Scattering

Adding a Solar-Radiance
Function to the Hošek-Wilkie
Skylight Model
Lukáš Hošek and Alexander Wilkie ■ Charles University in Prague

O ne prerequisite for realistic renderings of

outdoor scenes is the proper capturing of
the sky’s appearance. Currently, an ex-
plicit simulation of light scattering in the atmo-
sphere isn’t computationally feasible, and won’t
and use a solar-radiance function that exactly
matches our recently published analytical skylight
model.2

Scattering in the Atmosphere

be in the foreseeable future. Captured luminance As light travels from the sun to the earth’s surface,
patterns have proven their usefulness in practice the only thing that can cause attenuation and de-
but can’t meet all user needs. (For more informa- flection is scattering—on either air molecules or
tion, see the “Representing the Sky and Atmo- larger aerosol particles. These scattering events are
sphere” sidebar.) why the sky isn’t black. So, any attempt to simulate
To fill this capability gap, com- light propagation in the atmosphere must handle
puter graphics technology has these two types of scattering.
The more sophisticated a
employed analytical models of To do this, we can construct a Monte Carlo sim-
skylight model is and the
sky-dome luminance patterns ulation (that is, a path tracer) of light transport.
more subtly it can represent for more than two decades. For With solid objects, we can describe a scene in terms
different atmospheric technical reasons, such models of scene geometry and surface bidirectional reflec-
conditions, the more the deal with only the sky dome’s tance distribution functions. Analogically, we can
direct solar radiance should appearance, though, and exclude describe a participating medium’s scattering prop-
exactly match the skylight’s the solar disc. The widely used erties by free flight paths (how long a ray must
conditions. Toward that end, model proposed by Arcot Pree­ travel before scattering) and phase functions (the
researchers propose a solar- tham and colleagues employed directional pattern in which light is scattered).
radiance function that exactly a separately derived analytical
matches a recently published formula for adding a solar emit- Rayleigh Scattering
ter of suitable radiant intensity.1 Rayleigh scattering describes the interaction be-
high-quality analytical skylight
Although this yields reasonable tween electromagnetic radiation and particles
model.
results, the formula is derived significantly smaller than the radiation’s wave-
in a manner that doesn’t exactly length3—in this case, visible light and the atmo-
match the conditions in their sky-dome model. sphere’s gas phase constituents. Rayleigh scatter-
But the more sophisticated a skylight model is and ing is elastic: no energy is transferred between the
the more subtly it can represent different condi- interacting particles. Translated into path-tracing
tions, the more the solar radiance should exactly terms, this means a ray’s energy doesn’t change,
match the skylight’s conditions. although its direction can.
In this article, we give an overview of analytical Rayleigh scattering is the main reason the sky is
sky-dome functions and show how to construct blue. (It’s also why the setting sun has a red tint, as

44 May/June 2013 Published by the IEEE Computer Society 0272-1716/13/$31.00 © 2013 IEEE
 Representing the Sky and Atmosphere

In a renderer, three basic options exist for including atmo-

spheric effects and a realistic sky dome.
Explicit Simulation
The first option is explicit simulation of light transport in
the entire atmosphere. This option isn’t currently practical
owing to excessive computational requirements and will
likely remain impracticable for the foreseeable future.
HDR Capture
The second option, high-dynamic-range (HDR) capture
of real outdoor environments, is very practical. Such envi-
ronment maps are sometimes called sky boxes because in
some renderers the HDR panoramas are mapped to a box
environment. However, genuinely hemispherical environ-
ments, often captured with fish-eye lenses, are probably
more common. Such environments have many uses, can
provide very high-quality image backdrops—for example,
for “beauty shot” architectural renderings—and are widely
used in industry.
Three problems can arise with this technology. First,
because the data is usually captured with a digital camera,
it isn’t spectral.
Second, cameras often can’t capture extreme lumi-
nance differences that well—the area around the sun is
often washed out, even in composite HDR images. For
the sun to be captured correctly, the environment map
would have to have a dynamic range of about 18 f-stops.
Finally, animations that involve the sun moving over
the sky (for example, to illustrate how the illumination of a
planned room changes throughout a typical winter’s day)
require an entire series of such images to be available. Of
course, you can capture such sequences at a given location
and specific time of the year if you know in advance that
they’ll be needed. However, finding such data on short
notice for specific conditions that might be of interest can
be fairly hard.
Analytical Models
For many projects and usage scenarios (especially artistic
ones), none of the previous problems is a concern. But
for scenarios in which one or more of those problems
matter, the third option—analytical models of sky-dome
luminance—is a welcome addition to the arsenal. Such
models have weaknesses compared to captures of reality;
most noticeably, inclusion of clouds is much harder to
achieve. However, for many usage cases in architecture,
this isn’t an issue.

