I like introducing the concept of all-sky
photometry with the above example. It captures the notion that the
concepts involved are simple, and easily understood. However, it
is misleading because
to do all-sky photometry right there are many details that have to
be
managed carefully and systematically. A good starting point for
illustrating
subtle effects is what I'll refer to as the "generic photometry
equation":
Magnitude = Z - 2.5 × LOG10
( Flux / g ) - K' × AirMass + S × StarColor
+
S2 × AirMass × StarColor
(1)
where Z is a zero-shift constant,
specific to each telescope system and filter (which should
remain the same for many months),
Flux is the star's flux (sum of counts
associated with the star). It's called "Intensity" in MaxIm DL,
g is exposure time ("g" is an engineering
term meaning "gate time"),
K' is zenith extinction
(units of magnitude per air mass),
S is "star color sensitivity." S
is specific to each telescope system (and should remain the same
for many months),
StarColor can be defined using any two filter
bands. B-V is in common use; I use 0.57 × (B-V) - 0.33,
S2 is a second-order term that is
usually ignored because it is only important for high air mass
and extremely
blue or red stars.
This general equation is true for all filter bands (even
unfiltered), though there are different values for the constants
for each filter.
For example, the magnitude equation for V-band (omitting the
last term
in Eqn 1) is:
V = Zv - 2.5 ×
LOG ( Flux / g ) - Kv' × AirMass +
Sv × StarColor
(2)
Similar equations exist for bands B, Rc, Ic, g', r' etc.
Notice that in a "generic photometry equation" for a specific
filter there are two terms associated with a telescope system
that should remain constant (provided there are no hardware
configuration changes). This is illustrated for V-band:
V = Zv
- 2.5 × LOG ( Flux / g ) - Kv' ×
AirMass + Sv ×
StarColor
(3)
Zv and Sv are highlighted in green, and the
task of "photometrically calibrating a telescope system" amounts
to
evaluating these two constants, as well as their counterparts
for any
other filter band of interest. Since extinction is different
each
night the K' terms have to be established each observing
session;
at least this is a common initial response to a review of the
above equation.
I've chosen to define StarColor = 0.57 × (B-V)
- 0.33 because it is zero for typical stars. This is an
arbitrary choice but it is convenient for an iteration procedure
I employ (described below).
NAIVE
ALL-SKY PROCEDURE
When I began to perform all-sky photometry I endeavored to
measure all the unknowns in Eqn (1) each all-sky observing
session. I would do this for the 4-filter set B, V, Rc and Ic.
This entailed observing standard star fields at many elevations
and solving for Zj, Kj', Sj and S2j,
for filters j = B, V, Rc and Ic. This permitted me to observe
the target star at any elevation. Since I also solved for Kj'
trends I could observe the target at any time of the night and
minimize the effect of slow, monotonic extinction trends. This
procedure involved lots of
work; usually a full night of manual observations and several
days of
analysis.
It became apparent to me that the second-order term, S2×AirMass×StarColor,
was always close to zero and its coefficient, S2,
couldn't be established accurately enough to justify its use, so
a few years ago I
discontinued using this term. I also noticed that the star color
sensitivity coefficient, Sj, didn't vary with observing
date - provided the telescope configuration wasn't changed. I
slowly adopted the habit of combining a night's Sj
determination with an average of those from previous all-sky
observing sessions.
Later I noticed that whenever my extinction coefficient was
measured accurately I could count on the zero shift offsets, Zj,
to be the same for many months. But when extinction wasn't
measured accurately there was a relation between Zj and
Kj', wherein a parameter relating the two coefficients
was well-established but neither could be determined by itself.
It was apparent that if the target and standard stars were at
the same air mass the "degeneracy" of the Z-K'
parameter pair wouldn't affect the results. But it was
observationally difficult to observe all standard star fields at
the same elevation.
These seveal learning experiences guided me to a variety of
short-cut observing
and analysis procedures. I'll now describe the next-to-last
short-cut that
I've used, or "partial all-sky" procedure.
PENULTIMATE ALL-SKY PROCEDURE
Since zenith extinction is usually different on each bserving
date, and is time-consuming to measure, a commonly recommended
strategy
is to observe a set of standard stars and the target star when
they are
at the same air mass and at the same approximate time. The
standard stars
are used to evaluate the sum of the two terms
Zj and
Kj'
×AirMass.
