IMPROVING THE DISPLAY OF WIND PATTERNS AND OCEAN CURRENTS

The theory of human pattern perception is applied to the portrayal of winds, waves, and
ocean currents, resulting in significant improvements in interpretation.
while a great deal of effort has gone into
building numerical weather and ocean
prediction models during the past 50 years,
less effort has gone into the visual representation
of output from those forecast models and many of
the techniques used are known to be ineffective. In
particular, the representation of vector fields (winds,
currents, or waves) is almost always done using grid
patterns of small arrows or wind barbs (Fig. 1) despite
studies showing other methods to be far superior
(Laidlaw et al. 2005; Pilar and Ware 2013). Other
methods such as streamlines have been available for
decades (e.g., Saucier 1955) but are rarely used. Yet
representation is critical; ultimately, visualization
is the only viable method for interpreting complex
patterns of winds and currents as well as for scalar
fields such as pressure and temperature. The need for
improved visualization methods becomes even more
significant with the continued increases in the spatial
resolution and data density of numerical ocean and
weather forecast models.
Most prior research into the visualization of
ocean currents and winds has focused on applying
new visualization techniques to the model output.
Examples include methods for creating equally
spaced streamlines (Turk and Banks 1996; Jobard
and Lefer 1997), applying the technique of line inte-
gral convolution (Jobard et al. 2002) whereby digital
noise patterns are advected in the direction of flow,
and applying volume rendering methods (Max et al.
1993; Kniss et al. 2002) to 3D flow model output.
Others have examined the technical problem of
dealing with nested grids commonly used with flow
models (Treinish 2000) or time-varying gridded data
(Doleisch et al. 2005). There have also been tour de
force designs such as Baker and Bushell’s (1995) care-
fully crafted, unique representation of a storm cloud
done in consultation with Edward Tufte (Tufte 1997).
In contrast, relatively little effort has gone into
formal evaluation aimed at comparing different por-
trayal methods. An exception is Laidlaw et al.’s (2005)
study of six different alternative representations of the
same flow pattern. Among other things, this revealed
equally spaced streamlines to be more effective than
arrow grids. Another study by Martin et al. (2008)
showed systematic biases in the perception of wind
direction when grids of wind barbs were used to show
spiral patterns around a hurricane center. These stud-
ies provide invaluable insights and are first steps in
placing flow representation on a scientific footing.
But their findings are difficult to generalize. There
are many ways that arrows can be used to show flow,
for example. The arrows can have different spacing,
lengths, widths, and colors. Perhaps the result of
Laidlaw et al.’s (2005) study would have been differ-
ent with different kinds of arrows. The solution to
this problem is to develop a theory of effective flow
representation. Such a theory can guide designs, and
it can be tested and refined by means of experiments
with human participants. We contend that such a
theory should be based primarily on the science of
human perception.
The effectiveness of a data display depends on
how well critical patterns can be perceived and vision
science can tell us something about this. In this
paper, we outline the perceptual principles for what
makes a good representation of a 2D vector field and
show how these principles can be used in design. We
present a number of displays that we have designed
according to these principles and evaluated in vari-
ous ways using forecast guidance from the National
Oceanic and Atmospheric Administration (NOAA)
and U.S. Navy operational numerical weather and
oceanographic forecast modeling systems.
PERCEPTUAL PRINCIPLES FOR REPRE-
SENTING VECTOR FIELDS. From the point of
view of understanding what makes a flow visualiza-
tion effective, the most important part of the brain
is the primary visual cortex (V1). Fortunately, more
than 50 years of neuroscience research has investi-
gated the operation and function of this part of the
brain. It is here that the incoming signal from the
optic nerves of the two eyes is processed in parallel by
several billion neurons (there are only a million fibers
in each optic nerve) (Fig. 2a). Figure 2b illustrates a
slice through a small section of cortical tissue in V1
showing the functional structure. This is a synthesis
from hundreds of experiments (Livingstone and
Hubel 1988) revealing regions that process the signal
in three distinct ways; some areas break down the
incoming information into local color difference in-
formation, other areas contain neurons that respond
preferentially to moving patterns, and a third type of
area breaks down the information into local orienta-
tion and size information providing the elements of
both form perception and texture perception. The
regions provide different visual channels that separate
different kinds of visual information. It is important
to emphasize the parallel nature of this processing,
whereby every part of the image falling on the retina
is simultaneously decomposed through these mecha-
nisms; the image is broken down in terms of color
differences, moving elements, and the basic form and
shape perception (local orientation and size).
