RMLANDS is a stochastic landscape model that simulates disturbance and post-disturbance recovery of vegetation within a heterogeneous landscape. The landscape structures produced by this model were analyzed with FRAGSTATS, which summarizes landscape structure by means of numerous quantitative metrics, and HABIT@, which summarizes wildlife habitat capability by means of species-specific models for selected indicator species. We applied these models to the problem of characterizing the range of variability in landscape structure and wildlife habitat within the Uncompahgre Plateau Landscape in the southern Rocky Mountains, USA. The model was parameterized on the basis of our best empirical understanding of the pre-1900 disturbance regime in this region. The period of several centuries prior to 1900 represents a time when broad-scale climatic conditions were generally similar to those of today, but Euro-American settlers had not yet introduced the sweeping ecological changes that now have greatly altered many Rocky Mountain landscapes -- through fire suppression, grazing, road-building, timber cutting, recreation, and other activities (Knight et al. 2000). Thus, the pre-1900 period provides a suitable reference condition against which we can compare current landscape structure and dynamics (Swetnam et al. 1999, Landres et al. 1999). In addition, an understanding of natural landscape structures and variability during this reference period also provides a basis for forest management policies that seek to mimic natural disturbance patterns in our logging, grazing, and other activities involving commodity production from public forest lands (Romme et al. 2000, Buse and Perera 2002).

Scope & Limitations (Sources of Uncertainty)

      The results of our HRV analyses must be interpreted within the scope and limitations of this study - this is of paramount importance. Most importantly, our analyses were designed to simulate vegetation dynamics under a chosen historic reference period. We chose the period from about 1300 to the late 1800s, representing the period from Ancestral Puebloan abandonment to EuroAmerican settlement, as the reference or benchmark. This period is often referred to as the period of indigenous settlement, in contrast to the period of EuroAmerican settlement that began in the mid to late 1800s (Romme et al. 2003). The period of several centuries prior to 1900 represents a time when broad-scale climatic conditions were generally similar to those of today, but Euro-American settlers had not yet introduced the sweeping ecological changes that now have greatly altered many Rocky Mountain landscapes -- through fire suppression, grazing, road-building, timber cutting, recreation, and other activities (Knight et al. 2000). Thus, the pre-1900 period provides a suitable reference condition against which we can compare current landscape structure and dynamics (Swetnam et al. 1999, Landres et al. 1999). An understanding of natural landscape structures and variability during this reference period also provides a basis for forest management policies that seek to mimic natural disturbance patterns in our logging, grazing, and other activities involving commodity production from public forest lands (Romme et al. 2000, Buse and Perera 2002).

      It cannot be emphasized too strongly that the chosen reference period was not a time of stasis, climatically, ecologically, or culturally. For example, the “Little Ice Age” occurred during this time, and there were small shifts in the position of the upper timberline and in the elevational breadth of the forest zone on the middle slopes of the mountains (Petersen 1981). Local human inhabitants obtained horses and new technology and were affected by disease and displacement of other tribes brought about by European colonization farther to the east (Whitney 1994). Nevertheless, compared with some other periods in history, the period from about 1300 to the late 1800s was a time of relatively consistent environmental and cultural conditions in the region, and a time for which we have a reasonable amount of specific information to enable us to model the system. It should be emphasized, however, that our choice of reference periods does not suggest that it should be our goal in management to recreate all of the ecological conditions and dynamics of this period. Complete achievement of such a goal would be impossible, given the climatic, cultural, and ecological changes that have occurred in the last century. It also would be unacceptable socially, economically, and politically. Nor do we suggest that the reference period was completely “natural” or preferable in all ways to today’s landscape. However, the period of indigenous settlement does provide a good benchmark for evaluating current conditions, because it appears to have been a time when the ecosystem apparently supported rich biodiversity, conserved soils and nutrients, and ran sustainably on solar energy (Kaufmann et al. 1994).

      Because our study relied on the use of computer models, it is imperative that the limitations of these models be understood before applying the results in a management context. Here, we discuss several important limitations, some general to the modeling approach employed here and some specific to how we parameterized these models for application on the UPL.

      First, this report (and RMLANDS) focuses on the effects of two major natural disturbances: fire and insects/diseases. Other kinds of natural disturbances also occur, including wind-throw, ungulate and beaver herbivory, avalanches, and other forms of soil movement, but the impacts of these other disturbances tend to be localized in time or space and have far less impact on vegetation patterns over broad spatial and temporal scales than do fire and insects.

      Second, this report (and RMLANDS) devotes more attention to upland vegetation types than to riparian or aquatic types; indeed, riparian and aquatic vegetation are covered only briefly. There are two reasons for this emphasis on upland vegetation in RMLANDS: (1) riparian and aquatic vegetation cover only a small (but ecologically critical!) portion of the total landscape, and (2) vegetation patterns and dynamics of riparian and aquatic vegetation are more complex, more variable, and more difficult to model in a straightforward fashion than are patterns and dynamics of upland vegetation. Additional research is needed to fully characterize the range of variability in riparian and aquatic ecosystems in this landscape. We note, however, that the patterns and ecological processes of surrounding upland vegetation have profound influences on aquatic ecosystems; thus, our results for terrestrial vegetation provide a partial basis for future assessments of aquatic HRV.

      Third, it is important to realize that RMLANDS and HABIT@ models require substantial parameterization before they can be applied to a particular landscape and species, respectively. To the extent possible, we have utilized local empirical data. However, due to the paucity of local data, we also drew on relevant scientific studies, often from other geographic locations, and relied heavily on expert opinion when scientific studies and local empirical data were not available. The quality of the empirical data - which provides the best means of parameterizing the model - for this application varied considerably. In general, we had very little local empirical data on insect/pathogen disturbance regimes for the reference period. Consequently, our model parameterization largely reflects the opinions of several local and regional experts, supported in part by sparse empirical data often from other landscapes. By comparison, we had excellent empirical data on historical fire disturbance regimes, but its existence and quality for our application varied considerably among cover types. In general, we had good to excellent data for mountain shrublands and several forest cover types (especially spruce-fir forest, pure aspen forest, warm dry mixed-conifer forest, ponderosa pine forest), principally from Romme et al. (2003), from which to estimate the mean fire return interval (or rotation period) and/or the stand age distribution for the reference period. For other cover types, in particular the semi-desert grassland and savannah communities and mesic sagebrush cover type, we had no data at all and were forced to rely exclusively on expert opinion. Unfortunately, there was no objective way to verify our model findings for cover types without empirical data, so our model outcomes for these types must be viewed with extreme caution, or even discarded altogether if you are not willing to accept the expert opinion. The source of information used to parameterize the models is fully documented and subject to review. Thus, our results should not be viewed as definitive, but rather as an informed estimate of the HRV based on our current scientific understanding. In this regard, it is important to understand that our estimate of the HRV is subject to change as new scientific understanding or better data become available.

