Home > Getting Started > Resources >
|
![[Text graphic] Belief Network Templates for use in CRAFT](../assets/Titles_HumanstFont/Title_BNTemplates.gif)
 |
Building belief networks can be a challenge for beginners. Here we present an assortment of belief networks that are structured using different types of spatial and temporal models. These example templates cover a diverse range of management problems. Before selecting any particular belief network, it is worthwhile to examine the different ways that problems can be modeled by examining the templates shown below. You may use and modify any of these templates or build your own according to your specific project needs.
SIMPLE TEMPLATES Simple relationships can be readily expressed probabilistically to improve assessments of project effects. In some well-constructed belief networks, multiple simple relationships can be combined as "modules" in a larger model framework.
TEMPORAL TEMPLATES Time can be expressed in a number of ways in belief network modeling depending on management needs on one hand, and on the intrinsic characteristics of the phenomenon being modeled. Time step modeling is commonly used to predict successional conditions at explicit periods in the future, yet disturbances are often contingent on the life history characteristics of the species or community being modeled. Vegetation dynamics that result from a low-severity and sustaining fire regime may be better modeled with a time-implicit model than with a time-step model, while seral successional change may be well emulated by the latter model structure. The templates presented here include a range of different structural types of models and are provided to demonstrate the range of ways that management problems can be modeled. Adaptation of these templates to specific management problems may result in substantial changes.
SPATIAL TEMPLATES Many projects have effects at broad spatial scales. The management concerns that drive belief network construction may involve questions of movement or area. In the section below, we present several belief network templates that describe how these two types of questions can be incorporated into CRAFT. A wide range of elements in a natural system move. For example, sediment flows from project areas downstream, invasive species and fire spreads across the landscape, and wildlife movement is essential for sustaining many populations. Given this range of management problems, the effects of “movement” may be intentional or not. Fire and fuel managers deliberately design fuel breaks to alter landscape scale fire spread and behavior. In other cases, movement may have undesirable cumulative effects on watersheds or habitat. Many forest managers are concerned with maintaining areas at critical levels, often for habitat or silvicultural reasons. Belief networks can be constructed around these management concerns as well.
In some cases, it is possible to combine broad spatial scales and long term processes within a single model.
Simple Template Site model: Stand level fire behavior Local fire behavior is modeled based on the slope, fuel model and fire weather. Fire weather will vary probabilistically during a wildfire or prescribed fire event. In this example, the “individual case” for training the belief network is a single weather condition that could be derived from historical weather and fuel moisture data (processed in FireFamilyPlus). Weather conditions may be derived from monthly or hourly data. The probability of a given fire behavior can then be calculated using multiple runs of software, such as BehavePlus. Use this belief network to better describe the likelihood of a given fire behavior characteristic for a given fire event. Open Netica using this basic belief network structure
This template is structured around discrete periods of time and is useful for comparing the short and long term value of projects. Vegetation or habitat conditions are deterministically modeled through T1, but 10-year outcomes are influenced by the probability of wildfire occurring with a particular intensity (i.e., flame length). As a variant, transitional successional probabilities among overstory or understory classes can make outcomes even more conditional than is shown here. Use this template for projects that require comparisons of short and long term effects.
