Environ Model Assess DOI 10.

1007/s10666-006-9055-5

Evaluating forest management practices using a GIS-based cellular automata modeling approach with multispectral imagery
Christopher Bone & Suzana Dragievi & Arthur Roberts

Received: 12 September 2005 / Accepted: 10 May 2006 # Springer Science + Business Media B.V. 2006

Abstract The objective of this study was to develop an integrated geographic information system (GIS) cellular automata (CA) model for simulating insect-induced tree mortality patterns in order to evaluate the influence of different forest management activities to control insect outbreaks. High-resolution multispectral images were used to determine susceptibility of trees to attack, whereas the GIS-based CA model simulated the effectiveness of clearcuts and thinning practices for reducing insect-induced tree mortality. The results indicate that thinning susceptible forests should be more effective than clear-cutting for reducing tree loss to insect outbreaks. This study demonstrates the benefits of an integrated approach for understanding and evaluating forest management activities and expresses the need for spatial analysis and modeling for improving forest management practices. Keywords cellular automata (CA) . geographic information systems (GIS) . remote sensing (RS) . spatial modeling . forest management . forest insect outbreaks . mountain pine beetle

1 Introduction Remote sensing (RS) and geographic information systems (GIS) provide the opportunity to examine forest resources
C. Bone (*) : S. Dragievi : A. Roberts Department of Geography, Simon Fraser University, 8888 University Drive, Burnaby, BC, Canada V5A 1S6 e-mail: cbone@sfu.ca S. Dragievi e-mail: suzanad@sfu.ca A. Roberts e-mail: aroberts@sfu.ca

and obtain insight into appropriate methods for managing them. RS data can yield spatial information for monitoring forest characteristics such as species diversity [14, 38], stand density [34], and natural disturbances [25, 47, 39], among others, that are important for management decisions. GIS can facilitate data analysis of these characteristics through a host of spatial and statistical approaches. The effectiveness of RS and GIS has led to their use for developing forest management models for determining practical strategies. This includes combining RS and GIS with traditional knowledge of forest practices to adapt inventories for forest management planning [30] and for analyzing biophysical and social patterns in order to implement management practices [32]. However, although such analytical models are important for management, they are usually static representations applicable to a single moment in time. Considering the dynamic nature of forests, management decisions would benefit from being able to simulate various practices in a virtual environment to determine how management decisions affect forest structure and processes over time. A temporal component for forest management models can be provided by cellular automata (CA) modeling in a GIS environment using RS data. CA are spatially dynamic models where a set of simple transition rules govern changes in cell states that represent different landscape elements [1, 44]. These transition rules explain how the current states of cells in a defined area called the neighborhood influence the state of each cell at some future moment in time. CA have been employed for modeling a variety of geographic processes where land use changes over time. Examples include modeling urban growth [8, 9, 11, 49], land retirement [26], coastal-zone management [23], and socioenvironmental systems [15], among others.

Environ Model Assess

The integration of RS, GIS, and CA is beneficial for three main reasons. First, RS images are raster-based data sets in which a landscape is represented by a grid of cells. Each cell contains a value that corresponds to a specific characteristic of the landscape. CA models traditionally use a grid of cells to simulate dynamic processes because it provides convenience for neighborhood calculations. Thus, RS images can provide information in a format that is readily used by a CA model. Furthermore, GIS provide numerous spatial statistical tools that can be of great utility for CA models. The use of most of these tools requires the data to be in raster format. GIS analytical tools have been utilized in CA models for simulating the dynamics of urban processes [12, 43, 48] and land use change [24]. Second, CA add a temporal component to the otherwise static nature of RS and GIS. RS images only provide a snapshot of a landscape at a particular moment in time, which is a limitation for modeling phenomena that exhibit continual change. This problem is magnified by the high costs of RS images that make it difficult to continually obtain data at the same rate at which a phenomenon exhibits change. Furthermore, the tools provided in most GIS applications are ill-equipped for representing the dynamic nature of geographic phenomena. Therefore, CA provide a utility for RS and GIS, as they can simulate how information captured at a particular moment is likely to change over time. Third, the simple CA rules governing state transition facilitate computation efficiency [10]. High-resolution images are large in size and contain hundreds of thousands of pixels. As a result, the images can hinder complex applications due to the time it takes to process information. CA, however, typically use simple rules that evaluate local interactions between cells, and therefore avoid the application of complex equations applied to the entire data set; this in turn can significantly reduce computation time. Although these three benefits are evident in present research, the use of CA for evaluating forest management practices has only recently been explored. For example, Strange et al. [42] developed a CA model to evaluate different land use strategies to optimize afforestation (i.e., turning bare or harvested land into forest). Their model used land quality and cost measures to determine the benefits of planting different tree species as well as transforming land to pasture. With regard to human disturbance, a CA model, FORSAT, was developed for simulating the dynamics of areas co-dominated by forest and savanna that are heavily influenced by management activities such as fires caused by humans for clearing land [16, 17]. The results demonstrated how human influence can dictate the location of forestsavanna boundaries. Mathey et al. [27] provide the most evident attempt of using CA for forest management, as they constructed a CAdriven decision support tool for evaluating multiple

objectives for achieving sustainable forest management. The authors considered economic, social, and environmental objectives to test the effectiveness of using CA for simultaneously implementing different sustainability goals. The simplicity offered through the use of CA allowed for direct integration of expert knowledge into the simulations of forest dynamics. These studies demonstrate that a CA modeling approach is beneficial for understanding how anthropogenic influences affect forest processes; however, there remains a significant gap in the literature regarding the use of CA models for simulating natural influences such as insect infestations and the effectiveness of management practices for dealing with such disturbances. Furthermore, the integration of RS, GIS, and CA for forest research in general remains largely unexplored. The objective of this study was to integrate a GIS-based CA model with high-resolution RS data for evaluating forest management decisions for dealing with insect outbreaks. A case study of mountain pine beetle (MPB), Dendroctonus ponderosae Hopkins, outbreaks in lodgepole pine, Pinus contorta, forests in British Columbia, Canada, was used. Information was extracted from the thematically classified RS data and analyzed in a GIS. The resulting data were used in the CA model that was developed based on the premise that MPB-induced tree mortality is largely controlled by the susceptibility of trees to attack and the number of MPB present in a given area [36]. The model was first calibrated to simulate patterns of tree mortality over a 6-year period without management intervention, followed by implementing different strategies such as clearcutting and thinning to determine the effectiveness of specific strategies at reducing the loss of timber to MPB outbreaks.