      A good example of the evolving state of knowledge is our newly emerging understanding of disturbance regimes in pinyon-juniper woodlands. At the time of our modeling exercise, the only empirical data we had available was preliminary data on stand age distribution from the UPL and surrounding region (Eisenhart 2004). We parameterized the model to be consistent with our initial interpretation of this stand age distribution. However, recent work by Shinneman and Baker (in review) highlights several potential biases in our interpretation of the Eisenhart data. In particular, the fire rotation period for this cover type may be substantially greater than what we simulated. A longer simulated fire rotation period would have resulted in an older stand age distribution and a greater proportion of stands in the late-seral (tree-dominated) stage than we observed. Hence, in light of these new observations, our findings regarding pinyon-juniper woodlands must be viewed with extreme caution.

      Fourth, it is important to consider the major sources of uncertainty in the model as they relate to sensitive parameters, especially with regards to sensitivity in model outputs of key interest given the objectives of the study. Although it was not feasible to conduct a formal sensitivity analysis of the model parameters, given the shear number of parameters and their interactions, it was possible to qualitatively consider model sensitivity issues and identify important sources of uncertainty in model outcomes - these ultimately determine the level of confidence we can place in the model outcomes. Although what follows is not an exhaustive and detailed description of all known sources of uncertainty, it provides a qualitative summary of three major sources of uncertainty that should help guide the interpretation of our findings by managers as well as to help identify information needs which should serve as the focus for future research and/or model refinement.


    The vegetation transition models developed from expert input define the discrete stand conditions (seral stages) and the rates of successional development and transitions following disturbances for each cover type. These models ultimately determine the quantitative consequences of any specified disturbance regime and, in our case, influence the reported simulated historic range of variation in landscape composition - the model output of key interest in this application. A faster transition rate between two stand conditions, for example, would result in proportionately more area, on average, in the later seral stage. Thus, our reported HRV statistics are fundamentally influenced by these vegetation models, and any modifications to these models would be expected to produce a change in the HRV statistics. Continued improvement and refinement of these models should be a focus of any future simulation experiments.


    Due to the approach we took for implementing disturbance processes in RMLANDS, landscape context is inherently important, and even though landscape context is not a parameter(s) of the model per se, the output is sensitive to the way in which landscape context influences results. This is particularly important with regards to the effects of fire on vegetation dynamics. Fire disturbance events initiate and spread across the landscape (and across cover types) in relation to a number of factors. Cover type is just one of many factors potentially influencing fire behavior and effects. As a result, the quantified fire regime at a location is influenced by the local conditions (including cover type) as well as the landscape conditions around that location (i.e., its landscape context). Ultimately, all stands of the same cover type (often including widely varying landscape contexts) are combined together for reporting purposes. It is important to recognize that these global statistics can be misleading and may connote a single disturbance regime when in fact spatial variation is the trademark of these regimes. The problem lies not so much in the way RMLANDS treats disturbances or in the reporting by cover type, but the lack of empirical data on the role of landscape context. No local studies have been conducted that explicitly address the role of landscape context, so we have no way of objectively verifying our model results in this regard. Improving our understanding of the role of landscape context remains an important focus for future research.


    One of principal objectives of this study was to quantify the range of variation in landscape structure. Among the many factors that influence the range of variation (e.g., climatic variability, scale, etc.), the shape of the disturbance size distribution is one of the most important. In particular, the length and weight of the tail of the size distribution determine the maximum size and frequency, respectively, of largest disturbance events. These relatively infrequent large disturbances have a disproportionate impact on vegetation patterns and dynamics and can cause the landscape to fluctuate widely among states (Turner et al. 1993). Consequently, the range of variation in landscape structure is highly sensitive to the specified length and weight of the tail of the disturbance size distribution. Unfortunately, available empirical data do not allow us to precisely estimate the tail of the size distribution, even though they allow for a robust estimate of the remainder of the distribution (i.e., small and intermediate disturbances). For this application, we specified a maximum fire size of 100,000 ha and a probability of 0.0005 of any given fire being in the 10,000-100,000 ha size class. The weight of the tail was estimated from empirical data representing recent historical fires (1970 to present) within the ecoregion. The largest recorded fire was approximately 22,000 ha (Missionary Ridge fire, 2002). We reasoned that historical fires during the reference period had the potential to be much larger, but how much larger we did not know. There is evidence of contemporary fires exceeding 100,000 ha in other ecoregions (e.g., Yellowstone National Park, Romme and Despain 1989) and there is evidence of extensive fires of unknown extent that “burned across the San Juan landscape” during the reference period (Romme et al. 2003). Our final specification of 100,000 ha represented a reasonable estimate (in our opinion) based on available data, but was nevertheless imprecise. Unfortunately, it is unlikely that we will ever be able to precisely estimate the size distribution for fires during the historic reference period. Consequently, the best we can do is to quantify the sensitivity of the model in this regard - the subject of current investigation. In the interim, managers should exercise caution in interpreting the reported ranges of variation in landscape structure.

      In summary, our approach relies heavily on the use of computer models, and while it is important to recognize the many advantages of models, it is critical to understand that models are abstract and simplified representations of reality. RMLANDS, in particular, simulates several natural disturbance processes, but does not simulate all of the disturbance processes (as noted previously) or all of the complex interactions among them that characterize real landscapes. In addition, ultimately the results of a model are constrained by the quality of input data. While RMLANDS utilizes a rich database, the data layers themselves are not perfect. For example, the vegetation cover layer is subject to human interpretation errors and objective classification errors, and is further limited by the spatial resolution of the grid. Moreover, while we utilized a rich array of empirical studies to parameterize RMLANDS, our estimates are not perfect. For example, our estimate of the mean fire return interval in ponderosa pine forest does not take into account the potentially important role of landscape context. Thus, our results should not be interpreted as “golden”. Rather, they should be used to help identify the most influential factors driving landscape change, identify critical empirical information needs, identify interesting system behavior (e.g., thresholds), identify the limits of our understanding, and help us to explore “what if” scenarios.

Disturbance Processes & Dynamics

      Wildfire.--Numerous types of natural disturbance occurred in the study area during the reference period: fire, snow avalanches, windthrow, and a variety of tree-killing insects, fungi and other pathogens (Veblen et al. 1989, Lertzman and Krebs 1991, Veblen et al. 1991a, b, Roovers and Rebertus 1993, Veblen 2000, Romme et al. 2003). However, the most important and coarsest in scale of these natural disturbances was fire. Based on a number of fire history studies and relatively extensive empirical data from the region surrounding the UPL, Romme et al. (2003) concluded that the median fire return interval varied dramatically across the landscape along an elevational gradient in relation to fuels and moisture conditions, ranging from 10-30 years (for low-mortality fires) in the lower elevation ponderosa pine type, 20-50 years (for low-mortality fires) in the dry mixed-conifer type, 140-150 years (for high-mortality fires) in the aspen type, and 200-350 years (for high-mortality fires) in the spruce-fir type. Unfortunately, our empirical understanding of fire history in the lower elevation woodlands is much more limited, and it is essentially non-existent in the low elevation semi-desert communities. Recent studies in Mesa Verde (Floyd et al. 2004) and on the UPL (Shinneman and Baker, in review) suggest a long rotation period of perhaps 400 years or more for stand-replacing fires in pinyon-juniper woodlands. However, at the time of our simulation experiments these data were not available and we assumed a shorter return interval of 200-300 years (for high-mortality fires). Romme et al. (2003) also noted that many individual stands escaped fire for far longer than the median return interval and some burned at shorter intervals, creating a complex vegetation mosaic at the landscape scale. They further hypothesized that under these reference period conditions, stand replacement fires initiated stand development and maintained a coarse-grain mosaic of successional stages and cover types across the landscape, while non-replacement fires functioned to maintain communities in a particular condition (e.g., open canopy ponderosa pine forest) or accelerate the successional process of stand development.


