Content uploaded by Jerry Tessendorf

Author content

All content in this area was uploaded by Jerry Tessendorf on Sep 24, 2018

Content may be subject to copyright.

Simulating Ocean Water

Jerry Tessendorf

Copyright c

1999 – 2004, Jerry Tessendorf

1 Introduction and Goals

These notes are intended to give computer graphics programmers

and artists an introduction to methods of simulating, animating, and

rendering ocean water environments. CG water has become a com-

mon tool in visual effects work at all levels of computer graphics,

from print media to feature ﬁlms. Several commercial products are

available for nearly any computer platform and work environment.

A few visual effects companies continue to extend and improve

these tools, seeking to generate higher quality surface geometry,

complex interatctions, and more compelling imagery. In order for

an artist to exploit these tools to maximum beneﬁt, it is important

that he or she become familiar with concepts, terminology, a little

oceanography, and the present state of the art.

As demonstrated by the pioneering efforts in the ﬁlms Water-

world and Titanic, as well as several other ﬁlms made since about

1995, images of cg water can be generated with a high degree of

realism. However, this level of realism has been mostly limited

to relatively calm, nice ocean conditions. Conditions with large

amounts of spray, breaking waves, foam, splashing, and wakes are

improving and approaching the same realistic look.

Three general approaches are currently popular in computer

graphics for simulating ﬂuid motion, and water surfaces in partic-

ular. The three methods are all related in some way to the basic

Navier-Stokes equations at the heart of many applications. Com-

putational Fluid Dynamics (CFD) is one of the methods which has

recieved a great amount of attention lately. While many versions of

this technology have been in existence for some time, recent papers

by Stam [1], Fedkiw and Foster [2], and many others have demon-

strated that the numerical computation discipline and the computer

graphics discipline have connected well enough to produce useful,

interesting, and sometimes beautiful results. The primary draw-

back of CFD methods is that the computations are performed on

data structured in a 2D or 3D grid, called an Eulerian framework.

This gridded architecture limits the combined extent and ﬂow detail

that can be computed. At present for example, it is practical to sim-

ulate with CFD waves breaking as they approach a shallow beach.

However, that simulation is not able to simulate a bay-sized region

of ocean while also simulating the breaking down to the detail of

spray formation, simply because the number of grid points needed

is prohibitively large.

A second method that is shows great promise is called Smoothed

Particle Hydrodynamics (SPH). This is a completely different ap-

proach to solving the Navier-Stokes equations. SPH imagines that

the volume of a ﬂuid is composed of small overlapping regions, the

center of each region carrying some amount of mass and momen-

tum. The regions are allowed to move about within the ﬂuid and ex-

perience forces due to pressure, strain, gravity, and others. But now

the center of each region acts as a particle, and the Navier-Stokes

equations are converted into equations of motion for discrete par-

ticles. This approach is called a Lagragian framework (as opposed

to the Eulerian framework in CFD). The ﬂuid volume is not bound

to a grid geometry. SPH is a very useful method of simulation for

situations in which there is signiﬁcant splashing or explosions, and

has even been used to simulated cracking in solids [3]. Using im-

plicit methods to construct a surface for the ﬂuid, standard computer

graphics applications such as pouring water have been achieved [4].

Finally, the third method is the one that is the focus of these

notes, and unlike the CFD and SPH methods, this one is focused

on the more narrow goal of simulating the motion of the surface

of a body of water. Surface simulations are commonly generated

from the CFD and SPH methods, but the surface is generated by

algorithms that are added to those ﬂuid simulations, for example by

tracking a type of implicit surface called a level set. In choosing

to focus on the surface structure and motion, we eliminate much of

the computation and resolution limits in the CFD and SPH meth-

ods. Of course when we eliminate those computations we have lost

certain types of realistic motion, most notably the breakup of the

surface with strong changes in topology. In place of that compu-

tation we substitute a mixture of knowledge of oceanographic phe-

nomenology and computational ﬂexibility to achieve realistic types

of surfaces that cannot practically be achieved by the others. In that

sense, this phemonological approach should be considered comple-

mentary – not competitive – to the CFD and SPH methods, since

all three work best in different regimes of ﬂuid motion.

Broadly, the reader should come away from this material with

1. an understanding of the important physical concepts for ocean

surface propagation, most notably the concept of dispersion

and types of dispersion relationships.

2. an understanding of some algorithms that generate/animate

water surface height ﬁelds suitable for modeling waves as big

as storm surges and as small as tiny capillaries;

3. an understanding of the basic optical processes of reﬂection

and refraction from a water surface;

4. an introduction to the color ﬁltering behavior of ocean water;

5. an introduction to complex lighting effects known as caus-

tics and godrays, produced when sunlight passes through the

rough surface into the water volume underneath; and

6. some rules of thumb for which choices make nice looking

images and what are the tradeoffs of quality versus compu-

tational resources. Some example shaders are provided, and

example renderings demonstrate the content of the discussion.

Before diving into it, I ﬁrst want to be more concrete about what

aspect of the ocean environment we cover (or not cover) in these

notes. Figure 1 is a rendering of an oceanscape produced from mod-

els of water, air, and clouds. Light from the clouds is reﬂected from

the surface. On the extreme left, sun glitter is also present. The

generally bluish color of the water is due to the reﬂection of blue

skylight, and to light coming out of the water after scattering from

the volume. Although these notes do not tackle the modeling and

rendering of clouds and air, there is a discussion of how skylight

from the clouds and air is reﬂected from, or refracted through, the

water surface. These notes will tell you how to make a height-ﬁeld

displacement-mapped surface for the ocean waves with the detail

and quality shown in the ﬁgure. The notes also discuss several ef-

fects of the underwater environment and how to model/render them.

The primary four effects are sunbeams (also called godrays), caus-

tics on underwater surfaces, blurring by the scattering of light, and

color ﬁltering.

There are also many other complex and interesting aspects of the

ocean environment that will not be covered. These include break-

ing waves, spray, foam, wakes around objects in the water, splashes

from bodies that impact the surface, and global illumination of the

entire ocean-atmosphere environment. There is substantial research

underway on these topics, and so it is possible that future versions

of this or other lecture notes will include them. I have included a

brief section on advanced modiﬁcations to the basic wave height al-

gorithm that produce choppy waves. The modiﬁcation could feasi-

bly lead to a complete description of the surface portion of breaking

waves, and possibly serve to drive the spray and foam dynamics as

well.

There is, of course, a substantial body of literature on ocean sur-

face simulation and animation, both in computer graphics circles

and in oceanography. One of the ﬁrst descriptions of water waves in

computer graphics was by Fournier and Reeves[12] , who modeled

a shoreline with waves coming up on it using a water surface model

called Gerstner waves. In that same issue, Darwin Peachey[13]

2 RADIOSITY OF THE OCEAN ENVIRONMENT 3-3

Figure 1: Rendered image of an oceanscape.

presented a variation on this approach using basis shapes other than

sinusoids.

In the oceanographic literature, ocean optics became an inten-

sive topic of research in the 1940s. S.Q. Duntley published[17] in

1963 papers containing optical data of relevance to computer graph-

ics. Work continues today. The ﬁeld of optical oceanography has

grown into a mature quantitative science with subdisciplines and

many different applications. One excellent review of the state of

the science was written by Curtis Mobley[18].

In these lectures the approach we take to creating surface waves

is close to the one outlined by Masten, Watterberg, and Mareda[11],

although the technique had been in use for many years prior to

their paper in the optical oceanography community. This approach

synthesizes a patch of ocean waves from a Fast Fourier Transform

(FFT) prescription, with user-controllable size and resolution, and

which can be tiled seamlessly over a larger domain. The patch con-

tains many octaves of sinusoidal waves that all add up at each point

to produce the synthesized height. The mixture of sinusoidal am-

plitudes and phases however, comes from statistical, emperically-

based models of the ocean. What makes these sinusoids look like

waves and not just a bunch of sine waves is the large collection of

sinusoids that are used, the relative amplitudes of the sinusoids, and

their animation using the dispersion relation. We examine the im-

pact of the number of sinusoids and resolution on the quality of the

rendered image.

In the next section we begin the discussion of the ocean environ-

ment with a broad introduction to the global illumination problem.

The radiosity equations for this environment look much like those

of any other radiosity problem, although the volumetric character of

some of the environmental components complicate a general imple-

mentation considerably. However, we simplify the issues by ignor-

ing some interactions and replacing others with models generated

by remote sensing data.

Practical methods are presented in section 4 for creating realiza-

tions of ocean surfaces. We present two methods, one based on a

simple model of water structure and movement, and one based on

summing up large numbers of sine waves with amplitudes that are

related to each other based on experimental evidence. This sec-

ond method carries out the sum using the technique of Fast Fourier

Transformation (fft), and has been used effectively in projects for

commercials, television, and motion pictures.

After the discussion of the structure and animation of the water

surface, we focus on the optical properties of water relevant to the

graphics problem. First, we discuss the interaction at the air-water

interface: reﬂection and refraction. This leaves us with a simple

but effective Renderman-style shader suitable for rendering water

surfaces in BMRT, for example. Next, the optical characteristics of

the underwater environment are explored.

Finally, please remember that these notes are a living

document. Some of the discussion of the various top-

ics is still very limited and incomplete. If you ﬁnd a

problem or have additional questions, please feel free to

contact me at jerry@ﬁnelightvisualtechnology.com. The

latest version of this course documents are hosted at

http://www.ﬁnelightvisualtechnology.com

2 Radiosity of the Ocean Environment

The ocean environment, for our purposes, consists of only four

components: The water surface, the air, the sun, and the water

volume below the surface. In this section we trace the ﬂow of

light through the environment, both mathematically and schemat-

ically, from the light source to the camera. In general, the radiosity

equations here are as coupled as any other radiosity problem. To a

reasonable degree, however, the coupling can be truncated and the

simpliﬁed radiosity problem has a relatively fast solution.

The light seen by a camera is dependent on the ﬂow of light en-

ergy from the source(s) (i.e. the sun and sky) to the surface and

into the camera. In addition to specular reﬂection of direct sun-

light and skylight from the surface, some fraction of the incident

light is transmitted through the surface. Ultimately, a fraction of the

transmitted light is scattered by the water volume back up through

the interface and into the air. Some of the light that is reﬂected or

refracted at the surface may strike the surface a second time, pro-

ducing more reﬂection and refraction events. Under some viewing

conditions, multiple reﬂections and refractions can have a notice-

able impact on images. For our part however, we will ignore more

than one reﬂection or refraction from the surface at a time. This

not only makes the algorithms and computation easier and faster,

but also is reasonably accurate in most viewing conditions and pro-

duces visually realistic imagery.

At any point in the environment above the surface, including at

the camera, the total light intensity (radiance) coming from any di-

rection has three contributions:

LABOV E =rLS+rLA+tULU,(1)

with the following deﬁnitions of the terms:

ris the Fresnel reﬂectivity for reﬂection from a spot on the surface

of the ocean to the camera.

tUis the transmission coefﬁcient for the light LUcoming up from

the ocean volume, refracted at the surface into the camera.

LSis the amount of light coming directly from the sun, through

the atmosphere, to the spot on the ocean surface where it is

reﬂected by the surface to the camera.