2 MPB outbreaks and management strategies MPB is the most serious pest of pine forests in western North America, forcing forest managers to continually evaluate ways to maximize yields and minimize loss of timber revenues. The beetle attacks both lodgepole pine and ponderosa pine (Pinus ponderosa) in British Columbia and several states in the western United States. It was estimated in 2005 that the current epidemic of MPB that began in the mid1990s had killed approximately 283 million cubic meters of pine trees in British Columbia [19], which has serious economic and social implications for a region that greatly depends on timber as a source of revenue. This regionalscale outbreak is commonly linked to: (1) decades of fire suppression that has resulted in overmature, single-species stands that are highly vulnerable to MPB attack, and (2) the lack of significantly cold winters in the interior of British

Environ Model Assess

Columbia that control MPB populations. Furthermore, preventative methods such as stand density management, creating mixed age/species stands, and harvesting trees at maturity are seldom employed [46], which results in homogenous areas of susceptible pine. In 2002, the federal government of Canada responded to this issue by initiating the Mountain Pine Beetle Initiative, which includes reducing the risk of attack to noninfested areas and rehabilitation to federal and private forestlands that have be affected by the epidemic, through both land-based and research-based programs [46]. In 2005, the provincial government of British Columbia released the Mountain Pine Beetle Action Plan [19] with the goal of sustaining long-term economic, social, and environmental viability while dealing with the immediate implications of the current epidemic. These initiatives started to look beyond traditional direct management techniques that are of limited use once MPB outbreaks reach a certain level. These limited techniques include baiting trees with a chemical attractant to draw beetles toward a specific area, cutting infested trees and burning them on site, applying pesticides, and harvesting dead trees [46]. Efforts also began to focus on using various forms of technology for understanding the complex nature of MPB outbreaks, including the use of RS imagery for detecting infested areas [47]. However, a problem exists with using RS technology to manage MPB due to their life cycle characteristics and current practices: the timing of RS detection in relation to MPB life cycle. MPB typically leave their currently infested trees in late July to early August in search of a new host tree to attack. Beetles select trees based on different characteristics, which makes some trees far more susceptible to attack than others. Once the new tree is selected, it is attacked by mass numbers of beetles in order to overcome the trees defensive mechanism, and tree mortality proceeds in the subsequent weeks and months. The beetles lay eggs under the bark of the tree where the offspring develop over the winter and spring months until they are ready to emerge and search for a new host to attack. The significant barrier for forest management using RS to locate infested trees is that dead trees do not exhibit signs of mortality (i.e., turn fully red) until the next summer, at which point the insects that were born in the tree would have left those trees in search of a new host. Early detection is possible by late May to early June [31], but RS and forestry practices have not been operationally adapted. Therefore, current RS applications attempt to monitor MPB typically by detecting dead but MPB-vacant trees. This conundrum presents an opportunity to use CA for modeling annual MPB-induced tree mortality patterns. RS data and GIS can be utilized to provide information on susceptibility, and a GIS-based CA model can produce various scenarios that would allow forest management to

determine which areas are at greatest risk each year in the future. This would also provide them with an experimental environment to test different management strategies to suppress new infestations and minimize the overall loss of viable timber. Management to reduce tree mortality and consequentially to reduce MPB population levels is done through logging and is referred to as sanitation harvesting. Harvesting can be performed in different ways, but clear-cutting is most commonly employed [28]. Clear-cutting involves the removal of all trees from a given area, including both susceptible and nonsusceptible trees. The objective is to remove the MPB from the stand, which means harvesting trees that show signs of attack as well as adjacent trees that could become attacked in the near future. Therefore, clearcut practices attempt to remove infested and noninfested trees in the surrounding area. The advantages of clearcutting is that it is the fastest way to remove a specific volume of wood from the forest, it is the least expensive harvesting practice, it has the greatest operational experience and expertise, safety risks are better understood with clear-cutting practices, and sites that are clear-cut provide tolerable conditions for most commercial seedlings [40]. The disadvantages with clear-cutting as a sanitation harvesting method are that the amount of timber that is allowed to be cut is wasted on trees that are not susceptible, and MPB do not typically attack trees in a uniform pattern starting from an infestation and moving outward in concentric stages. Furthermore, from an ecological standpoint, clear-cuts lead to many problems such as increased instability and soil erosion. One way to improve the success of clear-cut methods for managing MPB outbreaks is to define areas that are most susceptible to MPB attack and design the shape of the clear-cut to reflect those areas. Therefore, instead of a cut that is symmetrically located around an infestation, the harvest will focus more in areas that are at a greater risk of attack. However, some low-susceptibility trees will still be cut, as stands are not homogeneous, and some detrimental ecological effects will persist. An alternative to clear-cuts for sanitation harvesting is a practice termed thinning, where only the most susceptible trees are removed from the stand. Removing highly susceptible pines leaves behind stronger trees and reduces the risk of major outbreaks [7]. The decrease in density because of thinning also reduces susceptibility by opening the stand and altering patterns of air, light, and temperature making it less favorable for beetles [45]. Furthermore, thinning ensures that the stand remains intact as lowsusceptibility trees are retained, which diminishes the ecological effects of harvesting that are apparent with clear-cutting. Although thinning susceptible stands seems logical, it is far more expensive to implement; therefore, the

Environ Model Assess

effectiveness of thinning for reducing tree mortality should be evaluated in comparison to clear-cutting.