There is general consensus that the orientation
detectors in V1 form part of a contour detection
mechanism that is critical for detecting the boundar-
ies of objects. Neurons that are smoothly aligned tend
to mutually reinforce one another, whereas those that
are not aligned are mutually inhibitory (Field et al.
1993). The result is a kind of winner-take-all effect
for aligned contour segments: they stand out clearly
whereas nonaligned segments are deemphasized.
This mechanism both reinforces the perception of
smoothly varying continuous contours and sharpens
up orientation tuning as illustrated in Fig. 2c.
Central to our theory is the following principle:
to show flow orientation clearly, a display should be
designed so that, as far as possible, all neurons that
encode orientation should signal orientations that are
tangential to the flow direction. Responses that are
not tangential to the flow direction will lead to incor-
rect judgments of flow orientation (Ware 2008). If we
consider the use of short line segments to represent
a vector field, then this theory predicts that certain
arrangements of the lines will be more effective than
others. Many visualizations of flow use a simple grid of
arrows or wind barbs to show the vec-
tor field (e.g., Trafton and Hoffman
2007). Figure 3a suggests that this
will not be as effective in stimulating
tangential responses as other solu-
tions. Also, an arrow grid will cause
a neural response to the grid itself,
an irrelevant distraction. Arranging
arrows so that they are smoothly
aligned (Fig. 3b) will be better, but
best of all will be a visualization
consisting of continuous streamlines
(Fig. 3c). This theoretical proposition
has been supported both by experi-
ments with humans and by models
that computationally simulate the
processing of contours in the human
visual cortex (Pineo and Ware 2010).SHOWING THE VECTOR SIGN. It is common
to decompose vectors into components of direction
and magnitude (speed in the case of flow pattern).
When discussing the visualization of vector fields it
is useful to further decompose direction into orienta-
tion and sign as shown in Fig. 4.
A continuous contour, such as a streamline, tan-
gential to flow direction may provide the best way
of showing flow orientation but it is still ambiguous
with respect to direction. To resolve the directional
ambiguity some form of asymmetry along the con-
tour is needed. Arrowheads are a common way of
providing along-contour asymmetry, but they are
likely not the best way. A neural mechanism that can
encode directionality is a type of V1 neuron called an
end stopped cell, and such cells respond best to linesthat terminate in the receptive field of the cell from
a particular direction. While a conventional arrow
will provide some asymmetry of response, because
the head will provoke a stronger response than the
tail, a stronger asymmetrical response will come
from other patterns. Examples are given in Fig. 5b.
Such glyphs are not new; Tufte (1983) reproduces a
map of North Atlantic currents drawn by Sir Edmund
Halley in 1686, using elongated teardrop streamlets
arranged head to tail in streamlines. However, use of
these types of glyphs is almost unknown in modern
practice.
STREAMLETS TO SHOW OCEAN
CURRENT PATTERNS. The theory we have
outlined suggests that in order to show a vector field
map, the best solution will be to use a dense pattern
of streamlines and along each streamline place
elements that have a much stronger head than tail.
To put this theory into practice, we implemented
Jobard and Lefer’s (1997) algorithm to create equally
spaced streamlines. Along the streamlines are placed
streamlets, graphical elements that have a much more
salient head than tail.