This is equivalent to re-arranging the above equations to the
following, for each filter band j:
{ Zj - Kj
' ×
AirMass } = Mj
+ 2.5 × LOG ( Flux / g
) - Sj ×
StarColor
(4)
Let's call the bracketed term on the left
Q for filter
band j. Let's assume for the moment that
Sj is known
(based on
many previous all-sky observing sessions; more on this below).
Notice
that all other items on the right side are known for standard
stars (magnitude Mj and StarColor) or can be measured (Flux and
exposure time, g). It is therefore possible to evaluate
Qj
using many standard stars.
Qj =
average
{ Mj
+ 2.5 × LOG ( Flux / g ) -
Sj
× StarColor } for many standard stars
(5)
The above
Q for filter band j can be determined from the
observation of a field of standard stars, and it is valid for one
air
mass value and one time of the night. Any target stars that are
also
observed at this air mass value and close in time can have their
magnitude
for filter band j evaluated using the following equation:
Mj
= Qj
- 2.5 ×
LOG10 ( Flux / g ) + Sj
×
StarColor
(6)
The parameter
StarColor in the above equation is the
target star's color. This, of course, is initially not known. My
approach to this is to employ an iterative procedure after data
for two bands is available, such as B and V. The iteration
converges very fast (2 or 3 iterations), so this is not a problem.
The above procedure requires that the standard stars be observed
at the same approximate air mass as the target star. If extinction
trends are suspected then it is possible to observe standard star
fields before and after the target star (at the same air mass),
and interpolate
Qj in time.
A method for quickly evaluating whether or not
Sj has
changed from the average of previously measured values is
described in a section below.
I liked this procedure until I experimented with a variant of it.
The next procedure, which I have now adopted as much better,
illustrates an important shortcoming of the procedure just
described: namely, the inability to identify the presence of
cirrus clouds.
ASSESSING SKY CONDITIONS FROM EXTINCTION
PLOT
Sometimes I'm concerned about the possibility that even though a
night begins with the appearance of being photometric (cloudless
and calm) I could not rule out that during the all-sky observing
session cirrus clouds could have drifted overhead. Regular cirrus
is easily noted in the daytime, and in theory it could be
monitored at night if there is moonlight. But
there's an even greater threat to an all-sky observing session:
subvisible cirrus. Even in daytime subvisible cirrus will usually
escape attention. Subvisible cirrus is located just below the
tropopause and defined as having an optical depth of < 0.05.
This optical depths correspond to 50 mmag, which is important when
trying to achieve an absolute photometry accuracy of half this
amount (which is my goal). From an airplane flying near the
tropopause these cirrus clouds are easy to detect visually (which
I've done many times during my atmospheric science career, before
retiring), but from the ground they are not usually noticeable -
even when they have spatial structure. At night they are even less
noticeable, especially from
inside an observing control room. It is true that if these
subvisible cirrus clouds lack spatial structure then they won't
matter, but no cloud is devoid of spatial structure.
In response to this concern I have developed exerimented with
another all-sky observing and analysis procedure designed to
detect the
presence of subvisible cirrus. It consists of an entire night's
monitoring
of a star field that has many standard stars within my FOV, with
occasional breaks for observations of a target star. The large
range of air mass that such an oibserving session provides allows
for a determination of extinction to high accuracy for each band.
I use a 10-position filter wheel, with
the following filters: B, V, Rc, u', g', r', i', z', CBB and NIR.
For each
filter it's possible to fit an extinction curve and identify when
clouds
were present, if they were. Any filter is adequate for this
purpose, so I
prefer the one with a high SNR (such as V or Rc). Here's an
example:
Figure 1. V-band extinction plot (left panel) and
"residual losses" (right panel). Clouds render this observing
session essentially useless, in spite of the evening's
"photometric" beginning.
The left panels shows a fit to the envelope of bright readings.
The night began looking photometric but after darkness episodes of
cirrus clouds apparently drifted over my site. The extinction is
well-established for clearings, and the right panel can be used to
determine when these
occurred (a plot with an expanded magnitude loss scale is used for
this
purpose). Even though this plot is based on observations of a
standard
star field (located at RA = 06:52) the extinction plot is made
from flux
readings of all stars in the field and the sum of fluxes is
plotted. Any
star field could be used for this purpose, since knowledge of star
magnitudes
is not needed for constructing these plots.