we’ll show.) The effective cross-section of Rayleigh
scattering is strongly wavelength dependent, about
16 times higher for the violet end of the spectrum
than for the red end. This means that as violet and
blue rays travel through the atmosphere, they have
a much higher chance of being scattered toward
the observer.
We calculate a scattering event’s mean free
flight path as
1
= ,
N ⋅ S macro
where N is the number of scatterers in a unit
volume and Smacro is the macroscopic cross-section.
For Rayleigh scattering, the cross-section is
128π5a02 (6 + 3rn )
SRmacro = ,
3l 4 (6 − 7rn )
where a0 is the volume polarizability (for example,
1.7 ⋅ 10 –24cm3 for terrestrial air), rn is the depo-
larization factor (0.0279 for naturally polarized
light), and l is the wavelength. We assume that
the particle density decays exponentially with the
altitude, with an exponential density half-life of
t half = 7,994 m.
Once a Rayleigh scattering event occurs, the ray
is recast in a new direction, with the phase func-
tion governing the directional distribution. This
is important for implementing a sky-dome model.
However, in our solar-radiance model, we’re con-
cerned only with direct rays; we discard all rays
that undergo scattering.
In a simulation with just Rayleigh scattering
enabled, the sky color outside the solar disc is re-
stricted to shades of blue, with the hue correspond-
ing to that of a real sky on a cloudless day. The blue
hue is a combined effect of the sun’s spectrum and
the fourth-power dependence of the Rayleigh scat-
tering free flight path on the wavelength. When
the sun is at altitudes lower than approximately
10°, this dependence’s effect becomes directly vis-
ible on the solar disc. Red light is significantly less
likely to be out-scattered, which produces the dis-
tinctive red tint of a setting sun and a slight red-
dening of the sky above the horizon. Our reference
model reproduces all these effects.
Mie Scattering
Lorenz-Mie theory is a solution to James Maxwell’s
equation for dielectric spherical particles with a
diameter-to-wavelength ratio close to 1. It gives
us a tool to predict the contribution of particles
of sizes comparable to the wavelength of visible
light—most notably, water droplets. The interac-
tion is again elastic. Its phase function is strongly

