E C O L O G I C A L I N F O R M A T I C S 1 ( 2 0 0 6 ) 1 7 9 – 1 8 7

ava i l ab l e a t www.sc i enced i r ec t . com

www.e l sev i e r. com/ loca te / eco l i n f

Real-time 4D visualization of migratory insect dynamicswithin an integrated spatiotemporal system

Yi Wua,⁎, Bronwyn Priceb, Daniel Iseneggerc, Andreas Fischlinb,Britta Allgöwerc, Daniel Nuescha

aRemote Sensing Laboratories, Department of Geography, University of Zurich, Winterthurerstr. 190, CH-8057 Zürich, SwitzerlandbTerrestrial Systems Ecology, Department of Environmental Sciences, ETH Zurich, Universitätstrasse 16, CH-8092 Zürich, SwitzerlandcGIScience Center, Department of Geography, University of Zurich, Winterthurerstr. 190, CH-8057 Zürich, Switzerland

A R T I C L E I N F O

⁎ Corresponding author. Tel.: +41 44 635 51 63E-mail addresses: yi.wu@geo.unizh.ch (Y. W

andreas.fischlin@env.ethz.ch (A. Fischlin), br

1574-9541/$ - see front matter © 2006 Elsevidoi:10.1016/j.ecoinf.2006.03.004

A B S T R A C T

Article history:Received 7 November 2005Received in revised form9 March 2006Accepted 12 March 2006

This paper presents a new approach of spatiotemporally visualizing the simulation outputof migratory insect dynamics and resultant vegetation changes in real-time. Thevisualization is capable of displaying simulated ecological phenomena in an intuitivemanner, which allows research results to be easily understood by a wide range of users. Inorder to design a fast and efficient visualization technique, a simplified mathematicalmodel is applied to intelligibly represent migrating groups of insects. In addition, impostorsare used to accelerate rendering processes. The presented visualization method isimplemented in an integrated spatiotemporal analysis system, which models, simulatesand analyzes ecological phenomena such as insect migration through time at a variety ofspatial resolutions.

© 2006 Elsevier B.V. All rights reserved.

Keywords:Real-time 4D visualizationSpatiotemporal analysisMigrationDispersionCloud modelingCloud rendering

1. Introduction

Nowadays, ecologists are able to model and analyze dynamicsof migratory insects such as larch bud moth (LBM) withinefficient simulation systems (Fischlin, 1982; Fischlin andBaltensweiler, 1979; Baltensweiler and Fischlin, 1988). As akind of migratory insects, LBM are often capable of migratinglong distances and causing conspicuous defoliation on AlpineLarch (Larix decidua) across extensive areas (Baltensweiler andFischlin, 1988). During peaks of population cycles, over-crowding of LBM larvae causes interrupted feeding behavior,which leads to partially eaten needles drying out such thatlarge-scale defoliation results during the summer months.The consequences directly impact on tourism and forestsuccession (Baltensweiler and Fischlin, 1988; Baltensweilerand Rubli, 1999). Therefore, knowledge of the migration

; fax: +41 44 635 68 46.u), bronwyn.price@env.eitta@geo.unizh.ch (B. Allg

er B.V. All rights reserved

dynamics of such insects is of value within ecology and forestmanagement.

However, current systems do not allow for the analysis anddisplay of simulation results in a visual manner that may beeasily interpreted by non-experts. As an example, Fig. 1 andTable 1 show two results froma LBM simulationmodel run. The2D illustration in Fig. 1 describes the temporal variations of LBMlarval density for the entire Upper Engadine valley aggregated toa single spatial unit, while Table 1 shows thenumbers of femalemoths that migrate from a particular site to other sites with theresulting defoliation percentages. It is very difficult for a layper-son to imagine the spatiotemporal dynamics of LBM migrationwhile looking at Fig. 1. Even if Table 1 contains all data to des-cribe thesedynamics inspaceand time, the figuresare detachedfrom their spatial locations. Therefore, interpretation of resultsfromecological simulationsystems isoften limitedtoexperts. In

thz.ch (B. Price), disen@geo.unizh.ch (D. Isenegger),öwer), nuesch@geo.unizh.ch (D. Nuesch).

.

Larval densities for 20 sites of the Upper Engadine valley

0.001

0.01

0.1

1

10

100

1000

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

1963

1964

1965

1966

1967

1968

1969

1970

1971

1972

1973

1974

Lar

val d

ensi

ty p

er k

g t

ree

bra

nch

es

Fig. 1 –LBM larval density variation in the Upper Engadine valley results from the LBM-M9 model.