LAis the (diffuse) atmospheric skylight

LUis the light just below the surface that is transmitted through

the surface into the air.

Equation 1 has intentionally been written in a shorthand way that

hides the dependences on position in space and the direction the

light is traveling.

While equation 1 appears to have a relatively simple structure,

the terms LS,LA, and LUcan in principle have complex depen-

dencies on each other, as well on the reﬂectivity and transmissivity.

3 MATHEMATICS, PHYSICS, AND EXPERIMENTS ON THE MOTION FOR THE SURFACE 3-4

There is a large body of research literature investigating these de-

pendencies in detail [19], but we will not at this point pursue these

quantitative methods. But we can elaborate further on the coupling

while continuing with the same shorthand notation. The direct light

from the sun LSis

LS=LT OA exp{−τ},(2)

where LT OA is the intensity of the direct sunlight at the top of the

atmosphere, and τis the “optical thickness” of the atmosphere for

the direction of the sunlight and the point on the earth. Both the

diffuse atmospheric skylight LAand the upwelling light LUcan be

written as the sum of two terms:

LA=L0

A(LS) + L1

A(LU)(3)

LU=L0

U(LS) + L1

U(LA)(4)

These equations reveal the potential complexity of the problem.

While both LAand LUdepend on the direct sunlight, they also

depend on each other. For example, the total amount of light pene-

trating into the ocean comes from the direct sunlight and from the

atmospheric sunlight. Some of the light coming into the ocean is

scattered by particulates and molecules in the ocean, back up into

the atmosphere. Some of that upwelling light in turn is scattered in

the atmosphere and becomes a part of the skylight shining on the

surface, and on and on. This is a classic problem in radiosity. It

is not particularly special for this case, as opposed to other radios-

ity problems, except perhaps for the fact that the upwelling light

is difﬁcult to compute because it comes from volumetric multiple

scattering.

Our approach, for the purposes of these notes, to solving this

radiosity problem is straightforward: take the skylight to depend

only on the light from the sun, since the upwelling contribution

represents a “tertiary” dependence on the sunlight; and completely

replace the equation for LUwith an empirical formula, based on

scientiﬁc observations of the oceans, that depends only on the di-

rect sunlight and a few other parameters that dictate water type and

clarity.

Under the water surface, the radiosity equation has the schematic

form

LBELO W =tLD+tLI+LSS +LM,(5)

with the meaning

tis the Fresnel transmissivity for transmission through the water

surface at each point and angle on the surface.

LDThe “direct” light from the sun that penetrates into the water.

LIThe “indirect” light from the atmosphere that penetrates into

the water.

LSS The single-scattered light, from both the sun and the atmo-

sphere, that is scattered once in the water volume before ar-

riving at any point.

LMThe multiply-scattered light. This is the single-scattered light

that undergoes more scattering events in the volume.

Just as for the above water case, these terms are all related to each

other is relative complex ways. For example, the single scattered

light depends on the direct and indirect light:

LSS =P(tLI) + P(tLD)(6)

with the quantity Pbeing a linear functional operator of its argu-

ment, containing information about the single scattering event and

the attenuation of the scattered light as it passes from the scatter

Figure 2: Illustration of multiple reﬂections and transmission

through the air-water interface.

point to the camera. Similarly, the multiply-scattered light is de-

pendent on the single scattered:

LM=G(tLI) + G(tLD).(7)

The functional schematic quantities Pand Gare related, since mul-

tiple scattering is just a series of single scatters. Formally, the two

have an operator dependence that has the form

G∼P⊗P⊗n1 + P+1

2! P⊗P+1

3! P⊗P⊗P+...o

∼P⊗P⊗exp(P).(8)

At this point, the schematic representation may have outlived its

usefullness because of the complex (and not here deﬁned) meaning

of the convolution-like operator ⊗, and because the expression for

Gin terms of Phas created an even more schematic view in terms

of an exponentiated P. So for now we will leave the schematic

representation, and journey on with more concrete quantities the

rest of the way through.

The formal schematic discussion put forward here does have a

mathematically and physically precise counterpart. The ﬁeld of

study in Radiative Transfer has been applied for some time to wa-

ter optics, by a large number researchers. The references cited are

excellent reading for further information.

As mentioned, there is one additional radiosity scenario that can

be important to ocean rendering under certain circumstances, but

which we will not consider. The situation is illustrated in ﬁgure 2.

Following the trail of the arrows, which track the direction light

is travelling, we see that sometimes light coming to the surface

(from above or below), can reﬂect and/or transmit through the sur-

face more than once. The conditions which produce this behavior

in signiﬁcant amounts are: the wave heights must be fairly high,

and the direction of viewing the waves, or the direction of the light

source must be nearly grazing the surface. The higher the waves

are, the less grazing the light source or camera need to be. This

phenonmenon has been examined experimentally and in computer

simulations. It is reasonably well understood, and we will ignore it

from this point on.

3 Mathematics, Physics, and Experi-

ments on the Motion for the Surface

In this section we take a look at the mathematical problem we are

trying to solve. We simplify the mathematics considerably by ap-

plying a series of approximations. How do we know these approx-

imations are any good? There are decades of oceanographic re-

search in which the ocean surface motion has been characterized

3 MATHEMATICS, PHYSICS, AND EXPERIMENTS ON THE MOTION FOR THE SURFACE 3-5

by measurements, simulations, mathematical analysis, and experi-

mentation. The approximations we apply in this section are not per-

fect, and there are many circumstances in the real world in which

they break down. But they work extraordinarily well for most con-

ditions at sea. To give you some idea of how well they work, we

show some experimental work in section 4 that has been done which

clearly shows these approximations at work in the real world.

3.1 Bernoulli’s Equation

The starting point of the mathematical formulation of ocean surface

motion is the incompressible Navier-Stokes equations. Chapter 2

of [14] provides a thorough derivation of Bernoulli’s equation from

Navier-Stokes. We provide here the short version, and refer you to

Kinsman’s or some other textbook for details.

The incompressible Navier-Stokes equations for the velocity

u(x, t)of a ﬂuid a any position xat any time tare

∂u

∂t +u· ∇ u=−∇p+F(9)

∇ · u(x, t)=0 (10)

In these equations, p(x, t)is the pressure on the ﬂuid, and F(x, t)

is the force applied to the ﬂuid. In our application, the force is

conservative, so F=−∇Ufor some potential energy function

U(x, t).

For ocean surface dynamics due to conservative forces like grav-

ity, it turns out to be worthwhile to restrict the type of motion to a

class called potential ﬂow. This is a situation in which the velocity

has the form of a gradient:

u(x, t) = ∇φ(x, t)(11)

This restriction has the important effect of reducing the number de-

grees of freedom of the ﬂow from the three components of velocity

uto the single one of the potential velocity function φ. In fact, us-

ing this gradient form, the four Navier-Stokes equations transform

into the two equations

∂φ

∂t +1

2(∇φ)2=−p−U(12)

∇2φ(x, t)=0 (13)

Equation 12 is called Bernoulli’s equation. As a fully nonlinear

reduction of the Navier-Stokes equations, Bernoulli’s equation is

capable of simulating a variety of surface dynamics effects, includ-

ing wave breaking in shoaling shallow water (the bottom rises from

deep water up to a beach). For more detail on numerical simulations

of Bernoulli’s equation in 3D, see [23].

3.2 Linearization

For our purposes, we want to reduce the complexity of Bernoulli’s

equation even further by applying two restrictions: linearize the

equations of motion, and limit evaluation of the equations to just

points on the surface itself, ignoring the volume below the sur-

face. This may seem like an extreme restriction, but when com-

bined with some phenomenolical knowledge of the ocean, this re-

strictions work very well.

The ﬁrst restriction is to linearize Bernoulli’s equation. This is

simply the task of removing the quadratic term 1/2(∇φ)2. Elimi-

nating this term means that we are most likely restricted to surface

waves that are not extremely violent in their motion, at least in prin-

ciple. So Bernoulli’s equation is reduced further to

∂φ

∂t =−p−U(14)

All of the quantities φ,p, and Uare still evaluated at 3D points x

on the surface and in the water volume.

The second restriction is to evaluation quantities only on the wa-

ter surface. To do this we have to ﬁrst characterize what we mean by

the surface. We will take the surface to be a dynamically changing

height ﬁeld, h(x⊥, t), that is a function of only the horizontal posi-

tion x⊥and time t. For convenience, we deﬁne the mean height of

the wate surface as the zero value of the height. With this deﬁnition

of wave height, the gravity-induced potential energy term Uis

U=g h (15)

and gis the gravity constant, usually 9.8m/sec2in metric units.

Restricting to just the water surface has several important conse-

quences. One of the ﬁrst consequences is for mass conservation. In

the incompressible Navier-Stokes equation, mass is conserved via

the mass ﬂux equation

∇ · u(x, t) = 0 (16)

When we chose to consider only potential ﬂow, this mass conserva-

tion equation became

∇2φ(x, t) = 0 (17)

If we label the horizontal portion of the position vector as x⊥, so

that x= (x⊥, y), and yis the coordinate pointing down into the

water volume, then the mass conservation equation restricted to the

surface looks in more detail like

∇2

⊥+∂2

∂y2φ(x⊥, t) = 0 (18)

Now, when you look at this equation and see that φnow depends

only on the x⊥on the surface, you might be tempted to throw out

the ∂2/∂y2part of the equation, because there does not appear to

be a dependence. That would produce useless results. Instead, what

works better in this odd world of partial differential equations is to

allow φto be an arbitrary function (at least with respect to this mass

conservation equation) and to deﬁne the y-derivative operator to be

∂

∂y =±p−∇2

⊥(19)

so that the operator ∇2is zero. We will use this approach for any

quantity evaluated on the water surface whenever we need a verti-

cal derivative. Of course, this introduces an unusual operator that

contains a square root function.

Another consequence of restricting ourselves to just the surface

is that the pressure remains essentially constant, and we can choose

to have that constant be 0. With this and the rest of the restrictions,

Bernoulli’s equation has been linearized to

∂φ(x⊥, t)

∂t =−g h(x⊥, t)(20)

There is one ﬁnal equation that must be rewritten for this situ-

ation. Recall that the velocity potential φis used to compute the

3D ﬂuid velocity as a gradient, u=∇φ. The vertical component

of the velocity must now use equation 19. In addition, the vertical

velocity of the ﬂuid is the same at the speed of the surface height.

Combining these we get

∂h(x⊥, t)

∂t =p−∇2

⊥φ(x⊥, t)(21)

These last two equations, 20 and 21, are the ﬁnal equations of

motion that are needed to solve for the surface motion. They can

3 MATHEMATICS, PHYSICS, AND EXPERIMENTS ON THE MOTION FOR THE SURFACE 3-6

also be converted into a single equation. For example, if we take a

derivative with respect to time of equation 21, and use 20 to sub-

stitute for the time derivative of the velocity potential, we get the

single equation for the evolution of the surface height.

∂2h(x⊥, t)

∂t2=−gp−∇2

⊥h(x⊥, t)(22)

This still involves the unusual operator p−∇2

⊥. However, tak-

ing two more time derivatives converts it to a more normal two-

dimensional Laplacian for the equation

∂4h(x⊥, t)

∂t4=g2∇2

⊥h(x⊥, t)(23)

This form is frequently the starting point for building mathematical

solutions to the surface wave equation.