The target star was observed at ~ 5.5 UT, which the right panel
shows was probably affected by cloud losses. Upon inspection of
the extra
losses plot I concluded that this observing session shouldn't be
used
for determining magnitudes for the target star. This is dramatic
illustration
of the fact that even when the sky looks perfect at dusk it would
be foolish
to assume that it will remain "photometric" during the rest of the
night.
I view this to be strong evidence for use of the "Simple All-Sky
Procedure
#2" instead of #1!
Five nights after the one just described another all-sky observing
session was started at dusk during conditions that, again,
appeared to be photometric. Here is the result of deriving V-band
extinction for this observing session:
Figure 2.
V-band extinction plot (left panel) and
"residual losses" (right panel) for another date, 2011.01.30,
when the observing session began during "photometric" skies.
Subvisible cirrus appear to have been present at ~ 6.0 to 7.5
UT, and maybe 10.3 UT. At
other times the sky appears to be photometric.
Again, the merits of relying upon "Simple All-Sky Procedure #2"
instead of #1 are illustrated by the damatically better extinction
and
extra losses plot for Fig. 2 compared with Fig. 1; both nights
began
under photometric conditions but the first night deteriorated ~ 2
hours
after sunset.
Here's an expanded magnitude scale version of the previous
figure's right panel:
Figure 3. Expanded magnitude scale of "extra losses"
plot for 2011.01.30, showing which observing cycles were
well-behaved (low
noise and no losses). Target stars were obseved at 3 UT, 5.3 UT
and 11
UT.
The above figure shows that the first target star observation (~ 3
UT) was unaffected by clouds at the level of ~ 5 mmag. The second
target star observations (~ 5.3 UT) may have been affected by
losses of 10 or
20 mmag. These extra losses will render that target star's all-sky
magnitude
solution uncertain by at least 10 mmag, so it might not be wise to
accept
the magnitude solutions. However, since the observations with B
and V filters
are close together in time for each observing cycle it should be
possible
to rely upon the B-V star color solution. The third target star
was observed
after 10.6 UT, and since it was observed for 40 minutes (3 cycles)
it might
be possible to determine if any or all of its observations are
cloud free
based on the consistency of these observations.
Since the extinction plot makes use of the envelope of bright
readings it is possible to obtain zenith extinction even for
nights that are occasionally cloudy. A similar analysis of images
for other filters allowed the following broadband extinction
spectrum to be constructed.
Figure 3. Broad-band spectrum of atmospheric
extinction at zenith for the night 2011.01.30. The 4
components of atmospheric extinction were adjusted to achieve a
fit to the measurements (described in a section below). The
Rayleigh component is essentially identical to that suggested by
Hayes and Latham, 1975. (The u' extinction is left over
from an earlier all-sky session, and it should be fairly
constant for a specific observing site.)
The model fit extinction spectrum is statistically compatible with
the measurements. The fact that the extinction measurements can be
readily fit with this 4-component model suggests that the
extinction measurements are accurate, and that a slightly better
extinction value for each filter band can be obtained by reading
the model fit (since it is influenced
by all the measurements). This, in turn, means that it should be
possible
to derive an all-sky magnitude for target stars observed at air
mass values not sampled by the standard star field(s), such as
near zenith. This translates to flexibility in scheduling a
night's observing session. For subsequent analyses it is possible
to adopt extinction readings made of the model fit of all filters
instead of the measurement for that filter. This may
be advisable when extinction measurements are noisy, but on this
occasion
the extinction measurements are not noisy so I have adopted them
instead
of readings from the model fit.
SAFER ALL-SKY PROCEDURE (CURRENTLY-USED)
When plots of extra losses, such as Fig. 3 above, show that an
observing session is found to be useable for determining target
star magnitudes I proceed to process clear sky observing cycles of
the standard star field for photometry measurement and analysis.