      Our simulations largely confirmed these observations and provided a detailed quantitative summary of the wildfire disturbance regime. In particular, we recorded a similar elevational gradient in mean fire return intervals (or rotation periods) among the forested cover types, ranging from 50-years (for low-mortality fires) in the lower elevation ponderosa pine (without and with aspen) type, 71-79 years (for low-mortality fires) in the warm dry mixed-conifer (without and with aspen) type, 144 years (for high-mortality fires) in the pure aspen type, 183-189 years (for high-mortality fires) in the cool moist mixed-conifer (with and without aspen) type, and 181-182 years (for high-mortality fires) in the spruce-fir (with and without aspen) type (Table-rotation). Similarly, we recorded mean fire return intervals ranging from 153-216 years (for high-mortality fires) in pinyon-juniper woodland types. The differences between our findings and those previously reported in pinyon-juniper woodlands largely reflect differences between the data we used to parameterize the model and more recent data not available at the time of our experiments. Otherwise, the differences can be attributed to biases associated with the approaches used to estimate return intervals in each study, and these biases are important to understand in order to properly interpret our results.

      First, return intervals in cover types with surface fire regimes (e.g., ponderosa pine forests) can be calculated in a number of ways that influence the computed statistics (Baker and Ehle 2001). For example, our estimate of a 50-year mean return interval for low-mortality fires in ponderosa pine-oak forests was based on the computed rotation period for low-mortality fires in this cover type. The rotation period is equivalent to the cell-specific mean return interval for the cover type, and is a nonspatial representation of the disturbance regime because it does not depend on the explicit spatial distribution of disturbances; rather, it depends only on the total area disturbed each time step in relation to the total eligible area. In our simulations, the rotation period is equivalent to the average interval between fires in a 25-m cell. In contrast, most published return intervals are derived as a composite mean return interval, which represents the frequency in which disturbance occurs anywhere within a sampling area of fixed size - usually corresponding to a forest “stand”. The composite mean return interval is profoundly influenced by the size of the sampled area: a larger area will encompass more disturbances and will therefore have a shorter composite mean return interval. In addition, because disturbances often only disturb a portion of the specified sampling area, the actual return interval to any particular location is usually much longer than the return interval to the sampling area. To help resolve this uncertainty, and to help reduce the effect of small, localized disturbances on the composite mean return interval, some investigators also compute a filtered composite mean return interval, based only on disturbances that are recorded on, say, >25% of recorder trees. These are the disturbances that presumably affect a greater proportion of the stand. The filtered composite mean return intervals are often twice as long as the unfiltered, but filtering does not resolve the fundamental issue of what fraction of a sampling area actually was affected by each disturbance event. Alternatively, the mean return interval can be computed by taking an average across all recorder trees of all recorded intervals. In this case, each recorded interval represents a separate observation, as opposed to a plot of ground in the previous approaches. Again, return intervals calculated in this manner are often much shorter than the actual rotation period (or mean return interval to a particular location). When we computed the mean return interval between low-mortality fires in ponderosa pine-oak forest using this approach (which we refer to as the sample mean return interval), in which each recorded interval between low-mortality fires in a cell was treated as an independent observation, the median and modal return interval was somewhat shorter (30 and 20 years, respectively) and within the range reported by Romme et al. (2003). In this report, we report all results using the population (i.e., cell-specific) mean return interval approach; readers should exercise extreme caution when comparing our results to published return intervals.

      Second, the differences between our mean fire return intervals and those previously reported for several cover types can be attributed to biases associated with the landscape context of stands sampled in the field. This is particularly true for the forested cover types sampled by Romme et al. (2003). In general, the stands sampled in the field were restricted to large contiguous stands, whereas our estimates from the simulation were based on an exhaustive summary of every stand regardless of landscape context. For example, in our simulations, large contiguous stands of ponderosa pine exhibited a much shorter fire return interval than the average across the entire population of ponderosa pine stands; the latter included a rather large proportion of stands interspersed with less flammable cover types at the lower and upper elevational limits of the cover type distribution (e.g., pinyon-juniper woodlands and warm, dry mixed conifer forest, respectively). Hence, it is not surprising that our return intervals were somewhat longer than those reported by Romme et al. (2003). Unfortunately, we have no way of determining the magnitude of the bias to be expected in this regard.

      Third, some of the discrepancy between our results and those previously reported can be attributed to our regional parameterization of the model. Specifically, we parameterized the model based on empirical data (and other sources) derived from different areas within the broader regional landscape - although generally within the South Central Highlands Section. In particular, we used data sources from both the San Juan National Forest (SJNF) and the UPL to parameterize and calibrate the model for application anywhere in the region. For this purpose, we utilized the “best” data available for simulating disturbances in each cover type. In general, the model was calibrated to produce the desired behavior in each cover type separately using the landscape containing the greatest extent of that cover type. For example, the SJNF contains proportionately more spruce-fir forest than the UPL (18.2% versus 1.2% of the landscape). Thus, we calibrated the model to produce the desired disturbance regime (e.g., fire rotation period) and stand age distribution on the SJNF and then simply applied these parameters to the UPL. We used a similar process for each of the major cover types. Hence, the simulation results in a particular landscape reflect the impact of the local landscape structure on the disturbance regime and vegetation dynamics. For example, the high-elevation forests on the UPL are less extensive and more interspersed with relatively flammable cover types that have shorter fire return intervals (e.g., mountain shrubland) than on the SJNF. Thus, the observed rotation periods for high-mortality fires in these cover types on the UPL were considerably shorter than we observed on the SJNF, where they were consistent with the empirical data from Romme et al. (2003).

      Overall, taking the biases mentioned above into account, our results are generally consistent with those reported by Floyd et al. (2000, 2004), Romme et al. (2003), Eisenhart (2004), Shinneman and Baker (in review) and others based on local (or nearby) empirical field studies. The most notable exception is the possible simulation of too much fire in the low and mid elevation woodland communities. If this be the case, then we may be reporting a much younger seral-stage and stand age distribution than is realistic for this landscape.