180 E C O L O G I C A L I N F O R M A T I C S 1 ( 2 0 0 6 ) 1 7 9 – 1 8 7

this paper, we take LBM dynamics as an example to demon-strate how a real-time 4D visualization in an integratedspatiotemporal analysis system brings easily understandableillustration of LBM migration and resultant vegetation changesto a wide range of users.

To improve the understanding of the ecological processesof migratory insects in both the temporal and the spatialdimensions, an integrated system, Interactive, Process Oriented,Dynamic Landscape Analysis and Simulation System (IPODLAS), isbeing developed. Therein, a temporally explicit ecologicalmodeling simulation system (Temporal Simulation) is coupledwith a spatially explicit geographic information system (GIS),and a real-time 4D visualization. This integrated systemIPODLAS allows simulation and investigation of spatiotempo-ral ecological phenomena such as LBM migration andrepresentation of simulation results in an intuitive manner.

For this case study, real-time visualization in IPODLASplays the role of representing results from ecological systems.To interactively visualize the simulation results in a 4D virtualenvironment, large data sets such as a digital elevation modeland terrain texture of study area must be additionallyrendered in real-time. Moreover, for the visualization of insectmigration dynamics, there are two alternating computingprocesses: 1) insects' migrating paths must be calculated

Table 1 – Data table of LBM female migration number anddefoliation percentage of every site in 1-year results fromthe LBM-M9 model

Site 1 2 … 20

1 4765 126 … 2862 0 64 … 1033 324 0 … 32… … … … …20 712 21 … 124Defoliation 0.00364 0.03211 … 0.04321

according to the simulation output and terrain and 2) livelyinsects in the air have to be rendered dynamically. In addition,once insects have landed, the new defoliation level resultingfrom a change of local insect density need to be visualized bychanging the appearance of vegetation on the terrain.

However, memory size and computing speed of today'scommon computers are insufficient to compute all thesechanges and photo-realistically visualize individual insect inreal-time (Roettger and Ertl, 2003; RoettgerHeidreich, 1998). Forthe convenience of awide range of users, efficientmethods areexplored in this paper for adapting existing computing andrendering algorithms to visualize both insect dynamics andresultant vegetation changes in real-time. The resultingspatiotemporal visualization achieves easier understandingof interactive and explorative simulation results, wherebysimulation results depicting insect dynamics and vegetationchanges are visually attached to the relevant 3D spatiallocations. In the following sections, we use LBM dynamics asan example to illustrate the potential of our methods.

2. Material and methods

2.1. Study area

In this paper, the Upper Engadine valley, located in the Swisspart of the European Alps, forms the study area. Herein,quantitativedata onhost trees, LBM larval densities, defoliation,and natural enemies has been collected for more than 50 years(Baltensweiler and Fischlin, 1988; Baltensweiler andRubli, 1999).

For the purposes of this study, larch bud moth populationdynamics, numbers of annually migrating females anddefoliation levels are simulated with a coupled local dynamicsand migration within the Upper Engadine valley model,known as LBM-M9 (Fischlin, 1982; Fischlin and Baltensweiler,1979). During development of the LBM-M9 model, the Upper

Fig. 2 –The distribution of larch forest sites in the Upper Engadine valley.

Visualization GUI

Storage

Data flow

Control flow

Mappings between subsystems

Temporal Simulation

System

IPODLAS kernel

GIS

Fig. 3 –The IPODLAS kernel is the central coordinationcomponent which controls communication between thesubsystems (GIS, Visualization, and Temporal Simulation),GUI, and the Storage. For performance reason, the designallows that the visualization subsystem can interact directlywith the GUI, and not through the IPDOLAS kernel asrequired for the other subsystems.

181E C O L O G I C A L I N F O R M A T I C S 1 ( 2 0 0 6 ) 1 7 9 – 1 8 7

Engadine valley was divided into 20 areas, considered to behomogeneous with respect to aspect and altitude for thepurposes of formulating the migration model (Fischlin, 1982).These areas, known as sites, are shown in Fig. 2. Field study-based LBM population census data cover a time series for eachof the 20 sites from 1949 to 1977, thus representing an idealstudy case or test bed for the dynamic visualization ofspatiotemporal processes.

