Digital Landscape Modeling and Visualization: A Research Agenda

Stephen M Ervin, Harvard Design School

48 Quincy St. Cambridge, Massachusetts, USA 02138

Abstract

Digital landscape models, whether made for purposes of 'visual inference', or for simulating and understanding behavior or other invisible aspects of the landscape, require abstractions and simplifications. Yet for many visual purposes, 'realistic' depictions are desirable. The conflicts between these two demands are substantial for landscape modelers. For the basic landscape elements -- terrain, vegetation, and water -- some standard techniques for convincing static visual representation have been developed, but many complicating questions and obstacles remain. In addition, landscapes are essentially dynamic, and digital techniques for representation of these dynamics are still in their infancy. Surveying these techniques, complications, and possibilities leads to some generalized comments about the promises and problems of landscape modeling, and to a handful of proposed research

topics to help pursue the landscape modeling agenda.

Keywords: Landscape visualization, terrain, vegetation, water, abstraction, dynamics, levels of detail

Introduction: Digital Landscape Modeling

The validity of the conclusions reached from any model is, of course, dependent upon the model's quality and characteristics. In many cases, the criterion of verisimilitude is imposed : "How 'good' is the model, and so how reliable are the inferences?" In the visual arena, this verisimilitude is often taken literally as realism, or realistic-ness, and photo-realistic images are sought after for all kinds of visualizations, visual preference surveys and other undertakings. A number of researchers have asked "Is a viewer’s response to a representation (photograph, or virtual model, etc.) of a landscape in some way equivalent to the same viewer’s (or more generally, 'experience-ers') response to the landscape in situ? And if so, or if not, what aspects of the representation (realism, color, navigability, etc.) are important?" This represents a vital and significant area of research into visualization. (Lange, 1999 , and Bishop, 1996)

But these kinds of visual questions are not the only ones that call for digital landscape models. GIS systems create maps of landscape suitability and visibility; CAD systems make three-dimensional renderings of road alignment geometry or estimates of cut-and-fill volumes; landscape ecologists use differential equations and particle systems to describe dynamic processes like vegetation succession. For these, vision and visualization -- and even the details of human perception -- are important to be sure, but the visual judgment in these cases is not limited to ‘looks like’ or ‘I like it’ (or not!) The judgments expected upon viewing the displayed results of a model are the inferential purpose for which the model is made; and as with all representations, these inferential -- and sometimes rhetorical -- purposes turn out to be critical in devising and evaluating models. (Arnheim, 1969)

While both of these kinds of models -- those meant for generation of images, and those meant for other kinds of inferential purpose -- have been extensively developed and explored in the past thirty years, and are capable of surprisingly informative, and sometimes even realistic simulations of landscapes and landscape elements, a number of vexing problems remain. Many of these have to do with realism of a sort, but are not limited to visual photo-realism. (Photographs, after all, are highly tailored abstractions too, with a number of carefully controlled variables.) In the effort to model and visualize landscapes, we need to seek a balance between 'looks like' and 'acts like'. Landscape modelers are continually confronted with tensions and trade-offs between these two attributes, and these tensions and trade-offs are the subject of the rest of this article.

The next section presents, in a manner hopefully accessible to any interested reader, an overview of some of the more common techniques of landscape modeling that are embodied in current systems and approaches; the final section highlights a number of remaining problems, formulated as a proposed research and development agenda which is intended for active (and would-be) digital landscape modelers.

Landscape Elements

Landscape(s) can be understood as usually composed of six essential elements, in combination:

The first four of these are the traditional palette of landscape design; the fifth must be included because of a broader ecological view, as well as in recognition of the fact that much modeling is done for the benefit of humans, one way or another; and the sixth is both pervasive and critical for visualizations and renderings of digital landscape models.

Each of these elements presents its own challenges in modeling, and fertile areas for research and development. Real as well as synthetic landscapes are almost always a combination of some or all of these elements together (see Figure 1, for example.)