IEEE Computer Graphics and Applications 45

Scattering
anisotropic, and Mie scattering strongly favors
light scattered in the forward direction, which
produces the characteristic glow around the sun.
Unlike Rayleigh scattering, in which the phase
function always has the same shape, the pattern of
Mie scattering and its macroscopic cross-section
varies with the droplet size and wavelength. Our
simulations used Ralph Eldridge’s droplet size
distributions.4 His paper also included meteoro-
logical range measurements, which we can use to
approximately calculate turbidity. Eldridge took
those measurements in a dense fog, corresponding
Our model yields different luminance
patterns based on the overall ground
albedo, which allows the sky-dome color to
match the overall surroundings’ ambience.
to a turbidity of approximately 60. For our simula-
tion runs with different turbidities, we scaled the
particle concentrations accordingly. (We discuss
turbidity more in the next section.)
Also unlike Rayleigh scattering, the free flight
path of Mie scattering isn’t strongly wavelength
dependent. For larger water droplets, the variation
in the phase function’s shape due to wavelength
changes can manifest itself as rainbow effects
caused by a slight shift in the secondary peak’s
position. However, this wasn’t an issue for the par-
ticle sizes we simulated (< 20 mm): our simulations
occurred in clear atmospheres containing no mac-
roscopic droplets.
Unfortunately, Mie scattering can’t be formu-
lated in closed form like Rayleigh scattering; both
effective cross-section and intensity can be de-
scribed only as an infinite expansion series. Our
simulations used precalculated tabulated values for
cross-sections. Again, we assumed density decays
exponentially with altitude, with t half = 1,200 m.
Using Turbidity to Describe
the Atmosphere
The Linke turbidity factor, or turbidity for short,
is a simple, intuitive measure of the air’s aerosol
content.5 It’s the ratio of the additional optical
thickness of the atmosphere in question, th, to the
optical thickness of an idealized atmosphere con-
sisting only of pure gas, t m:
t m + th
T= . continuity and brightness difference would pose a
tm
46 May/June 2013
This lets us easily define the sky’s appearance
without worrying too much about the intricacies
of meteorology. T = 2 yields a very clear, arctic-like
sky; T = 3 is a clear sky in a temperate climate; T =
6 is a sky on a warm, moist day; T = 10 is a slightly
hazy day; and T > 50 is dense fog. Of course,
flattening the diversity of all possible atmospheric
conditions into one number requires making
somewhat arbitrary decisions regarding particle
size distributions and the shape of the vertical
falloff of aerosol particle density. But because the
overall aerosol content is the dominant factor in
sky appearance, it’s still a highly useful parameter
that has emerged as the standard factor for
controlling sky-dome models.
The Hošek-Wilkie Sky-Dome Model
To derive the Hošek-Wilkie model for sky-dome
luminance, we first developed a path tracer.2 We
hard-coded the tracer to simulate light transport
in a simple scene with just one sphere that repre-
sents the earth and that’s surrounded by the par-
ticipating medium—the atmosphere. The reference
data obtained with the brute-force reference ren-
derer was fitted separately for each wavelength to
a modified sky-dome luminance formula based on
the approach used in the Preetham model.
To generate random free flight paths and scat-
tering events, the path tracer used delta scatter-
ing.6 This technique randomly shoots rays into
the hemisphere above an observation point on the
earth’s surface. Until a ray either gets absorbed
when hitting the earth or exits into free space,
a scattering event occurs on its path each time it
has traveled its free flight path distance. At each
scattering event, a new direction is generated ac-
cording to the scattering phase function, and the
ray is recast in that direction.
Using this approach, we generated a large set
of fish-eye view images of the sky. We used these
images as references in a nonlinear optimization
process to find the coefficients for the sky-dome
luminance functions.
We programmed the path tracer to discard rays
that don’t undergo any scattering events. In other
words, the reference images used for the sky-dome
luminance model didn’t include rays that directly
hit the sun. This, of course, means that the ana-
lytic model didn’t include the energy from these
direct hits on the sun, either. There’s a good rea-
son for this: the solar disk creates a very small
discontinuous area on the sky dome that can be up
to 18 stops brighter than its surroundings. The dis-
problem for the optimization’s stability. Also, the
 1.0

Limb-darkening attenuation
0.8

ψ 0.6

Sun center Ray

0.4 Wavelength
Short
γ Medium
0.2 Long
Camera

0.0 0.2 0.4 0.6 0.8 1.0

Solar-disc radius
(a) (b)

Figure 1. Limb darkening. (a) The geometry. y is the angle of the ray path with respect to the stellar radius;
g is the angle formed by the view direction and a vector pointing toward the sun. (b) The attenuation for
varying wavelengths across the solar disc’s radius. On the horizontal axis, 0 is the solar disc’s center and 1 is its
outer rim. Note the difference in residual brightness at the rim: blue light is attenuated twice as much as red
light. This leads to the orange tinge on the rim of the solar disc in Figure 2a.