      Consistent with Romme et al. (2003), we also noted the distinct variability in fire return intervals among locations within a single cover type. For example, in pinyon-juniper woodland, the most extensive cover type on the UPL (~20 % of landscape), the mean return interval between fires (of any mortality level) varied spatially across the forest from 24 years to >800 years, with a mean and median of 197 and 267 years, respectively, and roughly 5% of the area escaped disturbance altogether over the course of an 800-year simulation (Figure-return, Figure-map). In general, return intervals decreased for stands embedded in a neighborhood containing cover types with shorter return intervals (e.g., mountain shrublands and ponderosa pine forest at the higher elevations and semi-desert grasslands and savannahs at the lower elevations) and increased in areas containing extensive, contiguous pinyon-woodlands. These patterns of variation were remarkably consistent among all cover types, highlighting the importance of landscape context on fire regimes and demonstrating that no single statistic, such as mean fire interval (MFI), is adequate to characterize fire regimes, and that the widely used MFI actually may be quite misleading if taken literally - it may connote homogeneity and/or consistency when in fact spatial and temporal variability is the trademark of these regimes.

      Perhaps the single greatest insight gained from our simulations with regards to wildfire is the shear magnitude of wildfire disturbance that is required to produce the widely accepted return intervals for the reference period. On average, once every two decades, >10% of the area (~66,000 ha) was burned, and roughly once every 120 years, >20% (~132,000 ha) was burned (Figure-recurrence), inclusive of both high- and low-mortality affected areas. This is a tremendous amount of burning and perhaps a magnitude of burning that is poorly appreciated by land managers and the general public based on public reactions to recent “large” fires in the west. Note, the largest recorded fire in the region was the Missionary Ridge fire of 2002 on the SJNF, which burned a mere 20,000 ha, yet prompted significant reaction among managers and the public. As note above, it is quite possible that we simulated too much fire, especially in the low and mid elevations, but even if we were to double the fire rotation period for the landscape as whole, it would still result in a tremendous magnitude of wildfire.

      Insects & Disease.--Hundreds of species of insects, fungi, and other pathogens that cause tree death or damage also inhabited these forests during the reference period (Furniss and Carolin 1977). Any of them may have been locally important on occasion (Schmid and Mata 1996). However, it was not feasible to explicitly simulate more than a handful of insects and diseases in a complex landscape model like RMLANDS. Therefore, we identified four insect species and one insect/disease (pathogen) complex that have had the most frequent and widespread impact on vegetation in the SJNF region (Romme et al. 2003). The insects included mountain pine beetle (Dendroctonus ponderosae) and its affiliates, Douglas-fir bark beetle (Dendroctonus pseudotsugae), spruce bark beetle (Dendroctonus rufipennis), and western spruce budworm (Choristoneura occidentalis). The insect/disease complex that we treated is referred to as "pinyon decline" and includes a combination of black stain root rot (Leptographium wagneri) and the pinyon ips beetle (Ips confusus).

      In contrast to wildfire, there was comparatively little empirical data on insect/disease disturbance regimes for the reference period and almost no local data. Consequently, we were forced to draw heavily on contemporary observations of outbreaks from throughout the Rocky Mountain region (Schmid and Mata 1996) in combination with local and regional expert opinion. In addition, due to the paucity of empirical data available for model verification purposes, we were forced to calibrate the simulations based on the user-specified disturbance regimes. Not surprisingly, therefore, our simulations produced disturbance regimes consistent with our parameterization. While this may seem a bit circular, it was a necessary process for a complex model such as RMLANDS. Plus, our real emphasis was on quantifying the vegetation patterns and dynamics resulting from these disturbance processes. Hence, we gained few true insights from our simulations regarding the insect/disease disturbance regimes. Nevertheless, there were a couple of important observations worth noting.

      First, the overall rotation periods for insect/disease disturbances were generally much longer than wildfire, which had an overall rotation period of 100 years (Table-rotation). Spruce budworm had the shortest rotation period of any insect/disease agent at 104 years (Table-rotation), followed by pinyon decline at 197 years (Table-rotation), pine beetle at 222 years (Table-rotation), spruce beetle at 323 years (Table-rotation) and Douglas-fir beetle at almost 1,247 years (Table-rotation). Hence, taken individually, with the exception of spruce budworm, insect and disease disturbances had much less overall impact on the landscape than wildfire. However, taken collectively, insects and diseases clearly impacted more area per unit time than wildfire. We might conclude, therefore, that too little attention has been given to the potentially important role of insects/disease compared to wildfire.

      Second, the ecological impacts of insect/disease outbreaks were fundamentally different than wildfire in a couple of important ways with regards to vegetation patterns and dynamics. With few exceptions, insect/disease outbreaks resulted in proportionately very little stand replacement or none at all. Most disturbances were low mortality and either promoted successional development of younger stands by thinning the tree canopy and facilitating understory development or acted within older stands as a gap-scale disturbance processes that facilitated the development of the true old-growth, shifting mosaic stand condition, or in some cases (e.g., pinyon-juniper woodlands) caused retrogression to a previous stand condition (e.g., low-mortality pinyon decline shifting stands from the tree-dominated condition to the shrub-tree condition). Even high-mortality outbreaks in most cases did not cause stand replacement due to the host specificity of the insect/disease agent. For example, spruce beetle outbreaks were principally high-mortality disturbances, and in this way were similar to wildfires in high-elevation conifer forests. However, the high mortality was limited to the host species, Engelmann spruce. Consequently, high-mortality spruce beetle outbreaks did not cause stand replacement unless they were concurrent with a high-mortality spruce budworm outbreak - a rare occurrence. Thus, the more important affect of spruce beetle outbreaks was to thin stands and shift the species composition of the post-disturbance stand in favor of subalpine fir. In contrast, wildfire in most cover types resulted in proportionately more stand replacement and was therefore principally responsible for the maintenance of the coarse-grained mosaic of successional stages across the landscape. Insect/disease disturbances in most cases also exhibited notably different spatial patterns than wildfires. In general, wildfires produced relatively contagious disturbance patches (i.e., large, contiguous patches containing relatively few gaps - a spatial property known as low lacunarity, Plotnick et al. 1993), whereas most insect/disease outbreaks produced relatively non-contagious patterns characterized by a great deal of internal fine-grained heterogeneity (i.e., high lacunarity)(Figure-lacunarity). Although not universally true, in general, insect/disease agents were responsible for creating much of the fine-scale heterogeneity in vegetation patterns in our simulations.

Vegetation Patterns & Dynamics

      It is widely accepted that during the reference period natural disturbance processes, notably wildfire and a variety of insects/diseases, operated to create and maintain a complex vegetation mosaic of successional stages and cover types (Romme et al. 2003). It is less clear whether this vegetation mosaic was stable in structure (composition and configuration) or the degree to which it varied over time. With our simulations, we sought to quantify the range of variability in landscape structure during the reference period to help ascertain the degree of dynamism in landscape structure and to provide a benchmark for comparison with alternative future land management scenarios. To this end our simulations produced several important findings.


Click on the link below to view a movie depicting vegetation changes on the UPL over an 800-year (10-year time steps) simulation representing the reference period disturbance regime. NOTE, this is a large (57 Mb) Microsoft Media file (.avi) that requires appropriate movie viewing software (e.g., Quick Time Player).

