DOI: 10.1111/tgis.

12551

RESEARCH ARTICLE

A four‐dimensional agent‐based model: A case

study of forest‐fire smoke propagation

Alex K. Smith | Suzana Dragićević

Spatial Analysis and Modeling Laboratory,

Department of Geography, Simon Fraser Abstract
University (SFU), Burnaby, British Columbia, Dynamic geospatial complex systems are inherently four‐
Canada
dimensional (4D) processes and there is a need for spatio‐
Correspondence temporal models that are capable of realistic representation
Suzana Dragićević, Spatial Analysis and
Modeling Laboratory, Department of for improved understanding and analysis. Such systems
Geography, Simon Fraser University (SFU), include changes of geological structures, dune formation,
8888 University Drive, Burnaby, BC V5A
1S6, Canada. landslides, pollutant propagation, forest fires, and urban
Email: suzanad@sfu.ca densification. However, these phenomena are frequently an‐
Funding information alyzed and represented with modeling approaches that con‐
Natural Sciences and Engineering Research sider only two spatial dimensions and time. Consequently,
Council of Canada
the main objectives of this study are to design and develop
a modeling framework for 4D agent‐based modeling, and to
implement the approach to the 4D case study for forest‐fire
smoke propagation. The study area is central and southern
British Columbia and the western parts of Alberta, Canada
for forest fires that occurred in the summer season of 2017.
The simulation results produced realistic spatial patterns of
the smoke propagation dynamics.

1 | I NTRO D U C TI O N

Many geographical and environmental systems are constantly evolving in three‐dimensional space and time
and thus by nature they are four‐dimensional (4D). Examples of such systems include changes of geological
structures, dune formation, landslides, pollutant propagation through the atmosphere or in water environments,
nutrient cycles in the soil, forest fires, change of high‐rise buildings in the cities, and evacuation in urban settings.
However, these phenomena are often analyzed and represented with modeling approaches that consider only
two spatial dimensions and time. As examples, some spatio‐temporal models that can represent the dynamics
of change over two spatial dimensions and time are those representing land use change (Luus, Robinson, &

Transactions in GIS. 2019;23:417–434. wileyonlinelibrary.com/journal/tgis © 2019 John Wiley & Sons Ltd | 417
418 | SMITH and DRAGIĆEVIĆ

Deadman, 2011), human movement in a city (Crooks, Castle, & Batty, 2008; Zhu & Timmermans, 2011), soil
erosion (Heung, Bakker, Schmidt, & Dragićević, 2013), and wildfire spread (Clarke, Brass, & Riggan, 1994). This
simplification is largely determined by the nature of the data collection and availability that are primarily related
to two spatial dimensions such as satellite imagery, soil samples, land use or land cover states, all captured at
different temporal snapshots.
Furthermore, there are many different methods of spatio‐temporal modeling, including statistical and numer‐
ical models, that can also be grouped into the two general categories of equation (Rahmandad & Sterman,
2008) and process models (Alam & Dutta, 2012). Equation models are based on mathematical equations that
are developed to represent the effects of processes on the system (Parunak, Savit, & Riolo, 1998). With this
method, the inner mechanisms of the system are not well understood and only the output patterns are ana‐
lyzed (Murray‐Rust et al., 2014). In addition, these are top‐down modeling approaches that represent aggregate
behavior (Rounsevell, Robinson, & Murray‐Rust, 2012). Process models represent micro‐level interactions of
system elements to determine the macro‐level behavior (Luus et al., 2011), making them a bottom‐up approach
(Heppenstall, Malleson, & Crooks, 2016). The processes within a system are represented and are directly studied
to determine how they create spatial patterns that change over space and time (Parunak et al., 1998). In this
context, the observed patterns of the system emerge from the components of the system, allowing for unex‐
pected relationships and interactions among system parts and the environment to form the change (Heppenstall
et al., 2016). For example, an equation model of land use change may use regression analysis to determine the
impacts of specific variables such as economic growth, and then the equations can be applied to various regions
to determine how each will affect the land use change over time (He, Huang, & Wang, 2013). On the other hand,
the process model will determine how the decisions and interactions of different elements of the system change
the land under expected conditions and then forecast several possible scenarios of change process for the region
(Sun et al., 2014).
Given the bottom‐up local interactions and behavior of elements of the systems, most geospatial dynamic sys‐
tems can be viewed as complex systems. They behave in a nonlinear manner, and are characterized by evolution,
emergence, self‐organization, adaptation, feedback loops, and bifurcation, making them difficult to represent and
model (Manson, 2001). The most common mathematical approaches for complex systems modeling of geospatial
dynamic phenomena are geographic automata systems (GAS; Torrens & Benenson, 2005), particularly cellular
automata, voxel automata, and agent‐based modeling. These are bottom‐up methods that take account of the
interactions at the local or individual level, thereby allowing for the evolution of the spatial pattern of the system
at larger spatial scales.
Geographic space is typically represented in a cellular automaton by a regular grid of square cells also called a
raster and can be extended to 3D cubes called voxels (Greene, 1989), while irregular shapes or volumetric elements
can also be considered (Jjumba & Dragicevic, 2015). The process of change is represented by a set of transition rules
applied to each cell or voxel at a discrete time increment and within a defined neighborhood. When representing
individual behavior and the movement and interactions between elements of the system or the environment, agent‐
based modeling approaches are more appropriate (Huang, Parker, Filatova, & Sun, 2014). An agent‐based model
(ABM) operates using a series of mobile agents that are software routines moving and interacting with the landscape
using their own rule sets (Torrens & Benenson, 2005), allowing for higher degrees of heterogeneity than cellular
automaton or voxel automaton models.
While there are many uses of ABMs, one of the major geospatial applications is the modeling of land use
change where agents represent various stakeholders impacting the process such as households and homeowners
(Castella, Trung, & Boissau, 2005; Krebs, 2017; Rounsevell et al., 2012), and land developers (Dragićević & Hatch,
2018). These agents make decisions to change the land to meet their demand for urban resources such as employ‐
ment, and natural resources such as farmland (Castella et al., 2005). Ecological processes are also often modeled
using an ABM such as the spread of invasive insects (Anderson & Dragićević, 2015; Pérez & Dragićević, 2011),
SMITH and DRAGIĆEVIĆ | 419

