 for a total
of $103,073. This contract was intended to provide support for additional upgrades to the
system software, algorithms, and hardware, on an as-needed basis. The work was
supported by the Air Force Phillips Lab, later renamed Air Force Research Lab. The
Statement of Work is given in Section 2. The initial work, to provide software upgra
and hardware repairs, is discussed in Section 3. Much of the funding increment was used
to develop a first-generation night cloud algorithm, as discussed in Section 4. Other
software and hardware enhancements performed during the period when the algorithm
was under development are discussed in Section 5. The final funding under this contract
was used primarily to calibrate the camera in preparation for development of
transmittance algorithms. This calibration work is discussed in Section 6, which also
provides an overview of ongoing work continuing under other more recent contracts.
Although we normally do not have access to the site and experiments where the
instrument is located, our sponsor described the use of the instrument as follows: Air
Force Research Laboratory employs the Whole Sky Imager in the role of test support to
multiple, high-value experiments. WSI images are displayed on web pages in the test
control center to increase situational awareness. WSI data are provided for both daytime
and nighttime operations. During certain tests, the test director will use the Whole Sky
Imager to either continue, or cancel, the test based on the WSI cloud detection algorithm
and resulting picture. AFRL has supported continued algorithm development for
increased situational awareness during testing and collection of atmospheric data for post
processing.
2. Statement of Work
The Statement of Work from the proposal is as follows:
The contractor shall, unless otherwise specified herein, supply the necessary
personnel, facilities, services, and materials to accomplish the following tasks within a
two year period following receipt of funding.
a. Coordinate with the sponsor regarding the most appropriate tasks and estimated costs
for development.
b. Provide personnel trained in the Whole Sky Imager and its capabilities to address
these tasks (at MPL) to the limit of funding provided under the contract. These tasks may
include analysis, software development, documentation, minor hardware development,
and other tasks related to the WSI which are mutually agreed by the sponsor and by MPL
to be appropriate.
c. Provide a final written report and interim verbal reports to the sponsor regarding the
results of the above work.
This work was all completed in a timely manner, i.e. using a schedule which best fit the
needs of the sponsor. The tasks were discussed in phone communications and visits
during the period. The agreed-upon tasks are discussed in the following sections.
Interim reports were provided in the form of memos, phone discussions, and e-mail. This
report completes the final requirement. To the best of our knowledge, the funded
work has been completed to the satisfaction of the sponsor.
3. Initial Software Enhancements and Hardware Repairs, Jul 97 -- May 98
The initial funding increment of $10k was received approximately Jul 97, and followed
by $20k in approximately Oct 97. As agreed with the sponsor, these funds were used to
further develop the software and to repair the camera.
First, the program View16 was provided in order to enable easier viewing of the WSI
images. This program also allowed the user to save 16 bit images as a TIFF image. As
documented in Technical Memo AV98-011t (Kart 1998a), an upgrade to the software
t runs the WSI, RunWSI, was delivered in Feb 98. This upgrade fixed a subtle
problem between the system clock and the WWV clock that was occasionally causing
problems. In addition, three “hot keys” were added for more convenient changes in the
operational parameters during the program usage. Later upgrades funded under the later
funding increments are discussed in Section 5.
Late in 1997, the Photometrics Series 200 16 bit digital camera, which is at the core of
the Day/Night WSI, failed. We determined that this was caused inside the Photometrics
camera head, by corrosion of the metal surrounding the coolant path. This allowed
coolant to enter the chamber containing the camera CCD and the electronics, all of which
were destroyed. Under this contract, a new CCD chip was purchased, and CCD and
electronics were replaced by Photometrics. (This required separating the fiber optic taper
from the original CCD and bonding it to the new CCD.) The full camera assembly (built
by MPL to hold the camera and the optics) was then reassembled, and the camera was
recalibrated and returned to the field. This repair was documented in Technical Memo
AV97-051t (Shields 1997), which also includes the part numbers of the major
components.
The WSI camera housing was reinstalled by Air Force personnel, however some
remaining problems resulted in a trip to the site by MPL personnel. We were able to
trace the problem to one of the connectors installed by Photometrics during the repair
(Wertz, 1998).
4. Night Cloud Algorithm Development Work, May 98 - Mar 00
The next two funding increments, $20k received in May 98 and $40k received in May 99,
were intended and used primarily for the work toward a night cloud algorithm, with some
of the funding used for hardware support as needed.
The intent of the initial night cloud algorithm is to evaluate all regions of the sky and
identify the presence or absence of clouds. (Later algorithms which have been developed
or are in development address moonlight conditions and related issues, and should
provide optical depth or beam transmittance. Concepts for the development of a higher
spatial resolution algorithm have also been developed.) This first algorithm works
essentially by evaluating the radiometric signature of the image in the star field. We
initially wrote programs to analyze the character of the image near stars, in the presence
and absence of clouds. As reported informally in June and July 98 (via e-mail and
phone), we evaluated the contrast between a star and its background, and how this
contrast changed in the presence of thin clouds. This initial study was sufficiently
encouraging to continue the development.
4.1 Initial Star Mapping
In order to automate an algorithm based on star detection, it is necessary to have a means
of identifying the expected location of stars in the image. This requires the following:
a) a geometric calibration to convert from zenith and azimuth angle to x,y pixel location.
b) an automated star catalog to provide right ascension and declination of celestial
objects
c) a program to derive zenith and azimuth angle of these objects, given the time and
location of the WSI, and to convert them into pixel locations in the image
Initial geometric calibration equations were derived from in-house geometric calibrations
as documented in Memo AV98-041t (Burden, 1998b). These equations matched the
calibration data to within an average error of about .08 degrees or about .2 pixel. For
item b, the US Naval Observatory star catalog was chosen (Hoffleit 1991). (This catalog
is also known as the National Space Science Data Center star catalog.) This consists of
an ASCII file that lists the right ascension, declination, magnitude, and related
parameters. A series of programs were written in IDL' to derive the positions of the stars
in object space and image space, as documented in AV98-041t (Burden, 1998a).