3.3 Dispersion

It turns out that the equations we have built for surface height –

whether in the form of equations 20 and 21, or equation 22, or equa-

tion23 – reduce to one primary lesson about surface wave propaga-

tion. This lesson is embodied in a simple mathematical relationship

called the Dispersion Relation, which is the focus of this section.

Our goal here is to obtain that simple expression from the math-

ematics above, understand some of its meaning, and demonstrate

that, even though it appears unrealistically simplistic, the Disper-

sion Relation is in fact present in natural ocean waves and can be

measured experimentally.

Lets just use the version of the surface evolution equation in

equation 23 for convenience. The other two versions could be used

and arrive at the same answer using a slightly different set of ma-

nipulations. Note that the equation of motion for the surface height

is linear in the surface height. So as with any linear differential

equation, the general solution of the equation is obtained by adding

up any number of speciﬁc solutions. So lets ﬁnd a speciﬁc solution.

It turns out that all speciﬁc solutions have the form

h(x⊥, t) = h0exp {ik·x⊥−iωt}(24)

The 2D vector kand the numbers ωand h0are generic parameters

at this point. If we use this form of a solution in equation 23, it

turns into an algebraic equation like this:

h0ω4−g2k2= 0 (25)

( and kis the magnitude of the vector k). For this solution, there

are only two possibilities:

1. h0= 0. Then the surface height is ﬂat, and the solution is not

very interesting.

2. ω=±√gk and h0can be anything. This is the interesting

solution.

What we have found here is that the entire Navier-Stokes ﬂuid dy-

namics problem, reduced to an evolution equation for the water sur-

face and approximated to something that can be solved relatively

easily, amounts to a single equation imposing the constraint that the

temporal frequency ωof surface height movements is connected to

the spatial extent of a the propagating wave k=|k|. This relation-

ship, ω=±√gk is the Dispersion Relation mentioned earlier.

Note in particular that there is no constraint placed on the ampli-

tude h0. But if the Navier-Stokes equation does not have anything

to say about the amplitude, how do we give it a value? One way

is by imposing initial conditions on the height and on its vertical

speed. For ocean surface simulation in the next section, we will use

Figure 3: A slice through a 3D PSD showing that the observed

wave energy follows the deep water dispersion relation very well.

an alternate method, a statistical procedure to generate random real-

izations of the amplitude, guided by measurements of the variance

properties of wave height on the open ocean.

So the question remaining is just how reasonable is the disper-

sion relation for modelling realistic ocean surface waves? This is

where lots of experimental research can come into play. Although

there have been many decades of research on ocean wave proper-

ties using devices placed in the water to directly measure the wave

motion at a point, here we look at some relatively new research that

involves measuring wave properties remotely with a camera in a

plane.

The AROSS [24] is a panchromatic camera mounted in a special

hosing on the nose of a small airplane. Attached to the camera is

navigation and GPS instrumentation which allow the camera po-

sition, viewing direction, and orientation to be measured for each

frame. After the plane ﬂys a circular orbit around a spot over the

ocean, this data can be used to remap images of the ocean into a

common reference frame, so that the motion of the aircraft has been

removed (except for lighting variations). This remapping allows

the researchers to use many frames of ocean imagery, typically 1-2

minutes worth, in some data processing to look for the dispersion

relation.

The data processing that AROSS imagery is subjected to gen-

erates something called a 3D Power Spectral Density (PSD). This

is obtained by taking the Fourier Transform of a time series of im-

agery in time, as well as Fourier Transforms in the two spatial di-

rections of images. The output of these 3 Fourier Transforms is a

quantity that is closely related to the amplitudes h0(k, ω)for each

spatial and temporal frequency. These are then absolute squared

and smoothed or averaged in some way so that the output is a nu-

merical approximation of a statistical average of |h0(k, ω)|2.

But how does a 3D PSD help us decide whether the dispersion

relation appears in nature? If the imagery found only dispersion

constrained surface waves, then the 3D PSD should have the value

0 for all values of k,ωthat do not satisfy the dispersion relation. So

mostly we would expect the 3D PSD to only have signiﬁcant values

in a narrow set of k,ωvalues.

Figure 3 show a plot of the 3D PSD generated from AROSS im-

ages [24]. From the 3D volumetric PSD, this plot ﬁgure is a plane

sliced through the volume. When sliced like this, the dispersion

relation is a curve on the slice, shown as two dotted curves. The

data is plotted as contours of PSD intensity, color-coded by the key

on the right. PSD levels following the dispersion relation curves

are much higher than in other regions. This shows that the motion

of the surface waves on all scales includes a very strong dispersion

relation style of motion. There are other types of motion certainly,

which the PSD ﬁgure shows as intensity levels away from the dis-

4 PRACTICAL OCEAN WAVE ALGORITHMS 3-7

Figure 4: Site at which video data was collected in 1986, near Zuma

Beach, California.

persion curves. But the dispersion motion is the strongest feature

of this data.

Relatively simple experiments can be done by anyone with ac-

cess to a video camera and a hilltop overlook of an ocean. For

example, ﬁgure 4 is a frame from a video segment showing wa-

ter coming into the beach near Zuma Beach, California. The video

camera was located on hill overlooking the beach, in 1986. In 1993,

the region of video frames indicated in the ﬁgure was digitized, to

produce a time series of frames containing just water surface.

Figure 5 shows the actual 3D PSD from the image data. There

are two clear branches along the dispersion relationship we have

discussed, with no apparent modiﬁcation by shallow water affects.

There is also a third branch that is approximately a straight line

lying between the ﬁrst two. Examination of the video shows that

this branch comes from a surfactant layer ﬂoating on the water in

part of the video frame, and moving with a constant speed. Exclud-

ing the surface layer, this data clearly demonstrates the validity of

the dispersion relationship, and demonstrates the usefulness of the

linearized model of surface waves.

4 Practical Ocean Wave Algorithms

In this section we focus on algorithms and practical steps to build-

ing height ﬁelds for ocean waves. Although we will be occupied

mostly by a method based on Fast Fourier Transforms (FFTs), we

begin by introducing a simpler description called Gerstner Waves.

This is a good starting point for several reasons: the mathematics is

relatively light compared to FFTs, several important oceanographic

concepts can be introduced, and they give us a chance to discuss

wave animation. After this discussion of Gerstner waves, we go

after the more complex FFT method, which produces wave height

ﬁelds that are more realistic. These waves, called “linear waves” or

“gravity waves” are a fairly realistic representation of typical waves

on the ocean when the weather is not too stormy. Linear waves are

certainly not the whole story, and so we discuss also some meth-

ods by which oceanographers expand the description to “nonlinear

waves”, waves passing over a shallow bottom, and very tiny waves

about one millimeter across called capillary waves.

In the course of this discussion, we will see how quantities like

windspeed, surface tension, and gravitational acceleration come

into the practical implementation of the algorithms.

Figure 5: Slice from a 3D Power Spectral Density grayscale plot,

from processed video data.

4.1 Gerstner Waves

Gerstner waves were ﬁrst found as an approximate solution to the

ﬂuid dynamic equations almost 200 years ago. There ﬁrst appli-

cation in computer graphics seems to be the work by Fournier and

Reeves in 1986 (cited previously). The physical model is to de-

scribe the surface in terms of the motion of individual points on the

surface. To a good approximation, points on the surface of the water

go through a circular motion as a wave passes by. If a point on the

undisturbed surface is labelled x0= (x0, z0)and the undisturbed

height is y0= 0, then as a single wave with amplitude Apasses by,

the point on the surface is displaced at time t to

x=x0−(k/k)Asin(k·x0−ωt)(26)

y=Acos(k·x0−ωt).(27)

In these expressions, the vector k, called the wavevector, is a

horizontal vector that points in the direction of travel of the wave,

and has magnitude krelated to the length of the wave (λ) by

k= 2π/λ (28)

The frequency wis related to the wavevector, as discussed later.

Figure 6 shows two example wave proﬁles, each with a different

value of the dimensionless amplitude kA. For values kA < 1, the

wave is periodic and shows a steepening at the tops of the waves as

kA approaches 1. For kA > 1, a loop forms at the tops of the wave,

and the “insides of the wave surface are outside”, not a particularly

desirable or realistic effect.

As presented so far, Gerstner waves are rather limited because

they are a single sine wave horizontally and vertically. However,

this can be generalized to a more complex proﬁle by summing a

set of sine waves. One picks a set of wavevectors ki, amplitudes

Ai, frequencies ωi, and phases φi, for i= 1,...,N, to get the

expressions

x=x0−

N

X

i=1

(ki/ki)Aisin(ki·x0−ωit+φi)(29)

4 PRACTICAL OCEAN WAVE ALGORITHMS 3-8

-4

-3

-2

-1

0

1

2

3

4

-5 0 5 10 15 20 25

Wave Amplitude

Position

kA = 0.7

kA = 1.33

Figure 6: Proﬁles of two single-mode Gerstner waves, with differ-

ent relative amplitudes and wavelengths.

y=

N

X

i=1

Aicos(ki·x0−ωit+φi).(30)

Figure 7 shows an example with three waves in the set. Interest-

ing and complex shapes can be obtained in this way.

4.2 Animating Waves: The Dispersion Relation

The animated behavior of Gerstner waves is determined by the set

of frequencies ωichosen for each component. For water waves,

there is a well-known relationship between these frequencies and

the magnitude of their corresponding wavevectors, ki. In deep wa-

ter, where the bottom may be ignored, that relationship is

ω2(k) = gk . (31)

The parameter gis the gravitational constant, nominally

9.8m/sec2. This dispersion relationship holds for Gerstner waves,

and also for the FFT-based waves introduced next.

There are several conditions in which the dispersion relationship

is modiﬁed. When the bottom is relatively shallow compared to

the length of the waves, the bottom has a retarding affect on the

waves. For a bottom at a depth Dbelow the mean water level, the

dispersion relation is

ω2(k) = gk tanh(kD)(32)

Notice that if the bottom is very deep, the behavior of the tanh

function reduces this dispersion relation to the previous one.

A second situation which modiﬁes the dispersion relation is sur-

face tension. Very small waves, with a wavelength of about 1 cm or

less, have an additional term:

ω2(k) = gk(1 + k2L2),(33)

and the parameter Lhas units of length. Its magnitude is the scale

for the surface tension to have effect.

Using these dispersion relationships, it is very difﬁcult to create

a sequence of frames of water surface which for a continuous loop.

-8

-6

-4

-2

0

2

4

6

8

-5 0 5 10 15 20 25

Wave Amplitude

Position

Figure 7: Proﬁle of a 3-mode Gerstner wave.

In order to have the sequence repeat after a certain amount of time

Tfor example, it is necessary that all frequencies be multiples of

the basic frequence

ω0≡2π

T.(34)

However, when the wavevectors kare distributed on a regular lat-

tice, itis impossible to arrange the dispersion-generated frequencies

to also be on a uniform lattce with spacing ω0.