solar disk’s radiance pattern exhibits properties
that would be hard to capture with the framework
used for calculating the sky dome’s distribution
parameters.
The resulting model provides genuine spectral
data for the entire range from the near ultraviolet
through the visible range. It reliably supports tur-
bidities between 1 and 10, whereas the widely used
Preetham model (see the “Related Work in Sky-
Dome Models” sidebar) functions properly only
between 2 and perhaps 5. Also, our model yields
different luminance patterns based on the overall
ground albedo, which allows the sky-dome color to
match the overall surroundings’ ambience. How-
ever, because the model doesn’t include the solar
disc, it doesn’t completely describe an outdoor
scene’s radiance environment.
Adding a Solar-Radiance Function
If we analyze how the atmosphere affects the ap-
pearance of objects in general and the sun and
sky in particular, an object’s appearance as an ob-
server perceives it is
L = L0 ⋅ t + Lin ,(1)
where L is intensity of the light that reaches the
observer, L 0 is the intensity of the light emitted
and reflected at the power source, t is the par-
ticipating media’s transmittance, and Lin is the
sun’s position on the sky dome. Given a fixed con-
figuration of atmospheric conditions, transmit-
tance for a fixed ray elevation will stay the same,
regardless of solar elevation or ground albedo. The
same path-tracing routine we used to obtain the
reference sky-dome images can serve with only
minor modifications to estimate the atmospheric
transmittance. We can just cast rays in one direc-
tion and calculate the ratio of rays that weren’t
scattered before they escaped the atmosphere.
Limb Darkening
Obtaining L 0 from the NASA tabulated data7
gives us only the spectral radiance at the solar
disc’s center. This information isn’t yet sufficient
to generate reasonably accurate renderings of the
sun because the sun also exhibits limb darkening.
Light emitted from the sun’s surface must travel
through its atmosphere, which attenuates light.
Rays coming from the solar disc’s edge must travel
a longer path almost parallel to the sun’s surface
in its atmosphere (see Figure 1a). This results in
the solar disc’s edges being noticeably darker than
its center. As Figure 1b shows, this effect’s inten-
sity is wavelength dependent, which leads to the
solar rim being redder than the center.
In astronomy, it’s common to use a simple ap-
proximate model to describe the phenomenon:
I (y ) N

in-scattered light’s intensity. Lin is modeled by I (0)

= ∑a
k=0
k cosk (y) ,(2)

the Hošek-Wilkie sky-dome model. For most of

the sky, that’s all we need—L 0 is zero because no
rays come directly from the darkness of the empty
space. But that changes when the sun is in the
way, and the solar-radiance model aims to fill the
gap represented by the missing L 0 ⋅ t term.
It’s noteworthy that t doesn’t depend on the
where I(y) is the radiant intensity at a point on
the solar disc along the ray path forming angle y
with respect to the stellar radius. I(0) is the radi-
ant intensity at the center point.
The coefficients ak are available in tabulated
form for wavelengths in the entire visible part