To test the results, a program was written which extracted stars from the star catalog, and
plotted squares around the predicted location of stars of magnitude brighter than 4. These
were superimposed on a star image. We were reasonably pleased with the results,
however this initial geometric calibration was not as accurate as we would like, as
documented in Memo AV98-041t. The geometric calibration was refined as discussed in
the next section.
4.2 Refining the Star Mapping:
An image of a star in the WSI image plane is typically Gaussian in shape, with a width of
0.5 pixels. That is, the point source in object space is convolved with a roughly Gaussian
point spread function due to the atmosphere and the optics in the WSI. Using this
information in the image, one can develop more precise techniques for determining the
star location. We modified a technique developed by our colleague Dr. Tim Tooman
(verbal communication), using Gaussian best-fit routines to determine the sub-pixel
location of selected stars. An IDL program was written at MPL, GetStarInfo, which
allows one to interactively select stars distributed over the image and then associate each
star with a star in the US Naval Observatory catalog. Using this approach, a file is
created with the stars' predicted positions in object space and their measured positions in
image space.
The known azimuth and zenith angles of 80 to 100 user-selected stars, along with the
decimal pixel positions of the stars in the observed image, are then input into a
Mathematica' program Anglefit written by Dr. Tooman. This program determines an
empirical fit between the object space and image space parameters. Use of the program
is somewhat complex, and involves an iterative approach, including evaluating the
residuals. It is slightly subjective, in that it allows the user to estimate the center position
(actually the position on the image corresponding to the zenith), and reruns are made
until the user is satisfied with the residuals. Results are typically to within an average
error of 0.5 pixels or better. The resulting geometric equations are documented in
Technical Memo AV99-043t (Burden 1999a).
'IDL, which stands for "Interactive Data Visualization", is
produced by Research Systems, Inc., which was recently
purchased by Kodak. Their address is 4990 Pearl East
Circle, Boulder, CO 80301.
Mathematica is produced by Wolfram Research, Inc., whose
address is 100 Trade Center Drive, Champaign, IL 61820.
4.3 Development of the Algorithm
The next step in the algorithm development was to develop a program to detect stars in a
test image. The program uses the US Naval Observatory star catalog to determine the
stars within a selected magnitude range, and predicts their anticipated location in a test
image (in image space). The program then finds the largest pixel value within a two-
pixel radius of the predicted location. Using a Levenberg-Marquardt best-fit routine", the
program attempts to fit a two-dimensional Gaussian curve to the data (and in the process
determines the true pixel-center of the peak in image space). If the fitting routine gives a
reasonable result, the program counts this as a successful detection of the star in the
image. Using this technique, we were able to detect most of the stars from magnitude 1
to magnitude 7 in a sample clear image. For example, for zenith angles 30 degrees or
less, the program detected 100% of the 33 stars of magnitude 0 to 4, and 88% of the 99
stars of magnitude 0 to 7. For zenith angles of 70 degrees or less, the program detected
78% of the stars of magnitude 0 to 7. Many of the undetected stars were automatically
rejected by the algorithm, because it rejects any stars which are near other stars. A test of
random locations in an overcast image resulted in a 5% false detection rate.
On the basis of the above results, a first generation cloud algorithm was generated as
documented in Technical Memo AV99-044t (Burden, 1999b), which is included in
Attachment 1. This program divides the sky out to a 70 degree zenith angle into
segments which are approximately 5 degrees in zenith angle by 15 degrees azimuth
angle. (Near the center fewer azimuthal segments are made.) This results in 237 regions.
Then within each region the program determines whether there are any stars from the star
catalog, and if so, determines whether they can be detected simply by the presence of
their PSF. The program marks the region grey if there are no candidate stars in the star
catalog, blue more than 50% of the available stars in the region are detected, and red if
fewer than 50% of the available stars are detected.
This approach worked reasonably well, as shown in the memo. On the clear image, 90%
of the regions had candidate stars, and 83% of the regions with candidate stars were
identified as clear. The overcast image had a thin overcast; Orion could be seen through
the clouds, but most of the dimmer stars could not be seen in the image. In this image,
88% of the regions had candidate stars, and 3% of the regions with candidate stars were
identified as clear. The scattered cloud case was in between, with 66% identified as
clear.
For the LM best fit technique, we used a routine, MPFIT IDL
software, written by Craig Markwardt at NASA. It was
obtained from Craig B. Markwardt, NASA/GSFC Code 662,
Greenbelt, MD 20770, craigm@lheamail.gsfc.nasa.gov,
http://cow.physics.wisc.edu/~craigm/idl/idl.html. For a
further reference, consider
Press, W.H., Et. Al., 1992, Numerical Recipes In Fortran,
Cambridge University Press, p. 678-682.
U
(
At this point, there were several productive directions we might go in. The decision was
made to convert the program in its current (at that time) state to C code, in the hopes of
fielding it with the instrument at a later time. Other potential development are discussed
in Section 4.5.
4.4. Conversion of the Starlight Algorithm to C
The last step of the algorithm development funded under this contract was the conversion
of the code from the interactive IDL code developed as described in Section 4.3 into a
more automated C code. This version made more use of input files, as opposed to user
input, and was considerably faster (on the order of seconds as opposed to minutes to
process an image). This version of the code is documented in Memo AV99-054t (Burden
1999e).
4.5 Further Development of the Starlight Algorithm
Although the sponsor did not have sufficient funds to complete the algorithm in a version
which could be fielded with the instrument, this work has been completed under funding
from another sponsor. During this later work, the algorithm was also improved
considerably in sophistication, to enable it to handle moonlight and, to some extent,
nearby city lights.
At present the algorithm is based on uncalibrated raw data. There are several
enhancements which should be feasible if using calibrated data, and are invarious stages
of development. First, it should be quite feasible to provide estimates of earth-to-space
beam transmittance or optical depth. By evaluating the calibrated signal from a large
number of stars, and comparing it with the return expected from the stars in the absence
of clouds, we feel that we can develop a means for providing a map of the variance in
absolute earth-to-space beam transmittance. This work has been funded under a separate
contract.
In addition, better handling of moonlight and of city lights should be possible. Our
understanding from our sponsors who are using the automated algorithm
ho are using the automated algorithm is that they are
quite pleased with the results, but we have not yet done extensive analysis of the results
and the potential for improvement.