2.2. Data preparation

The local dynamics submodel of LBM-M9 is based on the foodquality hypothesis, which considers population dynamics oflarch bud moth to be driven by a relationship between the LBMand its host (Baltensweiler and Fischlin, 1988). The migrationsubmodel simulates LBM migration between sites and is depen-dant on wind conditions, forest cover and site quality (definedby degree of defoliation) (Fischlin, 1982). All outputs of LBM-M9are simulated within the Temporal Simulation subsystem,RAMSES (research aids for modeling and simulation of envi-ronmental systems), and then transferred via our integratedsystem kernel, shown in Fig. 3, to the visualization subsystemas an input.

In general, the following five data sets are necessary for thespatiotemporal visualization of LBM dynamics and resultantvegetation changes in the study area:

• A digital elevation model (DEM) [© Tydac AG, Bern] todescribe the geometry of the terrain in the Upper Engadine

valley. The used DEM is a raster-based elevationmodel withan original grid size of 50 m×50 m. It was resampled to10 m×10 m using a bilinear interpolation scheme.

Table 2 – Defoliation levels for categories of simulateddefoliation percentage and the corresponding visualizedforest color

Defoliation percentage Defoliation level Forest color

0.00–0.01 None Green0.011–0.33 Light Yellow0.331–0.66 Medium Orange0.661–1.0 Heavy Brown

182 E C O L O G I C A L I N F O R M A T I C S 1 ( 2 0 0 6 ) 1 7 9 – 1 8 7

• A satellite image [© ESA/Eurimage, CNES/Spotimage, Swis-stopo/National Point of Contact (NPOC)] to describe theappearance of the terrain in a photo-realistic manner. Amosaic of Landsat Thematic Mapper imagery with a rastersize of 25 m×25 m was used as a basic data set. Its spatialresolution was improved to 10 m×10 m by image fusionwith panchromatic SPOT data [15].

• Geographical locations and extents of the 20 spatiallyhomogenous LBM study sites within the Upper Engadinevalley. This information was generated within the GISsubsystem and has the same resolution as the DEM.

• Annual number of simulated female LBM migrating fromeach site to all other sites.

• Simulated defoliation percentage of the Larch trees follow-ing insect migration for each site each year. This simulationwas performed within RAMSES. Defoliation was categorizedto 4 different levels following Baltensweiler and Rubli (1999),colors for visualized forests were chosen accordingly torepresent a gradient from green, for no defoliation, tobrown, for heavy defoliation, shown in Table 2.

2.3. The integrated spatiotemporal system

The integrated spatiotemporal system, IPODLAS, uses dataand functionalities of three subsystems, GIS, Temporal Simu-lation, and Visualization in a transparent and seamless way toprofit from the individual strengths of the respective sub-systems. Fig. 3 shows the software architecture of IPODLAS,which supports the smooth interaction of the subsystems.

IPODLAS is designed as a user-driven system,whichmeansthat actions of a user in the graphical user interface (GUI)trigger the GUI to invoke activities, possibly in all othersubsystems. These activities are coordinated by a synchroni-zationmechanism, which is one of the main functionalities ofthe IPODLAS kernel. For example, when the IPODLAS kernelreceives a user request from the GUI, it gives primary feedbackto the user and breaks the request down into subtasks, whichare sent to the appropriate subsystems. After completion ofsubtasks, the subsystems send the results via the IPODLASkernel to the GUI and user.

The IPODLAS kernel not only supports the interoperabilityof the subsystems, it also allows a subsystem to be easilyexchanged. For example, if we wanted to use anotherTemporal Simulation system, it needs only to be registered inthe IPODLAS kernel. Since all the subsystems communicatethrough the IPODLAS kernel, different protocols map the datamodels of the subsystems to the data model of the IPODLASkernel. To support seamlessly accessing the required func-tionalities of the subsystems from the GUI, themapping of thefunctions of the subsystems to the IPODLAS kernel is also

mandatory (Leclercq et al., 1996). However, in order to speedupdata exchange, some interactions between the GUI andVisualization do not go via the IPODLAS kernel.

When using IPODLAS to model and simulate migratoryinsect dynamics, we use both public domain GRASS (geo-graphic resources analysis support system) and ESRI's ArcGIS/ArcInfo to obtain geographical locations and extents of thestudy area. Spatial data and functions from the GIS subsystemare used within the Temporal Simulation subsystem, RAMSESin this study, to simulate LBM spatiotemporal dynamics. Thenthe visualization subsystem functions as an interface to users,displaying the simulation results from RAMSES.