3 Methods The methods for this study consist of three parts. First, tree susceptibility was determined from the RS images to provide input for the model. Second, the tree mortality model was developed and calibrated to simulate MPBinduced tree mortality patterns. Third, four different management strategies were tested for reducing tree mortality. 3.1 Defining tree susceptibility Tree susceptibility to MPB attack was determined for a forest area in the central interior of British Columbia. The data for this forest area were provided by high-resolution multispectral aerial photographs with a pixel resolution of 15 cm. The RS images were collected at this resolution for initially studying the spectral response of water loss in trees that were attacked by MPB [31]. The high resolution of these images also provided a utility for this study, as individual trees can be distinguished from each other, as seen in Fig. 1. The high costs of ground truth data and obtaining high-resolution images limited the size of the study area to 750 750 m; however, the area selected was representative of forested environments in the region that

Fig. 1 High-resolution, four-band multispectral RS image of study site located in British Columbia, Canada

were susceptible to MPB infestations. The images were collected during the summer of 2001, and the ground truth data for the aerial photographs were collected in 2002 by the British Columbia Ministry of Forestry (BC MoF), and in 2002 by Simon Fraser University and BC MoF. The ground truth data were used to verify classification of tree species, tree size, and whether a tree had been attacked by MPB. The thematically classified high-resolution images were analyzed in a GIS and resampled so the spatial resolution corresponded to tree scale (i.e., each tree was one raster cell of a digital image). The images were analyzed to obtain information on four variables that describe the susceptibility of a tree to MPB attack as defined by the MPB susceptibility rating system developed by Shore and Safranyik [41]. The four variables were the proportion (P) of susceptible lodgepole pine in the stand of a given tree, a location factor (L) explaining a stands proximity to trees currently infested with MPB, a density factor (D), and a factor of the size (S) of the tree represented by diameter at breast height (DBH). The original rating system used the age of a tree instead of tree size; however, trees size was selected for this study because information on age was not available, and size can be representative of age because diameter increases as the tree gets older [41]. The proportion of susceptible pine (P) is an important indicator of susceptibility because, as Hopping [21] notes, MPB outbreaks seldom originate in mixed stands. This is because the likelihood of MPB locating and attacking a lodgepole pine would decrease as the number of nonsusceptible trees increases [41]. Therefore, trees located in a stand of pure lodgepole pine (i.e., P = 1) were considered more susceptible than those in a stand containing a high mixture of species. The location factor (L) explains that tree susceptibility increases the closer a tree is to an infestation and further increases if that infestation is relatively large. Table 1 demonstrates how the distance to and size of an infestation influences the value of L based on the Shore and Safranyik [41] rating system. Part (a) explains that the relative size of an infestation in a stand (e.g., Stand 1) can be classified as small, medium, or large depending on the number of infested trees within 3 km of Stand 1 and the number of trees infested within Stand 1. Part (b) uses this information in conjunction with the distance to the nearest infestation to provide a continuous value between 0 and 1 that represents the variable L. The use of this rating system results in 0.06 L 1, where a value of 0.06 would indicate a stand that is at least 4 km from a small infestation (i.e., under 900 attacked trees), whereas a value of 1 would indicate a stand where a large number of infested trees (i.e., greater than 9000) exist within the stand. Therefore, L increases as the infestation draws closer and/or becomes larger. The threshold

Environ Model Assess Table 1 (a) Parameters used to determine the relative size of a MPB infestation within 3 km of the stand and (b) the relative size of infestation used in conjunction with the distance to the nearest stand to determine the value of L. (a) No. of infested trees outside stand within 3 km No. of infested trees inside stand <10 <900 900 9000 >9000 (b) Relative infestation size Distance to nearest infestation (km) In stand Small Medium Large 0.6 0.8 1.0 01 0.5 0.7 0.9 12 0.4 0.6 0.7 23 0.3 0.4 0.5 3 4 0.1 0.2 0.2 4+ 0.06 0.08 0.1 Small Medium Large 10 100 Medium Medium Large >100 Large Large Large

values for the Shore and Safranyik [41] rating system were determined mainly by observations made during population and dispersal studies [37]. The density factor (D) of trees in a stand affects susceptibility in two ways. The first is that trees in a highly dense stand experience greater competition for light, water, and nutrition; therefore, trees become more stressed and more vulnerable to MPB attack [29]. Second, stands of lower density are believed to contain a less favorable microclimate for MPB to successfully attack a tree and reproduce [45]. The Shore and Safranyik [41] rating system defines the relationship between stand density and the susceptibility of the stand to attack by MPB. The rating system classifies all the stands in the forest into discrete classes, where each class receives a value describing its susceptibility. The dotted line in Fig. 2 illustrates the discrete susceptibility classes based on tree density. The limitation of this component of the Shore and Safranyik
Fig. 2 Crisp classification function of density classes versus fuzzy membership function for the degree of belonging to the set of (DS)

rating system is that it treats large ranges of stand densities as having identical susceptibility. This is especially problematic in areas where stands contain a variety of densities that fall into a single class. To overcome this limitation, a fuzzy set approach was used to interpolate the discrete classification rating system to produce continuous D values. Fuzzy sets are commonly used for spatial and temporal applications in GIS research to represent the nondiscrete nature of geographic phenomena [13]. Classifying soil types [6], individual trees [5], and regions of land [18] are some examples of the variety of spatial phenomena for which fuzzy sets have been employed; however, the use of fuzzy sets with identifying susceptibility to insect infestations remains minimal [3, 4]. A fuzzy sets approach was used here to produce a value explaining the degree (DS) to which a stand belongs to the set of dense stands, which is used to represent D. This was accomplished by defining a fuzzy membership function that corresponds to the Shore