Also, we have developed approaches for improving the resolution of the night algorithm
to potentially provide pixel-by-pixel resolution. While we have not been funded to
address this work, we feel that it is a definite possibile future application for the WSI
night-time data. An additional useful development would be the application to sunrise
and sunset time regimes.
5. Additional Work Performed during the May 98 - Mar 00 Period
Although the funding during this period was primarily directed toward night algorithm
development, as discussed in Section 4, additional work to support the ongoing
operational use of the WSI was completed as needed. Another software upgrade was
delivered in Oct 98, as documented in Technical Memo AV98-053t and 55t (Kart 1998b
and c). This upgrade included addition of a time sync operation that allows the display
computer to synchronize with the collection computer. Also, a SCRAMnet life check
feature was added to provide a way to check if the SCRAMnet collection is responding.
Additional features were added as documented in the tech memo. We also evaluated
changes required for Y2K handling and made the necessary changes. An additional
minor upgrade is documented in Memo AV99-051t (Karr, 1999).
The sponsor decided that they would like to use the new geometric calibration results to
improve the accuracy of the angle assignment to the pixels in the field images. Since
Unit 5 is at a classified site, we are restricted from viewing and processing images. For
this reason, we developed the geometric calibration procedures into a series of steps that
could be performed by the sponsor, and wrote and delivered two tutorial memos
documenting these processes. These memos are Technical Memoranda AV99-049t and
050t (Burden, 1999c and d).
A new occultor trolley motor was purchased during this interval, and installed at the site
in August 99, as documented in Technical Memo AV99-046t (Wertz and Shields 1999).
6. Preparation for Beam Transmittance Development Work, Mar 00 - Jun 01
The final funds transfer of $13,073 was received approximately Mar 00. At that time, the
request was that these funds be saved for camera repair, in case any repair was required.
No repairs were required initially. In the meantime, considerable development of the
night algorithm occurred under funding from other sponsors, with the result that the night
algorithm was fielded on other Day/Night WSI units. Following meetings with the
sponsor in October 00, the decision was made to fund the enhancement of the night
algorithm to determine earth-to-space beam transmittance. This task was funded under a
follow-on contract, N-00014-01-D-0043 DO #5. Development of the beam transmittance
algorithm would require calibration of the camera, so the remaining funds under the
existing contract were used to acquire the radiometric calibration data for the camera.
The camera was also repaired at this time, as documented in Memo AV01-040t (Baker,
2000). Initial work on the calibration processing was also begun, and although the
calibration processing was completed under the follow-on contract, the calibration
memos AV01-071 - 079 are referenced here. Work on the beam transmittance
development is in progress at this time, under the follow-on contract.
7. Summary
The raw WSI night imagery provides extremely useful images, which are easily
interpreted by human users. As may be seen from the sample images in the attachment,
both clouds and stars can be readily seen, and the general character of the clouds is
readily observed. To the best of our knowledge, there is no other instrument in existence,
which has this combination of features, allowing human interpretation of the cloud scene
during night-time, daylight, and in sunrise and sunset time regimes. In addition, because
1
the data have excellent dynamic range and low noise, and can be calibrated for absolute
radiance, they are very useful for further mathematical evaluation. The starlight
algorithm is an example of a first-generation algorithm which starts to utilize this
outstanding data.
During the period from July 97 through June 01, Contract N00014-97-C-0385 provided
funding to support the use of the Day/Night Whole Sky Imager by Air Force Phillips Lab
(later renamed Air Force Research Lab). The sponsors were very happy with the
instrument, reporting to us that their program “has been very happy with the WSI system
we purchased from MPL a few years back. It is an excellent source of data and is a very
reliable system.” (personal e-mail). During this time, repairs were performed as needed,
and software upgrades were delivered. A means of providing a very accurate assignment
f the angular location (in object space) of each pixel (in image space) was developed
and delivered. A first generation night algorithm, for identifying the presence of clouds
in a starlight image, was developed. Preliminary work toward extending these techniques
toward determination of earth-to-space beam transmittance was begun, and is continuing
under follow-on funding. This work has enabled the Air Force to continue many
productive years of using the WSI in a variety of test-support scenarios.
8. Acknowledgements
The personnel at the Air Force Research Laboratory have, without exception, been a
pleasure to work with. They have been very technically astute, and have enabled us to
work efficiently. We worked with Capt. Eric Claunch prior to the award, and Capt.
James Lee through approximately February 00. We then worked with Capt. Dan.
Gisselquist through the remainder of this contract. Capt. Craig Phillips is our contact for
the follow-on work. In addition, several individuals from Logican have worked with this
project, and again have been most helpful. These include Gerry Williams, Cliff Clint
Campbell, and Wil Graveman. Our thanks to the Air Force and the Air Force Research
Labs for funding this work.
9. References
In-house Technical Memorandums available at sponsor's request
Baker, J. G., (2001), “Delivery of Repaired Camera to Air Force Unit 5”, Atmospheric
Optics Group Technical Memorandum AV00-040t, 29 May 00.
Burden, A. R., (1998a), “IDL Software for Converting between Image Pixel Coordinates
and Equatorial Coordinates", Atmospheric Optics Group Technical Memorandum AV98-
040t, 5 Aug 98.
Burden, A. R., (19986), “Models from Geometric Calibration of WSI Unit 4”,
Atmospheric Optics Group Technical Memorandum AV98-041t, 5 Aug 98.
Burden, A. R., (1999a), “Application of ARM Image Geometric Calibration Technique to
SGP Data", Atmospheric Optics Group Technical Memorandum AV99-043t, 16 Aug 99.
Burden, A. R., (1999b), “Preliminary Results of Starlight-Based WSI Cloud Algorithm”,
Atmospheric Optics Group Technical Memorandum AV99-044t, 17 Aug 99.
Burden, A. R., (1999c), “Instructions for Using GetCalibStars: Geometric Calibration
Step I”, Atmospheric Optics Group Technical Memorandum AV99-049t, 20 Sep 99.