killer whales (Testa et al., 2012), animal movements (Tang & Bennett, 2010), and human disease propagation
(Crooks & Hailegiorgis, 2014; Perez & Dragicevic, 2009).
ABMs have been used to represent physical systems where agents represent objects or particles such as sand
dunes (Genois, Du Pont, Hersen, & Gregoire, 2013), overland flow of water (Reaney, 2008), chemical reaction
and diffusion (Agusdinata, Amouie, & Xu, 2015), and even ballistic fragment impacts (Bova, Ciarallo, & Hill, 2016).
The agent decision‐making is based on concepts from physics and empirical observations to calculate the correct
action (Azimi, Jamali, & Mofrad, 2011; Reaney, 2008) and affects the system state by changing attributes such as
particle location, volume or size (Genois et al., 2013).
However, many of these geospatial processes are inherently 4D, thus entailing a need for advancement of the
GAS models in four dimensions for more realistic representation of geographical and environmental phenomena.
Recently, a voxel automaton approach (Jjumba & Dragicevic, 2015) was developed to model phenomena such as
pollution propagation in four dimensions. The use of 4D ABMs has been developed to a lesser extent and focused on
systems such as building evacuations (Macatulad & Blanco, 2014) and urban earthquake evacuation (Torrens, 2014).
Consequently, these models are not modeling natural complex systems from a 4D perspective, which demands an
advancement of the development of 4D ABMs into natural systems that are not necessarily based on human deci‐
sion‐making processes. Consequently, the main objectives of this study are to explore the further advancement of
GAS models in four dimensions by proposing a theoretical modeling framework for 4D ABMs, and applying this
approach to a case study concerning forest‐fire smoke propagation in which two scenarios are considered. The study
area used in this research is central and southern British Columbia (BC) and the western parts of Alberta, Canada
where forest fires occurred in the summer of 2017.
Smoke propagation was selected as the focus of the case study because of its inherent four‐dimensionality
and also due to the need to mitigate adverse impacts on human health and the environment. Human exposure
to smoke has led to increased demand on medical services for respiratory problems (Dennekamp & Abramson,
2011). Advanced warning of smoke events through simulation of smoke propagation could help reduce exposure
to smoke as recommended, for example, by the BC Centre for Disease Control (Elliott, 2014). Two models of
smoke propagation are available, WRF‐Chem (Grell et al., 2005), based on the Eulerian perspective of motion, and
HYSPLIT (Rolph et al., 2009) based on the Lagrangian perspective of motion, both for space‐time smoke propa‐
gation. The Lagrangian motion perspective improves the representation of movement over the Eulerian motion
perspective by directly modeling the movement of smoke particles instead of the movement of smoke concen‐
trations past static points. The representation of smoke concentration as a continuous gradient in WRF‐Chem is
better than the parcel representation of smoke used in HYSPLIT model. However, the agent‐based modeling of
smoke propagation can encompass the advantages of both existing models by including moving agents with a
Lagrangian perspective of motion and by calculating a gradient of smoke concentration. An existing ABM of smoke
propagation has been proposed (Smith & Dragicevic, 2018), but it operates only in 2D space and time with a static
fire emitting the smoke. Therefore, in this study, the proposed 4D ABM builds upon the limitations of the existing
models by representing the dynamics of smoke propagation as a 4D process, considering moving forest fire and
smoke across the 3D landscape.

2 | TH EO R E TI C A L FR A M E WO R K O F A 4 D AG E NT‐ BA S E D M O D E L

An ABM is a GAS model consisting of mobile agents that transmit information as they move and interact (Torrens
& Benenson, 2005) and can be formulated as:

ABMt ∼ (St ,It ,Rt ), (1)

420 | SMITH and DRAGIĆEVIĆ

where ABM at time t represents the automaton composed of agents states S and transition rules R used by the agents
to change the state of the model using input data I. In a 3D ABM, the states S can be defined by:

S2D :S(x,y) , (2)

where (x, y) represents the horizontal position of the agent 2D state S2D. In a 4D ABM, the state must include the third
spatial dimension z and is represented by:

S3D :S(x,y,z) , (3)

where (x, y, z) represents the horizontal and vertical positions the 3D agent with state S3D. Similarly, the input data
I are also 3D and represented by:

I3D :I(x,y,z) , (4)

where (x, y, z) represents the horizontal and vertical positions of the data available to the agents. The 3D represen‐
tation of both S3D and I3D can be seen on the left‐hand side of Figure 1 for time t. While not all input data need to
be 3D, 2D data can be given z coordinates or can be stretched over the entire third dimension. Together with the
transition rules R, this results in a 4D ABM being defined at the current time t by:

ABMt4D ∼ St3D ,It3D ,Rt ,

( )
(5)

where the four dimensions are located in the three spatial dimensions of the 3D state S3D, and the one temporal
dimension of time t. This is represented twice in Figure 1, once at time t, and once at time t + 1. As time t pro‐
gresses, the 4D ABM moves from the current time t to the next time step t + 1:
( ) ( )
ABMt4D ∼ St3D ,It3D ,Rt → ABMt+1 ∼ St+1 ,It+1 ,Rt+1 → ⋯ → ABMt+n ∼ St+n ,It+n ,Rt+n ,
( )
4D 3D 3D 4D 3D 3D (6)

where the agents use their current state St3D, current input data It3D, and the current transition rules Rtto determine
their state in the next time step S3D
t+1
. At time t+1, the agents' have access to the data It+1
3D
for use with transition
rules Rt+1 that may or may not be different than the previous transition rules Rt. This progression of time in the
model is displayed in Figure 1 where the model evolves from initial time t through the next time t+1 and generally
for times t+n.

Zƚ Zƚнϭ ^ƚнŶ
/ƚнŶ ϭ ZƚнŶ
^ƚнϭ ^ƚнŶ ;dž͕LJ͕njͿϭ
/ƚнϭ Ŷ /ƚнŶ
^ƚнϭ ;dž͕LJ͕njͿŶ ;dž͕LJ͕njͿϮ Ϯ
^ƚ
/ƚ /ƚнϭ Ϯ
;dž͕LJ͕njͿϮ Ϯ ^ƚ ;dž͕LJ͕njͿϮ ^ƚнŶ
/ƚ /ƚнŶ
;dž͕LJ͕njͿŶ Ŷ Ŷ
;dž͕LJ͕njͿŶ
^ƚ ^ƚнϭ
/ƚ /ƚнϭ ϭ
;dž͕LJ͕njͿϭ ϭ ;dž͕LJ͕njͿϭ

Dϰƚ Dϰƚнϭ DϰƚнŶ

F I G U R E 1 Schematic representation of 4D ABM with the agents' states S, input data I, and transition rules R
for initial time t and iterations at t + 1 to t + n
SMITH and DRAGIĆEVIĆ | 421