The solution to that is to not use the dispersion frequences, but

instead a set that is close to them. For a given wavenumber k, we

use the frequency

¯ω(k) = ω(k)

ω0ω0,(35)

where [[a]] means take the integer part of the value of a, and ω(k)is

any dispersion relationship of interest. The frequencies ¯ω(k)are a

quantization of the dispersion surface, and the animation of the wa-

ter surface loops after a time Tbecause the quantized frequencies

are all integer multiples of ω0. Figure 8 plots the original disper-

sion curve, along with quantized dispersion curves for two choices

of the repeat time T.

4.3 Statistical Wave Models and the Fourier Trans-

form

Oceanographic literature tends to downplay Gerstner waves as a re-

alistic model of the ocean. Instead, statistical models are used, in

combination with experimental observations. In the statistical mod-

els, the wave height is considered a random variable of horizontal

position and time, h(x, t).

Statistical models are also based on the ability to decompose

the wave height ﬁeld as a sum of sine and cosine waves. The

value of this decomposition is that the amplitudes of the waves

have nice mathematical and statistical properties, making it sim-

pler to build models. Computationally, the decomposition uses Fast

Fourier Transforms (ffts), which are a rapid method of evaluating

the sums.

4 PRACTICAL OCEAN WAVE ALGORITHMS 3-9

0

0.5

1

1.5

2

2.5

3

3.5

012345678910

Frequency

Wavenumber

Quantizing the Dispersion Relation

Dispersion Relation

Repeat Time = 100 seconds

Repeat Time = 20 seconds

Figure 8: A comparison of the continuous dispersion curve ω=

√gk and quantized dispersion curves, for repeat times of 20 sec-

onds and 100 seconds. Note that for a longer repeat time, the quan-

tized is a closer approximation to the original curve.

The fft-based representation of a wave height ﬁeld expresses the

wave height h(x, t)at the horizontal position x= (x, z)as the sum

of sinusoids with complex, time-dependent amplitudes:

h(x, t) = X

k

˜

h(k, t) exp (ik·x)(36)

where tis the time and kis a two-dimensional vector with com-

ponents k= (kx, kz),kx= 2πn/Lx,kz= 2πm/Lz, and

nand mare integers with bounds −N/2≤n < N/2and

−M/2≤m < M/2. The fft process generates the height ﬁeld

at discrete points x= (nLx/N, mLz/M). The value at other

points can also be obtained by switching to a discrete fourier trans-

form, but under many circumstances this is unnecessary and is not

applied here. The height amplitude Fourier components, ˜

h(k, t),

determine the structure of the surface. The remainder of this sub-

section is concerned with generating random sets of amplitudes in

a way that is consistent with oceanographic phenomenology.

For computer graphics purposes, the slope vector of the wave-

height ﬁeld is also needed in order to ﬁnd the surface normal, angles

of incidence, and other aspects of optical modeling as well. One

way to compute the slope is though a ﬁnite difference between fft

grid points, separated horizontally by some 2D vector ∆x. While

a ﬁnite difference is efﬁcient in terms of memory requirements, it

can be a poor approximation to the slope of waves with small wave-

length. An exact computation of the slope vector can be obtained

by using more ffts:

(x, t) = ∇h(x, t) = X

k

ik˜

h(k, t) exp (ik·x).(37)

In terms of this fft representation, the ﬁnite difference approach

would replace the term ikwith terms proportional to

exp (ik·∆x)−1(38)

which, for small wavelength waves, does not well approximate the

gradient of the wave height. Whenever possible, slope computation

via the fft in equation 37 is the prefered method.

The fft representation produces waves on a patch with horizontal

dimensions Lx×Lz, outside of which the surface is perfectly peri-

odic. In practical applications, patch sizes vary from 10 meters to 2

kilometers on a side, with the number of discrete sample points as

high as 2048 in each direction (i.e. grids that are 2048 ×2048, or

over 4 million waves). The patch can be tiled seamlessly as desired

over an area. The consequence of such a tiled extension, however, is

that an artiﬁcial periodicity in the wave ﬁeld is present. As long as

the patch size is large compared to the ﬁeld of view, this periodicity

is unnoticeable. Also, if the camera is near the surface so that the

effective horizon is one or two patch lengths away, the periodicity

will not be noticeable in the look-direction, but it may be apparent

as repeated structures across the ﬁeld of view.

Oceanographic research has demonstrated that equation 36 is a

reasonable representation of naturally occurring wind-waves in the

open ocean. Statistical analysis of a number of wave-buoy, photo-

graphic, and radar measurements of the ocean surface demonstrates

that the wave height amplitudes ˜

h(k, t)are nearly statistically sta-

tionary, independent, gaussian ﬂuctuations with a spatial spectrum

denoted by

Ph(k) = D

˜

h∗(k, t)

2E(39)

for data-estimated ensemble averages denoted by the brackets h i.

There are several analytical semi-empirical models for the wave

spectrum Ph(k). A useful model for wind-driven waves larger than

capillary waves in a fully developed sea is the Phillips spectrum

Ph(k) = Aexp −1/(kL)2

k4|ˆ

k·ˆw|2,(40)

where L=V2/g is the largest possible waves arising from a con-

tinuous wind of speed V,gis the gravitational constant, and ˆwis

the direction of the wind. Ais a numeric constant. The cosine factor

|ˆ

k·ˆw|2in the Phillips spectrum eliminates waves that move perpen-

dicular to the wind direction. This model, while relatively simple,

has poor convergence properties at high values of the wavenumber

|k|. A simple ﬁx is to suppress waves smaller that a small length

`L, and modify the Phillips spectrum by the multiplicative fac-

tor

exp −k2`2.(41)

Of course, you are free to “roll your own” spectrum to try out

various effects.

4.4 Building a Random Ocean Wave Height Field

Realizations of water wave height ﬁelds are created from the prin-

ciples elaborated up to this point: gaussian random numbers with

spatial spectra of a prescribed form. This is most efﬁciently accom-

plished directly in the fourier domain. The fourier amplitudes of a

wave height ﬁeld can be produced as

˜

h0(k) = 1

√2(ξr+iξi)pPh(k),(42)

where ξrand ξiare ordinary independent draws from a gaussian

random number generator, with mean 0 and standard deviation 1.

Gaussian distributed random numbers tend to follow the experi-

mental data on ocean waves, but of course other random number

distributions could be used. For example, log-normal distributions

could be used to produce height ﬁelds that are vary “intermittent”,

i.e. the waves are very high or nearly ﬂat, with relatively little in

between.

Given a dispersion relation ω(k), the Fourier amplitudes of the

wave ﬁeld realization at time tare

˜

h(k, t) = ˜

h0(k) exp {iω(k)t}

+˜

h∗

0(−k) exp {−iω(k)t}(43)

4 PRACTICAL OCEAN WAVE ALGORITHMS 3-10

This form preserves the complex conjugation property ˜

h∗(k, t) =

˜

h(−k, t)by propagating waves “to the left” and “to the right”. In

addition to being simple to implement, this expression is also efﬁ-

cient for computing h(x, t), since it relies on ffts, and because the

wave ﬁeld at any chosen time can be computed without computing

the ﬁeld at any other time.

In practice, how big does the Fourier grid need to be? What

range of scales is reasonable to choose? If you want to generate

wave heights faster, what do you do? Lets take a look at these

questions.

How big should the Fourier grid be? The values of Nand M

can be between 16 and 2048, in powers of two. For many

situations, values in the range 128 to 512 are sufﬁcient. For

extremely detailed surfaces, 1024 and 2048 can be used. For

example, the wave ﬁelds used in the motion pictures Water-

world and Titanic were 2048×2048 in size, with the spacing

between grid points at about 3 cm. Above a value of 2048, one

should be careful because the limits of numerical accuracy for

ﬂoating point calculations can become noticeable.

What range of scales is reasonable to choose? The answer to this

question comes down to choosing values for Lx,Lz,M, and

N. The smallest facet in either direction is dx ≡Lx/M or

dz ≡Lz/N. Generally, dx and dz need never go below

2 cm or so. Below this scale, the amount of wave action is

small compared to the rest of the waves. Also, the physics

of wave behavior below 2 cm begins to take on a very differ-

ent character, involving surface tension and “nonlinear” pro-

cesses. From the form of the spectrum, waves with a wave-

length larger than V2/g are suppressed. So make sure that dx

and dz are smaller than V2/g by a substantial amount (10 -

1000) or most of the interesting waves will be lost. The se-

cret to realistic looking waves (e.g. ﬁgure 12 (a) compared to

ﬁgure 12 (c)) is to have Mand Nas large as reasonable.

How do you generate wave height ﬁelds in the fastest time? The

time consuming part of the computation is the fast fourier

transform. Running on a 1+ GHz cpu, 512 ×512 FFTs can

be generated at nearly interactive rates.

4.5 Examples: Height Fields and Renderings

We now turn to some examples of waves created using the fft ap-

proach discussed above. We will show waves in two formats: as

greyscale images in which the grey level is proportional to wave

height; and renderings of oceanscapes using several different ren-

dering packages to illustrate what is possible.

In the ﬁrst set of examples, the grid size is set to M=N= 512,

with Lx=Lz= 1000 meters. The wind speed is a gale force at

V= 31 meters/second, moving in the x-direction. The small-wave

cutoff of `= 1 meter was also used. Figure 9 is a greyscale rep-

resentation of the wave height: brighter means higher and darker

means lower height. Although produced by the fft algorithms de-

scribed here, ﬁgure 9 is not obviously a water height ﬁeld. It may

help to examine ﬁgure 10, which is a greyscale depiction of the

x-component of the slope. This looks more like water waves that

ﬁgure 9. What is going on?

Figures 9 and 10 demonstrate a consequence of water surface

optics, discussed in the next section: the visible qualities of the

surface structure tend to be strongly inﬂuenced by the slope of the

waves. We will discuss this in quantitative detail, but for now we

willl summarize it by saying that the reﬂectivity of the water is a

strong function of the slope of the waves, as well as the directions

of the light(s) and camera.

To illustrate a simple effect of customizing the spectrum model,

ﬁgure 11 is the greyscale display of a height ﬁeld identical to ﬁgure

Figure 9: A surface wave height realization, displayed in greyscale.

Figure 10: The x-component of the slope for the wave height real-

ization in ﬁgure 9.

4 PRACTICAL OCEAN WAVE ALGORITHMS 3-11

Figure 11: Wave height realization with increased directional de-

pendence.

9, with the exception that the directional factor |ˆ

k·ˆw|2in equation

40 has been changed to |ˆ

k·ˆw|6. The surface is clearly more aligned

with the direction of the wind.

The next example of a height ﬁeld uses a relatively simple shader

in BMRT, the Renderman-compliant raytracer. The shader is shown

in the next section. Figure 12 shows three renderings of water sur-

faces, varying the size of the grid numbers Mand Nand making

the facet sizes dx and dz proportional to 1/M and 1/N . So as we

go from the top image to the bottom, the facet sizes become smaller,

and we see the effect of increasing amount of detail in the render-

ings. Clearly, more wave detail helps to build a realistic-looking

surface.

As a ﬁnal example, ﬁgure 13 is an image rendered in the com-

mercial package RenderWorld by Arete Entertainment. This ren-

dering includes the effect of an atmosphere, and water volume scat-

tered light. These are discussed in the next section. But clearly,

wave height ﬁelds generated from random numbers using an fft pre-

scription can produce some nice images.