We carried out a human-in-the-loop optimization
study to determine values for the
remaining free parameters, such as
how to represent flow speed, how to
space the streamlines, the head and
tail size, and the head and tail trans-
parency of the streamlets (Mitchell
et al. 2009). Our interface had a set
of interactive sliders enabling study
participants to adjust each of the 22
parameters controlling the map-
ping of the data to a display, starting
from random values. Participants
were instructed to produce the best
representation they could. Some
participants were designers and
others meteorologists. A portrayal
method based on the best results
from the study were integrated into our FlowVis2D software, a package written in C++
and OpenGL for rendering currents from ocean
model output or winds from weather model output.
The result is shown in Fig. 6, in this instance illus-
trating surface water currents from the U.S. Navy
Operational Coastal Ocean Model (NCOM). It is also
used to portray forecast guidance from the NOAA/
National Ocean Service’s (NOS) estuarine and Great
Lakes oceanographic forecast modeling systems
and has been available on NOS’ nowCOAST portal
(http://nowcoast.noaa.gov) since 2009.
MULTIVARIATE METEOROLOGICAL
DISPLAY. In meteorological displays, a major chal-
lenge is to simultaneously show scalar variables, such
as atmospheric pressure and surface air temperatures,
together with wind patterns. To meet this challenge,
we took advantage of the perceptual channel theory
outlined in our introduction. The key design idea
is to use different visual channels to show differ-
ent types of information and thereby reduce visual
interference between the layers. We began with the
following mappings:
Temperature  color channel
Atmospheric pressure  texture channel
Wind speed and direction  motion channel
Wind patterns are represented using a pattern of
10,000 animated streamlets. To represent pressure, we
used a series of graduated textures in addition to con-
tours. To represent surface air temperatures we used a
fairly conventional color sequence with different color
bands every 5 degrees. The result is shown in Fig. 7
(without animation). In our evaluation, we measured a subject’s ability to accurately judge temperature,
pressure, and wind speed and direction, comparing
our new solution with a more conventional alterna-
tive (Fig. 8a), a glyph-based alternative (Fig. 8b), and
a nonanimated version of Fig. 7 (Ware and Plumlee
2012). The results showed our animated design to
be perceptually more accurate than the others in
the representation of wind direction and speed. It
was also judged to be greatly superior in terms of
how well specific wind patterns could be seen, such
as weather fronts and the circulation around a low
pressure center.
A BETTER WIND BARB. A common graphical
device for showing wind speed and direction is the
wind barb. Wind barbs were originally designed to represent wind speed and direction at observing
platforms on a surface weather map in a way that
can be directly read by someone familiar with sta-
tion model symbology. However, winds barbs are
not well designed to show patterns of winds such as
those produced by meteorological models or analysis
systems. The perceptual problem with wind barbs is
that only the very tip of the barb is tangential to the
wind direction, and therefore most of the contours
in the glyph are oriented in nontangential directions.
We undertook a study to both design and evaluate
alternatives to the classic wind barb with the goal of
combining the best feature of the wind barb, display-
ing speed in a readable form, with the best feature
of streamlines, showing wind patterns clearly. Two
of our most successful designs are shown in Fig. 9.
Our first solution used the
Jobard and Lefer (1997)
algorithm to generate
equally spaced stream-
lines and placed wind barbs
with curved shafts along
streamlines. This improves
the ability to see patterns,
but the feathers of the barb
still produce significant
visual interference because
they are not tangential
to the flow direction. In
a more radical redesign
(Fig. 9c), arrowheads of
different sizes and styles
are used to replace the barb
bars, showing 5-, 10-, and
50-kt increments (1 kt =
0.51 m s –1 ). This has the ad-
vantage of symmetry about
the streamlines producing
less systemic visual bias to the flow direction. It also
allows for streamlines to be placed somewhat closer
together, allowing for more details to be shown. Our
evaluation showed the new designs to be superior in
their ability to accurately show the wind speed and
direction (Pilar and Ware 2013). We also evaluated
the new designs in their ability to represent wind
patterns. To do this, we artificially created six differ-
ent wind patterns (two are shown in Fig. 10) embed-
ded in a larger-scale smooth flow. Study participants
were required to guess which of the six they were
seeing under various display conditions, including
the four that are illustrated. The results showed errors
reduced by about 70% in comparison with the grid
of wind barbs.