3 | S M O K E PRO PAG ATI O N M O D E L

3.1 | Study area and data

The study area for this case study is British Columbia (BC) and the western half of Alberta, Canada, including the
metropolitan regions of Edmonton and Calgary. This area experiences an annual forest‐fire season that fills the skies
with smoke that disperses to major urban centers such as Vancouver, Calgary, and Edmonton in Canada and Seattle
in the U.S. Forest fires in BC burned a record‐breaking area of 12,000 km2 in 2017 (BC Wildfire Service, 2018a), and
the record was broken again in 2018 with more than 13,000 km2 burned (BC Wildfire Service, 2018b). The location of
the hypothetical fire used in the case study to initialize the geosimulation is on the west coast of BC, near Khartoum
Lake, approximately 100 km northwest of Vancouver (49.86°N, 124.19°E). This location was chosen because of the
westward wind patterns during the selected dates between August 10 and 25, 2017. On the modeled dates, real fires
were burning across BC, with a majority of the fires further inland, near Kamloops.
Two data sets are used for the model input: a digital elevation model (DEM; US Geological Survey, 2012) and an
atmospheric conditions data set (National Centers for Environmental Prediction et al., 2005). The DEM represents the
landscape of the study area and is not updated during model processing. It has a horizontal and vertical resolution of
1 km and 1 m, respectively. The atmospheric conditions data have wind speeds and directions at a horizontal resolution
of 32 km, and a variable vertical resolution based on 29 pressure levels with pressure differences of 25 or 50 pascals.
Because elevation in the atmospheric conditions data is based on pressure levels, there is no fixed distance for vertical
resolution. In some areas these pressure levels can be closer together while in others they may be further apart.

3.2 | Model overview

The 4D ABM represents smoke as a 4D process by modeling in the three spatial dimensions (x, y, z) and the tem‐
poral dimension (t). By including three spatial dimensions, the model has the ability to propagate smoke vertically
through the atmosphere and over terrain where different wind directions and velocities may be present.
Using the theoretical formulation of the 4D ABM, the model of smoke propagation in the four dimensions,
ABMsmoke, can be written as:

ABMsmoke ∼ (A;SF ,IF ,RF ;SK ,IK ,RK ), (7)

where the types of agents A are represented in two agent sets. The first set represents the fire agents F with their states
SF, input data IF, and transition rules RF. Smoke agents K also have their own states SK, input data IK, and transition rules RK.
The flowchart presented in Figure 2 shows the processes the two agent types and model control use to model the
smoke propagation. Fire agents represent the location of the fires being modeled. The fire agents release smoke by creat‐
ing a specified number of smoke agents every iteration. This number is determined through calibration for the available
processing power and relative size of fires. Two scenarios were developed for fire agents, scenario 1 representing the
baseline model with static fire and scenario 2 representing a moving fire agent starting at the same location as the first
one. While full modeling of the forest fires is a large and complicated undertaking, some basic movement of fire agents
has been included. Scenario 2 assumes that fire agents move at 2% of the speed of the ground‐level wind that results in
an average speed of 3.8 m/min, within the range of observed forest‐fire speeds (Perrakis, Lanoville, Taylor, & Hicks, 2014).
Smoke agents are based on two methods of movement through the atmosphere, the processes of passive dif‐
fusion and active dispersion by wind. The movement from these two methods is calculated into separate vectors
which are then added together to form a movement vector. This reduces the number of times the agents move,
increasing the efficiency of the model. The diffusion process occurs as the smoke agents move from areas of high
concentration of smoke agents to areas of low concentration to realistically represent smoke behavior in the envi‐
ronment. In order to model the diffusion process, the smoke agents use a 3D cube to find the neighboring voxel
with the most agents and move directly away from it. The agents randomize the order in which they search for the
422 | SMITH and DRAGIĆEVIĆ

&ŝƌĞŐĞŶƚƐ ^ŵŽŬĞŐĞŶƚƐ

&ŝƌĞŐĞŶƚ /ƚĞƌĂƟŽŶdнϭ

/ƚĞƌĂƟŽŶdнϭ
ƚŵŽƐƉŚĞƌŝĐ ^ŵŽŬĞŐĞŶƚ
ŽŶĚŝƟŽŶƐ
ĂůĐƵůĂƚĞ D
ƐƉƌĞĂĚďLJ ŝīƵƐŝŽŶ tŝŶĚ
ǁŝŶĚ sĞĐƚŽƌ sĞĐƚŽƌ

ŵŝƚ ^ĞĂƌĐŚ;dž͕LJ͕njͿ
^ŵŽŬĞ ǁŝŶĚŽǁŝŶƚŚĞ tŝŶĚ^ƉĞĞĚ͕ tŝŶĚsĞƌƟĐĂů
ŐĞŶƚƐ ŐƌŝĚĨŽƌƚŚĞ ŝƌĞĐƟŽŶ͕ sĞůŽĐŝƚLJ͕
ĚĞŶƐĞƐƚĐĞůů WƌĞƐƐƵƌĞ͕ WƌĞƐƐƵƌĞ͕
ůĞǀĂƟŽŶ ůĞǀĂƟŽŶ
tƌŝƚĞůŽĐĂƟŽŶ
ƚŽĮůĞ

ĂůĐƵůĂƚĞ ĂůĐƵůĂƚĞ
/ƐƚŚĞĂŐĞŶƚ ŚŽƌŝnjŽŶƚĂů ǀĞƌƟĐĂů
EŽ ůŽĐĂƚĞĚŝŶ zĞƐ ĐŽŵƉŽŶĞŶƚƐ ĐŽŵƉŽŶĞŶƚ
ƚŚĞĚĞŶƐĞƐƚ
ĐĞůů͍ ĂůĐƵůĂƚĞ
ĂůĐƵůĂƚĞǀĞĐƚŽƌ ǀĞĐƚŽƌŝŶĂ ŽŵďŝŶĞƚŽĨŽƌŵ
ŽƉƉŽƐŝƚĞĨƌŽŵ ƌĂŶĚŽŵ tŝŶĚsĞĐƚŽƌ
ĚĞŶƐĞƐƚĐĞůů ĚŝƌĞĐƟŽŶ
D