4.6 Choppy Waves

We turn brieﬂy in this section to the subject of creating choppy

looking waves. The waves produced by the fft methods presented

up to this point have rounded peaks and troughs that give them the

appearance of fair-weather conditions. Even in fairly good weather,

and particularly in a good wind or storm, the waves are sharply

peaked at their tops, and ﬂattened at the bottoms. The extent of this

chopping of the wave proﬁle depends on the environmental condi-

tions, the wavelengths and heights of the waves. Waves that are

sufﬁciently high (e.g. with a slope greater than about 1/6) eventu-

ally break at the top, generating a new set of physical phenonema

in foam, splash, bubbles, and spray.

The starting point for this method is the fundamental ﬂuid dy-

namic equations of motion for the surface. These equations are ex-

pressed in terms of two dynamical ﬁelds: the surface elevation and

the velocity potential on the surface, and derive from the Navier-

Stokes description of the ﬂuid throughout the volume of the water

and air, including both above and below the interface. Creamer

Figure 12: Rendering of waves with (top) a fairly low number of

waves (facet size = 10 cm), with little detail; (middle) a reasonably

good number of waves (facet size = 5 cm); (bottom) a high number

of waves with the most detail (facet size = 2.5 cm).

Figure 13: An image of a wave height ﬁeld rendered in a commer-

cial package with a model atmosphere and sophisticated shading.

4 PRACTICAL OCEAN WAVE ALGORITHMS 3-12

-0.5

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0123456

Surface Height

Surface Position

Displacement Behavior

Figure 14: A comparison of a wave height proﬁle with and without

the displacement. The dashed curve is the wave height produced by

the fft representation. The solid curve is the height ﬁeld displaced

using equation 44.

et al[16] set out to apply a mathematical approach called the ”Lie

Transform technique” to generate a sequence of ”canonical trans-

formations” of the elevation and velocity potential. The beneﬁt of

this complex mathematical procedure is to convert the elevation and

velocity potential into new dynamical ﬁelds that have a simpler

dynamics. The transformed case is in fact just the simple ocean

height ﬁeld we have been discussing, including evolution with the

same dispersion relation we have been using in this paper. Start-

ing from there, the inverse Lie Transform in principle converts our

phenomenological solution into a dynamically more accurate one.

However, the Lie Transform is difﬁcult to manipulate in 3 dimen-

sions, while in two dimensions exact results have been obtained.

Based on those exact results in two dimensions, an extrapolation

for the form of the 3D solution has been proposed: a horizontal dis-

placement of the waves, with the displacement locally varying with

the waves.

In the fft representation, the 2D displacement vector ﬁeld is com-

puted using the Fourier amplitudes of the height ﬁeld, as

D(x, t) = X

k−ik

k˜

h(k, t) exp (ik·x)(44)

Using this vector ﬁeld, the horizontal position of a grid point of

the surface is now x+λD(x, t), with height h(x)as before. The

parameter λis not part of the original conjecture, but is a conve-

nient method of scaling the importance of the displacement vector.

This conjectured solution does not alter the wave heights directly,

but instead warps the horizontal positions of the surface points in

a way that depends on the spatial structure of the height ﬁeld. The

particular form of this warping however, actually sharpens peaks in

the height ﬁeld and broadens valleys, which is the kind of nonlin-

ear behavior that should make the fft representation more realistic.

Figure 14 shows a proﬁle of the wave height along one direction

in a simulated surface. This clearly shows that the “displacement

conjecture” can dramatically alter the surface.

The displacment form of the this solution is similar to the algo-

rithm for building Gerstner waves [12] discussed in section 4. In

-40

-35

-30

-25

-20

-15

-10

-5

0

5

-5 0 5 10 15 20 25

"gnuplotdatatimeprofile.txt"

Figure 15: A time sequence of proﬁles of a wave surface. From top

to bottom, the time between proﬁles is 0.5 seconds.

that case however, the displacement behavior, applied to sinusoid

shapes, was the principle method of characterizing the water sur-

face structure, and here it is a modiﬁer to an already useful wave

height representation.

Figure 15 illustrates how these choppy waves behave as they

evolve. The tops of waves form a sharp cusp, which rounds out

and disappears shortly afterward.

One ”problem” with this method of generating choppy waves

can be seen in ﬁgure 14. Near the tops of some of the waves, the

surface actally passes through itself and inverts, so that the outward

normal to the surface points inward. This is because the amplitudes

of the wave components can be large enough to create large dis-

placements that overlap. This is easily defeated simply by reducing

the magnitude of the scaling factor λ. For the purposes of computer

graphics, this might actually be a useful effect to signal the pro-

duction of spray, foam and/or breaking waves. We will not discuss

here how to carry out such an extension, except to note that in order

to use this region of overlap, a simple and quick test is needed for

deciding that the effect is taking place. Fortunately, there is such a

simple test in the form of the Jacobian of the transformation from

xto x+λD(x, t). The Jacobian is a measure of the uniqueness of

the transformation. When the displacement is zero, the Jacobian is

1. When there is displacement, the Jacobian has the form

J(x) = JxxJyy −Jxy Jy x ,(45)

with individual terms

Jxx(x) = 1 + λ∂Dx(x)

∂x

Jyy (x) = 1 + λ∂Dy(x)

∂y

5 INTERACTIVE WAVES FROM DISTURBANCES 3-13

-1.5

-1

-0.5

0

0.5

1

1.5

0 20 40 60 80 100

Height

Position

Water Surface Profiles

Basic Surface

Choppy Surface

Folding Map

Figure 16: Wave height proﬁle with and without the displacement.

Also plotted is the Jacobian map for choppy wave proﬁle.

Jyx(x) = λ∂ Dy(x)

∂x

Jxy(x) = λ∂ Dx(x)

∂y =Jyx

and D= (Dx, Dy). The Jacobian signals the presence of the over-

lapping wave bacause its value is less than zero in the overlap re-

gion. For example, ﬁgure 16 plots a proﬁle of a basic wave without

displacement, the wave with displacement, and the value of Jfor

the choppy wave (labeled ”Folding Map”). The ”folds” or overlaps

in the choppy surface are clearly visible, and align with the regions

in which J < 0. With this information, it should be relatively easy

to extract the overlapping region and use it for other purposes if

desired.

But there is more that can be learned from these folded waves

from a closer examination of this folding criterion. The Jacobian

derives from a 2×2matrix which measures the local uniqueness

of the choppy wave map x→x+λD. This matrix can in general

be written in terms of eigenvectors and eigenvalues as:

Jab =J−ˆe−

aˆe−

b+J+ˆe+

aˆe+

b,(a, b =x, y)(46)

where J−and J+are the two eigenvalues of the matrix, ordered

so that J−≤J+. The corresponding orthonormal eigenvectors are

ˆe−and ˆe+respectively. From this expression, the Jacobian is just

J=J−J+.

The criterion for folding that J < 0means that J−<0and

J+>0. So the minimum eigenvalue is the actual signal of the on-

set of folding. Further, the eigenvector ˆe−points in the horizontal

direction in which the folding is taking place. So, the prescrip-

tion now is to watch the minimum eigenvalue for when it becomes

negative, and the alignment of the folded wave is parallel to the

minimum eigenvector.

We can illustrate this phenomenon with an example. Figures 17

and 18 show two images of an ocean surface, one without choppy

waves, and the other with the choppy waves strongly applied. These

two surfaces are identical except for the choppy wave algorithm.

Figure 19 shows the wave proﬁles of both surfaces along a slice

through the surfaces. Finally, the proﬁle of the choppy wave is

plotted together with the value of the minimum eigenvalue in ﬁgure

20, showing the clear connection between folding and the negative

value of J−.

Incidentally, computing the eigenvalues and eigenvectors of this

matrix is fast because they have analytic expressions as

J±=1

2(Jxx +Jyy )±1

2(Jxx −Jyy )2+ 4J2

xy1/2(47)

for the eigenvalues and

ˆe±=(1, q±)

p1 + q2

±

(48)

with

q±=J±−Jxx

Jxy

(49)

for the eigenvectors.

5 Interactive Waves from Disturbances

The Fourier based approach to water surface evolutiondescribed in

the section 4 has several limitations that make in unworkable for

really interactive applications. Fundamentally, the Fourier method

computes the wave height everywhere in one FFT computation.

You cannot choose to compute wave height in limited areas of a

surface grid without completely altering the calculation. You either

get the whole surface, or you don’t compute it. This makes it very

hard to customize the wave propagation problem from one location

to another. So it you want to put some arbitrarily-shaped object

in the water, move it around some user-constructed path, the FFT

method makes it difﬁcult to compute the wave response to the shape

and movement of the object. So if you want to have an odd looking

craft moving around in the water, and maybe going up and down

and changing shape, the FFT method is not the easiest thing to use.

Also, if you want to have a shallow bottom, with a sloping beach or

underwater sea mound or some other variability in the depth of the

water, the FFT method again is a challenge use.

Fortunately in the last few years an alternative scheme has

emerged which allows fast construction of a water surface in re-

sponse to interactions with objects in the water and/or variable bot-

tom depth. The heart of the method is an approach to computing

the propagation which does not use the FFT method at all. The fun-

damental mathematics of this interactive method is described in the

reference [22], and is refered to as iWave. Here we very quickly run

through the approach and show some examples in action.

At its heart the iWave method returns to the linearized equation

for the wave height, 22. We can turn the time derivatives into ﬁnite

time differences for a time step ∆t, and get an explicit expression

for the wave height:

h(x, t + ∆t)=2h(x, t)−h(x, t −∆t)

−g(∆t)2p−∇2h(x, t)(50)

This form can be used to explicitly advance the surface wave height

from one frame to the next. The hard part of course is ﬁguring out

how to calculate the last term, with the square root of the Laplacian

operator.

The solution is to use convolution. The wave height is kept on

a rectangular grid of points (the dimensions of which do not have

to be powers of two), and we make use of the fact that any linear

operation on a grid of data values can be converted into a convolu-

tion of some sort. The details of the numerical implementation are

contained in [22].

But the real, amazing, property of iWave is that wave interaction

with objects on the water surface is evaluated with a very simple

5 INTERACTIVE WAVES FROM DISTURBANCES 3-14

Figure 17: Simulated wave surface without the choppy algorithm

applied. Rendered in BMRT with a generic plastic shader. Figure 18: Same wave surface with strong chop applied. Rendered

in BMRT with a generic plastic shader.

5 INTERACTIVE WAVES FROM DISTURBANCES 3-15

-2

-1

0

1

2

0 5 10 15 20

Height

Position

Water Surface Profiles

Basic Surface

Choppy Surface

Figure 19: Proﬁles of the two surfaces, showing the effect of the

choppy mapping.

-2

-1

0

1

2

0 5 10 15 20

Height

Position

Water Surface Profiles

Choppy Surface

Minimum E-Value

Figure 20: Plot of the choppy surface proﬁle and the minimum

eigenvalue. The locations of folds of the surface are clearly the

same as where the eigenvalue is negative.

linear operation that amount to an image masking. For this pur-

pose, an object in the water is described as a grid of mask values,

zero meaning the object is present, 1 if the object is not present,

and along the edges of the object values between 0 and 1 are use-

ful as an antialiasing. Applying this mask to the wave height grid,

the waves are effectively removed at grid points that the object is

located at. With just this simple procedure, waves that are incident

on the object reﬂect off of it in a realistic way. Figure 21 is a frame

from a sequence, showing the wave height computed using iWave

with several irregularly shaped objects.

