CCD Cameras Stefan Ivanov

This note is geared towards students preparing for IOAA, but I think some bits will be interesting
for a general audience too. Before going in, you should feel comfortable with stellar magnitudes and
telescope optics. You’ll also need to know how a p-n junction works, which in turn requires learning
about doped semiconductors. Check out this video for an excellent explanation.

1. CCD Operation
First we’ll touch on the photodiode, which is the main component of charge-coupled devices (CCD).
Consider a p-n junction where the p-type material is connected to the negative terminal of a voltage
source, and vice versa. This junction is reverse-biased: the electric field is such that the charge carriers
at both ends of the depletion region are pulled away, so the depletion region will widen. There won’t
be any substantial current in the junction by default, but that’s not our aim at the moment anyway!
In the depletion region the charge carriers on opposite sides of the junction have recombined,
leaving a net charge density associated with the lattice. Thus, locally there is a large electric field
which serves to block any further diffusion between the p-type and the n-type material in the steady
state. This is useful. We can choose to work with a material (e.g. silicon) for which photons of a
given wavelength are able to knock off the bound electrons in the depletion region, resulting in a free
electron and a hole. Usually this pair would recombine and that would be the end of the story. But
here the large electric field will send our hole-electron pair in opposite directions, which, if you squint
your eyes, is just a net current.

depletion
region electron
hole carriers
carriers

p-type n-type

Figure 1

Notice that the number of electrons accumulated this way indicates the number of photons that
have struck the depletion region. In reality, not every photon is converted to a free charge, but the
conversion rate is still impressively high. This quantity is called the quantum efficiency η. It varies
with the wavelength of the incident light; at some wavelengths CCD cameras can have η > 90%.
A CCD is essentially a grid of photodiode-like detectors. Each such detector is one pixel. When
you take a photo with some fixed exposure time, the pixels will absorb some photons and the resulting
electrons will accumulate locally. After we’re done with the exposure, we want to read out the number
of electrons at each pixel. The electrons are moved along the grid by applying appropriate voltages.
They finally accumulate at a capacitor, where we get a voltage signal proportional to the charge. The
signals from all the pixels are then combined to give you the full image. This chain of events is similar
to human vision. Photons strike the photoreceptive cells in the retina, giving rise to neural signals.

1
The brain then processes these signals to create a first-order approximation of the outside world.
Now, the CCD gives you just a single number per pixel – the counts. This is the number
of recorded electrons times some arbitrary constant factor (called the gain) which depends on the
internals of the device. The information from the counts is sufficient to recover only a black-and-white
image. The way to get a colour image is to stack together shots of the object taken under a few
different filters. Indeed, if any given pixel has more counts in band B (450 nm) compared to band V
(550 nm) or band R (650 nm), you’d know to expect a bluish colour there.
There’s another approach available if you want your colour image immediately and don’t mind
cutting corners on image quality – it is to use a filter mosaic. Perhaps the most popular arrangement
is the Bayer filter, shown on Figure 2. Cover a quarter of the pixel array with tiny red filters, a
quarter with blue, and the remaining half with green. The space under any little filter will then carry
information for a single band only. However, we can still guess its expected counts in the other bands
using extrapolation from those neighbouring bits that do carry legitimate information for those bands.

Figure 2

That’s how everyday cameras work, and incidentally, this is also similar to what the eye does.
The cone cells in the retina come in three types, which are sensitive to red, green, and blue light,
respectively. The signals from these cells are weighted in the brain, and this gives us a sense of colour.

Example 1. Digitisation (Russia 2025). A telescope with aperture D = 20 cm and a focal

ratio A = 1/10 was used to take a picture of a star. The table below shows the counts at each
pixel of the CCD array.
(a) Find the pixel size d of the CCD array.
(b) Find the apparent magnitude of the star m.
(c) Find the night sky brightness µ (in mag/arcsec2 ), assuming that 60% of the background
counts are due to noise.
Assume that the detector’s gain is 1 e− /count. The quantum efficiency of the detector is η = 86%
and the exposure time is τ = 10 s. The losses in the optics are q = 10%. The astronomical
seeing (the diameter of the disk from a point source blurred due the atmosphere) is ϕ = 1.1′′ .
The photon flux from a star of magnitude zero is Φ0 = 106 cm−2 s−1 .