My observing sequence is gggrrrriiiiizzzzzRcRcRcVVVVVBBBBBBB, for
exposures of 10 second each. This cycle requires 15 minutes to
complete. For target stars I usually complete 3 cycles, but for
the standard star field there will be ~ 30 cycles. A V-band image
set is calibrated (bias, dark, flat), a hot pixel filter is
applied to each image, and the images are aligned so that all
stars are at the same pixel location. For example, when this is
done for a cycle's V-band images there will be 5 calibrated and
star-aligned images. Because some of the standard stars are faint
I average this set of images (5 for V-band). (Note: median
combining is totally inappropriate for all-sky photometry, only
averaging is appropriate.) The averaged image is then measured by
the photometry tool, which creates a CSV-file for later import to
a spreadsheet.
When choosing a photometry aperture radius it is important to
measure the fraction of flux captured by that aperture size. I do
this by measuring a bright star in an uncrowded part of the image
using a large aperture and one that is smaller by an amount that
leads to a magnitude change of ~ 20 mmag. Typically, this
aperture radius is ~ 3 times FHWM. When the CSV-file produced by
my image measuring program (MaxIm DL, v 5.12) is imported to a
spreadsheet the magnitude readings are corrected for the lost flux
fraction (e.g., the ~ 20 mmag correction).
When the above procedure is performed on several clear sky cycles
for the standard star field, for one filter band, it is possible
to begin the process of solving for the zero shift constant and
star color sensitivity coefficient for that filter band. This is
done by importing the CSV-files (created by MaxIm DL's photometry
tool) to a template spreadsheet designed for this purpose. The
spreadsheet uses JD and FOV coordinates to calculate air mass.
Special parameters are calculated that have the feature that
regardless of air mass when they are plotted their slopes
correspond to
the desired magnitude equation constants:
Z,
K'
and
S.
I manaully search for fits by minimizing chi-square. The remainder
of this
section illustrates results of this analysis for the 2011.01.30
all-sky observing session.
The following plot illustrates a solution fit for determing
S
for Rc-band.
Figure 4. Plot of a parameter whose slope is S
for Rc-band, based on 2011.01.30 observing session. The star
field includes 26 calibrated stars in one FOV that were measured
on multiple observing cycles (i.e., different air masses). The
parameter plotted in the left panel takes into account air mass
and corrects for it.
The data plotted in this figure can be at any air mass
and the slope of this parameter with star color is the
coefficient I refer to as S. Since this plot is for
Rc-band images the slope corresponds to S(Rc). Notice
the good correlation and small slope value. The small
slope value means that my system's Rc-band response function
resembles the
standard Landolt telescope response function.
When
S is derived this way it is compared with values from
previous all-sky observing sessions to make sure it has not
changed. The next
figure shows
S values determined from all-sky observing
sessions
during the past 8 months.
Figure 5. Plot of Star Color
Sensitivity for B and Rc bands, versus date (from 2010.06.04
to 2011.01.30), showing their stability during this 8-month
period.
There is no evidence for a trend in either
S(B) or
S(Rc).
It makes little difference whether subsequent analyses use the
S
value from the observing date or an average from a plot of past
values when
they are this well-behaved.
Since the star color
sensitivity
coefficients appear to remain constant (provided I don't
change the hardware
configuration), it may be possible to simply adopt the average
value from
past all-sky sessions. However, I have decided to monitor
these coefficients
since I now have a calibrated star field with a sufficient
number of stars
within one FOV (26) that it is not difficult to verfiy that
the coefficent
value as not changes.
It is worth mentioning that even when an observing session can't
be
used for deriving target star magnitudes it is possible to derive
star
color sensitivity,
S, because thin cirrus should not
affect magnitude differences between stars of different color
(since subvisible cirrus consists of ice particles large compare
with visible wavelengths, and scattering is therefore Mie type).
Even the zero-shift parameters,
Z, for all filters can be
determined by confining analysis to the clear images.
The next figure shows verification of the extinction value adopted
from an earlier analysis (Fig. 2).
Figure 6. Plot of a parameter that
has a slope equal to zenith extinction. The extinction
value, and its temporal trend, is set by the solution from
another spreadsheet (see Fig. 2 for an
example).
I have the option of solving for zenith extinction (and its
temporal trend) using slide bars in this spreadsheet. Although
I used to solve for extinction this way I now use another
spreadsheet designed for a long sequence of images for doing
this more accurately (e.g., Fig. 2).
Once
S and
K' have been determined for a filter
band it is possible to determine
Z using a spreadsheet
slide bar to minimize chi-square.