Ordinarily, simulating waves that interact with an object on the

water surface should involve a careful treatment of boundary con-

ditions, matching water motion to the object on the boundary of

the geometry, and enforcing no-slip conditions. This is a complex

and time consuming computation that involves a certain amount of

numerical black art. It is suprising that a simple process like mask-

ing the wave height, as described above, should not be sufﬁcient.

Yet, with the iWave procedure, correct-looking reﬂection/refraction

waves happen in the simulation. It is not yet clear just how quanti-

tatively accurate these interactions are. Figure 22 shows a rendered

scene with a high resolution calculation of waves interacting with

the hull of the ship.

5.1 Modiﬁcations for shallow water

Wave simulations based on the FFT method can simulate shallow

water effects by using the dispersion relationship in equation 32.

This only applies to a ﬂat bottom. It would be nice to simulate wave

propagation onto a beach, or pasta shallow subsurface sea mount, or

over a submarine that is just below the surface. The iWave method

is a great way of doing those things. Recall that the iWave method

converts the mathematical operation

gp−∇2h(51)

into a convolution, so that effectively g√−∇2becomes a convolu-

tion kernel. A shallow bottom with depth Dchanges this term to

(compare with the dispersion relationship)

gp−∇2tanh p−∇2Dh(52)

This operation can also be converted into a convolution, with

g√−∇2tanh √−∇2Dbecomes a convolution kernel.

How is this applied to a variable-depth bottom? The convolution

kernel is a 13x13 matrix[22]. So we could build a collection of

kernels over a range of depth values D. At each point on the grid, a

custom kernel is constructed for the actual depth at that grid point

by interpolating from the set of prebuilt values. Figure 23 is the

wave height from a simulation using this technique. The bottom

slopes from deep on the right hand side, to a depth of 0 on the left

edge. In addition, there is a subsurface sea mount on the right.

There are three important behaviors in this simulation that occur

in real shallow water propagation:

1. Waves in shallow regions have large amplitudes that in deep

regions.

2. As waves approach a beach, they pile up together and have a

higher spatial frequency.

3. The subsurface sea mount causes a diffraction of waves.

In addition to these capabilities, iWave can also compute other

quantities, such as cuspy waves and surface velocity.

5 INTERACTIVE WAVES FROM DISTURBANCES 3-16

(a)

(b)

(c)

Figure 21: A sequence of frames from an iWave simulation, show-

ing waves reﬂecting off of objects in the water. The objects are the

black regions. In the upper left there is also diffraction taking place

as waves propagate the narrow channel and emerge in the corner

region.

Figure 22: Frame from a simulation and rendering showing waves

interacting with a ship.

Figure 23: Frame from an iWave simulation with a variable depth

shallow bottom.

6 SURFACE WAVE OPTICS 3-17

6 Surface Wave Optics

The optical behavior of the ocean surface is fairly well understood,

at least for the kinds of quiescent wave structure that we consider

in these notes. Fundamentally, the ocean surface is a near perfect

specular reﬂector, with well-understand relectivity and transmis-

sity functions. In this section these properties are summarized, and

combined into a simple shader for Renderman. There are circum-

stances when the surface does not appear to be a specular reﬂector.

In particular, direct sunlight reﬂected from waves at a large distance

from the camera appear to be spread out and made diffuse. This

is due to the collection of waves that are smaller than the camera

can resolve at large distances. The mechanism is somewhat similar

to the underlying microscopic reﬂection mechnanisms in solid sur-

faces that lead to the Torrance-Sparrow model of BRDFs. Although

the study of glitter patterns in the ocean was pioneered by Cox and

Munk many years ago, the ﬁrst models of this BRDF behavior that

I am aware of were developed in the early 1980’s. At the end of this

section, we introduce the concepts and conditions, state the results,

and ignore the in-between analysis and derivation.

Throughout these notes, and particularly in this section, we ig-

nore one optical phenomenon completely: polarization. Polariza-

tion effects can be strong at a boundary interface like a water sur-

face. However, since most of computer graphics under considera-

tion ignores polarization, we will continue in that long tradition. Of

course, interested readers can ﬁnd literature on polarization effects

at the air-water interface.

6.1 Specular Reﬂection and Transmission

Rays of light incident from above or below at the air-water interface

are split into two components: a transmitted ray continuing through

the interface at a refracted angle, and a reﬂected ray. The intensity

of each of these two rays is diminished by reﬂectivity and trans-

missivity coefﬁcients. Here we discussed the directions of the two

outgoing rays. In the next subsection the coefﬁcients are discussed.

6.1.1 Reﬂection

As is well known, in a perfect specular reﬂection the reﬂected ray

and the incident ray have the same angle with respect to the surface

normal. This is true for all specular reﬂections (ignoring roughen-

ing effects), regardless of the material. We build here a compact

expression for the outgoing reﬂected ray. First, we need to build up

some notation and geometric quantities.

The three-dimensional points on the ocean surface can be la-

belled by the horizontal position xand the waveheight h(x, t)as

r(x, t) = x+ˆ

yh(x, t),(53)

where ˆ

yis the unit vector pointing straight up. At the point r, the

normal to the surface is computed directly from the surface slope

(x, t)≡ ∇h(x, t)as

ˆ

nS(x, t) = ˆ

y−(x, t)

p1 + 2(x, t)(54)

For a ray intersecting the surface at rfrom direction ˆ

ni, the direc-

tion of the reﬂected ray can depend only on the incident direction

and the surface normal. Also, as mentioned before, the angle be-

tween the surface normal and the reﬂected ray must be the same

as the angle between incident ray and the surface normal. You can

verify for yourself that the reﬂected direction ˆ

nris

ˆ

nr(x, t) = ˆ

ni−2ˆ

nS(x, t) (ˆ

nS(x, t)·ˆ

ni).(55)

Note that this expression is valid for incident ray directions on either

side of the surface.

6.1.2 Transmission

Unfortunately, the direction of the transmitted ray is not expressed

as simply as for the reﬂected ray. In this case we have two guid-

ing principles: the transmitted direction is dependent only on the

surface normal and incident directions, and Snell’s Law relating the

sines of the angles of the incident and transmitted angles to the in-

dices of refraction of the two materials.

Suppose the incident ray is coming from one of the two media

with index of refraction ni(for air, n= 1, for water, n= 4/3

approximately), and the transmitted ray is in the medium with index

of refraction nr. For the incident ray at angle θito the normal,

sin θi=p1−(ˆ

ni·ˆ

nS)2=|ˆ

ni×ˆ

nS|(56)

the transmitted ray will be at an angle θtwith

sin θt=|ˆ

nt×ˆ

nS|.(57)

Snell’s Law states that these two angles are related by

nisin θi=ntsin θt.(58)

We now have all the pieces needed to derive the direction of trans-

mission. The direction vector can only be a linear combination of

ˆ

niand ˆ

nS. It must satisfy Snell’s Law, and it must be a unit vector

(by deﬁnition). This is adequate to obtain the expression

ˆ

nt(x, t) = ni

nt

ˆ

ni+ Γ(x, t)ˆ

nS(x, t)(59)

with the function Γdeﬁned as

Γ(x, t)≡ni

nt

ˆ

ni·ˆ

nS(x, t)

±1−ni

nt2

|ˆ

ni×ˆ

nS(x, t)|21/2

.(60)

The plus sign is used in Γwhen ˆ

ni·ˆ

nS<0, and the minus sign is

used when ˆ

ni·ˆ

nS>0.

6.2 Fresnel Reﬂectivity and Transmissivity

Accompanying the process of reﬂection and transmission through

the interface is a pair of coefﬁcients that describe their efﬁciency.

The reﬂectivity Rand transmissivity Tare related by the constraint

that no light is lost at the interface. This leads to the relationship

R+T= 1 .(61)

The derivation of the expressions for Rand Tis based on the elec-

tromagnetic theory of dielectrics. We will not carry out the deriva-

tions, but merely write down the solution

R(ˆ

ni,ˆ

nr) = 1

2sin2(θt−θi)

sin2(θt+θi)+tan2(θt−θi)

tan2(θt+θi)(62)

Figure 24 is a plot of the reﬂectivity for rays of light traveling down

onto a water surface as a function of the angle of incidence to the

surface. The plot extends from a grazing angle of 0 degrees to per-

pendicular incidence at 90 degrees. As should be clear, variation of

the reﬂectivity across an image is an important source of the “tex-

ture” or feel of water. Notice that reﬂectivity is a function of the

angle of incidence relative to the wave normal, which in turn is di-

rectly related to the slope of the surface. So we can expect that a

strong contributor to the texture of water surface is the pattern of

slope, while variation of the wave height serves primarily as a wave

hiding mechanism. This is the quantitative explanation of why the

6 SURFACE WAVE OPTICS 3-18

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30 40 50 60 70 80 90

Incidence Angle (degrees)

Reflectivity from Above

Incident from Above

Incident from Below

Figure 24: Reﬂectivity for light coming from the air down to the

water surface, as a function of the angle of incidence of the light.

surface slope more closely resembles rendered water than the wave

height does, as we saw in the previous section when discussing ﬁg-

ure 10.

When the incident ray comes from below the water surface, there

are important differences in the reﬂectivity and transmissivity. Fig-

ure 25 shows the reﬂectivity as a function of incidence angle again,

but this time for incident light from below. In this case, the re-

ﬂectivity reaches unity at a fairly large angle, near 41 degrees. At

incidence angles below that, the reﬂectivity is one and so there is no

transmission of light through the interface. This phenomenon is to-

tal internal reﬂection, and can be seen just by swimming around in

a pool. The angle at which total internal reﬂection begins is called

Brewster’s angle, and is given by, from Snell’s Law,

sin θB

i=nt

ni

= 0.75 (63)

or θB

i= 48.6 deg. In our plots, this angle is 90 −θB

i= 41.1 deg.

6.3 Building a Shader for Renderman

From the discussion so far, one of the most important features a ren-

dering must emulate is the textures of the surface due to the strong

slope-dependence of reﬂectivity and transmissivity. In this section

we construct a simple Renderman-compliant shader using just these

features. Readers who have experience with shaders will know how

to extend this one immediately.

The shader exploits that fact that the Renderman interface al-

ready provides a built-in Fresnel quantity calculator, which pro-

vides R,T,ˆ

nr, and ˆ

ntusing the surface normal, incident direction

vector, and index of refraction. The shader for the air-to-water case

is as follows:

surface watercolorshader(

color upwelling = color(0, 0.2, 0.3);

color sky = color(0.69,0.84,1);

color air = color(0.1,0.1,0.1);

float nSnell = 1.34;

float Kdiffuse = 0.91;

string envmap = "";

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30 40 50 60 70 80 90

Incidence Angle (degrees)

Reflectivity from Below

Incident from Above

Figure 25: Reﬂectivity for light coming from below the water sur-

face, as a function of the angle of incidence of the light.