101 98 97 98 99 100 99
99 98 101 170 100 99 97
97 96 766 777 771 102 100
101 190 795 809 802 202 99
100 102 780 791 771 100 95
104 105 98 196 100 99 97
105 104 101 104 102 104 103

2
 Solution. (a) There are actually two reasons why the star isn’t seen as a point source. One
is the seeing, and the other is diffraction at the aperture. The diameter of the Airy disk is
θ = 2.44λ/D = 1.4′′ . Seeing and diffraction can be treated as two independent sources of
random error. Hence, the angular diameter of the star’s image can be estimated as
p
α = ϕ2 + θ2 = 1.8′′ .

The focal length of the telescope is F = D/A = 2 m. Looking at the counts, it seems that the
3A = 5.7 µm.
stellar disk’s diameter α corresponds to about 3 pixels. Thus F α = 3d, and d = Dα

(b) We need to sum the counts at the pixels that contain signal from the star, subtracting
the background (100 counts) from each pixel. We get N ′ = 70 + 666 + 677 + 671 + 90 +
695 + 709 + 702 + 102 + 680 + 691 + 671 + 96 = 6520 counts. The number of photons is then
N = N ′ /(1 − q)η = 8424. From this we can obtain the photon flux Φ = N/(πD2 /4)τ , and all
that is left is to compare with the standard source:
 
Φ
m − 0 = −2.5 lg ⇒ m = 13.9.
Φ0

(c) We’re told that 40% of the background is due to the night sky, meaning that there are
(0.4 × 100)/(0.9 × 0.86) = 52 incident photons per pixel. One pixel corresponds to (d/F )2 =
8.1 × 10−12 sr = 0.35 arcsec2 , so there are about 148 photons from 1 arcsec2 of the sky. We
divide this by πD2 /4 and τ to get the photon flux φ = 470 cm−2 s−1 arcsec−2 . Thus
 
φ
µ − 0 = −2.5 lg ⇒ µ = 18.3 mag/arcsec2 .
Φ0

2. Signal and noise

Given that the detector’s output depends on the photon count per pixel, it’s fairly intuitive that what
we call signal correlates with the number of incident photons rather than their total energy. Denoting
the photon flux per pixel by N (with units [s−1 ]), the total signal1 on n pixels after an exposure time
t is ηnN t.
However, there are a few effects because of which the counts registered by the detector won’t
match this expression exactly. First and foremost, the number of incident photons is subject to
deviations from nN t. When we say that the observed source has a constant photon flux (say, 4/s),
we mean that this is what we expect on average if we could sample the results of a trillion observing
sessions. In reality, any two photon emissions are uncorrelated, meaning that they don’t know about
each other, and they cannot possibly coordinate. So, the notion that you receive a photon, you wait
exactly 0.25 s, then you receive another photon, then wait 0.25 s again, is completely false. On average,
within a given second, you expect to receive 4 photons – that’s just how the source is – but it may occur
that you get 6 photons, or even 8, though this gets increasingly unlikely. Processes with independent
events and a constant mean rate obey the so-called Poisson distribution. Here, the probability that
you get k occurences when the value expected from the mean is λ will be given by the expression
λk e−λ
P (x = k) = .
k!
A plot of this is shown on Figure 3. This is all rather unintuitive and requires further explanation –
for example, where did that e even come from? If you’re interested in the nitty-gritty, I direct you to
Chapter 30.8 of [1], but now we need to make haste.
√ For our purposes the important thing is that Poisson distributions have a standard deviation of
λ. The standard deviation is a measure of the expected variation from the mean value, and it is
1
This is in electrons rather than counts because we don’t account for the gain here. But the gain just multiplies
everything by a constant factor, so it doesn’t really matter for what we’ll cover next.

3
used to indicate your error when you declare that the output of a Poisson process is indeed the mean.
If
√ there’s a Poisson process with a rate of 16/s running for one second, the standard deviation is
16 × 1 = 4. Hence we claim that we expect to observe 16 ± 4 events within this one second. This is
good enough – provided that λ is large,
√ we can show√ that there’s about 70% chance that the number
of observed events falls between λ − λ and λ + λ (here, 12 and 20).