In IPODLAS, an efficient technique is essential for terrainrendering. To visualize spatiotemporal migratory insectdynamics, namely allowing extension of existing functions,the source code and access to the low level rendering appli-cation programming interface (API) must be available. Thevirtual terrain project (VTP) satisfies these requirements and isselected as a software foundation (Roettger and Heidreich,1998; Doellner and Kersting, 2000; Losasso and Hoppe, 2004).

VTP is a creative project for easy construction of virtualterrain in an interactive, 3D digital format. One of its mainadvantages is that its run-time environment makes use ofcontinuous level of detail (CLOD) algorithms for terrain rende-ring (Roettger and Heidreich, 1998). The work of Biegger (2004)shows that VTP can act as a suitable base for implementationof a visual system dedicated to simulation and analysis of adynamic process, i.e. glacier fluctuations. Therefore, visuali-zation of migratory insect dynamics and resulting vegetationchanges canbedevelopedas anembeddedmodulewithinVTP.

2.4. Visualization of LBM dynamics

According to the simulation results from RAMSES, manythousands and even millions of LBM may migrate from onesite to others (Baltensweiler and Rubli, 1999; Fischlin, 1982,1983). Since the number of migratory female LBM from eachsite each year is given statistically, we simply assume thatmoths from each site within the Upper Engadine valley movetogether and thus canbe thoughtof as a ‘cloud’with each smallgroup of moths representing an internal particle of the cloud.

LBM researchers found that LBM migrating from each sitenormally land in several sites separately. Hence, duringmigration, some LBM may land in a particular site, while theothers continue flying. Accordingly, LBM cloud can beconsidered split into several small clouds flying to differenttarget sites after taking off from each site. If a user views avisualization of migration from several different sites in agiven time period, he may find it difficult to distinguish fromwhich site a migrating cloud originates. For convenience,during visualization a different color is assigned to individualinsect groups originating from different sites.

However, visualization of animated clouds remains achallenge. On one hand, the difficulties in computing velocityfields are apparent. Velocity fields governing the dynamics ofclouds are influenced heavily by wind and other disturbanceforces in the air. Our LBM clouds need to be visualized asdynamic clouds because they are composed of dynamiccreatures that do not keep to the relative static positionswithin the clouds during migration. On the other hand,

Fig. 4 –Two clouds shaped by 4000 and 1000 blobs,respectively, in 3 ellipsoids.

183E C O L O G I C A L I N F O R M A T I C S 1 ( 2 0 0 6 ) 1 7 9 – 1 8 7

rendering a cloud requires consideration of both the effects ofoptical properties along light paths through the cloud volumeand the complex multiple scattering of light within themedium before it reaching the viewer. Previous work relatedto cloud visualization within the computer graphics fieldcould be divided into two parts: cloud modeling and cloudrendering (Dobashi et al., 2000). Cloud modeling deals withmethods used to represent construction and cloud dynamics,while cloud rendering deals with the techniques of efficientand realistic rendering.

2.4.1. LBM cloud modelingSince our goal is to render real-time dynamic LBM clouds inlarge numbers and at the same time the main task of centralprocessing unit (CPU) and graphics processing unit (GPU) ofcomputer is to render a large terrain data set with a high-resolution satellite image, a simple and efficient modelingmethod is required to construct and simulate the dynamics ofLBM clouds in a visually convincing way. Fluid dynamics is astraightforwardmethod to calculate velocity fields for realisticdynamic clouds with an arbitrary initial structure (Fedkiw etal., 2001; Foster and Metaxas, 1997; Stam, 1999). However, thisis impractical since it is computationally too expensive. Theheuristic approach is computationally inexpensive and mucheasier to implement (Ebert, 1997; Roettger and Ertl, 2003). Athird approach which lies between the above two approachesuses cellular automation to simplify dynamic cloud motion(Dobashi et al., 2000; Harris and Lastra, 2001).

Since the primary task for this study is to render LBMclouds and terrain, the heuristic approach is applied toconstruct and simulate dynamic LBM clouds in order to savecomputation time. In this paper, a two-level model is appliedto construct the initial shapes: the cloud macrostructure andthe cloudmicrostructure (Ebert, 1997). Implicit primitives suchas ellipsoids are assigned locations, sizes, and weights toconstruct the macrostructure. A Gaussian distribution com-bined with a random noise function is applied to create themicrostructure, which describes the distribution of particleswithin each ellipsoid. Moreover, the Gaussian center isirregularly located inside the ellipsoid to introduce morerandomness to the particle distribution. Each internal particleis represented by a sphere (blob) with a density functionidentifying the distribution of a LBM subgroup within thesphere. To reduce the computation cost and acceleraterendering process, blob centers are supposed to have thelocal highest density of LBM and the density decreasesaccording to a Gaussian distribution.