ĚĚƚŽĨŽƌŵ
/ƐƚŚĞĂŐĞŶƚ zĞƐ
DŽǀĞŵĞŶƚsĞĐƚŽƌ
ďĞůŽǁ DŽǀĞƚŽƐƵƌĨĂĐĞ
ŐƌŽƵŶĚ͍
DŽǀĞŵĞŶƚ
EŽ

ŽŶƚƌŽůDŽĚƵůĞ

ŚĞĐŬKƵƚƉƵƚ ůů^ŵŽŬĞŐĞŶƚƐ tƌŝƚĞ

/ƚĞƌĂƟŽŶdнϭ

ůŽĐĂƟŽŶƚŽ
ĮůĞ
hƉĚĂƚĞ
ŚĞĐŬĂƚĂhƉĚĂƚĞ
ĂƚĂ

FIGURE 2 Flowchart of 4D ABM for smoke propagation process with information on each agent type

highest concentration to avoid a bias in direction when two voxels have an equal amount of agents, forming the
first of three sources of heterogeneity in the model. When the agent is in the most populated voxel, it will move in
a random direction, representing the second source of heterogeneity.
Movement by the wind process is operationalized by the smoke agents reading the horizontal wind speed, hori‐
zontal direction, and vertical movement from the atmospheric data at their locations. The wind movement vector is
determined by the distance the wind would blow in the specified horizontal direction and repeating the process for
the vertical component. When both movements are made the agent's new location is checked to ensure it is above
ground, and the agent is moved to the surface if it is below ground. The order in which the agents are processed is
determined randomly by the modeling environment and represents the third source of heterogeneity in the model.
The control module (Figure 2) dictates data updates and model outputs. Atmospheric data are updated period‐
ically, depending on the frequency of the available data. For example, if new atmospheric data are available every
3 hours, the model performs an update that corresponds to 3 hours of real time. Output from the model at regular
intervals is provided when the smoke agents and fire agents output their locations to external files containing all
the agent locations for the current output. These files can be used to view agent locations or for post‐processing
in suitable geographic information system software.
SMITH and DRAGIĆEVIĆ | 423

3.3 | Model components

The variation of the Overview, Design, and Details (ODD) protocol (Grimm et al., 2006, 2010) adapted for this
study was used to describe the 4D ABM of smoke propagation to improve communication of the model develop‐
ment process. The ODD protocol has been modified for this study to better represent the different components
of the developed 4D ABM by omitting unused sections and adding a section for output (Table 1).

4 | R E S U LT S A N D D I S CU S S I O N

4.1 | Geosimulation outputs

The geosimulation model was implemented by development of coding routines in Repast Simphony using Java
(North et al., 2013). The model has a temporal resolution of 1 minute represented by one single iteration. Every
180 iterations, one output is generated that represents 180 minutes of time in reality. In total, 120 outputs are cre‐
ated representing the 15 days between August 10 and 25, 2017. Outputs from the model are in the form of *.txt
files but differ between the agent types. Fire agents output their location and time to a single file for all iterations.
Smoke agents output their location to a file specific to the current iteration, to create a series of files each of which
contains the locations of all agents during the one iteration it represents. These files are processed in Esri's ArcGIS
suite of software (Esri, 2018) to create a GIS layer of point locations for each output. For select simulation outputs,
ArcMap was used to create 2D maps and ArcGlobe was used to create 3D representations.
The simulation results for two‐day intervals are presented in Figures 3a and b, representing the first 12 days of
smoke propagation for both scenario 1 (Figure 3a) and scenario 2 (Figure 3b). The simulations indicate that smoke
agents stay local to the fire in the Strait of Georgia, BC prior to day 2 before being blown north into the interior
of BC by day 4. For the remainder of the model, the smoke generally propagates east into Alberta, reaching as far
south as the metropolitan area of Calgary between days 4 and 6, and leaving the study area in northern Alberta
after day 8 before returning to the metropolitan area of Edmonton by day 10. Figure 3c compares both scenarios
on day 12, showing two main differences. The first is the difference in the smoke near the coast where the fires
are located, mostly due to the fires being in different locations. The second is the location of the next plume in
each scenario, where the static fire (scenario 1), in yellow, has not traveled as far north as the related plume from
the moving fire (scenario 2) toward the north, in red. In all other locations the simulated pattern of smoke propa‐
gation from scenario 2 is very similar to that obtained for scenario 1. The path of the moving fire agent is shown in
Figure 3d with a general eastward movement.
Outputs for scenario 2, as more realistic with representation of moving forest fire locations, are used to pres‐
ent the first 12 hours of smoke propagation in Figure 4. These outputs are presented in the form of 3D perspec‐
tives with an oblique view on the left of each pair and the matching vertical view on the right with elevation
exaggerated by a factor of 5. In addition, the output at 24 hours is presented at the top from a low viewpoint.
The results indicate that the smoke agents form two main plumes, with the second forming around the second
output (Figure 4c) when the wind patterns change. The second and larger plume gets blown towards the Strait
of Georgia, BC, the large body of water to the west. In terms of height, the simulations indicate that the second
plume is much taller than the first, getting its height from converging and uplifting winds occurring between 9 and
12 hours (Figures 4d and e).
The summary of the vertical distribution of the smoke agents in scenario 2 is presented in Figure 5. The maxi‐
mum number of smoke agents is reached in the output for 8,100 minutes, or 5 days and 15 hours. The peak in the
distribution of smoke reaches a maximum of around 2,600 m on day 7, remaining close to this height until day 10
when it begins to drop, rising again toward the end when another peak of approximately 2,600 m is reached on
day 15. However, days 1, 2, 5, 8, and 13 show more smoke under 1,000 m than the other days. This was shown in
Figure 3 where the simulation outputs for these days contain the most smoke in the Strait of Georgia, which has
an elevation of 0 m. For comparison, the elevation of Vancouver ranges from ~0 to ~350 m, Blackcomb Peak, the
TA B L E 1 The ODD protocol for the 4D ABM
424
|