Figure 3

In photography, deviations from the mean photon flux are problematic. What you actually want
to probe is the mean photon flux from each object, because that’s what really tells you how objects
compare in brightness (or, how the world really looks). But you can’t avoid the inherent variations.
We call these Poisson noise, or shot noise. If the expected signal is ηN nt, then the Poisson noise is
√
ηN nt. The relative error is thus
√
noise ηN nt 1
= =√ .
signal ηN nt ηN nt

This decreases with time, so a longer exposure means a more accurate image (and, of course, brighter
too). Here’s a comparison of different integration times:

Figure 4

4
 Now we’ll discuss another source of noise, namely dark current. Even when a photodiode isn’t
pointed to some specific source, it is still influenced by the background. As the device is kept at some
fixed temperature T , hole-electron pairs can form spontaneously due to thermal excitations. These
still lead to counts and can’t be distinguished from the electrons generated due to source photons.
We call the current due to such pairs dark current. We’ll denote it by D, the units being electrons
per second per pixel (quantum efficiency doesn’t play a part here). √ We see that on top of the useful
signal ηN nt we get an extra signal Dnt with standard deviation Dnt. The dark current grows
exponentially with temperature, so we want to keep our sciencey CCDs cool. This is done with liquid
nitrogen or a Peltier element. Conversely, in everyday CCD use exposure times are shorter and there
are lower demands on quality, so we don’t bother with cooling.
It is possible to remove the extra signal from the dark current in a picture by taking another pic-
ture with a closed shutter at the same CCD temperature. This is called a dark frame. By subtracting
the counts of the dark frame from the counts of the original picture, we get a “cleaned-up” signal.
However, the dark frame doesn’t exactly match the√dark currents in the original, because this is still
a Poisson process. Therefore, there’s an error term Dnt that we can’t do much about even with the
subtraction.
Apart from that, the CCD also has read-out noise. The electronics are not perfect, and they will
miscount the number of electrons with some standard error R (a fixed number per pixel). Generally,
faster read-out will bring about an increase in R. We’ll conclude with a couple of examples.

Example 2. Counting Photons (IPhO 2022 2B). The absorption of a photon by a CCD
camera leads to the emission of an electron within the apparatus. This occurs only if the photon
has sufficient energy to excite an electron across an energy gap ∆Eg . Assume that every photon
with sufficient energy succeeds. There is also leakage of electrons across the gap caused by the
temperature of the CCD camera; this is the dark current id , measured in number of electrons
per second. It depends on the temperature T according to

id = i0 e−|∆Eg |/6kB T ,

where i0 is a constant. The graph below shows how dark current varies with temperature for
the James Webb Space Telescope.

(a) Using the dark current graph, provide an order of magnitude estimate for the temperature
Ts of a distant source of thermal photons that would just be capable of exciting an electron
on the pixel.

5
The electrons are collected in a capacitor, and after an exposure time τ , the electrons are
counted. There are three main sources of uncertainty in the process: a fixed uncertainty in the
counting process called read-out noise; a Poisson distribution error associated with the dark
current, and a Poisson distribution error associated with the detected incoming photons.
(b) Write an expression for the total count uncertainty σt if there is a read-out noise σr , a
dark current id , an incoming photon rate p, and an exposure time τ .
The measured photon count is equal to the number of electrons in the capacitor, minus the
number of electrons associated with the dark current. From now on, assume the exposure time
is τ = 104 s and the read-out noise is a fixed σr = 14.
(c) Assume an operating temperature of Tp = 7.5 K. Calculate the minimum photon rate p
so that the photon count is ten times the count uncertainty.
(d) Assuming all photons are barely capable of exciting an electron across the band gap,
what is the intensity of the source of photons found in (c) on the primary mirror, which
has an area A = 25 m2 ? Express your answer in W/m2 .

Solution. (a) If you look on the exponents on the y-axis, you’ll notice that this is a graph of
lg id against 1/T . We know that
 
ln id ln i0 |∆Eg | 1
lg id = = − ,
ln 10 ln 10 6kB ln 10 T
|∆E |
so the slope of the graph corresponds to − 6kB lng10 ≡ a = −100 K. The photons from the distant
source have a typical energy of kB Ts , which should equal |∆Eg |. Thus

Ts = −6a ln 10 = 1380 K.

(b) We’re told in the problem statement that the quantum efficiency is 100%. There are three
√ √
separate sources of error – the Poisson noise pτ , the dark noise id τ , and the read-out noise
σr . The total error is obtained by adding them in quadrature, i.e.
p
σt = σr2 + (p + id )τ .

(c) Using the standard jargon, we want a signal-to-noise ratio (SNR) of 10:
pτ
SNR = p = 10.
σr2 + (p + id )τ

We use the graph to find that at Tp = 7.5 K the dark current is id = 5 s−1 . From then on,
finding pτ is a matter of solving the quadratic equation

(pτ )2 = 100(σr2 + (id τ ) + (pτ )).

We reach p = 0.23 s−1 .

(d) The energy of each photon is |∆Eg | = 1.9 × 10−20 J. The power on the mirror that gets
sent to the CCD is p|∆Eg |, and the intensity is

p|∆Eg |
I= = 1.8 × 10−22 W/m2 .
A
Here we assumed no losses in the optics.