                Uncompahgre Plateau Landscape Movie

      Overall, the vegetation mosaic was remarkably variable in structure over time (see the movie link above). For example, across the 63 dynamic patch types (i.e., unique combinations of cover type and stand condition), the average coefficient of variation in percentage of the landscape comprised of the corresponding patch type was 186% (range 54-655%). Hence, while the landscape could be characterized as a shifting mosaic of successional stages and cover types, it was not a steady-state shifting mosaic (sensu Bormann and Likens 1979). In other words, the composition of the mosaic was not constant. This finding is consistent with evidence from a wide variety of other coniferous forest landscapes in North America, including boreal forests of western Canada (Johnson 1992) and the Great Lakes region (Baker 1992a,b), and subalpine forests of the Yellowstone Plateau (Turner et al. 1993). The dynamism we noted can be attributed to two major sources of variation in the model: (1) stochasticity in the disturbance parameters associated with initiation, spread, and mortality and (2) the climate modifier. The stochastic nature of disturbance initiation (e.g., representing random lightning strikes associated with local storm systems) and subsequent spread (e.g., representing uncertain weather conditions during the hours, days, etc. following a wildfire initiation) introduced uncertainty into which stands were disturbed, when they were disturbed, and how severely they were disturbed. Thus, as a matter of chance, a large stand-replacing disturbance might have affected a large proportion of a cover type and thereby altered its seral-stage distribution. The climate modifier had a similar destabilizing effect on vegetation patterns by influencing the rate of disturbance in each timestep (decade) of the simulation. For example, drought cycles (as implemented via the climate modifier) had a substantial influence on the frequency and extent of wildfire (Figure-initiations, Figure-extent). This pattern of variation in climate and fire is consistent with findings from tree-ring studies in Colorado and throughout the Southwest (e.g., Swetnam and Betancourt 1998, Veblen 2000). In combination, varying climate conditions and the stochastic occurrences of disturbances acted to keep the system in a constant state of change.

      Although the vegetation mosaic was not in a steady-state equilibrium, the mosaic was generally in dynamic (or bounded) equilibrium (sensu Turner et al. 1993). That is to say, while the structure of the landscape varied over time, it generally fluctuated within bounds about a stable mean (e.g., Figure-equilibrium). This behavior is essential to our objective of describing the range of variability in landscape structure, because the concept of a “range of variability” implies that the range is stable. If the landscape is not in dynamic equilibrium and, for example, exhibits a trend, then the measured range of variation will vary with the specific period of measurement (Figure-equilibration scale). In our simulations, most (but not all) metrics achieved a stable, bounded equilibrium within a 100-300 year simulation period - although it took twice that length of time to verify that the range of variation was in fact stable. Not surprisingly, the period required for equilibration in landscape composition varied somewhat among cover types. In general, cover types that experienced shorter disturbance return intervals and/or faster rates of succession equilibrated in the shortest period. For example, the seral-stage distribution in mountain shrubland, which experienced an 76-year rotation period for stand-replacement wildfire disturbances and succeeded to the latest seral stage in as little as 50 years following stand-replacement, reached stable bounds within a 100-year period (Figure-mts condition). Conversely, the seral-stage distribution in spruce-fir forest, which experienced a 182-year rotation period and took at least 300 years to succeed to the latest seral stage, reached relatively stable bounds within roughly a 300-year period (Figure-sf condition).

      Our results demonstrate that the range of variability in landscape structure cannot be expressed in a single metric - at least not effectively - because the metrics associated with different aspects of landscape composition and configuration exhibit varying degrees and patterns of dynamism. In our simulation, landscape composition metrics exhibited five-times greater dynamism (on average) overall than landscape configuration metrics (Table-hrv). Thus, while the composition of the vegetation mosaic - specifically, changes associated with seral-stage distributions - fluctuated dramatically over time, the spatial pattern of the mosaic was relatively stable. The variability in configuration was principally associated with changes in the size and continuity of the large patches in the landscape. Overall, this suggests that large, severe disturbance events, those that occurred relatively infrequently but that substantially altered the seral-stage distribution and created a coarse-grained mosaic of vegetation patches, were disproportionately important in regulating the dynamism in landscape structure. In contrast, the relatively frequent small disturbances had little impact on overall landscape pattern or change through time. This finding is consistent with increasing evidence from a wide variety of disturbance-dominated landscapes in North America that large fires are often more severe than smaller fires and have a stronger influence on long-term ecological structure and function, for example, by introducing successional trajectories that differ from the expected (Moritz 1997, Romme et al. 1998, Turner et al. 1997). In particular, landscapes resulting from large fires are often complex mosaics comprised of low severity surface burns where soils are largely intact, high severity areas of crown fires with extensive tree mortality and consumption of soil organic layers, moderate or mixed effects, and islands of unburned vegetation. Important consequences of the spatial patterns of burn severity include resulting patterns of surviving organisms that dictate initial succession patterns (Turner et al. 1998) and differential responses of animal species in relation to post-burn habitats (Kotliar et al. 2002).

      The role of large disturbances is especially noteworthy when considered in relation to climate change. Small climatic changes have potentially significant effects on disturbance regimes, and dramatic changes in ecological communities related to fire regime dynamics are possible (Clark 1988, Swetnam et al. 1990, Sprugel 1991, Turner et al. 1997, Meyer and Pierce 2003). One predicted outcome of global climate change includes the increased frequency of severe disturbance events, including large fires (Ryan 1991, Torn and Fried 1992). Indeed, the possibility that recent high-severity large fires are part of a longer-term trend has raised alarm in many areas, including the Rocky Mountain West, because the implications for key environmental processes and biological responses are highly uncertain (McCarthy and Yanoff 2003). Our results suggest that if climate change results in an alteration in the frequency and extent of large disturbances, then one of the principal impacts may be to alter the dynamics in landscape structure.

      Based on these findings, it is easy to reach the conclusion that the landscape was in equilibrium, albeit a dynamic one, during the reference period. However, this conclusion warrants careful consideration given the important ramifications. In particular, our simulation model treated disturbance and succession as a stationary process (i.e., stochastic process that does not change in distribution over time or space; Loucks 1970) with random perturbation. For example, all other things being equal, the rate (probability) of succession transition from one stand condition to another was held constant for the entire simulation. Given stationary parameters, it is a certainty that the model will eventually reach equilibrium if given enough time. Similarly, given the stochastic implementation of these stationary processes, it is just as likely that the landscape will not achieve equilibrium over a very short period of time (e.g., over a few timesteps). Hence, the equilibrium concept is ultimately simply a matter of scale (Turner et al. 1993). The more relevant question is: given the length of time it takes for the landscape to demonstrate equilibrium dynamics, is it likely that the factors governing disturbance and succession processes (and hence vegetation dynamics), such as climate, were stationary for that length of time during the reference period? As stated above, most metrics exhibited a stable, bounded equilibrium within a 100- to 300-year period, which is well within the length of our reference period (1300 to late 1800's). Moreover, while we recognize that Rocky Mountain climates have varied during the last 600 years at scales of decades and centuries (e.g., Millspaugh et al. 2000, Veblen 2000) and that the reference period was not a time of complete stasis, climatically, ecologically, or culturally (e.g., Petersen 1981, Whitney 1994), the period of several centuries prior to 1900 was a time of relatively consistent environmental and cultural conditions in the region (Romme et al. 2003). Thus, we believe that it is likely that the landscape exhibited dynamic equilibrium conditions during the reference period.