Both ellipsoids and blobs in a cloud have different sizes anddensities. The volume of total ellipsoids for a cloud decides itssize. Meanwhile, the number of the blobs also increases ordecreases accordingly to fill in the volume. As an example, Fig.4 shows two cloudmodels constructed by 4000 and 1000 blobs,respectively, in 3 ellipsoids.

With the above approach, every LBM cloud has differentstructure. Then to simulate dynamic LBM clouds, we assumemoths randomly flight inside a LBM cloud. A straightforwardmethod is to store every blob's position and continuouslyupdate it according to some random rules in real-time.However, this approach can cause a large burden on CPUand main memory since each LBM cloud may have hundreds

or thousands of blobs. Hence, we only store macrostructureand Gaussian center of blob distribution for each LBM cloud.Consequently, blobs are dynamically generated in each framewith random function under constraints of Gaussian distribu-tions for LBM clouds. Reader may think that LBM clouds willchange appearance unreasonably quickly since every blob hasa new position in each frame. However, this is not true sincestatistically blob distributions have been decided by macro-structures and Gaussian centers. After rendering, discontinu-ous changes can only be distinguished at places having fewblobs. Therefore, the above simulationmethod exactly reflectsour assumption of the unordered flight of moths inside LBMclouds.

2.4.2. LBM cloud rendering“Rendering clouds is difficult because realistic shadingrequires the integration of the effects of optical propertiesalong paths through the cloud volume, while incorporatingthe complex scattering within the medium” (Harris andLastra, 2001). In recent research within the computergraphics field, one of the most common approaches is touse 3D textures to render amorphous phenomena. Fedkiw etal. successfully rendered smoke with 3D texture mappinghardware (Fedkiw et al., 2001). Although it is simple to beused by programmers, GPU does many background compu-tations to support it. Moreover, to use this technique,standard APIs such as OpenGL have strict constraints onthe size of 3D texture data. Therefore, we use 2D texturemapping to render half-transparent property of LBM clouds.However, it is important to note that the algorithms basedon 2D texture mapping are applicable to 3D texture mappinghardware as well.

In order to display visually convincing cloud images, weneed to take into account light absorption and multiplescattering caused by blobs inside clouds. Absorption is decidedby extinction coefficient, which varies according to theproperties of internal blobs. In this studywe choose a constantfor LBM in order to simplify the computation. Since clouds arecomposed of many tiny blobs, multiple scattering amongthem is too complicated to be computed in real-time.Fortunately, the work of Nishita et al. demonstrated that thecontribution of multiple scattering is dominated by the firsttwo orders (Nishita et al., 1996). Hence, multiple forwardscattering is often used nowadays to approximate multiplescattering for cloud rendering.

184 E C O L O G I C A L I N F O R M A T I C S 1 ( 2 0 0 6 ) 1 7 9 – 1 8 7

With the help of hardware acceleration method, Dobashiet al. applied an approximation of isotropic single forwardscattering to render clouds (Dobashi et al., 2000). Later, Harrisand Lastra extended Dobashi's method with multiple for-ward scattering and anisotropic first order scattering towardsthe eye to obtain more realistic results (Harris and Lastra,2001). Both methods considered shading and self-shadingwhen light passes through a cloud. Moreover, both workapplied hardware acceleration method to accomplish thecomputation of light intensities on blobs in their implemen-tations. Although photo-realistic results can be obtained withtheir methods, it is still too time-consuming for ourpurposes, especially during the crucial process of readingpixels back from the frame-buffer in GPU. Our method is asimple approximation of their work, which is fast enough toachieve real-time rendering without losing too much visualquality.

If we consider a light ray traversing a cloud, starting fromoutside, the intensity Ik on any blob pk is equal to lightintensity scattered to pk from other blobs plus the intensitytransmitted to pk through pk−1 (as determined by its transpar-ency), which is closer to light source than pk along raydirection. Hence, Blobs need to be queued according toascendant distance to light source. Multiple forward scatteringassumes the light intensity scattered by other blobs on pk isdetermined by the light intensities on the blobs closer to lightsource than pk with an isotropic albedo. Then, light intensityon each blob can be computed from the beginning of thequeue to the end.