Environ Model Assess

and Safranyik [41] discrete function. The solid line in Fig. 2 illustrates the fuzzy membership function. A sigmoidal fuzzy membership function was chosen to provide a generalized transition between values of D, and because it is commonly employed in GIS applications using fuzzy sets for determining continuous class boundaries [33]. The sigmoidal function is represented by the equation D DS 1 1 e0:0094x472:22 ; 1

where x represents the density of trees in the stand. The size factor (S ) of a tree describes the susceptibility of a tree based on its DBH. Larger trees are more susceptible to MPB attack because they are typically older with weakened defense mechanisms that cannot withstand a mass attack of beetles [20]. The ground truth data from the study site and information provided by [41] revealed that trees under approximately 12 cm DBH are seldom attacked, whereas trees over approximately 42 cm DBH receive the highest likelihood of being attacked. This information was used to construct a sigmoidal fuzzy membership function to determine the degree of membership to a set of large trees, (LT ), which is used to represent S. Trees with a DBH less than 12 cm received an S value of 0, whereas trees with a DBH greater than 42 received an S value of 1. Trees in the intermediate range received increasing values of S as DBH increases. Using these parameters for defining the sigmoidal function, the value of S was calculated by S LT 1 1 e0:1599x20:34 ; 2

the four variables. Some GIS-based studies suggest using either the minimum or maximum value when combining layers of continuous values; however, calculating the product of the four variables will ensure that the influence of each variable is included in TS. Furthermore, problems can occur when using only the minimum or maximum value. A tree, for example, may have a high value (i.e., close to 1) for only one of the four variables, whereas the other three variables are extremely close to 0. If the maximum value was used the tree would be classified as highly susceptible, although three of the four variables indicate that it has a very low susceptibility to attack. Therefore, the value of TS was calculated for each tree using the equation TS P L D S ; 3

which assumes that all variables are equally important for determining tree susceptibility. As a result, each tree was represented by a value between 0 and 1, representing minimum to maximum susceptibility, respectively. 3.2 Tree mortality model The objective of the tree mortality model was to simulate annual MPB-induced mortality patterns of lodgepole pine using CA and the TS values derived from the previous section. A group of N trees was selected to be hypothetically attacked by MPB at the onset of the model; this initial stage was represented by T0. These trees acted as seeds from which MPB would disperse to attack other trees on an annual basis (i.e., T1, T2, ... Tn). The CA used a regionalscale neighborhood that covered the entire study area. This neighborhood was selected so that all trees in the study area had the potential of being attacked each year. The CA transition rules explaining the process of MPB-induced tree mortality was governed by an allometric function that defines the number of MPB required in the neighborhood for a tree with a given level of susceptibility to become

where x represents DBH. The fuzzy membership function for S is illustrated in Fig. 3. Once the values for the variables P, L, D, and S were calculated, they were combined so that each tree was represented by a single value of tree susceptibility (TS). A suitable method for calculating TS is to take the product of
Fig. 3 Fuzzy membership function for the degree of belonging to the set of (LT )

Environ Model Assess

below the curve indicates scenarios of no attack. The equation governing the allometric function is given by y 0:05xa ; 4

Fig. 4 The allometric function describing the CA transition rules

attacked. An allometric function was chosen to follow the logic that trees of high susceptibility required low levels of MPB in the neighborhood to become attacked, and as MPB populations increase to higher levels they would begin to attack less susceptible trees [36]. Figure 4 illustrates a potential allometric function that would be used to govern the CA transition rules. The function explains the minimum proportion of MPB required in the neighborhood of a tree with a given TS value for the tree to become attacked. On or above the curve indicates scenarios of attack, whereas
Fig. 5 The tree mortality model

which explains that the proportion of beetles in the neighborhood required for a tree of x susceptibility to become attacked can be no less than 5% for trees of highest susceptibility, and the proportion increases as a function of the exponent a. The calibration of the allometric function is described below. At the completion of attack on lodgepole pine for a given summer, adult MPB nest and their offspring develop under the bark of a tree over the winter months. MPB are highly susceptible to cold temperatures during portions of this period [2, 22], which, during outbreaks, inflict an 80% insect mortality rate [35]. MPB winter mortality was simulated in this study by reducing the number of infested trees by 80% at the end of each winter. This meant that the trees were still dead from MPB attack, but they no longer contained MPB that could kill more trees during the following summer. After the simulation of winter mortality, the susceptibility of each tree was updated to reflect the change in the location factor L and subsequently the change in TS. Figure 5 illustrates one complete cycle of the tree mortality model, which is composed of the CA, MPB winter mortality, and updating TS. Tree mortality was simulated for six cycles to simulate MPB-induced tree mortality over a 6-year period, as this amount of time was considered suitable for evaluating insect outbreak management.

Environ Model Assess Fig. 6 (a) Classification of study site into eight different stands and (b) the TS values of each tree and initial infestation of MPB at time T = T0

The tree mortality model was calibrated to meet two objectives. The first objective was to have the rate of tree mortality follow a typical growth curve that is governed by a carrying capacity. This type of growth curve explains that the initial exponential rate of increase in tree mortality is eventually slowed due to a finite number of trees that exist in the study area. The model was calibrated to coincide with the growth curve by altering the number of time steps of the CA that was to represent a single year. The calibration consisted of running the CA for 1, 2, 3, 4, 5, and 6 time steps to determine the number that resulted in annual tree mortality that corresponds to the growth curve. The second calibration objective was to have the most susceptible trees attacked first, and the less susceptible trees become attacked in future years as the MPB population in the study area increased. This objective was set in order that the model provides results that are congruent with the MPB literature [36]. The exponent value of the allometric function that defines the CA transition rules was altered to meet this objective. Exponent values between 1.0 and 1.5 were tested to determine which one produced results that coincided with MPB attack behavior.