Burden, A. R., (1999d), “Instructions for Using AngleFit: Geometric Calibration Step
II”, Atmospheric Optics Group Technical Memorandum AV99-050t, 20 Sep 99.
Burden, A. R., (1999), “GetStarDatal – Conversion to C and Updated Algorithm”,
Atmospheric Optics Group Technical Memorandum AV99-054t, 15 Dec 99.
Burden, A. R., (2001a), “Basic Geometric Calibration of WSI Unit 5v3”, Atmospheric
Optics Group Technical Memorandum AV01-071t, 25 Jul 01.
Burden, A. R., (2001b), “Unit 5v3 Rolloff Calibration”, Atmospheric Optics Group
Technical Memorandum AV01-072t, 26 Jul 01.
Burden, A. R., (20010), “Unit 5v3 Flat Field Calibration”, Atmospheric Optics Group
Technical Memorandum AV01-073t, 26 Jul 01.
Burden, A. R., (2001d), “Unit 5v3 Linearity and Exposure Calibration Results”.
Atmospheric Optics Group Technical Memorandum AV01-074t, 19 Jul 01.
Burden, A. R., (2001e), “Unit 5v3 Effective Lamp Irradiance Computation”,
Atmospheric Optics Group Technical Memorandum AV01-075t, 19 Jul 01.
Burden, A. R., (2001f), “Unit 5 Version 3 Aperture Calibrations”, Atmospheric Optics
Group Technical Memorandum AV01-076t, 17 Jul 01.
Burden, A. R., (2001g), “Unit 5v3 Acrylic Dome Callibration", Atmospheric Optics
Group Technical Memorandum AV01-077t, 19 Jul 01.
Burden, A. R., (2001h), "Absolute Calibration Results for Unit 5, Version 3”,
Atmospheric Optics Group Technical Memorandum AV01-078t, 17 Jul 01.
Burden, A. R., (20011), “Unit 5v3 Calibration Summary”, Atmospheric Optics Group
Technical Memorandum AV01-079t, 26 Jul 01.
Karr, M. E., (1998a), “Air Force Run WSI updates, Version 5.8”, Atmospheric Optics
Group Technical Memorandum AV98-011t, 17 Feb 98.
Karr, M. E., (1998b), “Updates to Air Force Run WSI - Version 6.0”, Atmospheric Optics
Group Technical Memorandum AV98-053, 01 Oct 98.
Karr, M. E., (1998c), “AF RunWSI version 6.0 documentation", Atmospheric Optics
Group Technical Memorandum AV98-055t, 07 Oct 98.
Karr, M. E., (1999), “AF RunWSI version 6.1”, Atmospheric Optics Group Technical
Memorandum AV99-051t, 19 Oct 99.
Shields, J. E., (1997), “Delivery of Repaired Camera to Air Force Unit 5”, Atmospheric
Optics Group Technical Memorandum AV97-051t, 30 Dec 97.
Shields, J. E., (2000), “Visit Report: Capt. Dan Gisselquist, 27 Oct 00", Atmospheric
Optics Group Technical Memorandum AV00-042t, 31 Oct 00.
Wertz, J. L., (1998), “Trip report for Air Force Unit 5, January 5-6, 1998”, Atmospheric
Optics Group Technical Memorandum AV98-008t, 06 Feb 98.
Wertz, J. L., and J. E. Shields (1999), "Unit 5 Repair of Trolley Drive Trip Report,
August 1999”, Atmospheric Optics Group Technical Memorandum AV99-046t, 10 Aug
99/9 Jun 00.
General References
Hoffleit E.D., Warren Jr. W.H., 1991, The Bright Star Catalogue, 5th Revised Ed.
(Preliminary Version), Astronomical Data Center, NSSDC/ADC
Johnson, R. W., W. S. Hering and J. E. Shields (1989), “Automated Visibility and Cloud
Cover Measurements with a Solid-State Imaging System”, Marine Physical Laboratory,
Scripps Institution of Oceanography, University of California San Diego, SIO 89-7, GL-
TR-89-0061, NTIS No. ADA216906
Johnson, R. W., J. E. Shields, and T. L. Koehoer (1991), “Analysis and Interpretation of
Simultaneous Multi-Station Shole Sky Imagery", Marine Physical Laboratory, Scripps
Institution of Oceanography, University of California San Diego, SIO 91-3, PL-TR-91-
2214
Shields, J. E., R. W. Johnson, and T. L. Koehler, (1993), “Automated Whole Sky
Imaging Systems for Cloud Field Assessment”, Fourth Symposium on Global Change
Studies, 17 – 22 January 1993, American Meteorological Society, Boston, MA
1
Shields, J. E., R. W. Johnson, M. E. Karr, R. A. Weymouth, and D. S. Sauer, (1997a),
“Delivery and Development of a Day/Night Whole Sky Imager with Enhanced Angular
Alignment for Full 24 Hour Cloud Distribution Assessment”, Marine Physical
Laboratory, Scripps Institution of Oceanography, University of California San Diego,
Report MPL-U-8/97
10
Shields, J. E., M. Kart, and R. W. Johnson, (19976), “Service Support for the Phillips
Laboratory Whole Sky Imager”, Marine Physical Laboratory, Scripps Institution of
Oceanography, University of California San Diego, Report MPL-U-10/97
1
Shields, J. E., R. W. Johnson, M. E. Karr, and J. L. Wertz, (1998), “Automated
Day/Night Whole Sky Imagers for Filed Assessment of Cloud Cover Distributions and
Radiance Distributions”, Tenth Symposium on Meteorological Observations and
Instrumentation, 11 – 16 January 1998, American Meteorological Society, Boston, MA
MARINE PHYSICAL LABORATORY, 0701
of the Scripps Institution of Oceanography
San Diego, California 92152-6400
AV99-0440
17 August 1999
Technical Memorandum
To: Atmospheric Optics Group
From: A.R. Burden and J.E. Shields
Subject: Preliminary Results of Starlight-Based WSI Cloud Algorithm
This memo describes preliminary work toward developing a nighttime, starlight-based
cloud algorithm to be included in the day-night WSI system. Related information regarding
automatic detection of stars in nighttime imagery can be found in Tech. Memos AV98-040t,
AV98-042t, and AV99-043t.