6
Example 3. GOTO (IOAA 2017). The Gravitational-Wave Optical Transient Observer
(GOTO) aims to carry out searches of optical counterparts of any gravitational wave (GW)
sources within an hour of their detection by the LIGO and VIRGO experiments. The survey
needs to cover a large area on the sky in a short time so as to search all possible regions con-
strained by the GW experiments before the optical burst signal, if any, fades away. The GOTO
telescope array is composed of 4 identical reflective telescopes, each with 40-cm aperture and
an f -ratio of 2.5, working together to image large regions of the sky. For simplicity, we assume
that the telescopes’ fields-of-view (FoV) do not overlap with one another.
(a) Calculate the projected angular size per unit length at the focal plane, i.e. the plate
scale, of each telescope. Answer in arcmin/mm.
(b) If the zero-point magnitude (i.e. the magnitude at which the count rate detected at each
pixel is 1 count per second) of the telescope system is 18.5 mag, calculate the minimum
time needed to reach 21 mag at signal-to-noise ratio (SNR) = 5 for a point source. We’ll
initially assume that the noise is dominated by both the read-out noise (RON) at 10
counts/pixel and the CCD dark (thermal) noise (DN) rate of 1 count/pix/minute. The
CCDs used with GOTO have a 6-micron pixel size and a gain (conversion factor between
photo-electron and data count) of 1. The typical seeing at the observatory site is around
1.0 arcsec. Neglect diffraction.
(c) Normally, when the exposure time is long and the source count is high, Poisson noise
from the source is also significant. Determine the relation between SNR and exposure
time in the case where the noise is dominated by Poisson noise of the source.
Recalculate the minimum exposure time required to reach 21 mag with SNR = 5 from
(b) if Poisson noise is also taken into consideration apart from dark noise and read-out
noise. Note that the background sky brightness is also an important source of Poisson
noise, but you should neglect it in in your calculations.
(d) The typical localisation uncertainty of the GW detector is about 100 deg2 . We would like
to cover all possible locations of any candidate within an hour after the GW is detected.
Estimate the minimum side length of the square CCD needed for each telescope (in
number of pixels). You may assume that the time taken for the CCD read-out and the
aiming are negligible.

Solution. (a) The focal length of each telescope is f = 2.5 × 0.4 = 1 m. One mm on the CCD
1 mm 1
corresponds to an angle of 1000 mm = 1000 rad = 3.44 arcmin. The scale is s = 3.44 arcmin/mm.

(b) The seeing disk’s diameter ϕ = 1.0 arcsec corresponds to ϕ/s = 4.8 µm, which is less than
the size of a pixel (6 µm), so the star can fit within one pixel. Next, assume that the standard
source emits photons similar to those of our 21 mag source (so that the quantum efficiency of
the CCD is the same for both). Then, the count rate p, quantum efficiency included, can be
found from  
p
m − m0 = −2.5 lg ,
p0
 p 
21 − 18.5 = −2.5 lg ⇒ p = 0.1 s−1 .
1 s−1
Since we’re working with a single pixel, the total read-out noise is R = 10 and the total dark
current is D = 0.01667 s−1 . The equation for the SNR looks like this:
pτ
SNR = √ .
R2 + Dτ
Again, a quadratic equation with respect to τ . After some legwork, we get τ = 8.7 min.

7
 √
(c) The signal is pt, while the Poisson noise is pt. In the first part of this subtask there’s only
Poisson noise, so
pt √
SNR = √ ∝ t.
pt
We also need to redo part (b), this time accounting for Poisson noise as well:
pτ
SNR = p .
2
R + (D + p)τ

The answer is τ = 11.1 min.

(d) Each snapshot takes time τ = 11.1 min, and we have 60 min available in total. We can fit in
a maximum of 5 snapshots with each telescope. There are 4 telescopes, so we can take a total
2
of 20
p snapshots. Each of these needs to cover 5 deg . This corresponds to a picture with a side
of 5 deg2 = 2.2 deg = 134 arcmin. We divide this by the plate scale to find that the size of
the CCD should be at least 39 mm. Dividing by the pixel size, this is 6505 pixels.