The base layer of most landscapes is the landform, or terrain. In nature an undulating, sometimes roughly broken surface of varying slopes and convexities, landform may also be geometrically simple, composed of tilted planes and simple curved forms, especially when designed by landscape architects or earth sculptors. The covering of this surface varies from simple repetitive surfaces such as grass, asphalt, brick or sand, to complex mixes of forest litter, or randomly sized boulders in glacial fields. In most cases a simple continuous elevation function of the form z = f (x,y) is sufficient to describe landform surface (with a factor to account for the curvature of the earth when the area involved is large enough!). Multiple coincident z-values, as from truly vertical, overhanging, or tunneled aberrations to this surface are generally rare, especially in the cases of landscapes of inhabitation, except for man-made structures such as tunnels, bridges, etc. To capture these surfaces in numeric (digital) forms, a number of conventions have arisen, from spot elevations and contour lines to 3-D meshes, ruled surfaces, triangulated surfaces and arbitrarily curved surfaces (such as those formed by 'Non-Uniform Rational B-Splines', or NURB's.) Each different data structure supports more or less well different inferences, and different rendering and visualization tasks. (Grid meshes are easy to store and transmit, but present limitations in rendering and in manipulating surface forms; NURB surfaces can be rendered to create the most smoothly shaded surfaces, and are effective for modifying entire surfaces with a few control points, but are hard to represent in a simple data structure.)

In most cases, terrain is extensive, and so leads to large data sets, but this fact can be offset by the generally smooth nature of the surface, so that coarse sub-sampling is often sufficient. Nonetheless, landform typically requires large digital data sets, and so terrain and site modeling were historically associated with GIS, rather than CAD, and assumed to require larger computers, more memory and faster processors than other modeling tasks.

Two notable algorithmic techniques have been developed to try to reduce the sheer number of points required for terrain models: Triangulated Irregular Networks (TIN's) which strategically limit the number of points required to those marking specific high spots, low spots, ridges, etc.; and recursive fractal generation routines, which artificially create an appropriately roughened surface from a small number of control points. Both of these have made terrain modeling more tractable, but also introduce their own questions and problems. E.g. "What is the optimum algorithm for determining a triangulated surface from a given set of data points (x,y,z triplets)?" "What fractal roughness coefficients are best assigned to different kinds of terrain surfaces for automatic generation?"

Vegetation

Whereas landform may be characterized as requiring an extension from 2D data to 3-D, vegetation -- trees, shrubs, groundcovers -- raise the stake even higher. No 2D representation can adequately describe the branching, volumetric, complex forms. An ordinary building might be well represented with several thousand polygons and simple geometric primitive solids, but no part of a plant is flat , square or even really cylindrical. Millions of polygons -- or greater orders of magnitude yet -- are required to begin to capture an ordinary tree. And the shapes and subshapes required are fractally complex, with detail at every level of investigation.

Perhaps the most common state of the art for vegetation modeling is the simple technique of using photographic textures applied to flat billboards, or cut-outs. This technique gives immediate photo-realism to individual tree images, but in application in more complex 3D models has several limitations. When seen side-on, these billboards disappear (or are reduced to thin lines) so the usual technique is to place these billboards on several intersecting planes. Even this, of course, fails when the tree is viewed from overhead. Also, making convincing light and shadow conditions is difficult. (Muhar, 1996)

It is possible to create a three dimensional computer model of a tree, even by hand, but it’s a daunting challenge. More promising is the development of computer programs that alogithmically generate tree-like structures: branching, twisted, tapered cylinders, even with leaves attached at the tips. (See for example, Figure 3.)

Not only are trees and other vegetation large and complex in their basic structure, they represent another dimension altogether in the landscape: that of dynamics, change over time. Plants grow on an annual cycle, sometimes changing their form dramatically, and not just by simple scaling; many change character on a seasonal cycle, shedding, re-growing and re-coloring their foliage; and often plants move and change with the diurnal solar cycle, and shake and sway with the wind. On top of all this, plants introduce the qualities of sound into the landscape, with their leaves rustling in the breeze. Modeling a tree's basic geometry is a daunting challenge; making it grow and change over time, or blow in the wind (digitally) is even more so. Some software has begun to approach these problems, adding to fractal plant-form generation, plant growth, or wind motion, and these are already essential utilities for digital landscape modelers. (TREE-PRO, AMAP)

Some tree-modeling software has the capability of modeling leaves on branches, with considerable control over the shape, color, size and structure of the foliage. Some might even ripple in the wind, with appropriate inverse kinematics equations causing the branches and twigs to blow back and forth with convincing elasticity. But none of these branches ever break and fall, nor do the leaves ever change color with the seasons, much less fall and pile in depressions and or swirl about in wind-induced funnels.

Growth of trees or shrubs -- a simple fact of the landscape, and a major contributor to visual quality over time -- is hardly addressed at all in any modern software except in the most cartoonish way. Thus, it may be possible to substitute different tree symbols (circles of different sizes in plan, or texture mapped photographs of different growth stages in elevation) as function of time, but no tree in any CAD system ever died from crowding or natural causes; and none ever grew lopsided from environmental factors.