IEEE Computer Graphics and Applications 47

Scattering
Related Work in Sky-Dome Models
M odels of sky-dome luminance fall into two categories:
earlier luminance-only models and later models that
include color information.
Luminance-Only Models
As a first approximation, simpler models that describe only
the luminance distribution of a clear sky were devised, and
even standardized by the CIE (International Commission
on Illumination). These were intended for architecture and
lighting-design applications, such as assessing how much
light in addition to direct sunlight a given room would
receive if its windows faced a certain direction. Owing to
the lack of color information, the models can’t be directly
used for rendering, but they’re still useful for their original
purposes. They also served as starting points for some sky
models that include color information.
Richard Perez and his colleagues devised an analytic
formula for describing the luminance of clear skies with
low turbidity:1
FPerez (q , g ) = (1+ Ae B cos q )(1 = Ce Dg + E cos2 g ) .
The CIE adapted this formula in a slightly different form:
( ( ) )
FCIE2003 (q , g ) = (1+ Ae B cos q ) 1 = C e Dg − e D 2 + E cos2 g .
π
In these formulas, g is the angle formed by the view direc-
tion and a vector pointing toward the sun, q is the angle
formed by the zenith and view direction, and e is the Euler
number. (See Figure A.) The values returned by these formu-
las are normalized so that the value at the zenith is 1. Mul-
tiplying this value by the luminance at the zenith1 yields
the final luminance distribution function:
of the spectrum and the near ultraviolet and in-
frared parts.8 Our model uses a slightly adapted
form of those coefficients: we interpolated the
coefficients for the exact wavelengths used in the
Hošek-Wilkie model.
We calculate cos(y) as
sin2 g
cos (y) = 1 − , cent, compared to a solar disc that’s uniformly
sin2 0.25°
where g is the angle formed by the view direction
and a vector pointing toward the sun and 0.25° is
the sun’s angular radius as viewed from the earth.
As Figure 2 shows, the inclusion of limb darken-
ing has a perceptible and pleasantly realistic, if not
overly dramatic, impact on the solar disc’s appear-
ance. This might lead to the conclusion that it can
be safely omitted for images that don’t show close-
48 May/June 2013
S v
γ
Figure A. The coordinate system used in the Perez, CIE, and Hošek-
Wilkie sky-dome models. g is the angle formed by the view direction
(v) and a vector pointing toward the sun. q is the angle formed by the
zenith and v. qS denotes the angle formed by the sun and zenith.
F (q , g )
Y= ⋅ YZ ,
F (0, qS )
where qS denotes the angle formed by the sun and zenith
and YZ is the luminance at the zenith.
In the Perez and CIE formulas, the parameters A, C,
and E tune the luminance distribution and don’t directly
translate to any physical quantities. One way to obtain
these parameters is by using tabulated values. This was
done in the joint ISO/CIE standard, which comprises 16
models describing luminance distributions for atmospheric
conditions varying from clear sky to overcast with different
luminance turbidities.
ups of the solar disc. In practice, this means most
outdoor scenes illuminated by skylight because the
sun usually acts only as a light source and isn’t
directly visible.
But it’s crucial to realize that limb darkening
has an additional effect beyond the mere realism
of sunset images. It reduces the unattenuated so-
lar disc’s radiant power by approximately 23 per-
the center brightness provided by our analytical
formula. This has direct implications for usage of
our model, which we discuss later.
The Analytic Model
Like the sky-dome model, the solar-radiance model
provides independent data for each spectral band—
two ultraviolet bands (320 nm and 360 nm) and
nine bands in the visible spectrum (400 to 720 nm
in 40-nm increments).
 Models That Include Color Information
Tomoyuki Nishita and his colleagues proposed one of the
first colored skylight models that were directly useful for
rendering.2 It simulated the sky’s appearance due to single
scattering only and ignored interreflections between the
ground and air molecules. Nishita and his colleagues later
presented a revised model that used discretized precom-
putation to simulate multiple scattering.3
More recently, Jörg Haber and his colleagues presented
a physically based model.4 They followed the same basic
principle that Nishita and his colleagues used—subdivid-
ing the atmospheric body into discrete blocks and calcu-
lating the radiative transfer between them. Given the cor-
rect configuration of atmospheric constituents, this model
can realistically predict the sky’s appearance over a range
of atmospheric conditions. Originally, it didn’t account for
ground reflectance, although Haber and his colleagues
claimed that they could easily modify it to do so.
Oskar Elek and Petr Kmoch presented a model that
incorporates multiple scattering, including scattering in
water bodies, and that allows for real-time rendering by
using a set of lookup tables.5 It subdivides the atmospheric
body into discrete blocks and works in two phases. The
first, computationally intensive, step precomputes a 4D
lookup table that can be used to render images from
any position and with any solar position in real time. This
model is fully spectral and takes the ground albedo into
account. The ground albedo’s value must be determined
in advance because it’s baked into the lookup table.
The widely used Preetham model6 is based directly
on Perez and his colleagues’ model. But unlike the CIE
models, it calculates A, C, and E analytically. Arcot Pree­
tham and colleagues generated reference images using
Nishita’s model and then, by nonlinear optimization, fit
the Perez formulas to those images. The result was a set of
linear functions that take one parameter, turbidity, as an
argument. These functions return A, C, and E, as well as a
bicubic function of turbidity and the solar elevation angle
for calculating zenith luminance. The Preetham model also
provides two chroma channels that are calculated using
the same approach, and the authors provided a way to
convert the outputs to spectral-radiance data. The result-
ing model is fast and easy to implement and is considered
the de facto standard analytic model of spectral sky-dome
radiance.
References
1. R. Perez, R. Seals, and J. Michalsky, “All-Weather Model for
Sky Luminance Distribution—Preliminary Configuration and
Validation,” Solar Energy, vol. 50, no. 3, 1993, pp. 235–245.
2. T. Nishita et al., “Display of the Earth Taking into Account
Atmospheric Scattering,” Proc. Siggraph, ACM, 1993, pp.
175–182.
3. T. Nishita et al., “Display of Clouds Taking into Account Mul-
tiple Anisotropic Scattering and Sky Light,” Proc. Siggraph,
ACM, 1996, pp. 379–386.
4. J. Haber, M. Magnor, and H.-P. Seidel, “Physically Based
Simulation of Twilight Phenomena,” ACM Trans. Graphics,
vol. 24, no. 4, 2005, pp. 1353–1373.
5. O. Elek and P. Kmoch, “Real-Time Spectral Scattering in
Large-Scale Natural Participating Media,” Proc. 26th Spring
Conf. Computer Graphics (SCCG 10), ACM, 2010, pp. 83–90.
6. A.J. Preetham, P. Shirley, and B. Smits, “A Practical Ana-
lytic Model for Daylight,” Proc. Siggraph, ACM, 1999, pp.
91–100.