Model overview

Purpose The purpose of the model is to represent smoke propagation in four dimensions by providing an environment in which to run cus‐
tomized scenarios. The model can generate various scenarios to forecast possible 3D locations of smoke from existing fires and to
estimate the risks from fire emissions. The model also aims to demonstrate the use of agent‐based modeling in the representation
of a physical complex system to mitigate difficulties of process representation with often non‐spatial equation models.
Entities, state variables, and scales The model contains two agent types:
• 
Fire agents. These agents represent the source of the smoke, often forest fires. Fire agents have two variables. One variable
determines the number of smoke agents released every model iteration, for this study set to five agents per one‐minute iteration.
The other variable determines how fast the fire agent moves compared to the wind, set to 0% for scenario 1 to represent no
influence due to wind, and 2% of wind speed for scenario 2 to represent moving location for fire.
• 
Smoke agents. These agents represent the smoke released by the fire agents and are uniform in characteristics and transition
rules. Smoke agents move through the atmosphere by diffusion and wind processes.
Diffusion process. Smoke agents search a 3D neighborhood for the voxel with the most smoke agents. The agent will then move
directly away from this voxel or in a random direction if it is detected in the central voxel. Diffusion has two variables, cubic neigh‐
borhood size, and movement distance.
Wind process. Smoke agents move with the wind by reading the atmospheric conditions for their location and moving in the same
direction as the wind by the distance the wind moves in the time of one iteration. There are no agent variables in this process,
simply the agent location and input data.
Ground check.: If the smoke agent is below ground level as reported by the digital elevation model, the agent will move to the
surface.
The model also contains a control module with three components:
• 
Main control. This part of the module sets the model parameters such as scale, extent, temporal resolution, and the model time
and date, and defines the fire agents' starting locations. This module also manages the location of the input data for data control.
• 
Data control. This part of the module reads input data into the model for use by the agents and converts data into alternate
formats if required.
• 
Time control. This part of the module dictates when it is time to change to the next atmospheric data set and when to write agent
outputs.
The user is able to modify several key variables that affect agents (current values in parentheses): voxel size or spatial resolu‐
tion (100 m3), diffusion distance inner shell (0.36 voxels) outer shell (0.18 voxels), neighborhood size (5 × 5 × 5 voxels), output
frequency (180 minutes), data update frequency (180 minutes), iteration length or temporal resolution (1 minute), model duration
(15 days), as well as many other variables found in the “Details” section of this table.

(Continues)
SMITH and DRAGIĆEVIĆ
TA B L E 1 (Continued)

Model overview

Process overview and scheduling The model begins with a setup phase where input data are loaded into the appropriate location for access. Then the model creates
the desired fire agents.
Each time‐step or iteration follows the same series of steps:
1. Fire agents emit the desired smoke agents.
SMITH and DRAGIĆEVIĆ

2. Fire agents move, depending on the wind.

3. Smoke agents move using diffusion and wind.
4. Smoke agents checks if output is required. Writes location if required.
5. Control module checks if a data update is required. Updates if required
Design concepts
Emergence The emergence of smoke propagation is represented by the aggregated movement of smoke agents. The smoke patterns are mainly
affected by the measured atmospheric conditions.
Adaptation The smoke agents adapt to changing atmospheric conditions and nearby agents through their movement by moving with the current
wind and moving away from other smoke agents. The rules smoke agents use to determine movement are static for the entire run of
the model.
Interaction Smoke agents interact through the process of diffusion where agents move from areas of high smoke concentration to areas of low
concentration. To do this the smoke agents search their surrounding area and move directly away from high concentrations of
other smoke agents.
Stochasticity Three random processes occur within the model:
• The order in which agents are processed is random and controlled by the modeling environment.
• During the diffusion process, the smoke agents randomize the order they search the voxels within the neighborhood to avoid a
bias in any one direction caused by searching in the same order every time.
• During the diffusion process, when the smoke agent is in the voxel with the most smoke agents, it will move in a random
direction.
Collectives Collectives are actively avoided through diffusion to prevent the smoke agents from forming clumps. This allows smoke plumes to
emerge when wind patterns and terrain overpower diffusion instead of the smoke agents actively trying to form plumes.
Calibration The number of agents and the diffusion process need to be calibrated for each unique scenario. The number of agents is depend‐
ent primarily on the computational power available. While more agents will typically produce better results, they will also linearly
increase the processing time required, meaning that the model has computational complexity O(n). The number of fire agents used
will also be impacted by the number and size of fires being represented. Some larger fires may require two or more agents to be accurately
represented.
|

The diffusion process needs to be calibrated for the movement distance and the cubic neighborhood size. These variables are
adjusted together to match the results from Kellogg (1956) and is mainly dependent on the voxel size and number of smoke agents.
425

(Continues)
TA B L E 1 (Continued)
426
|

Model overview

Outputs Smoke agents' locations are collected at regular intervals and saved to external files where each agent writes its (x, y, z) coordinate to
one line of the file. These files can be opened for viewing and post‐processing using GIS software.
Details
Initialization Modeling Environment: Repast Simphony 2.4
Language: Java
Projection: Lambert conformal conic
Origin lower left (x, y, z) [m]: –1785776.725475, –376124.151374, 0
Voxel size (x, y, z) [m]: 100, 100, 100
Extent (x, y, z) [voxels]: 14000, 14000, 200
Iteration length [minutes]: 1
Output interval [minutes]: 180 (3 hours)
Time extent [minutes]: 21600 (15 days)
Simulation start date: corresponds to August 10, 2017, 0:00
Simulation end date: corresponds to August 25, 2017, 0:00
Number of fire agents: 1
Fire agent location (x, y, z) [voxels]: 5723, 5090, 10
Smoke agent spawn rate [agents/iteration]: 5
Fire agent speed, scenario 1: 0%
Fire agent speed, scenario 2: 2%
Smoke agent neighborhood size (x, y, z) [voxels]: 5, 5, 5
Smoke agent diffusion outer shell [voxels]: 0.18
Smoke agent diffusion inner shell [voxels]: 0.36
Smoke agent diffusion center [voxels]: 0.18

(Continues)
SMITH and DRAGIĆEVIĆ
TA B L E 1 (Continued)

Model overview

Input Data Atmospheric data source: National Center for Atmospheric Research's North American Regional Reanalysis, ds608.0
Link: https​://rda.ucar.edu/datas​ets/ds608.0/
Format: GRIB1
SMITH and DRAGIĆEVIĆ

Extent: North America, clipped to modeling extent

Horizontal resolution [km]: 32
Vertical resolution [Pa]: variable between 25 and 50
Vertical layers: 29
Dates available: January 1, 1979 to present, updated monthly
Dates used: 00:00 August 10, 2017 to 24:00 August 24, 2017
Layers used: Geopotential height [gpm], u wind [m/s], v wind [m/s], vertical velocity [Pa/s], pressure (obtained from layer name)
Digital elevation data source: USGS Global Digital Elevation Model (GTOPO30)
Resolution, horizontal [km]: 1
Resolution, vertical [m]: 1
Extent: North America, clipped to modeling extent
Format: GeoTIFF

gpm, geopotential meter.

|
427
428 | SMITH and DRAGIĆEVIĆ

F I G U R E 3 4D ABM simulation smoke outputs presented in 2D every 2 days up to 12 days for: (a) scenario 1
with static fire; (b) scenario 2 with moving fire; and (c) comparison of day 12 for both scenarios; and (d) the path
of the fire agent for scenario 2

alpine location of the 2010 Winter Olympics, is ~2,400 m, Kelowna, BC in the Okanagan Valley is ~350 m, Banff,
Alberta is ~1,400 m, and Calgary, Alberta is ~1,000 m. With smoke generally higher than even some mountain
peaks, it is these days with low smoke that pose the greatest threat to human health.