These last are specific aspects of the 'forest vs trees' problem. Individual plants have their characteristics and modeling complexities, but in aggregation they present another whole set of different problems. Sometimes modeling the forest may sacrifice individual tree models' verisimilitude, in favor of a better, or more tractable, forest model. Trees in the far background of a view have different modeling characteristics than those in the near foreground. Individual leaves on a tree have the same characteristic (massing at one scale, individual shapes at another), as do many aspects of vegetation modeling (blades versus fields of grass, etc.) ‘SmartForest’ (Orland 1994), ‘Vantage Point’ (VANTAGEPOINT), and some other research efforts are starting to implement some growth effects and model forest stand dynamics, to generate forest visualizations (even at the level of the individual trees, see figure 4), but these efforts too are really only just beginning. (House, et al.)

Water

Compared to the patent complexity of vegetation, water seems deceptively simple. On a still morning, a lake may be modeled as a flat reflective plane, just a mirror in the landscape. Look a little closer, though, and you see refractivity and transparency mixed in. Even these are relatively straightforward mathematically, and thick glass is handled well in modern rendering systems. But water in the landscape takes many other forms, from damp spots in shade to rushing streams and waterfalls, to fog, mist and clouds that take their form from complex interactions of landform, vegetation, atmospheric currents and temperature differentials. (Ebert, 1997)

Water also represents -- both stands for and manifests -- dynamics in the landscape. Changing in both physical state and appearance in the landscape, water is the medium for ripples, splashes, waves and a host of other surface deformations. These are beginning to be modeled in software, both by detailed physical simulations (GEORGIATECH), simple mathematical formulae, and special effects designed to look like water, by acting like water. (DIGIMATION)

Finally, even more than with plants, representations of water achieve greater verisimilitude, and expressive potential, when accompanied by sound -- splashing, trickling, tinkling, roaring. Representing the underwater landscape is one enterprise in which computer graphics may be more effective than on dry land, due to the already distorted visual character of the 'aqua-scape'. And modeling water in all of its dimensions may well be a problem better suited to multimedia computers than to even the best hand renderer with pencils and oil paints.

Structures(including infrastructure)

The role played by structures in many landscapes is significant, especially in designed and built landscapes, but treatment of all the architectural and engineering issues involved is beyond the present scope. Suffice it to say that mechanical and architectural digital modeling and visualization techniques are very highly developed, but not all of the techniques that work well for gears and buildings translate well to the problems of modeling the landscape.

Animals (including people);

Similarly, animals including people are essential elements of most natural or built landscapes, even indirectly or invisibly. Modeling the appearance of animals is just as hard as other landscape elements, for many of the same reasons: fuzzy, curved, complex, dynamic features. And their behavior is all the more difficult to represent, much less comprehend! Yet there are beginning to be digital modeling tools for realistic people and some work in behavior modeling is proceeding, largely in the 'Artificial Life' community (Langton, 1988)

Atmosphere (including sun, wind, etc.)

As much as every landscape model is made up out of the previous five elements, especially for visualization purposes they are all dependent upon the atmosphere within which they are situated. In many rendering programs there are rendering parameters that substitute for the (ethereal but real) atmospheric qualities of lighting, fog , haze, and others. In this category also are included the motion of the air (wind), and other elements which are present but so distant they must be substituted for in most landscapes (the sun, the moon, clouds, e.g.).

Likewise, many rendering systems include fog or atmospheric haze as an effect. But all of these simplified versions -- diminution of brightness and contrast, and increased blue-ness, with distance of objects from the viewer -- fail to capture the multi-layered, often patchy, truly three-dimensional and often dynamic nature of real fog (and none of them help to provide water to specialized plant life found in frequently foggy environments.) In fact, atmospheric haze may be a phenomenon requiring multiple input parameters, depending upon location, lighting, terrain, microclimate, and others. (See figures 5 and 6, examples of varying atmospheric parameters.)

Generalizations

Reviewing the assortment of specifics presented above, I discern a handful of recurrent key words and phrases, which seem to apply generally across some or all specific elements, and are characteristic of digital landscape modeling complications. Each of the following words or phrases denotes a cluster of sometimes interrelated problems, questions, and research areas and topics:

(Limited) knowledge of physical laws and systems

Sheer size / magnitude problems

So digital modelers of landscape have always been frustrated by limited amounts of memory, or processor speed, or disk size, and have developed an array of tricks, techniques and work-arounds to achieve impressive results with finite resources. Many of these involve substituting 'looks like' for 'works like'; that is, they are rendering-only techniques, clever methods for using internal data structures and representations to achieve satisfactory (sometimes 'realistic') external representations, or images. Using bit-mapped textures on arbitrary simple solids and planes, for example, is an obvious example. While it solves an instant problem, the technique introduces other problems (of viewing planes edge-on, for example, or containing mis-matched lighting effects, or casting inappropriate shadows) and doesn't serve for any other kinds of modeling purpose than producing images (texture-mapped trees cant grow, or interact, e.g.).