There are two more issues that we’ll touch on briefly. First there’s the bias, which, like the
read-out noise, comes from the electronics. However, the bias is there on purpose. When we’re doing
read-out at the capacitor, the read-out noise might lead to a net negative voltage. The circuitry is
designed to handle positive voltages only, and it can’t convert such a signal to counts. To avoid this,
we apply a constant voltage offset for each pixel, which of course will add a few extra counts. But
because the bias is systematic rather than random, it’s not too difficult to remove it entirely. One
way to do this is to take a number of pictures (called bias frames) with an exposure of zero seconds.
That way you don’t even get a dark current, only read-out noise and the bias. After averaging2 the
bias frames, we’re left with just the bias, which can be promptly subtracted from our raw CCD image.
And yet, this whole procedure is redundant for most use cases. The reason is that dark frames also
contain bias, so subtracting an averaged-out dark frame gets rid of the bias too.
Finally, there’s the problem that a uniform signal on all pixels simply doesn’t produce an uniform
output, even if the error sources listed above didn’t exist. There are multiple reasons for this. The
quantum efficiency will vary pixel-to-pixel due to fabrication flaws; additionally, the optics may cause
uneven illumination on the detector, which can create undesirable effects like vignetting. Luckily, this
is all systematic and can be corrected for. The trick is to perform what is called a flat-field correction.
This involves taking an image of a uniform source, for example the sky background at dusk or dawn.
The output is termed a flat field. It will present as nonuniform owing to the fixed error pattern. Next,
for each pixel we divide the counts in the raw image by the counts in the flat field. This essentially
calibrates the raw image, because pixels that consistently overshoot are normalised by larger count
values, and vice versa. We then take the resulting quotients and scale them back by a constant, usually
the average count number within the flat field – after all, we want the end product to resemble the
raw image. In more detail, the procedure for each pixel is as follows:
raw − dark
corrected = · constant.
flat − dark
The dark frame subtraction is there because it’s better to get rid of everything except the fixed-pattern
errors before applying the flat-field correction.
Even if you find the bookkeeping of errors tedious (many do!), remember that this is a science in
and of itself, and it has enabled some extraordinary feats of precision measurement. Data analysis is
perhaps the core skill of a professional astronomer.

Last updated: 8th August 2025

Send comments and corrections to [Link]@[Link]

2
The jargon for this is “stacking”.

8
Problems
Problem 1. The eye as a camera (USAAAO 2025). Derive the limiting magnitude for naked-
eye visibility of a star in a dark-sky environment, by considering an analogy to a CCD camera. Assume
√
that the dominant noise source is n photon shot noise, and that “visibility” requires a signal-to-noise
ratio (SNR) > 1. You’ll need the following information:
• The dark-adapted human pupil dilates to an aperture of 6 mm. The quantum efficiency of the
rod cells used for night vision is around 5% (averaged across the full visible spectrum), and the
effective “exposure time” is roughly 30 ms.
• For a convenient magnitude reference, the intensity of sunlight within the visible spectrum is
5 × 102 W/m2 , and the apparent magnitude of the Sun is −26.7.
Approximate all visible photons as having wavelength λ = 500 nm when calculating photon energy.
This is near the peak sensitivity for human night vision.

Problem 2. Cluster photography (IOAA 2024). An astronomer takes pictures, in the V -band,
of a faint celestial target, from a place with no light pollution. The selected target is the globular
cluster Palomar 4, which has an angular diameter of θ = 72.0′′ and a uniform surface brightness in
the V -band of m = 20.6 mag/arcsec2 . The observation equipment consists of one reflector telescope,
with diameter D = 305 mm and f -ratio f /5, as well as a prime focus CCD with quantum efficiency
η = 80 % and square pixels with size l = 3.80 µm.

(a) Calculate the plate scale (the angle of sky projected per unit length of the sensor) of the
observation equipment in arcmin/mm.
(b) Estimate the number of pixels np covered by the cluster image on the CCD.
(c) Using an exposure time of t = 15 s, the astronomer obtains a signal-to-noise ratio SNR = 225.
Compute the brightness of the sky at the observation site, knowing that the CCD has a read-out
noise (standard deviation) R = 5 counts/px and dark noise of D = 6 counts/px/min. Answer
√
in mag/arcsec2 . Note that the total read-out noise is σRON = np · R.

Work with the following data:

• V -band central wavelength: λV = 550 nm
• V -band bandwidth: ∆λV = 88.0 nm
• Photon flux for a zero-magnitude object in V -band: 10 000 photons/nm/cm2 /s

Solutions
1. You can find the solution here.

2. You can find the solution here.

References
[1] K.F. Riley, M.P. Hobson, and S.J. Bence, Mathematical Methods for Physics and Engineering: A
Comprehensive Guide, 3rd edition, 2006

[2] Michael Newberry, CCD Camera Gain Calibration, 1996, [Link]

note_ccdgain.php