3.3 Simulating forest management strategies The objective of simulating different management strategies was to determine how harvesting methods influence the persistence of MPB attack. The two main strategies compared in this study were clear-cutting and thinning. This comparison is intended to allow management to evaluate whether the high financial costs of thinning provide a significant reduction of trees attacked by MPB. In addition, clear-cutting practices were evaluated based on how the shape of the cut is defined. Square and circular clear-cuts around the center of the infestation were first evaluated to determine if different symmetrical cuts influence MPB outbreaks. Next, the symmetric square and circular clear-cuts were compared to a clear-cut that was shaped based on susceptibility of the trees. This involved selecting the shape of the clear-cut so that more highsusceptibility trees were removed than low susceptibility trees, which results in nonsymmetric clear-cut areas. This comparison was evaluated to determine if the time and money invested in defining a clear-cut area based on tree susceptibility results in significantly less killed trees than

Table 2 Information extracted from RS images to calculate the variables P, L, D, and S. Stand Percent of lodgepole pine Location Trees infested within stand 1 2 3 4 5 6 7 8 0.88 0.98 0.93 0.78 0.94 0.00 0.50 0.50 0 0 <10 10 100 10 100 0 0 0 Trees infested external to stand <900 <900 <900 <900 <900 <900 <900 <900 Distance to infestation <1 km <1 km Within Stand Within Stand Within Stand <1 km <1 km <1 km 250 450 450 550 600 150 150 150 2252 2252 2252 2252 2252 2252 2252 2252 Stand density (trees/ha) Range of tree size (DBH cm)

Environ Model Assess Table 3 Values of the variables P, L, D, and S for each stand calculated from the information provided in table 2. Stand 1 2 3 4 5 6 7 8 P 0.88 0.98 0.93 0.78 0.94 0.00 0.50 0.50 L 0.5 0.5 0.6 0.8 0.8 0.5 0.5 0.5 D 0.30 0.70 0.70 0.87 0.92 0.13 0.13 0.13 S 0.56 0.99 0.56 0.99 0.56 0.99 0.56 0.99 0.56 0.99 0.56 0.99 0.56 0.99 0.56 0.99

the symmetric clear-cuts. These four harvesting strategies (i.e., square clear-cut, circular clear-cut, selective clearcutting based on susceptibility, and thinning) were each simulated removing seven different areas of trees.

4 Results The study area was classified into eight different stands, as shown in Fig. 6a, based on the density and diversity of trees. RS data interpretation and the ground truth data were used to obtain the information on the size distribution of trees, the diversity and density of each stand, and the information required to calculate the location of each stand relative to initial infestation (Table 2). This information was used to calculate the values of the variables P, L, D, and S for each tree represented by a 4 4 m pixel (Table 3), which was subsequently used to calculate the value of tree susceptibility, TS. The TS values between 0.0 and 1.0 for each tree are illustrated in Fig. 6b, along with the initial simulated infestation of MPB, which acted as the seeds for the tree mortality model.
Fig. 7 The percent of tree mortality generated at each cycle of the model

The tree mortality model was performed for six cycles while altering the number of time steps of the CA as well as altering the exponent value of the allometric function until the two calibration objectives mentioned above were met. Regarding the first calibration objective (i.e., to have the rate of tree mortality follow a typical growth curve that is governed by a carrying capacity), Fig. 7 demonstrates the tree mortality curves that were a result of using the different number of CA time steps against the expected rate of tree mortality based on a typical carrying capacity growth curve. The figure illustrates that three time steps were most appropriate for representing each year of tree mortality, as the curve most closely resembles the expected curve. The second objective (i.e., simulating an attack on the most susceptible trees first followed by attack on less susceptible trees in subsequent years) was satisfied by altering the exponent of the allometric function defining the CA transition rules. Figure 8 demonstrates how changing the exponent value alters the height of the function. After a heuristic evaluation of exponent values between 1.0 and 1.5, it was determined that a = 1.3 was most suitable for satisfying the second objective. Therefore, the allometric function was governed by the equation y 0:05x1:3 : 5

Figure 9 shows the annual cumulative percentage of MPBinduced tree mortality (thick line) that was derived using the selected allometric function at three time steps. The graph illustrates the MPB-induced tree mortality over time based on equal interval classes of tree susceptibility when using the same allometric function. The graph shows that trees of highest susceptibility (i.e., 0.761.00) experienced attack at a faster rate than the other susceptibility classes. The lowest susceptibility class (0.010.25) did not experi-

Environ Model Assess

Fig. 8 The influence of the exponent value on the allometric function

ence any attack until the third time step when population levels of MPB have increased enough to overcome the defensive mechanisms of low-susceptibility trees. Figure 10 provides the simulated results of tree mortality at time steps 1, 3, and 6 using this equation. The two calibration objectives can be observed through these results as follows: (1) Tree mortality increases rapidly over the first few iterations then declines as the number of nonattacked trees becomes limited, and (2) areas containing high-susceptibility trees are attacked by the first time step, and trees of lower susceptibility are attacked as each time step passes.

The calibrated model was subjected to the four different forest management harvesting strategies. For each strategy, seven different areas of trees were removed: 1.00, 2.25, 4.00, 6.25, 9.00, 12.25, and 16.00 ha. These seven areas of trees were determined by increasing each side of the square clear-cut by 50 m. Each strategy is illustrated in Fig. 11 for the minimum (1 ha) and maximum (16 ha) area of trees removed. The square and circle clear-cuts were located based on having the infestation as the center of the clearcut. The location of the selective clear-cut was focused on the susceptibility of trees. Therefore, instead of a symmetric clear-cut, the shape of the cut followed the contours of the stands that contained the trees of highest susceptibility. Harvesting was thus focused in stands 4 and 5, whereas minimal trees were removed from stand 3 (see Fig. 6). The thinning management strategy involved harvesting a quantity of the most susceptible trees that was equivalent in area to the clear-cut strategies. Therefore, an equivalent of 1 ha of the most susceptible trees were removed for the 1-ha harvesting scenario, whereas an equivalent of 16 ha of the most susceptible trees were removed for the 16-ha harvesting scenario. The tree mortality model was performed using each management strategy for a total of six time steps. Figure 12 shows how tree mortality was affected by implementing each scenario. The graph shows the percent of trees remaining after harvesting that were attacked and killed by MPB. The results from the tree mortality model simulations displayed in Fig. 12 conclude that the choice of harvesting practice has significant consequences on the number of trees that are attacked by MPB. The most pertinent

Fig. 9 MPB-induced tree mortality over time using the calibrated allometric function