Resolution and level-of-detail problems (LOD)

These issues are not just limited to visualization systems, or the generation of images. Many other physical / natural systems display differing effects as a function of distance and aggregation; gravity effects and other non-linear systems are found throughout the landscape.

Confronting this problem head on leads inevitably to the root consideration for any modeling enterprise: "What is the purpose, or inferential intent, of the model in the first place?" and a similarly basic question: "What constitutes the landscape, anyway?" If a visual simulation is being developed that has a fixed, or limited number of, viewer position(s), then the background can be determined a-priori. If a virtual reality model is to allow navigation anywhere, then any element may be at times in the foreground, mid-ground or background. (The VRML modeling language includes specific syntax for specifying the varying appearance of objects at different distances from the viewer -- a simplistic, but useful, first approximation of the complexity of continuously changing levels of detail and resolution required in any model.) If a model is made for other analytic purposes than just visualization, then the rules governing appropriate levels of detail are quite changed (for example, in a model of bird-nesting habitat preference, individual pine needles may not need to be represented, so much as foliage density, or some other element or abstraction.) The perceived landscape really does include elements at differing resolutions; and so does the landscape considered through other lenses than purely visual.

The mathematics of fractals has been applied to describing the recursive depths of detail which are inherent in natural phenomena, enabling an automatic generation of appropriate levels of detail. These algorithmic approaches have been used to characterize terrain, stream-bed courses, vegetation outlines, clouds and other all sorts of other landscape phenomena. In this case, numbers like parameters and coefficients take the primary role of model representation; polygons or pixels are generated only as needed. L-systems and other plant generation techniques have developed using similar techniques. All of these techniques trade off time and representational efficiency for accuracy, but they also offer a kind of verisimilitude that may be appropriate and useful, given an appropriate inferential purpose ("What might a meandering stream look like?", vs. "Where does the stream actually meander?")

Complexity / interrelationship / interaction problems

Various algorithmic and data-structuring techniques are available for organizing and representing complexity (object-oriented techniques, for example, support 'multiple-inheritance' and 'part-of' relationships, enabling the construction of complex interrelationships), but these have only just begun to be used or appreciated by landscape modelers. Similarly, dynamic relationships between elements can be modeled using 'cellular automata' and related parallel-processing techniques, but these too have not yet been well integrated into landscape modeling efforts.

Dynamics problems

1. movement through the landscape
2. movement of the landscape
3. interaction with the landscape

As for movement through the landscape, modern rendering/animation programs allow for the creation of paths, but these are for the most part unconvincing, especially for pedestrian movement. (For helicopters, trains and automobiles, fast, smooth splines make sense; on foot, they don't.) User-controlled movement, as in VRML, suffers from similar problems. Walking up or down stairs is a good test of the verisimilitude of virtual movement through landscapes, which too often feels like a friction-less escalator ride.

Dynamic motion of the landscape comes in many scales, magnitudes, and frequencies. Almost none of them are successfully modeled at present (first-approximations of leaves or palms blowing in the wind is about the extent of it.) Many texture-mapping and rendering systems provide a 'brick' pattern and texture, which can be applied to any surface. But none has a weathering function, which might change the color of the brick over time, much less a colonization function which would introduce moss on bricks in the shaded garden path. (There are appearing now some 'dirt' modeling components, to provide for -- at least visually -- roughened up textures.) (DIGIMATION)

Interaction with the landscape is barely addressed in any digital model. We are still far from having haptic interfaces and tactile feedback mechanisms available for most modelers, and we don’t have a good theoretical or empirical basis for determining the kinds of interactions we need to model. In this regard, virtual landscapes offer a quite different (and possibly far wider) range of kinds of interactions to model and explore. (VRML, JAVA3D)

'Objects vs. field' problems

Abstraction level problems

(Limited) knowledge of human perceptual systems

Of course, all these models and representations are made for, and by, humans with complex and specialized -- and still only barely understood -- perceptual systems. What difference does color, or lineweight, or screen resolution, make to a human evaluator? under what circumstances? What role do representational conventions -- as opposed to content -- play in visual preference assessments? There is a growing body of research in this area, but little has trickled through to the modelers. Most is focussed to date on visual artifacts and preferences, but there are other human perceptual systems, too. Smell, touch and hearing are part of the landscape experience, and barely incorporated yet into digital modeling. Some virtual reality experiments have begun, but there are many more to do.