Although the above process mimics light traversing cloudwell, it is computationally expensive to achieve real-time LBMclouds rendering with a common computer. Hence, instead ofprecise computation, we imitate the result by assuming lightintensity for each big ellipsoid in a cloud can be mapped ontoan elliptic figure through its centre and facing the viewer. Theintensities on blobs inside this ellipse have an anisotropicGaussian distribution. That means if we place a Gaussiancenter at the centre of the ellipse, light intensities willisotropically increase from the centre of the ellipse to itsborder. Here, light intensities have inverse relationship to LBMdensities in ellipsoids. It coincides with the assumption in thelast section.

Besides the approximation of light multiple forwardscattering inside clouds, single scattering towards the viewermust also be considered to re-adjust each blob color in eyes.We simplify light scattering towards the viewer by givingconstant values to the albedo and extinction coefficient.Assuming l

tis view direction from blob pk to the viewer, and

wtis light direction; the intensity Ek scattered from the position

of pk towards the viewer is computed:

Ek ¼ apð lt;wtÞIk þ ð1� eÞEk�1; 1VkVN ð1Þ

where α is albedo and e is extinction coefficient, pð lt;wtÞ is thephase function, which determines the extent of light incidentfrom direction wt scattered by cloud to direction l

t, N is the

number of blobs. This equation simply says that lightscattered from the position of pk towards the viewer is equalto the intensity scattered to the viewer by pk, plus the intensityscattered by other blobs to the viewer through pk but notabsorbed by pk. Then cloud image is rendered by the portion of

light passed through cloud towards the viewer and notabsorbed by internal blobs. Here, the simple Rayleigh scatter-ing phase function is used:

pðhÞ ¼ 3=4ð1þ cos2hÞ ð2Þ

where θ is the angle between the incoming and scattered lightdirections. This phase function was proved to be efficient andpractical for cloud rendering by Harris and Lastra (2001).

In rendering process, 2D polygons texturedwith a Gaussianfunction are used to represent blobs. The colors of polygonsare determined by light intensities on blobs, albedo and anglesbetween incident light and view directions. Polygons areblended according to Eq. (1) with alpha equaling to extinctioncoefficient using alpha blending technique. This technique is astandard capability of today's common GPU. Hence, allblending processes are executed automatically and interme-diate results are saved as image in GPU, which largely alleviatethe storage burden of the main memory.

Since all the blobs of a LBM cloud are finally blended intoone cloud image, dynamic impostor is used to insert theLBM cloud to the 3D virtual landscape. An impostor is asemi-transparent quadrangle textured with an image of theobject which is replaced by the impostor in the virtual world.Since our virtual landscape allows free navigation, impostorsmust always be placed on planes facing the viewer. Thedirection from the centre of an impostor to the viewer isalways the normal direction of the impostor. Finally, weachieve rendering of LBM clouds in real-time with promisingresults.

2.4.3. LBM migration pathMigration paths of LBM in the real world are complicated anddifficult to determine. To simplify the computation, weassume they migrate in the following manner:

Whenmoths start migrating from each site, they graduallyfly to a certain height directly above terrain. As the number ofmigrating LBM varies largely in different sites and years, wesuppose that migration starts from a region near the centre ofa site. As migrating LBM number becomes larger, the regionextends gradually to the whole site.

Once moths reach the assumed height above terrain, theysplit into smaller LBM clouds, which fly directly to theirrespective destination sites. To enhance visual effects, weassume migration paths moving up and down according toterrain. Since every LBM cloud heads to its own destinationsite, there is a computable shortest path on terrain betweenthe projection point of a LBM cloud center and the center ofits destination site. We further assume that every LBM cloudflies along the shortest path and keeps the height aboveterrain. As soon as the projection point of a LBM cloudcenter exceeds the centre of its landing site, the LBM cloud isready to land.

The landing process is also a dispersive process. During thelanding of a LBM cloud, its density is reduced, while its verticalextension becomes shorter and its cross section becomesgradually larger. Finally, the LBM cloud covers a certain regionof the landing site and moths disappear from the users view.Assumptions about the landing region are the same as theassumptions for the take-off region described above. Thelanding paths are implemented as individual straight lines

Fig. 5 –Clouds fast rendered by simplified method.

185E C O L O G I C A L I N F O R M A T I C S 1 ( 2 0 0 6 ) 1 7 9 – 1 8 7

between some sample points in the LBM cloud and pointsinside the landing region.