      Lastly, the existence of dynamic equilibrium does not imply that the mean condition of the landscape (with respect to any particular metric) is an adequate descriptor of the reference period - only that it is stable. Indeed, the actual mean condition of the landscape, as given by any metric, was rarely, if ever, realized. Instead, given the high variability over time in landscape structure, the range (or bounds) of variation should be emphasized, especially when using these results as a benchmark for comparison with other disturbance scenarios.

Wildlife Habitat Patterns & Dynamics

      Policy mandates require National Forest managers to maintain viable populations of all native wildlife species found on National Forest lands. Additional research is urgently needed to evaluate the current status of wildlife species, and to identify management strategies that will help ensure their long-term persistence. A combination of empirical, theoretical, and modeling studies can best contribute to our understanding of past, present, and future status of overall biodiversity and individual species in National Forests. In this study, we have used a modeling approach to characterize the baseline range of variability in habitat capability for a suite of species representing a diversity of habitat requirements. An important next step will be to compare the range of variation that we have described with potential habitat under alternative land management scenarios. It is important to note that we did not simulate wildlife populations per se. Rather, we simulated habitat conditions, and thereby implied potential population distributions and dynamics as a function of habitat conditions. It is important to realize that other factors (e.g., predator pressure) also play an important role in determining actual population densities. Empirical studies are urgently needed to test the predictions of population distribution and dynamics generated by this study.

      We demonstrated that habitat capability varies over time and space for all species – something that has long been recognized intuitively but rarely quantified. Not surprisingly, the magnitude and pattern of variation differed somewhat among species (Table-LC index). We initially predicted that habitat characteristics for generalist species (represented by elk in this study) would exhibit the least variation over time. Our reasoning was that as one kind of suitable habitat became reduced in availability through normal landscape dynamics, other kinds would become more available. Thus, long-term variability in the total amount and configuration of suitable habitat would be small. In contrast, we predicted that habitat for specialist species (pine marten, three-toed woodpecker, and olive-sided flycatcher in this study) would fluctuate more widely over time, because alternative kinds of habitat did not exist to compensate for natural fluctuations in the preferred habitat. To our surprise, one of the specialists, the olive-sided flycatcher, exhibited the least variation in habitat capability over time. This was likely due to the relatively consistent supply of high-contrast edges - the preferred habitat - bordering permanent openings such as meadows, barren areas, and lakes and ponds that acted like a buffer against major fluctuations in available habitat. Not surprisingly, the three-toed woodpecker exhibited the greatest variation in habitat capability over time - almost twice that observed for pine marten and elk. The dramatic fluctuations in tree-toed woodpecker habitat capability reflected the periodic pulses of high quality habitat following large-scale disturbance events in the mid- and high-elevation conifer forests. In addition, as expected, the peaks in habitat capability were usually short-lived (i.e., one time step) and followed by a rapid return to the pre-disturbance habitat capability level. These patterns held at the finer watershed scale.

      We selected these four wildlife indicator species based on differences in life history and habitat associations. Not surprisingly, therefore, each species exhibited somewhat unique patterns of variation through time and space that made generalizations difficult. For example, olive-sided flycatcher habitat was well-distributed throughout the landscape at all times and consisted of a mixture of relatively persistent high-contrast edges bordering permanent openings such as meadows, barren areas, and water bodies, in addition to transient edges associated with the heterogeneous pattern of tree mortality following both wildfire and insect outbreaks. Thus, there were both persistent local sources of high-quality habitat and transient but extensive sources of high-quality habitat that followed episodic disturbances. Overall, the spatial distribution of high-quality olive-sided flycatcher habitat was quite distinct (Figure-osfl) when compared to, say, the more coarse-grained and contagious distribution of pine marten habitat (Figure-marten). In addition, in contrast to the very transient nature (i.e., 10-20 years) of three-toed woodpecker habitat following disturbances, high-quality olive-sided flycatcher habitat tended to degrade much more slowly over several decades in response to gradual succession processes operating in the disturbance opening. Ultimately, as these few examples illustrate, each indicator species exhibited a unique spatial and temporal pattern of variability in capable habitat that reflected differences in life history (e.g., home range size) and habitat affinities (e.g., preferences for edges, forest interiors, or post-disturbance environments).

HRV Departure

      One of the principal purposes of gaining a better quantitative understanding of the historic reference period is to know whether recent human activities have caused landscapes to move outside their historic range of variability (Landres et al.1999; Swetnam et al. 1999). To this end, land managers have largely adopted a single approach based on Fire Regime Condition Class (FRCC) determination. FRCC is a categorical classification of the degree to which the current fire regime and composition and structure of a vegetation community deviates from its natural or historic range of variability (HRV) under a designated reference period (FRCC website). FRCC has become a driving force behind current land management activities and is widely being used as the primary (or even sole) basis for identifying and prioritizing areas for ecological restoration, including the reduction of wildfire risks associated with hazardous fuels.

      Our simulations were designed to provide a quantitative description of landscape structure dynamics for the pre-1900 reference period against which to compare landscape trajectories under alternative future land management scenarios. We strongly believe that the most appropriate and defensible use of our quantitative findings is in the context of evaluating the relative impacts of alternative scenarios (Romme et al. 2000, Buse and Perera 2002). It was not our original intent to compare the HRV in landscape structure against the current landscape. However, HRV departure has become a (the) critical management issue on many national forests, as noted above. Thus, we modified the approach for FRCC determination to one better suited to our modeling environment, and possessing other distinct advantages over FRCC (see Methods for a detailed description), in order to examine the magnitude of departure of the current landscape condition from the simulated HRV. Briefly, our approach is based on a spatially-explicit model of disturbance and succession (instead of a nonspatial model), incorporates multiple disturbance processes (not just fire), explicitly incorporates the measured range of variation in each metric (instead of using the mean), results in a continuously-scaled departure index (instead of a 3-class categorization of departure level), adopts a truly multivariate perspective on vegetation departure by incorporating multiple composition and configuration metrics (instead of a bivariate summary), and allows for an explicit assessment of the effects of scale on departure. Like the FRCC approach, however, our approach is not without significant limitations; therefore, the findings discussed below must be considered within the scope and limitations that follow.

      Our chief limitation in estimating HRV departure was the paucity (or absence) of spatially explicit data on current stand age and/or condition (i.e., seral stage) for the non-forested cover types. Unfortunately, these cover types have not received the same attention in this regard as the forested cover types; collecting this data should be a priority for future work. Without a reliable estimate of current age and/or seral-stage distribution, it is impossible to provide a reliable estimate of HRV departure. Consequently, at the class level (in which each cover type is considered independently), we restricted our estimate of departure to the forested cover types - where we had confidence in the results. For the landscape as a whole, we included only forested cover types in our estimate of composition departure, for the reasons mentioned above. However, our estimate of configuration departure was inclusive of all cover types, because it was not practical or meaningful to measure landscape configuration for only a portion of the patch mosaic. This has important implications because forested cover types represent only 20% of the area comprised of dynamic patch types on the UPL. Thus, our estimate of overall landscape departure must be viewed with extreme caution.