4.2 | Model validation

The lack of 3D smoke location data prevents full model validation. Only two sources of data were found, satel‐
lite imagery and ground reports, but both include smoke from anthropogenic sources such as factories. Satellite
imagery also compresses all smoke into a 2D layer which has limited capabilities to determine accurate concentra‐
tion values. Ground reports, such as weather stations, only report on ground‐level smoke, and do not measure
smoke higher in the atmosphere.
In order to proceed with 4D ABM validation, simulated spatial patterns of smoke propagation are visually
compared to satellite images from NASA (Figure 6), making several key points of comparison. While similarities
between the modeled smoke propagation and satellite imagery of smoke from different fires do not fully validate
the model, the similarities in the observed patterns and distribution still form a partial validation. The validation
SMITH and DRAGIĆEVIĆ | 429

F I G U R E 4 Geosimulation outputs for scenario 2 with moving fire agent depicting (a) an oblique view of
smoke agents after 24 hours from a low view point; oblique views (left) and vertical views (right of smoke agents
forming a plume from the moving fire location with outputs after (b) 3 hours, (c) 6 hours, (d) 9 hours and (e)
12 hours from the initialization of the model. Vertical distances scaled 5 × to exaggerate elevation differences
430 | SMITH and DRAGIĆEVIĆ

of patterns can show that the model produces a realistic representation of the smoke propagation as complex
system.
The first two images (Figures 6a and b) are from August 11 and 14, 2017, corresponding to approximately
1.5 days and 4.5 days from the start of the model at midnight on the morning of August 10. These images show
similar spatial patterns to those presented in Figures 3a and b where the smoke enters the interior plateau, moves
north, and on day 4 begins to exit the plateau to the northeast, north of the Cariboo Mountains. This shows that
the model follows wind patterns and terrain properly. The main difference with these images is that the smoke
in the imagery fills the entire plateau instead of a single plume in the model. This can be attributed to the model
containing only a single source of smoke instead of many small fires geographically distributed.
Figure 6c shows the smoke moving to the east through Alberta, and while it is approximately 20 days after
the time period that was modeled, it shows similarities with the simulated spatial patterns. This shows that the
prevailing wind patterns will cause the smoke to take a similar path. However, because the imagery is from a later
date that is not included in the simulation, the patterns are not an exact match.
The final comparison shows the smoke trapped over the Strait of Georgia (Figure 6d) and can be seen occur‐
ring to a smaller extent in the simulations presented in Figures 3a, b and 4e. In the area between the mainland and
Vancouver Island, the tall mountains can trap smoke as both a physical barrier and by influencing wind patterns.
This is significant in the study area due to the urban centers of Vancouver, Victoria, and Seattle being located in
this region. This smoke pattern was visible in many satellite images, including Figures 6a and b, however Figure 6d
is the most extreme in terms of amount of smoke trapped in this location. The obtained simulation outcomes
demonstrate trapped smoke over the Strait of Georgia, indicating that the proposed 4D ABM model can generate
non‐imposed patterns emerging from the represented processes.
In correspondence to the actual simulation dates, it was observed that on August 14, 2017 at 04:00 smoke was
reported in the metropolitan area of Calgary, Alberta with an Air Quality Health Index of 10 as reported by the
Government of Canada's real‐time air quality website (Environment and Natural Resources, 2017). The metropoli‐
tan area of Calgary is on the far east side of the study area, and 04:00 on August 14 corresponds to 5,940 minutes
of model simulation, or in reality 4 days and 3 hours, after adjusting for time zones. The model shows smoke pres‐
ent in the metropolitan area of Calgary after 7,200 minutes (5 days) of model simulations, or in reality at midnight
of August 14. Due to limited reports, it is difficult to identify how long the smoke event lasted in Calgary. The delay
in the simulated smoke reaching Calgary may be due to the distance of travel after the change in wind patterns.
The real smoke originated from the nearby Rocky Mountains, while the simulated smoke traveled from the coastal
area. The overall obtained spatial patterns of smoke propagation generated by the proposed 4D ABM are similar
to the real smoke propagation patterns.

dŝŵĞ;ĚĂLJƐͿ
ϭ Ϯ ϯ ϰ ϱ ϲ ϳ ϴ ϵ ϭϬ ϭϭ ϭϮ ϭϯ ϭϰ ϭϱ
ϭϯϬϬϬ ϭϯϬϬϬ

ϭϮϬϬϬ ϭϮϬϬϬ

ϭϭϬϬϬ ϭϭϬϬϬ

ϭϬϬϬϬ ϭϬϬϬϬ
ůĞǀĂƟŽŶ;ŵͿ

ϵϬϬϬ ϵϬϬϬ

ϴϬϬϬ ϴϬϬϬ

ϳϬϬϬ ϳϬϬϬ

ϲϬϬϬ ϲϬϬϬ

ϱϬϬϬ ϱϬϬϬ

ϰϬϬϬ ϰϬϬϬ

ϯϬϬϬ ϯϬϬϬ

ϮϬϬϬ ϮϬϬϬ

ϭϬϬϬ ϭϬϬϬ

Ϭ Ϭ
Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ Ϭ ϱϬϬ ϭϬϬϬ ϭϱϬϬ

EƵŵďĞƌŽĨ^ŵŽŬĞŐĞŶƚƐ

F I G U R E 5 Quantification of simulation distribution of all smoke agents in 100 m bins of vertical distance at a
daily time‐step for scenario 2
SMITH and DRAGIĆEVIĆ | 431

F I G U R E 6 Examples of smoke propagation patterns for British Columbia and Alberta, Canada, on: (a) August
11, 2017; (b) August 14, 2017; (c) July 18, 2017; and (d) July 6, 2015 (NASA, 2019)

5 | CO N C LU S I O N S

The 4D ABM developed for the representation of forest‐fire smoke propagation is a proof of concept demonstrat‐
ing the potential of ABMs to model 4D physical complex systems. The model produced smoke propagation pat‐
terns that correspond to the location and distribution patterns observed in satellite imagery of forest‐fire smoke.
This indicates the 4D ABM is capable of representing the geospatial complex systems of a physical process where
agents characterize matter such as particles of smoke or dirt with interactions and relatively simple decision‐mak‐
ing process. Agents used in a model representing a physical process only need to use information from their sur‐
roundings with simple equations to influence their decisions and replicate the actual geospatial process.
432 | SMITH and DRAGIĆEVIĆ