)

{float reflectivity;

vector nI = normalize(I);

vector nN = normalize(Ng);

float costhetai = abs(nI . nN);

float thetai = acos(costhetai);

float sinthetat = sin(thetai)/nSnell;

float thetat = asin(sinthetat);

if(thetai == 0.0)

{reflectivity = (nSnell - 1)/(nSnell + 1);

reflectivity = reflectivity * reflectivity;

}

else

{float fs = sin(thetat - thetai)

/ sin(thetat + thetai);

float ts = tan(thetat - thetai)

/ tan(thetat + thetai);

reflectivity = 0.5 * ( fs*fs + ts*ts );

}

vector dPE = P-E;

float dist = length(dPE) * Kdiffuse;

dist = exp(-dist);

if(envmap != "")

{sky = color environment(envmap, nN);

}

Ci = dist * ( reflectivity * sky

+ (1-reflectivity) * upwelling )

+ (1-dist)* air;

}

There are two contributions to the color: light coming downward

onto the surface with the default color of the sky, and light coming

upward from the depths with a default color. This second term will

be discussed in the next section. It is important for incidence angles

that are high in the sky, because the reﬂectivity is low and transmis-

sivity is high.

6 SURFACE WAVE OPTICS 3-19

Figure 26: Simulated water surface with a generic plastic surface

shader. Rendered with BMRT. Figure 27: Simulated water surface with a realistic surface shader.

Rendered with BMRT.

7 WATER VOLUME EFFECTS 3-20

This shader was used to render the image in ﬁgure 27 using the

BMRT raytrace renderer. For reference, the exact same image has

been rendered in 26 with a generic plastic shader. Note that the

realistic water shader tends to highlight the tops of the the waves,

where the angle of incidence is nearly 90 degrees grazing and the

reﬂectivity is high, while the sides of the waves are dark, where

angle of incidence is nearly 0 that the reﬂectivity is low.

7 Water Volume Effects

The previous section was devoted to a discussion of the optical be-

havior of the surface of the ocean. In this section we focus on the

optical behavior of the water volume below the surface. We begin

with a discussion of the major optical effects the water volume has

on light, followed by an introduction to color models researchers

have built to try to connect the ocean color on any given day to un-

derlying biological and physical processes. These models are built

upon many years of in-situ measures off of ships and peers. Fi-

nally, we discuss two important effects, caustics and sunbeams, that

sometimes are hard to grasp, and which produce beautiful images

when properly simulated.

7.1 Scattering, Transmission, and Reﬂection by

the Water Volume

In the open ocean, light is both scattering and absorbed by the vol-

ume of the water. The sources for these events are of three types:

water molecules, living and dead organic matter, and non-organic

matter. In most oceans around the world, away from the shore lines,

absorption is a fairly even mixture of water molecules and organic

matter. Scattering is dominated by organic matter however.

To simulate the processes of volumetric absorption and scatter-

ing, there are ﬁve quantities that are of interest: absorption coef-

ﬁcient, scattering coefﬁcient, extinction coefﬁcient, diffuse extinc-

tion coefﬁcient, and bulk reﬂectivity. All of these coefﬁcients have

units of inverse length, and represent the exponential rate of atten-

uation of light with distance through the medium. The absorption

coefﬁcient ais the rate of absorption of light with distance, the

scattering coefﬁcient bis the rate of scattering with length, the ex-

tinction coefﬁcient cis the sum of the two previous ones c=a+b,

and the diffuse extinction coefﬁcient Kdescribes the rate of loss

of intensity of light with distance after taking into account both ab-

sorption and scattering processes. The connection between Kand

the other parameters is not completely understood, in part because

there are a variety of ways to deﬁne Kin terms of operational mea-

surements. Different ways change the details of the dependence.

However, there is a condition called the asymptotic limit at very

deep depths in the water, at which all operational deﬁnitions of K

converge to a single value. This asymptotic value of Khas been

modeled in a variety of ways. There is a mathematically precise

result that the ratio K/c depends only on b/c, the single scatter

albedo, and some details of the angular distribution of individual

scattering distributions. Figure 28 is an example of a model of

K/c for reasonable water conditions. Models have been gener-

ated for the color dependence of K, most notably by Jerlov. In

1990, Austin and Petzold performed a revised analysis of spectral

models, including new data, to produce reﬁned models of Kas a

function of color. For typical visible light conditions in the ocean,

Kranges in value from 0.03/meter to 0.1/meter. It is generally true

that a < K < c.

One way to interpret these quantities for a simulation of water

volume effects is as follows:

1. A ray of sunlight enters the water with intensity I(after los-

ing some intensity to Fresnel transmission). Along a path un-

derwater of a length s, the intensity at the end of the path is

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1

K/c

b/c

1-x + sqrt(mu*x*(1-x))

1-x

Figure 28: Dependence of the Diffuse Extinction Coefﬁcient on the

Single Scatter Albedo, normalized to the extinction.

Iexp(−cs), i.e. the ray of direct sunlight is attenuated as fast

a possible.

2. Along the path through the water, a fraction of the ray is scat-

tered into a distribution of directions. The strength of the scat-

tering per unit length of the ray is b, so the intensity is propor-

tional to bI exp(−cs).

3. The light that is scattered out of the ray goes through poten-

tially many more scattering events. It would be nearly im-

possible to track all of them. However, the sum whole out-

come of this process is to attenuate the ray along the path from

the original path to the camera as bI exp(−cs) exp(−Ksc),

where scis the distance from the scatter point in the ocean to

the camera.

A ﬁfth quantity of interest for simulation is the bulk reﬂectivity

of the water volume. This is a quantity that is intended to allow

us to ignore the details of what is going on, treat the volume as a

Lambertian reﬂector, and compute a value for bulk reﬂectivity. That

number is sensitive to many factors, including wave surface condi-

tions, sun angle, water optical properties, and details of the angular

spread. Nevertheless, values of reﬂectivity around 0.04 seem to

agree well with data.

7.2 The Underwater POV: Refracted Skylight,

Caustics, and Sunbeams

Now that we have underwater optical properties at hand, we can

look at two important phenomena in the ocean: caustics and sun-

beams.

7.2.1 Caustics

Caustics, in this context, are a light pattern that is formed on sur-

faces underwater. Because the water surface is not ﬂat, groups of

REFERENCES 3-21

Figure 29: Rendering of a caustic pattern at a shallow depth (5

meters) below the surface.

light rays incident on the surface are either focussed or defocussed.

As a result, a point on a ﬁctitious plane some depth below the ocean

surface receives direct sunlight from several different positions on

the surface. The intensity of light varies due to the depth, orig-

inal contrast, and other factors. For now, lets write the intensity

of the pattern as I=Ref I0, with I0as the light intensity just

above the water surface. The quantity Ref is the scaling factor that

varies with position on the ﬁctitious plane due to focussing and de-

focussing of waves, and is called a caustic pattern. Figure 29 shows

an example of the caustic pattern Ref. Notice that the caustic pat-

tern exhibits ﬁlaments and ring-like structure. At a very deep depth,

the caustic pattern is even more striking, as shown in ﬁgure 30.

One of the important properties of underwater light that produce

caustic patterns is conservation of ﬂux. This is actually a simple

idea: suppose a small area on the ocean surface has sunlight passing

through it into the water, with intensity Iat the surface. As we

map that area to greater depths, the amount of area within it grows

or shrinks, but most likely grows depending on whether the area

is focussed or defocussed. The intensity at any depth within the

water is proportional to inverse of the area of the projected region.

Another way of saying this is that if a bundle of light rays diverges,

their intensities are reduced to keep the product of intensity time

area ﬁxed.

Simulated caustic patterns can actually be compared (roughly)

with real-world data. In a series of papers published throughout the

1970’s, 1980’s, and into the 1990’s, Dera and others collected high-

speed time series of light intensity[21]. As part of this data collec-

tion and analysis project, the data was used to generate a probability

distribution function (PDF) for the light intensity. Figure 31 shows

two PDFs taken from one of Dera’s papers. The two PDF’s were

collected for different surface roughness conditions: rougher wa-

ter tended to suppress more of the high magnitude ﬂuctuations in

intensity.

Figure 32 shows the pdf at two depths from a simulation of the

ocean surface. These two sets do not match Dera’s measurements

because of many factors, but most importantly because we have

not simulated the environmental conditions and instrumentation in

Dera’s experiments. Nevertheless, the similarity of ﬁgure 32 with

Figure 30: Rendering of a caustic pattern at great depth (100 me-

ters) below the surface.

Figure 31: PDF’s as measured by Dera in reference [21].

Dera’s data is an encouraging point of information for the realism

of the simulation.

7.2.2 Godrays

Underwater sunbeams, also called godrays, have a very similar ori-

gin to caustics. Direct sunlight passes into the water volume, fo-

cussed and defocussed at different points across the surface. As the

rays of light pass down through the volume, some of the light is

scattered in other directions, and a fraction arrives at the camera.

The accumulated pattern of scattered light apparent to the camera

are the godrays. So, while caustics are the pattern of direct sun-

light that penetrates down to the ﬂoor of a water volume, sunbeams

are scattered light coming from those shafts of direct sunlight in

the water volume. Figure 33 demonstrates sunbeams as seen by a

camera looking up at it.

References

[1] Jos Stam, “Stable Fluids,” Proceedings of the 26th annual

conference on Computer graphics and interactive techniques,

pp 121-128, (1999).

REFERENCES 3-22

-12

-10

-8

-6

-4

-2

0123456

PDF

Intensity / Mean Intensity

"shallowcausticdata.txt"

"deepcausticdata.txt"

Figure 32: Computed Probability Density Function for light inten-

sity ﬂuctuations in caustics. (upper curve) shallow depth of 2 me-

ters; (lower curve) deep depth of 10 meters.

Figure 33: Rendering of sunbeams, or godrays, as seen looking

straight up at the light source.

[2] Nick Foster and Ronald Fedkiw, “Practical animation of liq-

uids,” Proceedings of the 28th annual conference on Com-

puter graphics and interactive techniques, pp 23-30, (2001).

[3] B.A. Cheeseman and C.P.R. Hoppel,“SIMULATING THE

BALLISTIC IMPACT OF COMPOSITE STRUCTURAL

ARMOR,” http://www.asc2002.com/summaries/h/HP-08.pdf

[4] Simon Premoze, “Particle-Based Simulation of Fluids,” Euro-

graphics 2003, Vol 22, No. 3.

[5] Jeff Odien, “On the Waterfront”, Cinefex, No. 64, p 96,

(1995)

[6] Ted Elrick, “Elemental Images”, Cinefex, No. 70, p 114,

(1997)

[7] Kevin H. Martin, “Close Contact”, Cinefex, No. 71, p 114,

(1997)

[8] Don Shay, “Ship of Dreams”, Cinefex, No. 72, p 82, (1997)

[9] Kevin H. Martin, “Virus: Building a Better Borg”, Cinefex,

No. 76, p 55, (1999)

[10] Simon Premoˇ

ze and Michael Ashikhmin, “Rendering Natu-

ral Waters,” Eighth Paciﬁc Conference on Computer Graphics

and Applications, October 2000.

[11] Gary A. Mastin, Peter A. Watterger, and John F. Mareda,

“Fourier Synthesis of Ocean Scenes”, IEEE CG&A, March

1987, p 16-23.