Background
The day/night WSI has test software capable of detecting clouds at night under full-moon
conditions, but no software is available for detecting clouds at night under moonless conditions.
We are currently working on development of a moonless night cloud-detection algorithm for use
with the WSI. Here, we discuss some initial findings and present a preliminary cloud-detection
algorithm written in IDL. The current capabilities of the algorithm are discussed, as well as
issues that we feel need further development.
The approach of the initial study has been to rely on the difference in the optical
characteristics of a star under cloud-free conditions compared to cloudy conditions. When an
image is formed by an optical system, the image is normally slightly defocused. Mathematically,
the image is convolved with the point-spread function (PSF) of the optics. For example, the
image of a point source, such as a star, may be spread out over several pixels. The PSF for the
WSI optical system can be modeled by a Gaussian distribution. Using images of stars, we have
found that the width of the PSF, which is equal to the standard deviation of the Gaussian
distribution, is typically 0.3-0.8 pixels. It does not appear to vary systematically over the image.
An image of a star centered on a pixel is characterized by a peak signal at the center pixel
surrounded by pixels exhibiting signals that decrease to the background value further from the
peak. A star with a point-spread function of 0.5 is characterized by a signal that is reduced to
background within three pixels of the center. If the same star is obscured by opaque cloud, most
of the signal from the star will not be transmitted through the cloud to the camera. The signal
from the cloud will not exhibit Gaussian behavior, and will most likely exhibit only small
variations in signal. If the cloud covering the star is semi-transparent, a reduced signal from the
star may still show through. The background signal may be increased by the presence of the thin
cloud.
Using a star catalog to determine the locations of bright stars within a WSI image, we can
search for the stars using tests based on the differences in optical characteristics just described.
The presence of a Gaussian-like peak at the expected star location can be used as an indicator of
cloud-free conditions, while the absence of the peak can be used as an indicator of cloud
obscuration. This basic feature, along with other tests based on differences in signal for stars and
clouds, is incorporated into the first-generation cloud-detection algorithm describ
scheme represents one possible method for identifying clouds in WSI imagery. Other possible
schemes may be considered as funding permits.
The algorithm described here has not been written in a form that is ideal for real-time use
with the WSI. Rather, it was written in IDL to facilitate the development of such an algorithm
that could later be written in C and incorporated into the field software. The basic structure of
the current algorithm can be divided into two parts: subroutines that obtain information about the
catalog stars and the image pixels that correspond with these stars, and subroutines that use this
information in tests that differentiate cloud-free and cloudy regions.
The first group of subroutines is called by the main routine GetStarData. The IDL
version of GetStarData can be very time consuming to run, as it stores data for thousands of stars
in an image with magnitudes in the range 0-7. The information from GetStarData is saved in a
file for each image, and available for future analysis using the second part of the algorithm. At
this time, the cloud decision part of the algorithm is found in a routine called StarAlgorithm1.
As new methods of cloud detection are developed, they can be incorporated into the algorithm
and applied to GetStarData output files without having to rerun the entire set of procedures.
Data
In order to develop and test the new nighttime algorithm, images have been obtained
from an archive of data collected with the WSI located at the SGP site operated by ARM. These
images were obtained by contacting the ARM archive and requesting data from several nights
without moonlight during the years 1997-1999. The data were downloaded from the ARM
archive web site and extracted from compressed archive files. Each image file was then
calibrated using the IDL software Wsi SGP Data Ingest, written by Tim Tooman of Sandia
National Laboratory, using the radiometric calibration equations developed at MPL. The
resulting files are in the NetCDF data format. Files written in the NetCDF format can be read
using standard IDL routines. The MPL-written routines GetCDFData and GetCDFDate are used
to extract the image and header information for use in the current study.
GetStar Data
Locating Star Positions
In order to identify a star in WSI imagery, the theoretical location of the star and a
transformation from the theoretical coordinates to image coordinates are required. The first item
is easily obtained from one of many bright star catalogs available. Star catalogs, such as the
National Space Science Data Center (NSSDC) bright star catalog used here, contain vital
information about known stars, including right ascension, declination, magnitude, and reference
numbers. The NSSDC catalog used here is a formatted ASCII file containing over 9100 stars of
up to magnitude 8 (magnitude decreases with brightness).
The geometric calibration routine, Rad2Az, described in AV98-040t is used to convert
celestial coordinates (right ascension, declination) for a star to spherical coordinates (azimuth,
zenith) when the time of observation is known. Although coarse equations to derive pixel
locations from spherical coordinates are given in AV98-040t, we now use a more precise
geometric calibration based on an empirical best fit that is a function of azimuth and zenith
angles to obtain image pixel coordinates. This procedure, described in AV99-043t, is capable of
locating stars in the catalog to within half a pixel. The conversion routine that uses the empirical
fit equations is called GetStarXY.
The conversion equations obtained with the fitting procedure must be reworked for
different WSI cameras and each time a WSI camera is moved. In the case of SGP data, the
camera was moved on a couple of occasions during the period of study, December 18, 1997 to
January 17, 1999. Therefore, three sets of calibration equations were developed, and the
appropriate set is used according the date given in the filename of each image.
Identifying Stars in Cloud-Free Images
Once the position of a star in image pixel coordinates has been determined, the next step
is to determine whether a star-like peak exists in the image at that point. For this study, we use
an iterative method that searches the region surrounding the calculated star center for a peak.
First, the maximum signal for the pixels directly next to and including the calculated star center
is determined. If the pixel with maximum signal is the calculated star center, the search is
stopped. Otherwise, the pixel with maximum signal is labeled as the new center pixel and the
process is repeated. If no star center can be found within a certain pixel distance of the
calculated star center, the search is considered a failure. For this study, a maximum distance of
three pixels was used. Because crowded star fields may contain multiple star peaks in the same
region, a second attempt, more sophisticated attempt, is made to find a different peak near the
calculated star center when the first attempt fails. The IDL routine that searches for potential star
peaks is called FindCenterPixel.