Environ Model Assess Fig. 10 Results from the model simulation of MPB-induced tree mortality patterns for T1, T3, and T6

observation was the success of the thinning harvesting strategy over the clear-cut practices for reducing tree mortality. Whereas thinning of 1 ha and 2.25 ha areas showed no significant difference, thinning of 4 ha and beyond displayed a dramatic decrease in tree mortality. The reason for this was that the most susceptible trees in the study area were removed, which severely debilitated the rate at which the MPB population could increase. Furthermore, thinning also decreases the tree density of a stand, which in turn decreases the overall susceptibility to MPB attack. Thinning works better at reducing tree mortality than do clear-cuts because the latter strategy focuses all harvesting efforts in a centralized area around the infestation, whereas thinning will remove the same number of trees over a greater area, thus having a regional-scale effect. Therefore, as the results presented in Fig. 12 illustrate, less harvesting is required with thinning to minimize timber loss to MPB attack. With regard to forest management, these results suggest that the high costs of thinning are warranted if reducing tree mortality from MPB is a priority. However, thinning efforts must take into consideration the minimal number of trees that should be removed to prevent tree mortality. Figure 12 demonstrates that although thinning is most effective, it may not be useful if performed on a small volume of wood. Therefore, a challenge for future research is to determine the minimal effective thinning volume required for different levels of initial MPB attack. Examining the results from the three clear-cut methods illustrates that the shape of the clear-cut does not have an impact on tree mortality if the cut is symmetrically placed around the infestation. Both the square and circle clear-cut approaches exhibited minimal success with reducing timber loss to MPB even when increasing the size of the harvest. This was evident in that the loss of timber was only reduced by 30% for both methods when comparing a 1-ha harvest to a 16-ha harvest. Conversely, the other two methods reduced tree mortality by approximately 80% when comparing a 1-ha to 16-ha harvest. The lack of success for the square and circle clear-cuts was because trees were removed

regardless of their susceptibility to attack. Low-susceptibility trees were removed even though they were at minimal risk of attack during the initial MPB population level, which meant that less high susceptibility trees could be harvested. Since MPB have the ability to disperse distances

Fig. 11 Harvesting strategies simulated in the study area at time T = T0 for 1-ha harvest (left) and 16-ha harvest (right)

Environ Model Assess Fig. 12 Tree mortality at time T = T6 for each harvesting practice for different harvesting sizes

throughout the study area, the square and circle clear-cuts did not prevent them from attacking high-susceptible trees. Therefore, for this technology to be at all successful, management would have to develop large-scale clear-cutting plans to produce a significant decline in tree mortality. However, as clear-cuts become larger their detrimental effects to the ecology of the forest become magnified, which in turn could prevent the forest from being a viable source of timber. The selective clear-cut displayed intermediate results, as tree mortality was reduced when using the 9-ha harvest. This was an improvement over the other two clear-cut methods because the shape of the cut was selected based on where the most susceptible trees were located rather than being based on the distance to the center of the infestation. However, the method proved not as successful as thinning because clearcutting ensures that some low-susceptible trees will still be removed. Selective clear-cutting could provide an intermediate strategy that balances the affordability of simple clearcutting and the effectiveness of thinning. Some forest managers may not have the financial resources or the time to implement thinning harvesting, but would be able to benefit from clear-cutting areas based on susceptibility. Although selective clear-cutting is not as effective as thinning, Fig. 12 indicates that when selective cutting is applied over large enough areas it can have similar results.

5 Conclusion The results from this study indicate that the size or the shape of the harvest is not the most important factor when attempting to reduce tree loss to MPB attacks; instead, the

susceptibility of trees should be the focus of management activities. As tree susceptibility can vary at the local scale, practices such as clear-cuts that invoke a spatially homogeneous response to insect outbreaks are not as appropriate because they ignore local-scale variability. Thinning practices that acknowledge individual tree susceptibility are more useful, as the most susceptible trees can be identified and removed while retaining important ecological components of the forest. Thus, the high-resolution RS data were an essential component of this study as they enabled susceptibility to be defined at the tree level. The disadvantage of using high-resolution images is that they are costly, which can restrict the scale over which data are collected. This presents a conundrum for studies such as this where high-resolution is necessary for distinguishing individual trees, but the scale of the study site limits the quantity of information that is available for determining tree susceptibility. For example, the location factor L determined the susceptibility of a stand using the number of trees attacked at a given distance to the stand. The Shore and Safranyik [41] rating system explains that a stand is susceptible if trees are attacked within a distance of 3 km, but the study site covered only a portion of this distance. This can lead to underestimating the susceptibility of a stand to MPB attack because detecting infested trees within 3 km is limited by the size of the study area. For this study, the initial infestation occurred within the study area, which resulted in a relatively high susceptibility rating for some stands regardless of whether there were infestations outside the study area. However, the susceptibility rating for all stands could still be higher if a significant number of trees were infested beyond the boundaries of the study area but

Environ Model Assess

within 3 km. Although this issue presents a limitation for this study, it does not discredit the findings, as the annual increase in attack followed a logical exponential curve that is representative of MPB-induced tree mortality. Tree mortality increases exponentially until availability of susceptible trees becomes exhausted. Therefore, although tree susceptibility may have been slightly underestimated, the rate of attack was still a close representation of reality and allowed for effective evaluation of the objective of this study, which was to analyze the influence of different management strategies for controlling MPB outbreaks. This demonstrates that although larger study sites are optimal for providing adequate information, the model can be applied across a range of spatial scales for determining the implications of MPB attack and management activities. Thus, even when only large-scale images are available, simulating the dynamics of infestations should precede harvesting decisions. The use of GIS was beneficial to this study as it provided methods in this study for calculating the susceptibility of trees by measuring the susceptibility variables and effectively combining them to produce an output demonstrating the variability of susceptibility across the forest landscape. GIS are a utility for forest management because the spatial relationships between individual trees and stands can be evaluated and integrated into management decisions. However, GIS and RS data have limited predictive capabilities, as they usually provide static representations of the world. This study provided an opportunity to integrate the RS data and GIS with CA for simulating patterns of tree mortality and evaluating appropriate management decisions. The CA model proved useful for simulating insect propagation and evaluating management practices. The parameters were specified so that the model can be applied to a variety of scales and locations that are susceptible to MPB infestations. The model could be incorporated as a decision-support tool where interested stakeholders can use GIS as a virtual laboratory to change parameters, such as the initial size and location of the infestation, growth and mortality rates, and dispersal distances, to determine how life-cycle characteristics of MPB affect tree mortality. The model also provides the potential for creating different scenarios to evaluate appropriate practices for managing forests in anticipation of insect outbreaks. For example, species diversity or stand density can be altered in the GIS, and resulting data could be run in the CA to determine how important these variables are compared to one another for influencing MPB attack. CA are advantageous in this respect because numerous stakeholders with a variety of backgrounds can collaborate to determine a desirable outcome. Unlike typical models based on partial differential equations that are mathematically intensive, explicit knowledge of the