For each spectral band l and each integer tur-
bidity value T in the range [1, 10], we provide co-
efficients for a piece-wise cubic Hermite interpo-
lation polynomial function rλ,T(ϵ). This function
approximates the L0 ⋅ t term of Equation 1—that is,
transmittance for a given ray at elevation ϵ multi-
plied by the solar radiance from the NASA dataset
(solar radiance unattenuated by the earth’s atmo-
close to the horizon. At such elevations, the solar
radiance changes more dramatically, and an actual
radiance gradient can be seen on the solar disc.
Provided that ϵ falls into the ith segment (be-
tween xi–1 and xi), we read the corresponding coef-
ficients mλ,T,i,k from a table and calculate rλ,T(ϵ):
3
k

sphere or limb darkening). In other words, r(ϵ)

rλ,T (ϵ) = ∑m
k=0
λ,T,i,k ⋅ (ϵ − x i−1 ) .

gives us the direct solar radiance for a ray shot at

ϵ, assuming it would hit the sun’s center. To get a limb-darkened spectral-radiance value,
The function rλ,T(ϵ) consists of 45 segments with we multiply the output of r by the results of the
boundary points xi placed at formula from Equation 2. If we need values for
noninteger turbidities or wavelengths between the
π i provided channels, linear interpolation provides
xi = ⋅3 .
2 45 sufficient results.

The nonequidistant placement of segment bor- Model Usage in Practice

ders produces a finer resolution for ray elevations Because our solar-radiance function changes with

IEEE Computer Graphics and Applications 49

Scattering

(a) (b) (c) (d)

Figure 2. Direct views of the solar disc provided by our model for (a) a turbidity of 1 at an elevation of 42 degrees, (b) a turbidity
of 1 at 0.5 degree, (c) a turbidity of 6 at 0.5 degree, and (d) a turbidity of 7 at 0.25 degree (that is, when the sun touches the
horizon). The white solar disc in Figure 2a isn’t an anomaly: at higher elevations, the sun is practically not attenuated and has a
color that is close to D65, the white point of most imagery.