Figure 7. Plot of
differences between "true" Rc-band magnitudes and
measurements using the "generic
magnitude equation" with model fit values for S, K' and
Z (for
Rc-band).
The RMS difference in this figure is 14 mmag, which is very
good.
Typically the RMS for this and other bands is ~ 20 to 25 mmag.
I use an
algorithm for identifying outlier measurements that is based
on statistical
theory and a criterion that for the number of measurements
only the worst
25% of non-outlier measurements shall be rejected. This
assures that un-modeled systematic effects (such as an
imperfect flat field) won't influence the model fitting
solution. Notice that scatter is greatest for the faintest
stars.
At this point in the analysis of standard star images for one
band it is possible to cmpletely specify the magnitude
equation for the band. For the example just treated:
Rc = 21.115 ± 0.005
- 2.5 × LOG10 ( Rc
Flux / g ) - (0.110 ± 0.007) × AirMass -
(0.126 ± 0.011) × StarColor
(7)
The zero shift solution for this date is almost identical to
the solution for 2011.01.20 (21.115 versus 21.158). This 3
mmag repeatability is a good indication that the hardware has
not changed and the analysis is not seriously flawed for
either observing date.
For the clear sky V-band image sets for this date the value for
Z(V)
was determined to be 21.073 ± 0.003, with a RMS scatter of 27 mmag
(for 76 standard star flux readings).
The target star's color isn't known, so the V-band and B-band
tentative solutions (starting with zero for star color) are
iterated
until convergence is attained. This leads to final values for V-
and B-band
magntiudes, which mean that StarColor for thetarget star has been
determined,
and can be used for the processing of the other bands.
The analysis procedure just described is possible
because there is one star field with many calibrated stars
within my FOV. The next section describes how I've increased
the number of calibrated stars in a Landolt/Sloan star field
from 19 to 26.
ENHANCED STANDARD STAR FIELD
The all-sky observing and analysis procedure just described can be
performed best if just one standard star field has a large number
of standard stars. If two or more standard star fields were needed
to obtain a sufficent number of stars to establish a zero-shift
parameter, and to verify the star color coefficient for each
filter, then the observing session would be more complicated and
more time would be spent acquiring the standard star fields - and
this would require me to spend more time duing the night at the
telescope controls (instead of sleeping). My telescope system has
a FOV of 19.7 x 13.1 'arc, and an optimum placement of the FOV on
the various Landolt/Sloan standard star fields produces ~ 6
Landolt standard stars
and maybe one Sloan standard star, typically. The best
Landolt/Sloan star
field includes 19 Landolt stars and 4 Sloan stars, and it is
located at
06:52:06 -00:22:30 (which I refer to as L0652). I have used the
Landolt
stars to calibrate 10 more nearby, bright stars, and so far they
appear to
be stable. Thus, this star field affords 26 stars for establishing
zero-shift
parameters, and star color coefficients, for the filter bands B
and V. For
Rc-band there are 24 stars that I can use for this purpose.
Eventually
I may extend the 4 Sloan stars to some of the Landolt stars.
During winter nights L0652 rises shortly after dark and sets at
about sunrise, so for this observing season it is possible to
compare rising
with setting data portions to determine temporal trends. For
several months
on either sideof L0652's optimum observing date (Jan 3) there will
be sufficient air mass overlap to also determine temporal trends.
During summer months, however, there is no comparable calibrated
star field. One of my future projects is to enhance a
Landolt/Sloan star field, such as L1745 or L2042, to create new
standard stars within a single FOV.
FLAT FIELDING
I currently use either dusk or "dome flats." I've painted the
zenith portion of the sliding door (when it's in the closed
position) for this purpose. I use 4 lights (2 with blue coverage,
and 2 with the standard
yellow to red coverage). When taking dome flats I place a "double
T-shirt"
diffuser cover on the telescope aperture, and this reduces the
effects
of any non-uniformity of my illumination of the white spot atop
the dome. Part of my motivation for performing flat field images
using the dome
was to avoid any effects that might exist when using dusk flats
before
the telescope had cooled. After achieving the capability of
exposiing dome
flats at any time during the night I demonstrated that focus
setting doesn't
matter; so now the dome flat capability is merely a convenience,
and I do
not claim that it offers a superior quality flat.
FILTER BANDS
The following two plots show transmission functions for the two
filter sets, g'r'i'z' and BVRcIc.