2.5. Modeling dynamic vegetation change

Depending on LBM larval density, larch forests experiencedifferent degrees of defoliation. This results in significantcolor changes of the larch foliage during the summer,changing the appearance of the forested landscape withineach LBM site considerably (Baltensweiler and Fischlin, 1988;Baltensweiler and Rubli, 1999). However, satellite images arenot available for all the investigated years (1949–1977), andthus cannot fulfill our requirements. Moreover, since a largeraster-based satellite image is used as a texture layer for ourvirtual landscape, it is time-consuming to change someparts of the image and reload back to texture cache todisplay the yearly resultant vegetation changes of the wholestudy area in real-time. Therefore, we import geographicalinformation about the larch forest sites with the sameresolution as the DEM used to form the virtual terrain fromthe subsystem GIS in advance and build polygon-basedrepresentations of these sites slightly above the virtualterrain. By doing so, the highlighted resultant vegetationareas can easily be distinguished by the viewers.

With the alpha blending technique once again, thecolors of the sites can be combined with the underlyingsatellite image. Therefore, dynamic vegetation changes inreal-time is possible to be visualized by changing only thecolors of these sites. This method illustrates apparently

Fig. 6 –Several orange LBM clouds beginning from one siteand a landing purple LBM cloud beginning from another site.The height of the clouds is determined by the terrain.

annual resultant vegetation changes to users with lowcomputation costs.

3. Results

To present spatiotemporal simulation results in a visuallyintuitive manner, an approach for 4D real-time visualizationof migratory insects and resultant vegetation changes wasdeveloped. Standard digital terrain models and satelliteimages are used to render virtual pseudo-realistic land-scapes. Against this background, groups of migrating insectsare represented as continuously animated clouds using aGPU accelerated technique. A simplified mathematicalmodel is applied to make cloud modeling and renderingfast and efficient. Thus, the analysis outputs of LBMmigration, data tables which detached from spatial loca-tions, from the temporal ecological simulation system can bevisually approximately expressed by the dynamic processes ofascending, splitting, flying, descending and dispersing of LBM

Fig. 7 – (a, b) The appearances of forest sites following larvalfeeding in two successive years.

186 E C O L O G I C A L I N F O R M A T I C S 1 ( 2 0 0 6 ) 1 7 9 – 1 8 7

clouds. Simultaneously, resulting from varied larch forestdefoliations caused by annually different LBM densities, theappearance of the landscape is changed by dynamically blen-ding colors of influenced areas with the underlying terraintexture.

As an example, Fig. 5 shows three clouds, with a little morecomplicated macrostructure, fast rendered by our simplifiedmethod.

Fig. 6 shows LBM clouds in migration. All orange LBMclouds originate from a same site while the small purple onecomes from another site. At this moment, the purple LBMcloud is landing and the orange LBM clouds are still heading totheir target sites. Because LBM cloud migration paths aremoving up and down according to the terrain, the orangeclouds are flying at different heights.

Once each LBM cloud has landed, the simulated resultantlevel of defoliation in each larch forest site is displayed bychanging the site colors according to the output of thetemporal ecological simulation system. In Fig. 7(b), twooriginally green sites in the Upper Engadine valley turn intoyellow (light defoliation).

4. Discussion and future work

Using the simplified models presented in this paper, we canvisualize the simulation results of migratory insects in virtuallandscape in real-time. The 4D dynamic scenarios give users abetter understanding of insect migration and its influence onthe vegetation. Hence, simulation results are more under-standable and directly perceivable by users than a table or a 2Dpicture.

Animated clouds are rendered by taking advantages ofcolor blending capability of GPU and dynamic impostors. Itallows real-time visualization of a few animated clouds withour simple cloud models at the same time.

Assumptions are made in order to visualize migrationbehavior of LBM using limited computing time. Nevertheless,the goal of this visualization is to efficiently demonstratesimulation results of a modeled ecological system in real-timeso that they are easily understandable and perceivable by thelayperson, as opposed to other researchers or experts, whomight ask for high quality of realistic images, for exampleanimated movies. A more realistic visualization would allowfor greater visual feeling, however, the understanding ofsimulation outputs would not be improved at the cost ofcomputation efficiency.

Although the methods developed are tested in the UpperEngadine valley, they could be also applied to other regions.Moreover, they are suitable for the visualization of thedynamics of other migratory insects, which have similarmigration properties as LBM.

Our visualization approach treats the LBM simulationprocedure as a ‘self-contained’ process and does not takefurther physical parameters into account while visualizingLBM migration. For example, wind speed and direction couldchange cloud geometries and dynamics so heavily that LBMcloud models should be accordingly reconstructed for eachframe. Therefore a more efficient method is necessary tocomplete this task.