Fitting Potential Stars to 2-Dimensional Gaussian Distribution
If a potential star peak is found near the calculated star position, a best fit calculation is
made using the peak pixel and the surrounding pixels. The best fit subroutine uses the
Levenberg-Marquardt method (LMM), which is a robust technique commonly used for solving
non-linear least-squares problems (see Numerical Recipes, Press, et al, for instance). Additional
information on the method will be documented in a future memo describing software for
focusing the WSI camera. The array of pixel values that represents the star are fitted with a two-
dimensional Gaussian function with six parameters, as shown in Equation 1. An additional
parameter, the tilt, which can be used to determine the orientation of an asymmetrical Gaussian
about the axes, has been left out for now. While the asymmetric nature of a star distribution is
more realistic, model results to date have not been reasonable.
F(x, y) = Ag + Age ?!
with the following parameters:
Ao = Background value
A1 = Scaling factor
A2 = Width of Gaussian in the x-direction (point spread)
A3 = Width of Gaussian in the y-direction (point spread)
A4 = Center in x-direction
As = Center in y-direction
LMM finds the set of parameters that best fits the array of pixel values. The fit is "best"
in the least-squares sense; that is, the sum of the weighted, squared differences between the
model and data is minimized. While calibrated data were used in this case, the routines work
similarly for uncalibrated WSI images. The IDL subroutines that perform the best fit calculation
are called MPFit and MPFitFun. These subroutines were converted from Fortran routines
included in the MinPack statistical package by Craig B. Markwardt of NASA. Additional
information can be obtained at his website: http://cow.physics.wisc.edu/~craigm/idl.
As described in the Validation section, the LMM has several optional inputs that may
affect the results: an array may be passed to MPFitFun that contains weights for each pixel value;
the six parameters can be given constraints on their values, and a ceiling can be put on the
number of iterations performed. Additionally, multiple best fit calculations may be required
when the original, 5- by 5-pixel array does not provide a good fit. The 1-sigma errors produced
by the covariance matrix of the best fit are used as an indicator of good fit, and may be high
when the array size is not large enough. When a parameter error is higher than a certain
threshold, the fit is considered suspect. In this case, the array width is increased by 2 pixels, and
the fit is performed again. A maximum array size of 11 by 11 pixels is used. While somewhat
arbitrary, this method has shown to give better results than using either small arrays or large
arrays alone. Some stars require additional surrounding pixels to be distinguished from the
background, while others lie in crowded star fields where too large an array prohibits a good fit
with the Gaussian model.
For each best fit calculation, important information is stored regarding the results. As
stated in the previous paragraph, when a best fit calculation results in large error values, a new
calculation is made with a larger input array. Other factors that merit a new calculation include
iterations that exceed a specified maximum and star point-spread function widths that exceed
specified limits. For this study, calculations exceeding 300 iterations or point-spread function
widths outside the range 0-2 were performed again with larger arrays. Most best fit calculations
converge within a few iterations, especially for brighter stars. If no array width between 5 and
11 pixels results in an acceptable outcome, the star in question is flagged with an appropriate
code (Code 1 if maximum iterations or point-spread limits exceeded; Code 4 if large errors)
identifying the cause of the failure. Additional flags are set if the region near the calculated star
center is potentially crowded with stars (Code 2: multiple peaks nearby), or if the star peak
coincides with a previously identified star (Code 3: stars with identical center). All other
outcomes are classified as a success (Code 0). The list of MPL-written subroutines used in the
best-fit calculation includes CheckForPeak, StarPeak, RunMPFit, and Gauss2D Mod.
Statistics and Data Storage
The next step after the best fit has been determined is to calculate certain statistics for the
star, or lack thereof. The contrast is calculated by finding the integral of the two-dimensional
Gaussian model of the star and dividing the result by the background value for the star. The
background is just the first parameter resulting from the best fit. The peak is equal to the sum of
the background and scaling factor (i.e., Ao+A1). The full-width-half-maximum (FWHM) is
defined for a Gaussian distribution to be 8 ln(2)0 , or 2.360". Since in this case the distribution
is two-dimensional, we take the mean of the x and y FWHM values.
For cases where no satisfactory fit was found, the statistics are calculated differently.
The background is the median of the 9- by 9-pixel array surrounding the calculated star center.
The peak is the mean of the 3- by 3-pixel array surrounding the calculated star center. The
contrast is the peak minus the background divided by the background. The FWHM is set equal
to 0. The statistical calculation routines are StoreStats and Gauss2Dintegral.
Currently, for each image analyzed, information is recorded for all stars in the NSSDC
catalog with magnitudes in the range 0-7 and zenith angles less than 70°. The following
information is stored for each star: HR and SAO catalog numbers, right ascension, declination,
magnitude, multiple star designation, zenith, azimuth, expected x-pixel, expected y-pixel, image
x-pixel, image y-pixel, quality flag, background, peak, contrast, the x and y point-spread
functions, and a image cell number described in the next section.
StarAlgorithm1
Below is a description of the preliminary version of the nighttime, starlight-based cloud
algorithm based on initial study results. While simple in its application, it demonstrates that the
basic concepts described in the previous sections can yield reasonable results. The approach
used here is to define a number of regions, or ‘image cells', that are approximately the same size
and oriented around the image center, as shown in Figure 1. The grid is made up of a total of
237 image cells covering zenith angles up to 70°. Each image cell is examined to determine the
number of potential stars that exist there. If at least half the potential stars are identified as
exhibiting star-like qualities, the entire image cell is classified as cloud-free.
* The Full Width at Half Maximum is the width of the Gaussian distribution at the half-power point, which is equal
to 0.5 for a normal distribution. The value shown here is obtained by solving the one-dimensional Gaussian for x
and multiplying the result by 2.

Figure 1. WSI Image Overlayed with Cloud Decision Grid
Parameters that are considered important for identifying a star include: the contrast of the
star peak with the sky background; the width of the point spread functions associated with the
Gaussian model of the star; and the value of the quality flag based on the best fit determination
for the star. Additionally, the contrast of a star as well as the ability of the best fit algorithm to
successfully converge is diminished for higher magnitude stars. Therefore, the quality flag,
contrast, and point-spread functions of low magnitude stars within an image cell are examined
first. If none exist, then all other stars are examined. The magnitude threshold used here is 5.
