Discussion
The relative nature of our evaluation of fire danger has at least three important implications. First, the observed data value for each elementary topic in the logic model and for each attribute in the decision model was evaluated against reference conditions that were defined by the data themselves (Table: Definition of data inputs). As a result, basic evaluations at the lowest level of each model were relatively objective. A second consequence of defining reference conditions in this manner was that the models were maximally sensitive to the data, thus assuring a high level of discrimination among outcomes over the set of subwatersheds in map zone 16. Finally, this method of deriving reference conditions means that the values used depended on the spatial extent of the assessment area. For example, reference conditions appropriate to an assessment of the entire Southwestern United States would be at least somewhat broader than those for map zone 16 alone.
Evaluation outcomes and their underlying premises are affected by the scale of input data, whether they are at a relatively fine, (e.g., 30- to 90-m pixels) patch scale or, in the case of the PDSI data used here, the continental scale. For map zone 16, evaluating the likelihood that a subwatershed experienced drought in the past 20 years was derived from a 2.5-degree continental scale grid of reconstructed PDSI (Figure 1). Although there was wide variation in the probability of experiencing a long-term drought (PDSI < -2) for the continental United States (0 – 37 percent, Figure 1), map zone 16 exhibited a relatively narrow range of probabilities from 14 to 23 percent; or about 25 percent of the continental scale variation. Thus, one might be concerned that the contribution of long-term drought to the evaluation of ignition risk at the scale of a map zone may be neutral, as if adding a constant. This was not the case. Figures 2a and 2b illustrate the influence of including continental scale drought data in the map zone evaluation of fire danger. Differences can be seen among subwatersheds within evaluations of fire danger (Figure 2a) and ignition risk (Figure 2b) when comparing the same evaluations with and without PDSI. For map zone 16, PDSI does provide information on long-term drought that is beneficial to managers.
In addition to considering the scale of input data, the contributions of topics at each level to overall fire danger should be considered when interpreting an evaluation. For example, 10 subwatersheds that share a similar overall result for evaluation of fire danger, (i.e., moderate support, 0.56, for the proposition of low fire danger) are shown in Figure 3, but they differed by evaluation result at the primary topic and lower levels. Use of the union operator in the design of the knowledge base made it possible for relatively high fire hazard within a subwatershed to be offset by relatively low predicted fire behavior in the event of a wildfire, (e.g., see subwatershed 224, Figure 3). Similarly, subwatershed 339 (Figure 3) displayed evidence for low fire behavior but high ignition risk. An important strength of the logic model is that the full range of variability is expressed among subwatersheds at the level of an elementary topic, and each elementary topic contributes to evaluations of secondary and primary topics within a subwatershed and among subwatersheds. Thus, it is important to keep in mind that variability of support for a subwatershed at the elementary topic level in the hierarchy should be considered when interpreting a primary or secondary topic level evaluation result for any subwatershed and among subwatersheds.
The present study illustrates application of EMDS for evaluating wildland fire danger and prioritizing vegetation and forest fuels treatments at the spatial extent of a USGS map zone. When the national LANDFIRE mapping effort (www.landfire.gov) provides full coverage for the continental United States (CONUS), it will be technically feasible to conduct an analysis of fire danger for all subwatersheds in the CONUS in the same manner as we have illustrated here. Moreover, it is a relatively simple matter, given such a base analysis, to summarize such watershed-scale evaluations to various intermediate broader scales such as States, geographic regions, forest boundaries, or forest planning zones as a basic input to broad-scale planning and resource allocation.
At the other extreme, the present study provides a starting point for finer scale planning. We have examined the evidence for fire danger in subwatersheds of map zone 16, but this information, by itself, is not necessarily sufficient for fuels treatment planning. As shown above, subwatersheds that exhibit a similar moderate level of fire danger do not necessarily share the same evaluation results for primary topics (Figure 3). Thus, variability of support for propositions within a subwatershed at the level in the logic model where data are evaluated should be considered when interpreting an evaluation result among subwatersheds at the level of the primary or secondary topics.