There are some limitations in the current model, and further research in needed to address them. The first is
the inability to fully validate the 4D ABM with actual smoke data collected in 3D over time. One solution is to use
LiDAR ceilometers (van der Kamp, McKendry, Wong, & Stull, 2008), but their use is cost prohibitive at regional
scale. The second limitation is the resolution of the atmospheric conditions data. At the current spatial resolution,
entire valleys and mountains can be contained within one voxel horizontally, and only a few voxels vertically.
The low spatial and temporal resolution of the atmospheric data also causes the smoke agents to move in near
linear patterns because they will be blown in one constant direction until they reach the next atmospheric voxel.
The solutions to this are either to obtain higher‐resolution data, or to create an additional atmospheric model
to artificially increase the resolution. Moreover, there is a lack of methods to validate model outputs that fully
include both 3D space and time. Consequently, the next research stage includes the development of 4D model
validation approaches that can compare simulated patterns of voxels and inner processes of agents' behavior in
four dimensions. With the proposed model enhancement and full validation, there is potential for professionals
in the public health sector or in fire and rescue services who are in need of smoke propagation forecasting to use
the proposed model.
In summary, the 4D ABM developed demonstrates the strengths of the proposed geosimulation approach and
provides a foundation for further development of models and analysis methods that can operate in four dimen‐
sions and can represent physical processes as complex systems.

AC K N OW L E D G M E N T S

This study was fully funded by a Natural Sciences and Engineering Research Council (NSERC) of Canada Discovery
Grant awarded to the second author. Authors are thankful to valuable comments and feedback of the anonymous
reviewer.

C O N FL I C T O F I N T E R E S T

There is no conflict of interest.

ORCID

Suzana Dragićević https://orcid.org/0000-0003-4144-7530

REFERENCES

Agusdinata, D. B., Amouie, M., & Xu, T. (2015). Diffusion dynamics and concentration of toxic materials from quantum
dots‐based nanotechnologies: An agent‐based modeling simulation framework. Journal of Nanoparticle Research, 17,
26.
Alam, M. J., & Dutta, D. (2012). A process‐based and distributed model for nutrient dynamics in river basin: Development,
testing and applications. Ecological Modelling, 247, 112–124.
Anderson, T., & Dragićević, S. (2015). An agent‐based modeling approach to represent infestation dynamics of the emer‐
ald ash borer beetle. Ecological Informatics, 30, 97–109.
Azimi, M., Jamali, Y., & Mofrad, M. R. K. (2011). Accounting for diffusion in agent based models of reaction‐diffusion
systems with application to cytoskeletal diffusion. PLoS One, 6(9), e25306.
BC Wildfire Service. (2018a). Wildfire season summary. Retrieved from https​://www2.gov.bc.ca/gov/conte​nt/safet ​y/wildf​
ire-statu​s/about-bcws/wildf​ire-histo​r y/wildf​irese​ason-summary
BC Wildfire Service. (2018b). Current statistics. Retrieved from http://bcfir​einfo.for.gov.bc.ca/hprSc​ripts/​Wildf​ireNe​ws/
Stati​stics.asp
Bova, M. J., Ciarallo, F. W., & Hill, R. R. (2016). Development of an agent‐based model for the secondary threat resulting
from a ballistic impact event. Journal of Simulation, 10, 24–35.
SMITH and DRAGIĆEVIĆ | 433

Castella, J.‐C., Trung, T. N., & Boissau, S. (2005). Participatory simulation of land‐use changes in the northern mountains
of Vietnam: The combined use of an agent‐based model, a role‐playing game, and a geographic information system.
Ecology and Society, 10, 27.
Clarke, K. C., Brass, J., & Riggan, P. J. (1994). A cellular‐automaton model of wildfire propagation and extinction.
Photogrammetric Engineering & Remote Sensing, 60, 1355–1367.
Crooks, A., Castle, C., & Batty, M. (2008). Key challenges in agent‐based modelling for geo‐spatial simulation. Computers,
Environment and Urban Systems, 32, 417–430.
Crooks, A. T., & Hailegiorgis, A. B. (2014). An agent‐based modeling approach applied to the spread of cholera.
Environmental Modeling & Software, 62, 164–177.
Dennekamp, M., & Abramson, M. J. (2011). The effects of bushfire smoke on respiratory health. Respirology, 16, 198–209.
Dragićević, S., & Hatch, K. (2018). Urban geosimulations with the Logic Scoring of Preference method for agent‐based
decision‐making. Habitat International, 72, 3–17.
Elliott, C. (2014). Guidance for BC public health decision makers during wildfire smoke events. Vancouver, Canada: BC Centre
for Disease Control.
Environment and Natural Resources. (2017). Calgary: Air quality health index. Retrieved from https​://weath​er.gc.ca/airqu​
ality/​pages/​abaq-002_e.html
Esri. (2018). ArcGIS Desktop: Release 10.5. Redlands, CA: Esri.
Genois, M., Du Pont, S. C., Hersen, P., & Gregoire, G. (2013). An agent‐based model of dune interactions produces the
emergence of patterns in deserts. Geophysical Research Letters, 40, 3909–3914.
Greene, N. (1989). Voxel space automata: Modeling with stochastic growth processes in voxel space. Computer Graphics,
23, 175–184.
Grell, G. A., Peckham, S. E., Schmitz, R., McKeen, S. A., Frost, G., Skamarock, W. C., & Eder, B. (2005). Fully coupled “on‐
line” chemistry within the WRF model. Atmospheric Environments, 39, 6957–6975.
Grimm, V., Berger, U., Bastiansen, F., Eliassen, S., Ginot, V., Giske, J., … DeAngelis, D. L. (2006). A standard protocol for
describing individual‐based and agent‐based models. Ecological Modelling, 198, 115–126.
Grimm, V., Berger, U., DeAngelis, D. L., Polhill, J. G., Giske, J., & Railsback, S. F. (2010). The ODD protocol: A review and
first update. Ecological Modelling, 221, 2760–2768.
He, C., Huang, Z., & Wang, R. (2013). Land use change and economic growth in urban China: A structural equation anal‐
ysis. Urban Studies, 51, 2880–2898.
Heppenstall, A., Malleson, N., & Crooks, A. (2016). “Space, the final frontier”: How good are agent‐based models at simu‐
lating individuals and space in cities? Systems, 4, 9.
Heung, B., Bakker, L., Schmidt, M. G., & Dragićević, S. (2013). Modelling the dynamics of soil redistribution induced by
sheet erosion using the Universal Soil Loss Equation and cellular automata. Geoderma, 202–203, 112–125.
Huang, Q., Parker, D. C., Filatova, T., & Sun, S. (2014). A review of urban residential choice models using agent‐based
modeling. Environment and Planning B, 41, 661–689.
Jjumba, A., & Dragicevic, S. (2015). Integrating GIS‐based geo‐atom theory and voxel automata to simulate the dispersal
of airborne pollutants. Transactions in GIS, 19, 582–603.
Kellogg, W. W. (1956). Diffusion of smoke in the stratosphere. Journal of Meteorology, 13, 241–250.
Krebs, F. (2017). An empirically grounded model of green electricity adoption in Germany: Calibration, validation and
insights into patterns of diffusion. Journal of Artificial Societies and Social Simulation, 20(2), 10.
Luus, K. A., Robinson, D. T., & Deadman, P. J. (2011). Representing ecological processes in agent‐based models of land use
and cover change. Journal of Land Use Science, 8, 175–198.
Macatulad, E. G., & Blanco, A. C. (2014). 3DGIS‐based multi‐agent geosimulation and visualization of building evacuation
using GAMA platform. International Archives of Photogrammetry, Remote Sensing and Spatial Information Science, 40,
87–91.
Manson, S. M. (2001). Simplifying complexity: A review of complexity theory. Geoforum, 32, 405–414.
Murray‐Rust, D., Brown, C., van Vliet, J., Alam, S. J., Robinson, D. T., Verburg, P. H., & Rounsevell, M. (2014). Combining
agent functional types, capitals and services to model land use dynamics. Environmental Modeling & Software, 59,
187–201.
NASA. (2019). World fire image gallery. Retrieved from https​://www.nasa.gov/missi​on_pages/​fires/​main/world/​index.
html
National Centers for Environmental Prediction, National Weather Service, NOAA, & US Department of Commerce.
(2005). NCEP North American Regional Reanalysis (NARR), Research Data Archive at the National Center for Atmospheric
Research, Computational and information Systems Laboratory. Retrieved from http://rda.ucar.edu/datas​ets/ds608.0/
North, M. J., Collier, N. T., Ozik, J., Tatara, E. R., Macal, C. M., Bragen, M., & Sydelko, P. (2013). Complex adaptive systems
modeling with Repast Simphony. Complex Adaptive Systems Modeling, 1, 3.
434 | SMITH and DRAGIĆEVIĆ