Instead of static resultant vegetation geometries, newCLOD techniques should be applied to those irregular geom-etries to make the rendering process faster.

In addition, improvement in the interaction betweenvisualization and the other subsystems of IPODLAS couldsignificantly support users in studying the simulation results.For example, the geographical information of a selected areaon the virtual terrain will help the layperson know where theinsect migration takes place. Such problems could be avenuesof our future work.

We hope our work can also serve as an example for otherresearchers to apply computer graphics to illustrate analysisresults in ecological research.

Acknowledgements

This work is funded by the Swiss National Science Foundationwithin the National Research Program 48 ‘Landscapes andHabitats of the Alps’ (Grant No. 4048-064432/1).

R E F E R E N C E S

Baltensweiler, W., Fischlin, A., 1988. The larch bud moth in theAlps. In: Berryman, A.A. (Ed.), Dynamics of Forest InsectPopulations: Patterns, Causes, Implications. PlenumPublishingCorporation, New York, pp. 331–351. a.o.

Baltensweiler, W., Rubli, D., 1999. Dispersal: an important drivingforce of the cyclic population dynamics of the larch budmoth. Zeiraphera diniana Gn. Forest Snow Landsc. Res. 74 (1),3–153.

Biegger, S., 2004. A visual system for the interactive study andexperimental simulation of climate-induced 3D mountainglacier fluctuations. Remote Sensing Series, vol. 43. Universityof Zurich. 150 pp.

Dobashi, Y., Kaneda, K., Yamashita, Okita, H.T., Nishita, T., 2000. Asimple, efficient methos for realistic animation of clouds. Proc.Of SIGGRAPH’2000, pp. 19–28.

Doellner, J., Kersting, O., 2000. Dynamic 3D maps as visualinterfaces for spatiotemporal data. ACMGIS 2000. ACM Press,pp. 115–120.

Ebert, D.S., 1997. Volumetric modelling with implicit functions: acloud is born. Visual Proc. of SIGGRAPH'97, pp. 147–156.

Fedkiw, R., Stam, J., Jensen, H.W., 2001. Visual simulation ofsmoke. Proc. of SIGGRAPH'2001, pp. 15–22.

Fischlin, A., 1982. Analyse eines Wald-Insekten-Systems: Dersubalpine Lärchen-Arvenwald und der graue LärchenwicklerZeiraphera diniana Gn. (Lep., Tortricidae). Diss. ETH No. 6977,Swiss Federal Institute of Technology: Zürich, Switzerland,pp. 294–318.

Fischlin, A., 1983. Modelling of Alpine valleys, defoliatedforests, and larch bud moth cycles: the role of migration. In:Lamberson, R.H. (ed.), Mathematical Models of RenewableResources. Humboldt State University, Mathematical Mod-elling Group, University of Victoria, Victoria, B.C., Canada,pp. 102–104.

Fischlin, A., Baltensweiler, W., 1979. Systems analysis of the larchbud moth system: Part I. The larch–larch bud moth relation-ship. Mitt. Schweiz. Entomol. Ges. 52, 273–289.

Foster, N., Metaxas, D., 1997. Modeling the motion of a hotturbulent gas. Proc. of SIGGRAPH'97, 181–188.

Harris, J.M., Lastra, A., 2001. Real-time cloud rendering. Euro-graphics 20 (3), 76–84.

187E C O L O G I C A L I N F O R M A T I C S 1 ( 2 0 0 6 ) 1 7 9 – 1 8 7

Leclercq, E., Benslimane, D., Yetongnon, K., 1996. A distributedobject architecture for interoperable GIS. IEEE Knowledge andData Engineering Exchange Workshop, pp. 73–80.

Losasso, F., Hoppe, H., 2004. Geometry clipmaps: terrainrendering using nested regular grids. Proc. of SIGGRAPH'2004,pp. 769–776.

Nishita, T., Dobashi, Y., Nakamae, E., 1996. Display of clouds takinginto account multiple anisotropic scattering and sky light.SIGGRAPH, pp. 379–386.

Roettger, S., Ertl, T., 2003. Fast volumetric display of naturalgaseous phenomena. Comput. Graph. Int. 74–86.

Roettger, S., Heidreich, W., 1998. Ph. Slusallek and H. Seidel, Real-time generation of continuous levels of detail for height fields.Proc. WSCG'98, pp. 315–322.

Stam, J., 1999. Stable fluids. Proc. of SIGGRAPH'99, pp. 121–128.