[12] Alain Fournier and William T. Reeves, “A Simple Model of

Ocean Waves”, Computer Graphics, Vol. 20, No. 4, 1986, p

75-84.

[13] Darwyn Peachey, “Modeling Waves and Surf”, Computer

Graphics, Vol. 20, No. 4, 1986, p 65-74.

[14] Blair Kinsman, Wind Waves, Their Generation and Propaga-

tion on the Ocean Surface, Dover Publications, 1984.

[15] S. Gran, A Course in Ocean Engineering,De-

velopments in Marine Technology No. 8, El-

sevier Science Publishers B.V. 1992. See also

http://www.dnv.no/ocean/bk/grand.htm

[16] Dennis B. Creamer, Frank Henyey, Roy Schult, and Jon

Wright, “Improved Linear Representation of Ocean Surface

Waves.” J. Fluid Mech, 205, pp. 135-161, (1989).

[17] Seibert Q. Duntley, “Light in the Sea,” J. Opt. Soc. Am., 53,2,

pg 214-233, 1963.

[18] Curtis D. Mobley, Light and Water: Radiative Transfer in

Natural Waters, Academic Press, 1994.

[19] Selected Papers on Multiple Scattering in Plane-Parallel At-

mospheres and Oceans: Methods, ed. by George W. Kattawar,

SPIE Milestones Series, MS 42, SPIE Opt. Eng. Press., 1991.

[20] R.W. Austin and T. Petzold, “Spectral Dependence of the Dif-

fuse Attenuation Coefﬁcient of Light in Ocean Waters: A Re-

examination Using New Data,” Ocean Optics X, Richard W.

Spinrad, ed., SPIE 1302, 1990.

[21] Jerzy Dera, Slawomir Sagan, Dariusz Stramski, “Focusing of

Sunlight by Sea Surface Waves: New Measurement Results

from the Black Sea,” Ocean Optics XI, SPIE Proceedings,

1992.

8 APPENDIX: SAMPLE CODE FOR INTERACTIVE WATER SURFACES 3-23

[22] Jerry Tessendorf, “Interactive Water Surfaces,” Game Pro-

gramming Gems 4 , ed. Andrew Kirmse, Charles River Media,

(2004).

[23] St´

ephan T. Grilli home page,

http://131.128.105.170/ grilli/.

[24] AROSS - Airborne Remote Optical Spotlight System,

http://www.aross.arete-dc.com

8 Appendix: Sample code for interactive

water surfaces

The code listed in this appendix is a working implementation of the

iWave algorithm. As listed below, iwave paint looks like a crude

paint routine. The user can paint in two modes. When iwave paint

starts up, it begins running an iWave simulation on a small grid

(200x200 with the settings listed). This grid is small enough that

the iWave simulation runs interactively on cpus around 1 GHz and

faster. The running of the simulation is not apparent since there are

no disturbing waves at start up. The interface looks like ﬁgure 34.

In the default mode at start up, the painting generates obstructions

that show up in black and block water waves. Figure 35 shows

an example obstruction that has been painted. Hitting the ’s’ key

changes the painting mode to painting a source disturbance on the

water surface. As you paint the disturbance, iwave paint propagates

the disturbance, reﬂecting off of any obstructions you may have

painted. Figure 36 shows a frame after source has been painted

inside the obstructed and allows a brief period of time to propagate

inside the obstruction and exit, creating a diffraction pattern at the

mouth of the obstruction.

A few useful keyboard options:

oPuts the paint mode into obstruction painting. This may be se-

lected at any time.

sPuts the paint mode into source painting. This may be selected

at any time.

//------------------------------------------------

//

// iwave_paint

//

// demonstrates the generation and interaction of

// waves around objects by allowing the user to

// paint obstructions and source, and watch iwave

// propagation.

//

// author: Jerry Tessendorf

// jerry@finelightvisualtechnology.com

// August, 2003

//

// This software is in the public domain.

//

//-------------------------------------------------

//-------------------------------------------------

//

// usage:

//

// iwave_paint is an interactive paint program

// in which the user paints on a water surface and

// the waves evolve and react with obstructions

// in the water.

//

// There are two paint modes. Typing ’o’ puts the

// program in obstruction painting mode. When you

// hold down the left mouse button and paint, you

// will see a black obstruction painted. This

// obstruction may be any shape.

//

// Typing ’s’ puts the program in source painting

// mode. Now painting with the left mouse button

// down generates a source disturbance on the water

// surface. The waves it produces evolve if as you

// continue to paint. The waves bounce of off any

// obstructions that have been painted or are

// subsequently painted.

//

// Typing ’b’ clears all obstructions and waves.

//

// Typing ’=’ and ’-’ brightens and darkens the display

// of the waves.

//

// Pressing the spacebar starts and stops the wave

// evolution. While the evolution is stopped, you

// can continue painting obstructions.

//

//

// This code was written and runs under Linux. The

// compile command is

//

// g++ iwave_paint.C -O2 -o iwave_paint -lglut -lGL

//

//

//-------------------------------------------------

#include <cmath>

8 APPENDIX: SAMPLE CODE FOR INTERACTIVE WATER SURFACES 3-24

Figure 34: The iwave paint window at startup.

Figure 35: The iwave paint window with obstruction painted.

Figure 36: The iwave paint window after painting source inside the

obstruction and letting it propagate out.

#ifdef __APPLE__

#include <GLUT/glut.h>

#else

#include <GL/gl.h> // OpenGL itself.

#include <GL/glu.h> // GLU support library.

#include <GL/glut.h> // GLUT support library.

#endif

#include <iostream>

using namespace std;

int iwidth, iheight, size;

float *display_map;

float *obstruction;

float *source;

float *height;

float *previous_height;

float *vertical_derivative;

float scaling_factor;

float kernel[13][13];

int paint_mode;

enum{ PAINT_OBSTRUCTION, PAINT_SOURCE };

bool regenerate_data;

bool toggle_animation_on_off;

float dt, alpha, gravity;

float obstruction_brush[3][3];

float source_brush[3][3];

int xmouse_prev, ymouse_prev;

//--------------------------------------------------------

//

// Initialization routines

//

//

// Initialize all of the fields to zero

void Initialize( float *data, int size, float value )

{for(int i=0;i<size;i++ ) { data[i] = value; }

}

// Compute the elements of the convolution kernel

void InitializeKernel()

{double dk = 0.01;

double sigma = 1.0;

double norm = 0;

for(double k=0;k<10;k+=dk)

{norm += k*k*exp(-sigma*k*k);

}

for( int i=-6;i<=6;i++ )

{

for( int j=-6;j<=6;j++ )

{double r = sqrt( (float)(i*i + j*j) );

double kern = 0;

for( double k=0;k<10;k+=dk)

{kern += k*k*exp(-sigma*k*k)*j0(r*k);

}kernel[i+6][j+6] = kern / norm;

}}

}

void InitializeBrushes()

{obstruction_brush[1][1] = 0.0;

obstruction_brush[1][0] = 0.5;

obstruction_brush[0][1] = 0.5;

obstruction_brush[2][1] = 0.5;

obstruction_brush[1][2] = 0.5;

obstruction_brush[0][2] = 0.75;

obstruction_brush[2][0] = 0.75;

obstruction_brush[0][0] = 0.75;

obstruction_brush[2][2] = 0.75;

source_brush[1][1] = 1.0;

source_brush[1][0] = 0.5;

source_brush[0][1] = 0.5;

source_brush[2][1] = 0.5;

source_brush[1][2] = 0.5;

source_brush[0][2] = 0.25;

source_brush[2][0] = 0.25;

source_brush[0][0] = 0.25;

source_brush[2][2] = 0.25;

}

void ClearObstruction()

{for(int i=0;i<size;i++ ){ obstruction[i] = 1.0; }

}

void ClearWaves()

{for(int i=0;i<size;i++ )

{

8 APPENDIX: SAMPLE CODE FOR INTERACTIVE WATER SURFACES 3-25

height[i] = 0.0;

previous_height[i] = 0.0;

vertical_derivative[i] = 0.0;

}

}

//----------------------------------------------------

void ConvertToDisplay()

{for(int i=0;i<size;i++ )

{display_map[i] = 0.5*( height[i]/scaling_factor + 1.0 )*obstruction[i];

}

}

//----------------------------------------------------

//

// These two routines,

//

// ComputeVerticalDerivative()

// Propagate()

//

// are the heart of the iWave algorithm.

//

// In Propagate(), we have not bothered to handle the

// boundary conditions. This makes the outermost

// 6 pixels all the way around act like a hard wall.

//

void ComputeVerticalDerivative()

{// first step: the interior

for(int ix=6;ix<iwidth-6;ix++)

{for(int iy=6;iy<iheight-6;iy++)

{int index = ix + iwidth*iy;

float vd = 0;

for(int iix=-6;iix<=6;iix++)

{for(int iiy=-6;iiy<=6;iiy++)

{

int iindex = ix+iix + iwidth*(iy+iiy);

vd += kernel[iix+6][iiy+6] * height[iindex];

}

}

vertical_derivative[index] = vd;

}}

}

void Propagate()

{// apply obstruction

for( int i=0;i<size;i++ ) { height[i] *= obstruction[i]; }

// compute vertical derivative

ComputeVerticalDerivative();

// advance surface

float adt = alpha*dt;

float adt2 = 1.0/(1.0+adt);

for( int i=0;i<size;i++ )

{float temp = height[i];

height[i] = height[i]*(2.0-adt)-previous_height[i]-gravity*vertical_derivative[i];

height[i] *= adt2;

height[i] += source[i];

height[i] *= obstruction[i];

previous_height[i] = temp;

// reset source each step

source[i] = 0;

}

}

//------------------------------------------

//

// Painting and display code

//

void resetScaleFactor( float amount )

{scaling_factor *= amount;

}

void DabSomePaint( int x, int y )

{int xstart = x - 1;

int ystart = y - 1;

if( xstart < 0 ){ xstart = 0; }

if( ystart < 0 ){ ystart = 0; }

int xend = x + 1;

int yend = y + 1;

if( xend >= iwidth ){ xend = iwidth-1; }

if( yend >= iheight ){ yend = iheight-1; }

if( paint_mode == PAINT_OBSTRUCTION )

{for(int ix=xstart;ix <= xend; ix++)

{for( int iy=ystart;iy<=yend; iy++)

{int index = ix + iwidth*(iheight-iy-1);

obstruction[index] *= obstruction_brush[ix-xstart][iy-ystart];

}}

}

else if( paint_mode == PAINT_SOURCE )

{for(int ix=xstart;ix <= xend; ix++)

{for( int iy=ystart;iy<=yend; iy++)

{int index = ix + iwidth*(iheight-iy-1);

source[index] += source_brush[ix-xstart][iy-ystart];

}}

}

return;

}

//----------------------------------------------------

//

// GL and GLUT callbacks

//

//----------------------------------------------------

void cbDisplay( void )

{glClear(GL_COLOR_BUFFER_BIT );

glDrawPixels( iwidth, iheight, GL_LUMINANCE, GL_FLOAT, display_map );

glutSwapBuffers();

}

// animate and display new result

void cbIdle()

{if( toggle_animation_on_off ) { Propagate(); }

ConvertToDisplay();

cbDisplay();

}

void cbOnKeyboard( unsigned