REPRESENTING WAVE PATTERNS. Our
final example is an extension of the alternative wind
barb design. As illustrated in Fig. 11, we developed
a quantitative glyph to show wave height and the
direction of travel as forecast by a NWS wave model.
This encodes information in a manner similar to a
wind barb using symbolic bars and triangles. Wind
information is also shown using our redesign of wind
barbs. Two perceptual methods are used to minimize
the visual interference between wave information
and wind information. Because mariners are often
interested in the angle of wave fronts, as opposed
to the direction of travel, we draw contours that are
tangential to wave fronts, and orthogonal to direction
of travel. This tends to minimize visual interference between wind and wave
patterns because, most of
the time, wave fronts are
roughly orthogonal to wind
direction. Also, perceptual
research shows that graphi-
cal elements that are coun-
terphase to the background
in terms of lightness can be
easily separated (Theeuwes
and Kooi 1994), so we use
black for the wave informa-
tion and white for the wind
information.
CONCLUSIONS AND
RECOMMENDATIONS.
Our experience suggests
that an understanding of
basic perceptual processes
can help in the design of
clear and effective visual-
ization of meteorological
and oceanographic analyses
and forecast model guid-
ance. But perceptual theory
can only motivate promis-
ing approaches; it cannot
be used to specify detailed
design solutions. A kind
of cognitive task analysis
is required for a successful
design. This involves determining the goals of the
visual data representation. What patterns are likely
to be most important for the user? Visualizations are always tradeoffs. If only wind patterns were
important, then a much denser mesh of animated
particle traces could have been used in Fig. 7. If it were only necessary to see atmospheric pressure and not
temperature too, then color might have been used to
represent pressure. The relative salience of different
features must be carefully tuned in the design process
to meet the design goals.
Ideally evaluation will also be part of the design
process. Both formal and informal experiments with
human participants are useful both to set parameter
values for the mapping from data to display and to
compare a new design against existing alternatives.
For most of our studies we took the most basic tasks to
be judgments of wind orientation, direction, or speed.
Additional research is needed to discover the optimal
way of bringing out patterns such as wind shear at a
front, or the branching of the jet stream. However, so
long as tasks can be understood and defined, percep-
tual theory can be applied to the problem.
One of the more difficult problems in designing
effective wind, current, or wave visualizations is
dealing with the scale of the map. A great advantage
of color coding values such as wind speed or surface
temperature is that color tends to scale well. This
is because with a well-chosen color scheme, even if
certain details cannot be seen, their colors will blend
in the visual receptors to something approximating
an average. This is not the case for vector representa-
tions, whether conventional arrows are used or one
of the methods advocated here. There is an optimal element spacing for show-
ing the greatest amount of
detail; if the spacing is too
small, patterns will become
invisible, if too large, detail
will be lost. The representa-
tion must therefore change
with scale.
The development, main-
tenance, and operation of
weather and oceanographic
forecast modeling systems
and their underlying nu-
merical three-dimensional
models is a hugely expen-
sive undertaking but the
cost is justified by the high
value of the data. Some of
the resulting products are
only viewed by specialists,
whereas others are seen by
millions who have a more
casual interest. In either
case the visual portrayal of
the output of these models
deserves substantial effort because this is usually the
only way model results can be interpreted. With a
poor visualization, much of the information may be
lost and a proportionate amount of modeling effort
and operational costs wasted. Designing and evaluat-
ing representations for perceptual efficiency is not a
trivial undertaking, but it is worth the effort.
ACKNOWLEDGMENTS. Funding for this project was
provided by NOAA Grant NA05NOS4001153 and by NSF
ITR Grant 0324899. The authors thank Jason Greenlaw,
Matthew Plumlee, and Roland Arsenault for their help
in improving FlowVis2D software in NOS’ nowCOAST
GIS-based web mapping portal. We also thank Matthew
Plumlee, Peter Mitchell, and Daniel Pineo, who partici-
pated in the research.