Computer science and algorithmics

Finally, no representation is free of its medium. Digital models depend for their efficacy, and certainly for their future development, on non-domain specific techniques and advances in computer science. Overcoming the first brute-force thresholds for modeling terrain and vegetation was a combination of hardware (larger disks, more memory, faster processors) and software (TIN algorithms, procedural generation) technologies. Those will certainly break more barriers (billions-of-polygons, immersive displays) in the future. But there will always be tensions, trade-offs and compromises, as between clarity and efficiency, for example, or space efficiency and speed. Landscape modeling presently represents the kinds of complexities and complications that all digital modeling will confront as techniques get better, consumers more sophisticated and modeling demands more exacting.

Just as the best oil painters know about brushes and pigments, so digital modelers need to know about algorithms, data structures and display devices; one more discipline to stay abreast of!

Conclusion

A central theme in all of the above discussion has been the distinction between 'looks like' and 'acts like' -- whether or not 'real' or 'realistic' enter into the equation. For digital landscape models, some inferential and rhetorical purposes are served by 'looks like'. But many others purposes, including those involved in exploratory creative design processes as well as analytic scientific ones, demand more. For even the more straightforward generation of convincing images, dynamics and procedural models have value; for the latter, they are essential.

From these parallel strands of topics, inter-related as they are, and with an eye towards the future of research and development in the field of landscape modeling, a prioritized agenda of sorts emerges:

1. Continue with experiments to further understand human perception, in response to a wide range of visual stimuli, both real and virtual. At the same time, work to gain greater operational understanding of the roles of different kinds of models, at differing levels of abstraction, for different purposes; knowing their limitations, potentials and benefits is the only way to choose between competing modeling and representational strategies.

2. Research and develop immersive display techniques and haptic interfaces (the essentials of so-called virtual reality) to further engage the whole array of human senses into digital landscape models

3. Research and develop specific digital modeling techniques for the essential landscape elements, from shadows of trees in forests to wind effects on sand particles, in collaboration with appropriate scientists and others.

4. Research and develop techniques for handling level-of-detail variations in virtual worlds, (beyond the primitive 'use representation x at distance y' method) and field modeling

5. Research and develop techniques for encoding, expressing, manipulating and evaluating dynamics in landscape models, from choreographed walk-throughs to interactive multimedia gardens.

6. Research and develop procedural techniques, using modern algorithmic methods (data structures, programming constructs, etc.) to aid in 3.) 4.) and 5.) above, by an active and informed collaboration with computer and information scientists.

Conclusion

Is all this concern over representations and simulations which represent -- or don't really represent -- 'reality' justified? Can't we simply make the best visual simulations we can to compare alternatives and judge visual quality, accepting the limitations of our representational systems for what they are? Certainly, as a culture we have been working with idealized artist's conceptions for a long time. And even knowledgeable and sensitive landscape architects use abstract circles for tree symbols in plans, without thereby losing their appreciation of the complexity, individuality and variety of arboreal beauty. But it seems to me that in digital models we do often homogenize and distill the qualities of landscape, to a degree which may once have been necessary -- just as engineers accepted first order approximations of structural strength, until we learned to do better -- but is no longer necessary, with modern computing tools, properly tuned.

Why make a digital model of a dune which really is generated by the coevolution of wind, surf, sand particles and dune grasses? Or represent the dry leaves of fall plastered by the wind up against the chain link fence of a children's playground in the winter?

These details represent the many visual and other qualities of the ordinary landscape which is so important to the environment in which we live, and whose omission makes for a certain sterility of virtual landscapes (as observed, for example, by Stilgoe, 1999). And the most convincing, and often the most effective way of achieving these details, especially in their interactions, is by modeling -- and therefore, of necessity, understanding -- their generative processes, rather than just 'painting them on'.

At the same time, we landscape modelers must also remember the valuable roles of abstraction in both cognition and communication, and not believe that 'photo-realism' -- or even 'physical realism' -- is the be-all, end-all of digital modeling. We make models to make explorations or to convey messages, and the infinite variety of explorations and messages will surely yield an equally boundless variety of digital landscape models.

----------

References

-------

Software referred to