(a) (b) (c) (d) (e)

(f) (g) (h) (i) (j)

Figure 3. Sunrise: solar elevations of (a) 1, (b) 2, (c) 3, (d) 6, (e) 9, (f) 12, (g) 18, (h) 24, (i) 32, and (j) 42 degrees for a turbidity of
3. We tone-mapped the images to yield a sequence that approximately follows the brightness development of the sunrise. We
performed no white balance adjustment on the images, so that the changing sunlight color can be properly seen.

elevation, we must invoke it separately for the
varying elevations on a given solar disc in order
for a gradient to appear.
It’s also vitally important to add the output of
the original Hošek-Wilkie sky-dome model to this
function. This function captures only the direct
illumination, to which we must add the energy
from all photons that have scattered at least once.
Effectively, this is the in-scattered light in the area
of the solar disc, which is exactly the information
the original Hošek-Wilkie sky-dome model pro-
vides. Failing to take this into account (that is,
using only the direct solar radiance) results in a
much-too-dark solar disc at lower elevations, es-
pecially for higher turbidities.
As we mentioned earlier, owing to the nonuni-
form brightness distribution on the sun even at
high solar elevations, we must take care when de-
termining the solar disc’s overall radiant power.
In practice, we must average a sufficiently large
number of random samples on the disc to get rea-
sonable results. Owing to limb darkening, and at
lower solar elevations also to the nonlinear pro-
gression of attenuation, using just one sample at
the solar disc’s center yields significantly wrong
results. On average across wavebands, limb dark-
ening by itself induces an error of approximately
23 percent.
Owing to the solar-radiance function’s nonlin-
earity across elevations and wavelengths, finding a
simple linear correction factor for this effect isn’t
feasible. In future versions of the model, we plan
to include a specially fitted function for the energy
the solar disc emits for each solar elevation. In the

50 May/June 2013
 3.5

3.0
meantime, the only reliable way to determine the

Spectral radiance (×104)

solar disc’s energy is to take and average multiple 2.5
samples. However, because we only need to do this Sun
2.0 60°
once for a given solar elevation, it doesn’t create 24°
significant runtime overhead.
1.5 7°
Figure 3 shows the results of adjusting our
model for different solar elevations. 1.0

Results 0.5 1°

Figure 4 shows two examples of the progression

of the direct solar-radiance function we obtained 0
300 350 400 450 500 550 600 650 700 750
for each turbidity level. The color shift induced
(a) Wavelength
by Rayleigh scattering is clearly visible, as is the
expected reduction in direct solar radiance for de-
creasing solar elevations. Figure 5 illustrates how 3.5
changing turbidity affects the ratio between direct
3.0
sunlight and sky-dome radiance. Spectral radiance (×104)
Figure 6 shows close-up views of the sun for 2.5
the complete model—that is, combining the di- Sun
rect radiance and the luminance. Adding the two 2.0 60°
components provides a high-quality image of 24°
1.5
the solar disc with the proper horizontal gradi-
ent near the horizon. For higher solar elevations, 1.0 7°
the model also behaves as expected. However—in
accordance with nature—it’s less visually promi- 0.5
nent, in that it “just” provides a very bright solar
1°
disc with no discernible gradient. For high eleva- 0
tions, the model’s main benefit is that it automat- 300 350 400 450 500 550 600 650 700 750
ically provides a solar disc of correct brightness (b) Wavelength
for daytime outdoor skies.
You can download the fitted coefficient data and Figure 4. The direct solar radiance for turbidities of (a) 1 and (b) 4, for
a reference implementation of the solar-radiance different solar elevations. The spectrum labeled “Sun” is the original
function as part of the Hošek-Wilkie sky-dome unmodified solar radiation that hits the top of the atmosphere
model reference code package at http://cgg.mff. according to NASA data,1 with suitable data added for the ultraviolet
cuni.cz/projects/SkylightModelling. part of the spectrum. The color shift’s overall tendency remains the
same, but the solar radiance for the setting sun with low turbidity is