system is not required for creating a valid CA model because the necessary information for the modeled process is included in the form of rules rather than mathematical equations. This allows for direct incorporation of knowledge from experts that is not necessarily restricted to hard data, and is particularly useful when attempting to model problems that are complex. Furthermore, the outputs from the model simulations provide management with a visual understanding of how MPB are most likely to disperse through the forest and the related effectiveness of different practices they may wish to implement. Collectively, RS, GIS and spatial modeling with CA provide forest and other natural resource managers with the opportunity to investigate natural processes in both space and time and facilitate the evaluation of management practices to determine effective measures for dealing with ecological problems such as insect infestations.
Acknowledgments The authors are thankful to the Natural Sciences and Engineering Research Council (NSERC) of Canada for full support of this study under the Discovery Grant awarded to the second author. Acquisitions of high-resolution data sets used in this study are funded from BC Forestry Innovation and Forestry Investment Account grants awarded to the third author.

References
1. Batty, M., & Xie, Y. (1994). From cells to cities. Environment and Planning. B, Planning and Design, 21, 531548. 2. Bentz, B. J., Logan, J. A., & Vandygriff, J. C. (2001). Latitudinal variation in Dendroctonus ponderosae (Coleoptera: Scolytidae) development time and adult size. Canadian Entomologist, 133, 375387. 3. Bone, C., Dragievi, S., & Roberts, A. (2005). Integrating high resolution RS, GIS and fuzzy set theory for identifying susceptibility areas of forest insect infestation. International Journal of Remote Sensing, 26, 4809 4828. 4. Bone, C., Dragievi, S., & Roberts, A. (2006). A fuzzyconstrained cellular automata model of forest insect infestations. Ecological Modelling, 192, 107125. 5. Brandtberg, T. (2002). Individual tree-based species classification in high spatial resolution aerial images of forests using fuzzy sets. Fuzzy Sets and Systems, 132, 371387. 6. Burrough, P. A. (1989). Fuzzy mathematical methods for soil survey and land evaluation. Journal of Soil Science, 40, 477 492. 7. Canadian Forest Service (2001). Commercial thinning of mature lodgepole pine: Results of beetle proofing research in the East Kootenays. Victoria: Natural Resources Canada. 8. Clarke, K. C., Hoppen, S., & Gaydos, L. (1997). A self-modifying cellular automation model of historical urbanization in the San Francisco Bay area. Environment and Planning. B, Planning and Design, 24, 247261. 9. Couclelis, H. (1997). From cellular automata to urban models: New principles for model development and implementation. Environment and Planning. B, Planning and Design, 24, 165174. 10. Darwen, P. J., & Green, D. G. (1996). Viability of populations in a landscape. Ecological Modelling, 85, 165171. 11. Deadman, P., Brown, R. D., & Gimblett, H. R. (1993). Modeling rural residential settlement patterns with cellular automata. Journal of Environmental Management, 37, 147160.