Temporal Template Time-step model - Habitat dynamics, multiple species
For a given landscape, successional transitions of stands can be described according to a probability distribution. Such transitions can be calculated based on either (1) analysis of past trends or (2) stand by stand modeling using software such as FVS. The chance effects of whether a wildfire occurs and/or what fire weather occurs during a wildfire can further model uncertainties. Longer term forecasting is possible by adding more time steps. A diverse array of wildlife habitat parameters can be described probabilistically for each time period using habitat relations models. Use this template to compare the uncertainties regarding short and long term conditions within a landscape. Open Netica using this basic belief network structure Temporal Template Cohort model - threats to a shrubland bird population
Time can be represented in a belief network in a number of ways. Rather than model change according to calendar time (e.g., Year 1, Year 5, Year 10, etc.) as with the Time Step Model, distinct phases in the life history of a key element of interest can be used to define time units. Models can be built around the life history stages of individual trees or a wildlife species, a cohort of even-aged trees or wildlife, or seral developmental stages of vegetation after severe disturbance. As with other models, weather is a useful variable to express probabilistically, but the uncertainties that surround mortality from any cause can be expressed similarly. Open Netica using this basic belief network structure
Temporal Template Cohort model - Sustaining whitebark pine
In this more complex cohort model, a variety of factors cause the mortality of whitebark pine as trees age. This justifies using a cohort model in which time is expressed in terms of the life history stages of the cohort. The chance of a beetle outbreak and wildfire of a given intensity affect the probability of achieving the goal of sustainable recruitment. Initial treatment options include burning and out-planting seedlings that are rust resistant, thereby changing future mortality. Open Netica using this basic belief network structure
Temporal Template Time-to-event model - Time to crown fire for a plantation
Plantations are common in forests having a high risk of wildfire and they pose special problems for managers due to their flammability. Better knowledge of their long-term viability in fire-prone environments can lead to more sustainable management. Under certain weather conditions, plantations burn with such high intensities that complete mortality is likely. This is especially problematic when logging residues were not treated or when plantations were not thinned as originally planned. In this belief network, “time to crown fire” is modeled based on how often fire is likely to occur at the site (i.e., the occurrence rate) and if a given fire will crown based on possible fuel conditions and chance fire weather. (The fire occurrence rate could be inferred from the regional fire rotation or from more sophisticated fire spread modeling). In this example, the plantation’s crown fire hazard is modeled probabilistically because sampling showed patchy fuel conditions that include a range of fuel models and canopy fuel conditions. Note that this model is not capable of altering fuel models after a non-lethal fire. This rare scenario is not considered in Time to crown fire calculations. That is, the embedded assumption is that a plantation can only burn once with any type of fire. Open Netica using this basic belief network structure
Temporal Template Time-to-event model - Time to old growth recovery
Restoration of forests takes time, but not time alone. When structural restoration results from the gradual progression through seral stages, the time required before the restoration goal is achieved depends on how fast succession progresses and whether or not the process is set back by severe disturbance. Growth rate depends on a number of conditions, including the effectiveness of thinning prescriptions. Open Netica using this basic belief network structure
Temporal Template Time-conditional model - Probability that a site is in a successional state over time
The Time-Conditional Model differs from other models in that time is specified upfront—in this case, as a decision node. The likelihood that a site will transition into a different state depends on the initial vegetation characteristics, growth rate and disturbance. Growth rate depends on both a fixed parameter (e.g., soil productivity, slope) and stochastic variation in climate. Disturbance is more difficult to accurately model. This example could be greatly simplified by modeling “wildfire occurrence” instead of fire intensity (as measured by “flame length”). Such a generalization would ignore the wide range of fire effects that are possible when wildfire does not crown. Fires of lower intensity may have little effect or they may even accelerate development of complex “old-growth” forest structure rather than “reset” the system to an earlier successional state. This feature imbues the model with a more realistic portrayal of fire’s diverse effects. The second decision node called “Treatment” allows decision makers to explore the system’s sensitivity to a range of treatment schedules. Open Netica using this basic belief network structure
Temporal Template Sustaining model - Perpetuating old growth ponderosa pine
When a management objective involves sustaining key ecosystem components indefinitely, successional models that emphasize severe disturbance and gradual recovery are not useful. When success is measured by maintenance of a state or condition, time implicit models may be most appropriate. In this template, the perpetuation of old growth legacy trees depends on keeping flame lengths below a threshold level. Management decisions involving both livestock grazing and fire frequency control fuel accumulation, along with the site’s productivity. Heavy grazing will influence fire patchiness and spread. In this example, stochastic uncertainty results from variable fire weather. Open Netica using this basic belief network structure
Spatial Template Movement-based model - Assessing spatial fire risks near communities