Before the clear and cloudy regions in the image are identified, all stars that were flagged
as being stacked on other, lower magnitude stars (Code 3), are removed from the dataset. Each
image cell is then categorized as cloud-free or cloudy. The first step of the decision process is to
identify any star in the image cell with Code 1, defined here as a best fit failure, as cloudy. Next,
all stars in the image cell with a contrast >0.18 and x- and y-point spread function widths in the
range 0.3-0.8 are identified as cloud-free. All other stars in the image cell are classified as
cloudy. If the number of cloud-free stars is at least half the total number of low magnitude stars
in the cell, the entire cell is identified as cloud-free. If no low magnitude stars are found in the
image cell, the same process is done for the high magnitude stars. At this time, the additional
information from Codes 2 and 4 is not used.
The thresholds used in the algorithm were obtained by examining the point-spread
function widths of stars in completely cloud-free images and by comparing the contrast of stars
in completely cloud-free images and completely overcast images. Figure 2 shows the
distribution of point-spread functions for stars identified in several cloud-free images. The data
were separated into low and high magnitude stars, and stars flagged as having poor best fit
convergence were removed from the data. Each distribution was fitted with a Gaussian
distribution to estimate the mean and standard deviation. In both cases, the point-spread function
widths were centered near 0.48, with most cases being in the range 0.3-0.8. The high magnitude
distribution exhibited a slightly lower mean, but skewed toward higher point-spread function
widths. This demonstrates the difficulty in performing best fit calculations on high magnitude
stars.
Figure 3 shows contrast distributions for stars under cloud-free conditions and stars under
cloudy conditions. Again, stars were divided into low and high magnitude datasets. In order to
highlight the differences in the distributions, only contrasts up to a value of 5 were plotted. In
-free, low magnitude case, most of the values were in the range 0.25-2.5. A few stars
had contrasts as high as 50 or more. The contrasts for the cloud-free, high magnitude case were
much lower, with most of the values lying in the range 0.05-1.0. Under completely overcast
conditions, the contrasts rarely increase above 0.2, and these cases may be due to optical
aberrations, bright stars being viewed through clouds with lower optical depth, or natural
variations in cloud radiance. Since some crossover exists between the cloud-free and overcast
distributions, the results of the cloud decision algorithm will be depen
is applied. In this case, a threshold of 0.18 will misidentify some potentially cloud-free stars as
being obscured by clouds. Since the number of cases that fall into that range appears small, the
results should not be drastically affected.
Algorithm Results and Validation
When GetCloudData is run, several optional parameters may be included in the
calculation of the LMM best fit. Results to date show that there is some dependence of the
results on how these parameters are initialized. For instance, an array input to MPFit contains
limits for each of the six parameters. Preliminary studies showed that limiting the point-spread
function widths had an adverse affect on the results. Therefore, requiring that the background
and scaling factors be greater than zero were the only limits used.
It was also found that the LMM subroutine works best when given initial parameters that
are close to the minimized values. The background value is set equal to the minimum pixel
value in the input array. The scaling factor is set equal to the maximum pixel value minus the
minimum pixel value. The x-and y-widths are equal to 0.5, since this is a typical point-spread
function width. Finally, the x- and y-centers are set to the central pixel location.
An additional parameter that needs to be investigated further is the use of weighted pixel
values. It has been found that leaving the pixel values unweighted is effective for low magnitude
Figure 2. Point-Spread Function Widths for Cloud-Free Images
Magnitudes < 5 (solid) and Magnitudes > 5 (dash)
0.20 TTTTTTTTTTTTTTT

0.15
Gaussian Fit (M<5):
Mean = 0.482
SDev = 0.083
L.-
Gaussian Fit (M>5):
Mean = 0.473
SDev = 0.103
Relative Frequency
.-.-.-.
0.05
..
.
0.00 LEFT
IIIIIIII
ESE10
0.0
0.5
1.0
1.5
Point-Spread Function Width (pixels)
2.0
Figure 3. Comparison of Star Contrast for Cloud-Free and Cloudy Conditions
Star Contrast under Cloud-Free Conditions
Magnitudes in Range 0-5
0.050 (TTTTTTTT T TTTTTTTTTTTTTTTTTTTTTTTTTTT)
Star Contrast under Cloud-Free Conditions
Magnitudes in Range 5-6
OSO
M agnus
0.050 [
M ITTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTT


LUI
0.040F
0.040F
0.030
@ 0.030
Relative Frequency
Relative Frequency
0.020
0.0201
0.0104
0.010
0.000 Lumumului Hrbatilih i
-1 0 1 2 3 4 5
Contrast
0.000 tuw
-1 0
tilluululuuluu
1 2 3 4
Contrast
5
Star Contrast under Overcast Conditions
Magnitudes in Range 0-5
0.40 PITTITYTTITI
TUTTIITTITTIPITIM
Star Contrast under Overcast Conditions
Magnitudes in Range 5-6
0.20 MMITTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTI


0.3
0.15
Relative Frequency
Relative Frequency
10
0.05
0
.
00
السبييييييييييييييليييييييييييييييلييياليليييييييييا
0.00 Culo
-1 0
dawulu
1 2
Contrast
3
m u
4
5
-1
0
1
1
Contrast
2
3
4
stars, but using weights that are directly proportional to each pixel value is more effective for
high magnitude stars. The point-spread functions are more realistic, and the algorithm converges
much quicker. One possible scenario is to use unweighted pixel values for low magnitude stars,
switching to weights beyond a certain magnitude value. For the data used in this memo, no
weights were used.
Despite the large number of dependencies, LMM is quite robust. CURVEFIT, the
nonlinear least-squares routine included with IDL, uses a gradient expansion algorithm. A
comparison of the two methods using modeled stars in the form of two-dimensional Gaussian
distributions showed LMM converged to the exact point-spread function values in most cases,
while the gradient expansion method did not. Results for the two methods were comparable
when an input array with a width of nine pixels was used, except for point-spread function
widths less than 0.4, where LMM results were superior. When the array width was reduced to
five pixels, the gradient expansion method performed far worse than LMM. Neither method
worked well for point-spread function widths much lower than 3.