Environ Model Assess 12. De Almeida, C. M., Monteiro, A. M. V., Soares, G., Cerqueira, G. C., Pennachin, C. L., & Batty, M. (2005). GIS and remote sensing as tools for the simulation of urban land-use change. International Journal of Remote Sensing, 26, 759774. 13. Dragievi, S. (2004). Fuzzy sets for representing spatial and temporal dimensions in GIS databases. In G. Bordogna (Ed.), Spatio-temporal databases: Flexibility querying and reasoning (pp. 1128). Berlin: Springer-Verlag. 14. Dymond, C. C., & Johnson, E. A. (2002). Mapping vegetation spatial patterns from modeled water, temperature and solar radiation gradients. ISPRS Journal of Photogrammetry and Remote Sensings, 57, 6985. 15. Engelen, G., White, R., Uljee, I., & Drazan, P. (1995). Using cellular automata for integrated modelling of socio-environmental systems. Environmental Monitoring and Assessment, 34, 203214. 16. Favier, C., Chave, J., & Fabing, A. (2004). Modeling forest savanna mosaic dynamics in man-influenced environments: Effects of fire, climate and soil heterogeneity. Ecological Modelling, 171, 85102. 17. Favier, C., & Dubois, M. A. (2004). Reconstructing forest savanna dynamics in Africa using a cellular automata model, FORSAT. Lecture Notes in Computer Science, 3305, 484 491. 18. Fisher, P. (1996). Boolean and fuzzy regions. In A. U. Frank (Ed.), Geographic objects with intermediate boundaries (pp. 8794). London: Taylor and Francis. 19. Government of British Columbia (2005). British Columbia_s mountain pine beetle action plan 20052010. Victoria, Canada: Government of British Columbia, p. 20. 20. Hindmarch, T. D., & Reid, M. L. (2001). Forest thinning affects reproduction in pine engravers (Coleoptera: Scolytidae) breeding in felled lodgepole pine trees. Environmental Entomology, 30, 919924. 21. Hopping, G. R. (1961). Damage agents. Victoria, Canada: Canadian Department of Forestry, pp. 7799. 22. Jenkins, J. L., Powell, J. A., Logan, J. A., & Bentz, B. J. (2001). Low seasonal temperatures promote life cycle synchronization. Bulletin of Mathematical Biology, 63, 573595. 23. Kok, J. L., Engelen, G., White, R., & Wind, H. G. (2001). Modeling land-use change in a decision-support system for coastal-zone management. Environmental Modeling and Assessment, 6, 123132. 24. Li, X., & Yeh, A. G. O. (2001). Zoning land for agricultural protection by the integration of remote sensing, GIS and cellular automata. Photogrammetric Engineering & Remote Sensing, 67, 471477. 25. Luther, J. E., Franklin, S. E., Hudak, J., & Meades, J. P. (1997). Forecasting the susceptibility and vulnerability of balsam fir stands to insect defoliation with Landsat Thematic Mapper Data. Remote Sensing of the Environment, 59, 7791. 26. Marshall, E. P., & Homans, F. R. (2004). A spatial analysis of the economic and ecological efficacy of land retirement. Environmental Modeling and Assessment, 9, 6575. 27. Mathey, A. H., Krcmar, E., & Vertinsky, I. (2005). Re-evaluating our approach to forest management planning: A complex journey. Forestry Chronicle, 81, 359364. 28. McGregor, M. D., Amman, G. D., Schmitz, R. F., & Oakes, R. D. (1987). Partial cutting lodgepole pine stands to reduce losses to the mountain pine beetle. Canadian Journal of Forest Research, 17, 1234 1239. 29. Mitchell, R. G., & Preisler, H. K. (1991). Analysis of spatial patterns of lodgepole pine attacked by outbreak populations of the mountain pine-beetle. Forest Science, 37, 1390 1408. 30. Polansky, C., & Heermans, J. (2004). Developing forest management plans with high-tech tools and traditional knowledge in Zambia. Journal of Forestry, 102, 46 51. 31. Roberts, A., Northrup, J., & Reich, R. (2005). Mountain pine beetle detection and monitoring: Replication trials for early detection and monitoring. In Proceedings of the Third International Workshop of Multi-temporal Remote Sensing, Biloxi, Mississippi, May 16 18. Institute of Electrical and Electronics Engineering, pp. 2024. 32. Robiglio, V., & Mala, W. A. (2005). Integrating local and expert knowledge using participatory mapping and GIS to implement integrated forest management options in Akok, Cameroon. Forestry Chronicle, 81, 392397. 33. Robinson, V. B. (2003). A perspective on the fundamentals of fuzzy sets and their use in geographical information systems. Transactions in GIS, 7, 330. 34. Roy, P. S., Sharma, K. P., & Jain, A. (1996). Stratification of density in dry deciduous forest using satellite remote sensing digital dataAn approach based on spectral indices. Journal of Biosciences, 21, 723734. 35. Safranyik, L. (2004). Mountain pine beetle epidemiology in lodgepole pine. In Proceedings of Mountain Pine Beetle Symposium: Challenges and solutions. Kelowna, British Columbia, October 3031. Canadian Forestry Service, pp. 33 40. 36. Safranyik, L., Barclay, H., Thomson, A., & Riel, W. G. (1999). A population dynamics model for the mountain pine beetle, Dendroctonus Ponderosae Hopk. (Coleoptera: Scolytidae). Victoria, Canada: Pacific Forestry Centre, p. 35. 37. Safranyik, L., Silversides, L. H., McMullen, L. H., & Linton, D. A. (1989). An empirical approach to modeling the local dispersal of the mountain pine beetle (Dendroctonus ponderosae Hopk.) (Col., Scolytidae) in relation to sources of attraction, wind direction and speed. Journal of Applied Entomology, 108, 498511. 38. Salajanu, D., & Olson, C. E. (2001). The significance of spatial resolutionIdentifying forest cover from satellite data. Journal of Forestry, 99, 3238. 39. San-Miguel-Ayanz, J., Ravail, N., & Kelha, V. (2005). Active fire detection for fire emergency management: Potential and limitations for the operational use of remote sensing. Natural Hazards, 35, 361376. 40. Scientific Panel for Sustainable Practices in Clayoquot Sound. (1995). Sustainable ecosystem management in Calayoquot Sound: Planning and practices. Victoria: Government of British Columbia. 41. Shore, T. L., & Safranyik, L. (1992). Susceptibility and risk rating systems for the mountain pine beetle in lodgepole pine stands. Victoria, Canada: Pacific Forestry Centre, p. 12. 42. Strange, N., Meilby, H., & Jellesmark, T. (2000). Optimization of land use in afforestation areas using evolutionary self-organization. Forest Science, 48, 543555. 43. Sui, D. Z., & Zeng, H. (2001). Modeling the dynamics of landscape structure in Asia_s emerging desakota regions: A case study in Shenzhen. Landscape and Urban Planning, 53, 3752. 44. White, R., & Engelen, G. (1997). Cellular automata as the basis of integrated dynamic regional modeling. Environment and Planning. B, Planning and Design, 24, 235246. 45. Whitehead, R. J., Safranyik, L., Russo, G. L., Shore, T. L., & Carroll, A. L. (2004). Silviculture to reduce landscape and stand susceptibility to the mountain pine beetle. In Proceedings of Mountain Pine Beetle Symposium: Challenges and Solutions, Kelowna, British Columbia, October 3031. Canadian Forestry Service, pp. 233244. 46. Wilson, B. (2004). An overview of the mountain pine beetle initiative. In Proceedings of Mountain Pine Beetle Symposium: Challenges and Solutions, Kelowna, British Columbia, October 3031. Canadian Forestry Service, pp. 39. 47. Wulder, M. A., & Dymond, D. (2004). Remote sensing technologies for mountain pine beetle survey. In Proceedings of Mountain Pine Beetle Symposium: Challenges and Solutions, Kelowna, British Columbia, October 3031. Canadian Forestry Service, pp. 146153. 48. Yeh, A. G. O., & Li, A. (2001). A constrained CA model for the simulation and planning of sustainable urban forms by using GIS. Environment and Planning. B, Planning and Design, 28, 733753. 49. Yeh, A., & Li, X. (2002). A cellular automata model to simulate development density for urban planning. Environment and Planning. B, Planning and Design, 29, 431 450.

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.