Movement based models require a consistent unit of analysis throughout. In the above example, a wide range of ignition locations and fire spread scenarios are used to determine if wildland fire reaches a single house (or clustered set of houses). Iterative modeling runs define the vulnerability of a single location, and multiple analyses would be required to generate a map of a broader landscape or dispersed community. Some fire spread models, such as FARSITE, can take considerable computational time, but iterative use of more simple models or even cursory analysis by a professional may be sufficient to exclude large number of runs. In other words, because the only question of concern is if wildfire reaches a site or not, many combinations of suppression, weather, start date and ignition location can not possibly result in fire arrival under any scenario. Hence, it may not be necessary to perform every analyses. Open Netica using this basic belief network structure
Spatial Template Area-based model - Probabilistic patterns of fire behavior
Landscape scale fire modeling is difficult due to great uncertainties in the accuracy of stand data. For example, FlamMap requires detailed information on canopy cover, crown bulk density, lower canopy height and fuel model. This knowledge is rarely available for even a few locations in a landscape. Forest exam data rarely collects all of this information and it is easily outdated. Integration of FIA data and regularly-updated remotely sensed vegetation data are arguably the best way to overcome local uncertainties. By identifying the range of FIA points that fall in a given vegetation type for a region, the parameters needed by FlamMap can be probabilistically assigned to similar polygons (vegetation strata). The spatially-explicit product that results from this process can help identify locations in the landscape that may provide fire refugia that are least prone to high severity fire and sites that are most prone to severe stand-replacing fire. More practically, this process can help identify the most probable fire behavior to inform fuel treatment design, logging and silvicultural decisions, and sustainable wildlife management. Open Netica using this basic belief network structure
Spatio-Temporal Template Complex model - Restoring old growth in hazardous surroundings This belief network uses information from broad spatial scales to model fire within a stand. Spatial modeling is used to predict the occurrence of fire and local stand-level information is used to predict fire behavior. Fire spread is modeled from random points to the stand. Of note, whether or not the stand burns at a future Time 2 depends on landscape scale successional dynamics. These are not shown modeled here directly, and may not need to be modeled with the same degree of reliability as the local stand of primary interest. In the surrounding area, successional change could be addressed by altering the fuel and vegetation maps used in fire spread modeling so that they include a range of possible successional scenarios based on assumptions derived from fire rotation data.
Spatio-Temporal Template Complex model - Invasive species spread risk over time
The spread of invasive species following disturbance is a major concern for forest managers. The likelihood that a given site will become infested following wildfire depend on the availability of seed, control efforts and suitable habitat for establishment. Many invasives spread along road corridors due to the high levels of disturbance associated with road maintenance. Wildfire can greatly increase the availability of bare ground to provide a “window of opportunity” for invasives. In this example, bare ground is probabilistically modeled based on modeling duff consumption across a range of burning conditions using Consume or FETM. Alternatively, duff consumption could be described from sampling. Control and dispersal are also modeled probabilistically. Weed control efforts are rarely entirely effective, and re-survey data can be used to inform conditional probability tables. Control is especially difficult with increased distance from roads and on less accessible slopes where bare mineral soil may be greatest. There are also substantial uncertainties associated with dispersal distances and the availability of seed sources. This too can be modeled probabilistically. Open Netica using this basic belief network structure
Spatio-Temporal Template Complex model - Sustaining native plants affected by patterns of recreational ORV activity
An agency’s off-road vehicle (ORV) policy could degrade a sensitive native plant of a given area. The policy chosen will influence the months that use is permitted and the pattern of ORV use within the landscape, both legal and illegal. On-trail ORV use will correlated with off-trail ORV use, and both may introduce invasive species that complete with native plants. In addition, off-trail ORV activity can degrade native plants by trampling/compaction. The leading reason the sensitive native plant populations are degraded other than through recreational activity is by fire suppression and the consequent increase in competitive woody plant cover. Spatio-Temporal Template Complex model - Predicting varied forest dynamics over a heterogeneous landscape
By modeling probabilistic fire behavior over a complex landscape, planners can describe how alternative locations of fuel treatments may affect broader vegetation outcomes and assess the longer-term viability of silvicultural features, such as plantations, and wildlife habitat conditions. This information is highly valuable for weighing landscape-scale tradeoffs among different managemet alternatives. Having a temporal dimension modeled as discrete time-steps (e.g., T1, T2, etc.) allows use of successional transition matrixes and variable rates of fire. For a similarly structured belief network, see the BayVeg model described by Lee and Irwin (2005).
Go to Fire Modeling page Go to the Vegetation Change Modeling page Go to Building Belief Networks Go to the Wizard |
|---|
In the belief networks provided below, variables are color coded depending on the function that they serve. Decision nodes (blue) are specified - these typically include a range of management prescriptions for a project, including the no action alternative. Measurable assessment endpoints (green) are derived from the the objectives hierarchy. Independent driver nodes (such as fire weather) and intermediary nodes are shown in white. 
Home ~ Getting Started ~ Resources ~ Wizard ~ Site Map
|
|---|