W
much stronger.
e plan to investigate the modeling of at-
mospheric perspective—how transmittance
and in-scattering affect the appearance of objects
in the scene. To do this, we might parameterize the
scattering and haze function that Preetham and
colleagues proposed.1 This parameterization’s goal
will be a modification that provides a haze color
that matches the surface albedo.
References
1. A.J. Preetham, P. Shirley, and B. Smits, “A Practical
Analytic Model for Daylight,” Proc. Siggraph, ACM,
1999, pp. 91–100.
2. L. Hošek and A. Wilkie, “An Analytic Model for Full
Spectral Sky-Dome Radiance,” ACM Trans. Graphics,
vol. 31, no. 4, 2012, article 95.
3. J.W. Strutt, “On the Scattering of Light by Small
Particles,” Philosophical Magazine, vol. 41, 1871, pp.
447–454.
4. R.G. Eldridge, “A Few Fog Drop-Size Distributions,”
J. Atmospheric Sciences, vol. 18, no. 5, 1961, pp.
671–676.
5. E.J. McCartney, Optics of the Atmosphere, J. Wiley &
Sons, 1976.
6. M. Raab, D. Seibert, and A. Keller, “Unbiased Global
Illumination with Participating Media,” Monte Carlo
and Quasi-Monte Carlo Methods 2006, A. Keller, S.
Heinrich, and H. Niederreiter, eds., Springer, 2008,
pp. 591–605.
7. K.L. Coulson, Solar and Terrestrial Radiation: Methods
and Measurements, Academic Press, 1975.
8. H. Neckel and D. Labs, “Solar Limb Darkening
1986–1990 (ll303 to 1099 nm),” Solar Physics, vol.
153, nos. 1–2, 1994, pp. 91–114.

IEEE Computer Graphics and Applications 51

Scattering

(a) (b) (c) (d) (e)

(f) (g) (h) (i) (j)

Figure 5. The results of our model for a solar elevation of 8 degrees and turbidities of (a) 1, (b) 2, (c) 3, (d) 4, (e) 5, (f) 6, (g) 7, (h)
8, (i) 9, and (j) 10. Images are tone-mapped to appear roughly equally bright. As with Figure 3, we performed no white-balance
computations.

Solar elevation
0° 0.1° 0.2° 0.3° 0.4° 0.5° 0.6° 0.7° 0.8° 0.9° 1°
5
6
Turbidity
7 8
9

Figure 6. Renderings of the region immediately around the rising or setting sun, done with the superposition of the Hošek-Wilkie
sky-dome model and the proposed solar-radiance function. The small images in the image set are at the same exposure level.
From left to right, the sun starts exactly at the horizon; the solar elevation progresses in 0.1. increments upward. The turbidity
ranges from 5 (top) to 9 (bottom). We omitted lower turbidities because at this exposure level, they would appear as a uniform
white circle. We omitted a turbidity of 10 because the sun wouldn’t be visible through the haze.

Lukáš Hošek is a computer graphics developer at Bohemia
Interactive Simulations and a PhD candidate at Charles
University in Prague. His research interests include real-
time computer graphics and atmospheric rendering. Hošek
received a master’s in computer science from Charles Uni-
versity. Contact him at hosek@cgg.mff.cuni.cz.
Alexander Wilkie is an associate professor of computer
graphics at Charles University in Prague. His research in-
terests are predictive rendering and color science. Wilkie
received a PhD in computer science from the Vienna Univer-
sity of Technology. He’s a member of Siggraph and a senior
member of IEEE. Contact him at wilkie@cgg.mff.cuni.cz.
Selected CS articles and columns are also available
for free at http://ComputingNow.computer.org.

52 May/June 2013