To that end, subwatersheds in the worst condition with respect to fuels may not be the best candidates for fuels treatment. In particular, additional strategic or logistical factors such as proximity to population centers, presence of endangered species, slope steepness, and road access all might be taken into account in selection of specific watersheds within a management area for fuel treatment. Such an approach was illustrated by Reynolds and others (2003) using the Priority Analyst component of EMDS, which uses a decision engine for such purposes. In that study, they considered the compositional and structural integrity of forests along with contemporary fire risks, and the technical and economic feasibility of restoration. Carefully designed decision models can not only assist with a more circumspect approach to selection of individual treatment units, but can also show which of several treatment options may be most suitable in a given unit, thus also providing support for the tactical level of planning.
Similarly, evaluation of treatment priorities related to fire danger is not necessarily limited to fuel and fire characteristics; it can also incorporate human impacts and social or economic, or other value considerations. One such consideration, when evaluating the context of fire danger, may be the pattern of wildland-urban interface in the study area (Figure 4b). Readers might fairly ask, “Given that the structures of the logic model for danger evaluation and the decision model for treatment priorities are so similar in this example, why bother with two separate models?” First, and perhaps most obviously, the two models produce very different interpretations of the data (compare fire danger in Figure 5 with treatment priority in Figure 4a). The logic model is a relatively objective interpretation of fire danger, given that parameters used to interpret observations (Table: Definition of data inputs) were derived from field data, and given that the logic is presented in a relatively pure form insofar as all topics (with the exception of fireline intensity and flame length) are equally weighted. Although weights can easily be applied to topics in a logic model, they also add an additional level of subjectivity that is more effectively managed within the context of decision models, such as those based on the analytic hierarchy process, for example, that are more specifically designed to deal with such issues (Reynolds and others 2003). Logic models also offer the opportunity to synthesize and summarize potentially complex information, thus simplifying the structure of a decision model. In this study, for example, the decision model used summarized information about the topics under fire hazard that would otherwise have been difficult to adequately represent in an intrinsically linear decision model (see, for example, the description of the CBD topic in Logic Model Design).
Finally, the two types of models are very complementary in the sense that the logic model focuses on the question, “What have I got?”, whereas the decision model focuses on the question, “Now that I know what I have, what should I do about it?” Notice that logistical issues are not pertinent to the first question, but they may be extremely important for the second. An important consequence of separating the overall modeling problem into these two complementary phases is that each phase is rendered conceptually simpler. The logic model evaluates and keeps separate the status of the components of each ecological system under evaluation; in this case, the components of wildland fire danger of each subwatershed in the map zone. The decision model takes the ecological status of each ecosystem and places it in one or more social contexts that are designed to further inform decision making. The decisions will be based only partially on the ecological status information. They will also be based on social context and human values, in this case, proximity to and amount of wildland-urban interface, which captures a measure of the potential risk of fire damage to people and their structures. After priorities have been derived by the decision model concerning what to do about the existing fire danger conditions, the decision maker can look back at the decision and see the relative contributions of the ecological states and their social context(s) to the overall decision. This transparent model design and structure aids in decision explanation, and it allows decision makers to consider, in the sense of scenario planning, the effects of alternative weightings of important decision criteria.
As George Box (1979) noted, “All models are wrong; some are useful.” Thus, as with any model intended to support significant management decisions, our model of fire danger requires both verification and validation because all models are necessarily simplifications of reality. The present model has, in fact, been substantially verified in the sense that it performs as expected based on our own analyses and has been vetted in several meetings over the past year involving substantial numbers of prominent fire managers and fire scientists who agree that the representation of fire danger is reasonable. In contrast to verification, validation is a more rigorous process in which model accuracy is objectively evaluated by comparing predicted and actual outcomes, ideally with statistical procedures. Readers unfamiliar with logic-based models may wonder if validation is even possible. However, models based on logic are no better or worse in this respect than their probabilistic counterparts. Although a detailed discussion of this assertion is beyond the scope of this report, it may be sufficient to note that metrics expressing strength of evidence have commonly been treated as subjective probabilities (Zadeh 1968). Finally, model validation was not feasible within the temporal scope of our study. Realistically, even a preliminary validation in this context would require 5 to 10 years. If the model for fire danger were to be adopted as a tool to support strategic planning for fuels treatment, then we certainly recommend that explicit provisions for validation be an integral part of any ongoing assessment process designed to support it.
Encyclopedia ID: p3645