      In the context of these limitations, the current landscape structure appears to deviate substantially from the simulated HRV (Table-hrv-summary). Many characteristics appear far outside the HRV. Indeed, the majority of both landscape composition metrics (21/40) and landscape configuration metrics (10/19) are completely outside their HRV (i.e., 100% departure)(Table-hrv). In general, the current landscape has fewer, larger, more extensive and less isolated patches with less edge habitat than existed under the simulated HRV. The larger patches tend to be geometrically less complex and contain proportionately more core area than existed under the simulated HRV. Overall, the current landscape is more contagious and less structurally diverse than existed under the simulated HRV. This can be interpreted as a more homogenous landscape, where the lack of any extensive disturbance during the past 100 years has led to large, mostly late-seral patches, with low contrast due to the paucity of younger seral stages.

      The patterns of departure were generally similar at the class level for each of the forested cover types with reliable data on current conditions, although there were some noteworthy variations. In particular, overall the high-elevation forest types exhibited much less departure in terms of both composition and configuration than the low-elevation forest types, and this pattern is consistent with the observations of Romme et al. (2003) and others that the disturbance regime in high-elevation forests has not been altered by twentieth century land use practices to the same extent as low-elevation forests. In addition, while the current seral-stage distribution of high-elevation spruce-fir forests appears to be generally within the simulated HRV, the mid- and low-elevation forest types, including aspen, mixed-conifer, and ponderosa pine types, appear to contain an overabundance of stands in the mid-seral stages (i.e., stem exclusion) compared to the simulated HRV. Moreover, these mid-seral stands appear to be larger, more extensive, geometrically more complex and less isolated than was typically observed under the simulated HRV. Perhaps the most notable departure we observed was the complete absence of stands in the fire-maintained open canopy condition in the low-elevation ponderosa pine and warm dry mixed-conifer forests. Historically, within these cover types this condition was quite prevalent in our project area (Romme et al. 2003) and throughout the southwest (Swetnam and Baisan 1996). In our simulations, low-elevation fire-maintained open canopy forest comprised anywhere from roughly 2% to 5% of the landscape over time and was therefore never a dominant feature of the landscape at any time, but this condition comprised anywhere from 20% to 80% of the corresponding cover types and was therefore a dominant feature of these cover types. The current departure is clearly due to the dearth of wildfires over the past century, which has had a couple of notable consequences. First, the prolonged absence of fire following a pulse of widespread regeneration in the early twentieth century has allowed many young stands to become overstocked through the successful establishment and growth of new stems. This has resulted in the preponderance of stands in the stem exclusion stage of development. Second, the absence of fire has allowed many stands to succeed to the understory reinitiation or shifting mosaic (i.e., late-seral) stages, instead of transitioning to the fire-maintained open canopy condition.

      Given the direct link between vegetation patterns and wildlife habitat, it is not surprising that the wildlife indicator species we analyzed exhibited varying degrees of departure from their simulated HRVs (Table-LC index). All of the species we considered are doing poorly in the current landscape compared to the simulated HRV. The pine marten is doing the least poorly in the current landscape. The moderately extensive late-seral conifer forest in the higher elevations is likely providing some ideal habitat for this species (Buskirk and Powell 1994, Hargis et al. 1999), although the total amount of ideal habitat is apparently somewhat less than was realized on average under the simulated HRV. The three-toed woodpecker, a species better adapted to exploit post-disturbance environments (Harris 1982, Hitchcox 1988, Hutto 1995), is doing slightly worse than pine marten owing to the paucity of recent large disturbances. Elk and olive-sided flycatchers are both disadvantaged the most in the current landscape. Both species benefit from edges between early- and late-seral vegetation patches. Specifically, elk benefit from the juxtaposition of forage (found in early-seral openings) and hiding cover (found in closed-canopy stands)(Reynolds 1966, Boyce and Hayden-Wing 1980, Thill et al. 1983), while olive-sided flycatchers use open areas as foraging habitat and use edges as nesting habitat (Finch and Reynolds 1988, Altman 1997). The paucity of disturbances over the past century has left the current landscape rather deprived of edge habitat and has reduced the overall interspersion and juxtaposition of the vegetation mosaic, with negative consequences on habitat capability for these two indicator species. This result with regards to current elk habitat may seem surprising given that current elk populations are thought to be substantially higher than historical populations. It is important to realize that we measured habitat characteristics, not population densities, and that other factors (e.g., predator pressure) also play an important role in determining actual population densities.


      Limitations.–Managers need to be cognizant of two important considerations when interpreting our HRV departure results. First, although it is clear that the current landscape structure is not within the modeled range of variability, the magnitude of the deviation is less clear. Specifically, inconsistencies in the spatial resolution of the initial cover type map (e.g., failure to delineate all small vegetation patches) may impose an artificial coarseness to the landscape structure that affects the computed values of most landscape metrics - at least the configuration metrics. In other words, the fine-grained heterogeneity in vegetation created by the disturbance processes in RMLANDS was probably not comparably represented in the Forest Service database due to human inconsistencies in mapping small vegetation patches. We took two precautions to guard against this problem. (1) In addition to analyzing the 25-m resolution maps of cover type and stand condition, we rescaled the output maps from RMLANDS to a 0.5-ha minimum mapping unit and analyzed these coarser-grained maps as well. (2) We included a number of area-weighted configuration metrics that are relatively insensitive to small patches. The rescaling was not completely effective in accounting for these discrepancies. Thus, we emphasized the area-weighted metrics when evaluating configuration departure. Note, the composition metrics (i.e., the percentage of the landscape in each class) are relatively immune to this issue. Given our reliance on the composition metrics and the emphasis we placed on the area-weighted configuration metrics, we feel that is safe to conclude that the current landscape structure is well outside the modeled range of variability. However, it is important to be aware that our reported HRV departure indices, except for the seral-stage departure index (class level) and landscape composition departure index (landscape level), are probably biased high (i.e., inflated).