Parunak, H. V. D., Savit, R., & Riolo, R. L. (1998). Agent‐based modeling vs. equation‐based modeling: A case study and
users' guide. In J. S. Sichman, R. Conte, & N. Gilbert (Eds.), Multi‐agent systems and agent‐based simulation: MABS 1998
(Lecture Notes in Computer Science, Vol. 1534, pp. 10–25). Berlin, Germany: Springer.
Perez, L., & Dragicevic, S. (2009). An agent‐based approach for modeling dynamics of contagious disease spread.
International Journal of Health Geographics, 8, 50.
Pérez, L., & Dragićević, S. (2011). ForestSimMPB: A swarming intelligence and agent‐based modeling approach for moun‐
tain pine beetle outbreaks. Ecological Informatcs, 6, 62–72.
Perrakis, D. D. B., Lanoville, R. A., Taylor, S. W., & Hicks, D. (2014). Modeling wildfire spread in mountain pine beetle‐
affected forest stands, British Columbia, Canada. Fire Ecology, 10, 10–35.
Rahmandad, H., & Sterman, J. (2008). Heterogeneity and network structure in the dynamics of diffusion: Comparing
agent‐based and differential equation models. Journal Management Science, 54(5), 998–1014.
Reaney, S. M. (2008). The use of agent based modelling techniques in hydrology: Determining the spatial and temporal
origin of channel flow in semi‐arid catchments. Earth Surface Processes and Landforms, 33, 317–327.
Rolph, G. D., Draxler, R. R., Stein, A. F., Taylor, A., Ruminski, M. G., Kondragunta, S., … Davidson, P. M. (2009). Description
and verification of the NOAA Smoke Forecasting System: The 2007 fire season. Weather Forecasting, 24, 361–378.
Rounsevell, M. D., Robinson, D. T., & Murray‐Rust, D. (2012). From actors to agents in socio‐ecological systems models.
Philosophical Transactions of the Royal Society of London Series B: Biological Sciences, 367, 259–269.
Smith, A. K., & Dragicevic, S. (2018). An agent‐based model to represent space‐time propagation of forest‐fire smoke.
ISPRS Annals of Photogrammetry, Remote Sensing and Spatial Information Science, 4(4), 207–212.
Sun, S., Parker, D. C., Huang, Q., Filatova, T., Robinson, D. T., Riolo, R. L., … Brown, D. G. (2014). Market impacts on land‐
use change: An agent‐based experiment. Annals of the Association of American Geographers, 104(3), 460–484.
Tang, W., & Bennett, D. A. (2010). Agent‐based modeling of animal movement: A review. Geography Compass, 4, 682–700.
Testa, J. W., Mock, K. J., Taylor, C., Koyuk, H., Coyle, J. R., & Waggoner, R. (2012). Agent‐based modeling of the dynamics
of mammal‐eating killer whales and their prey. Marine Ecology Progress Series, 466, 275–291.
Torrens, P. M. (2014). High‐resolution space‐time processes for agents at the built‐human interface of urban earthquakes.
International Journal of Geographical Information Science, 28, 964–986.
Torrens, P. M., & Benenson, I. (2005). Geographic automata systems. International Journal of Geographical Information
Science, 19, 385–412.
US Geological Survey. (2012). Global Digital Elevation Model (GTOPO30). Retrieved from https​://www.usgs.gov/cente​rs/
eros/scien​ce/usgs-eros-archi​ve-digit​al-eleva​tion-global-30-arc-second-eleva​tion-gtopo​3 0?qt-scien​ce_center_objec​
ts=0#qt-scien​ce_center_objects
van der Kamp, D., McKendry, I., Wong, M., & Stull, R. (2008). LiDAR ceilometer observations and modeling of a fireworks
plume in Vancouver, British Columbia. Atmospheric Environment, 42, 7174–7178.
Zhu, W., & Timmermans, H. (2011). Modeling pedestrian shopping behavior using principles of bounded rationality:
Model comparison and validation. Journal of Geographical Systems, 13, 101–126.

How to cite this article: Smith AK, Dragićević S. A four‐dimensional agent‐based model: A case study of
forest‐fire smoke propagation. Transactions in GIS. 2019;23:417–434. https​://doi.org/10.1111/tgis.12551​