To validate the preliminary star algorithm, several nighttime images were analyzed.
Figures 4 to 8 each show the original image and the image overlaid with a mask that is color-
coded to represent the cases of cloud-free (blue), overcast (red), and no usable stars (gray). The
first two examples are cloud-free images collected at 0427 UT on December 18, 1998 and 0220
UT on December 18, 1997, respectively. Figures 6 and 7 are of completely overcast images
collected at 0356 UT on November 18, 1998 and 0820 UT on January 17, 1999, respectively
Finally, Figure 8 shows an image partly covered with thin and opaque clouds collected at 1127
UT on December 18, 1998.
Table 1 summarizes the number of image cells out of 237 that fall into each category.
Also shown is the number of image cells that did not have any stars with magnitudes <5. While
the results are not perfect, the overall strategy appears to be effective.
Table 1.
Number of Image Cells in Each Category

#w/o Low
Mag Stars
Cloud-Free
Image
Cloud-Free 1
M
Cloud-Free 2
Obscured
by Cloud
8
7
234
235
206
/
228
230
3
1
28
No Stars
1
0
0
1
3
Overcast 1
/
/
L
L
35
34
41
37
1
L
Overcast 2
Partly-Cloudy
10
We can also examine the results by looking at the stars individually. Table 2 gives the
algorithm results for catalog stars as a function of magnitude and zenith angle. The data is
summarized from multiple cloud-free images. The first column lists the magnitude range for the
row. Results shown in the other columns are broken down by zenith angle: stars with zenith
angles less than 70, 50, and 30°, respectively. The second column shows the total number of
catalog stars in the given magnitude range. The third column gives the percentage of peaks
identified as catalog stars using the full algorithm logic, including contrast and point-spread
width thresholds. The last column shows the mean residual in pixels between the star
coordinates found from the Gaussian best fit and the image star coordinates corresponding to the
stars identified in column three.
-
-
. -
-
----
-
-
-
-
-
-
Table 2. Star Identification Results for Cloud-Free Scenes as a Function of
Star Magnitude and Zenith Angle
Magnitude Number of Stars % Success by Zenith Angle Mean Pixel Residual by
Range
(70/50/309)
(70/50/309)
Zenith Angle (70/50/309)
0.0-4.0 296 / 168 / 75
0.93 / 0.89 / 0.91
0.48 / 0.37 / 0.31
4.0-5.0 728 / 421 / 175
0.90 / 0.90 / 0.86
0.51 / 0.43 / 0.35
5.0-6.0 2093 / 1164 / 438
0.61 / 0.65 / 0.68
0.58 / 0.49 / 0.43
6.0-7.0 1 2104 / 1151 / 438
0.25 / 0.30 / 0.30
0.57 / 0.45 / 0.43

It is clear from the table that the number of available stars increases substantially with
magnitude. It can also be seen that the ability to successfully identify stars decreases with
magnitude. This is not all that surprising, since the fainter stars are not expected to stand out
from the background as well as brighter stars, and therefore less likely to be identified with the
best fit algorithm. The percentage of identified stars with magnitudes less than 5, 86-93%, is
quite good. Beyond that, the percentages drop off substantially, and for magnitudes 6 and
greater, the stars appear to be a useless indicator for the methods used here. Looking at the flags
that were set for these stars, 30% of magnitude 5-6 stars and 56% of magnitude 6-7 stars had
additional peaks nearby. These additional signals undoubtedly contributed to the failure of the
best fit algorithm for most of the cases. The residuals shown in the last column increase
somewhat with magnitude, but the change is not significant. All cases had mean residuals less
than 0.6 pixel.
While there is a finite probability that, given random pixel coordinates, a peak associated
with a star will be found nearby, examination of images showing the location of catalog stars and
the associated peaks shows that the method works well for low magnitudes. To test the
likelihood of randomly finding a peak resembling a star in an image, a routine called
GetRandomStars was written. This routine uses a random number generator to obtain dummy
star locations that are then passed to the peak search and best fit subroutines, just as with the
catalog stars. In the case of completely cloud-free images, there was about 20% chance that a
peak resembling a star was located using the star algorithm, given 500 random coordinates.
Examining the resulting peak locations showed that the majority of these peaks were indeed
stars. The same exercise with a completely overcast image resulted in less than 1% chance that a
peak resembling a star was located.
Conclusion and Future Work
This first generation algorithm appears to work well for skies with no moon. There are
several approaches to further optimization that we feel would be worth evaluating. One
approach that has showed reasonable results involves the ratio between the contrast for a star in a
cloud-free image and the contrast for the same star in an unknown image. This method appears
work well in distinguishing cloud-free, thin cloud and opaque cloud in the unknown image. In
this case, a library of cloud-free contrast values for specific stars would need to be created.
Unfortunately, light from man-made and other sources creates some dependence of the sky
background on zenith angle. This dependence must be removed before the contrast ratio method
could be applied. In addition, it may be possible to further optimize this algorithm with
techniques such as customizing the star list to remove stars that are close to a brighter star
Removing secondary peaks near the star center from the input array may aid in calculating the
best fit Gaussian, particularly with high magnitude stars. The algorithm also needs to be
evaluated using data from other sites and using raw data as opposed to calibrated data.
However, we feel that this first generation algorithm is sufficiently accurate to merit
converting to C-code that can be included in field software. Once a usable field version is
written, future developments could include determination of optical depth from the calibrated
images, evaluating the transition to partial moonlight, and increasing the angular extent and/or
angular resolution of the algorithm.

Figure 4. Cloud-Free Scene
Grey = No Stars, Red = >50% Failure, Blue = <50% Failure

Figure 5. Cloud-Free Scene
Red = >50% Failure, Blue = <50% Failure
14

Figure 6. Overcast Scene
Red = >50% Failure, Blue = <50% Failure
15

Figure 7. Overcast Scene
Grey = No Stars, Red = >50% Failure, Blue = <50% Failure
16

Figure 8. Partly-Cloudy Scene
Grey = No Stars, Red = >50% Failuré, Blue = <50% Failure
17