      Second, any conclusions regarding HRV departure depend on an accurate mapping of stand conditions in the current landscape. We note two related problems in this respect. First, in the forested cover types, stand inventory data did not allow us to consistently and reliably discriminate between stands in the understory reinitiation and shifting mosaic stages. In particular, recorded stand ages were based on the age of the oldest trees in the stands, not the age since that last stand-replacing disturbance. By definition, the age since stand origin is always greater than the age of the oldest trees once the stand reaches the true shifting mosaic stage of development, but the size of the oldest (largest) trees may not be appreciably different between these stages. Thus, it is likely that a significant portion of the stands classified as being in the understory reinitiation condition are actually in the shifting mosaic stage of development. This bias would result in an inflated seral-stage departure index at the class level and an inflated landscape composition departure index at the landscape level. We took one precaution to guard against this problem. For purposes of HRV departure calculations, where appropriate, we combined the understory reinitiation and shifting mosaic stages into a combined late-seral stage and reported these results. Second, as noted previously, we altogether lack reliable age and stand condition data for several cover types. In particular, we have inadequate data for most non-forested types (e.g., mountain shrublands, mesic sagebrush, pinyon-juniper woodlands). Consequently, our initial assignment of stands to condition classes (seral stages) was based on interpolation from sparse data or on a random assignment based on seral-stage distributions estimated by local experts. In either case, we are not confident that our current condition estimates are accurate. Unfortunately, there was no way to guard against potentially spurious results in these cover types. Therefore, we did not report HRV departure results for these cover types. Note, the unreliability of the current condition for the non-forested cover types does not affect our simulated HRV distribution, as we accounted for an equilibration period in the model.

      Management Implications.–Despite the severe limitations imposed by available data, our simulations nonetheless indicate that the current landscape structure deviates substantially from its historic range of variability. In general, the current landscape is dominated by mid- to late-successional forest and lacks the fire-dependent stand conditions and spatial heterogeneity in vegetation that was maintained by natural disturbances during the reference period. This landscape condition appears to be largely a legacy of the last century of land management practices, in particular, fire exclusion (Romme et al. 2003). Indeed, Euro-American activities have altered the disturbance regime of many western forest landscapes, resulting in substantial changes in landscape structure and function (e.g., Baker 1992; Wallin et al. 1996; Baisan and Swetnam 1997; Agee 1999, McGarigal et al. 2001). In the southern Rocky Mountains these effects have been less ubiquitous and less straightforward in high-elevation landscapes than in low-elevation landscapes (Romme et al. 2003). Lower elevations have been subject to substantially altered disturbance regimes for more than a century (Romme et al. 2003, Swetnam and Baisan 1996). Despite apparently little change in the natural disturbance regime in the high-elevation landscapes (e.g., Romme and Despain 1989; Bessie and Johnson 1995; Weir et al. 1995; Schmid and Mata 1996), other human activities since the late 1800s have clearly altered disturbance regimes and landscape structure (Hejl et al. 1995; Miller et al. 1996; Reed et al. 1996a,b; Tinker et al. 1997). These activities are related mainly to timber harvest and to the extensive network of roads constructed to support timber harvest, fire control, and recreation. In addition to these ubiquitous human impacts, the generally benign climate of the twentieth century also was a significant reason for the lack of large, stand-replacing disturbances, either by fire or spruce beetle (Romme et al. 2003).

      Our findings are particularly interesting in light of increasing concern over anthropogenic habitat loss and fragmentation (Rochelle et al. 1999; Knight et al. 2000). Forest fragmentation has received considerable research attention in many regions of North America (e.g., Whitcomb et al. 1981; Robbins et al. 1989; Lehmkuhl and Ruggiero 1991; McGarigal and McComb 1995; Schmiegelow et al. 1997; Trzcinski et al. 1999; Villard et al. 1999). However, we are in the earliest stages of understanding the patterns, processes, and ecological significance of forest fragmentation in the southern Rocky Mountain region (Knight et al. 2000). It is not clear, for example, how the native biota responds to anthropogenic changes in landscape patterns caused by logging and road-building and disruption of natural disturbance regimes (e.g., fire suppression). This difficulty is exacerbated because Rocky Mountain landscapes are inherently very heterogeneous – a result of steep natural gradients in elevation, topography, and substrate – and forests in this region tend to be somewhat patchy even in the absence of human alterations (Hejl 1992).

      Based on our results, it might be tempting for managers to reach the simple conclusion that the landscape is less fragmented today than during the reference period. However, this conclusion is not as straightforward as it might seem for the following reasons. First, fragmentation is a landscape-level process in which a specific habitat is progressively sub-divided into smaller, geometrically altered, and more isolated fragments as a result of both natural and human activities, and this process involves changes in landscape composition, structure, and function at many scales and occurs on a backdrop of a natural patch mosaic created by changing landforms and natural disturbances (McGarigal and McComb 1999). Of critical importance is the fact that fragmentation occurs to a specific habitat type, not the entire landscape mosaic, even though it happens at the landscape scale. Thus, landscapes don’t get fragmented, specific habitats do. In our study, we evaluated the spatial pattern - and by implication, the fragmentation - of many different patch types (defined by unique combinations of cover type and stand condition). Many of these patch types are indeed less fragmented in the current landscape than they were under the simulated HRV. This is true in general for most of the late-seral forest patch types. However, not all patch types are less fragmented in the current landscape. For example, many of the early-seral forest patch types are in fact much more fragmented in the current landscape than they were under the simulated HRV. Thus, conclusions about habitat fragmentation in the current landscape must be qualified with specific reference to one or more well-defined habitats.

      Second, we evaluated vegetation patterns in the current landscape after excluding roads (i.e., we removed roads from the land cover map by filling in those areas with the abutting cover type), in order to be consistent with our simulation of landscape structure changes during the reference period. Yet, of all the novel kinds of disturbances that humans have introduced in the forests of the southern Rocky Mountains during the last century, roads may be the most ubiquitous and significant long-term legacy of our activities (Romme et al. 2003). Roads are unprecedented features in the ecological history of these landscapes (Forman 1995), and potentially affect many ecological processes (Forman and Alexander 1998; Trombulak and Frissell 2000). In particular, roads are linear landscape features that can create high-contrast edges and bisect patches. Consequently, roads can cause greater fragmentation of habitats than the direct loss of habitat from associated land use activities (Reed et al. 1996b; Tinker et al. 1997, McGarigal et al. 2001). Given the ubiquitous nature of roads and their disproportionate influence on landscape structure and function, any conclusions regarding departure in relation to habitat fragmentation that does not consider road impacts should be viewed with extreme caution.

      Our simulations indicate that returning the landscape structure to a condition that falls within the simulated HRV would likely be a difficult and long-term undertaking if it were deemed desirable. We deduced this from the time it took the current landscape to equilibrate to the reference-period disturbance regime. The equilibration period in many ways provides a direct measure of HRV departure; it is defined as the period required to return the initial landscape condition to a stable range of variation. It is a function of not only how far outside the stable range of variation the current landscape is, but also the speed at which disturbance and succession processes interact to affect a change in the landscape trajectory. Thus, we can infer that if management activities were designed to emulate natural disturbance processes, then it would take a length of time equal to the equilibration period to return the landscape to its HRV. In our simulations, most landscape structure metrics equilibrated within 100 years, although some metrics equilibrated faster and others slower. In particular, the configuration of the high-elevation conifer forest mosaic took considerably longer (up to 300 years) to equilibrate owing to the long return interval between disturbances and the relatively slow rate of stand development. It must be emphasized, however, that this does not imply that it should be our goal in management to recreate all of the ecological conditions and dynamics of the reference period. Complete achievement of such a goal would be impossible, given the climatic, cultural, and ecological changes that have occurred in the last century. Moreover, the extent and intensity of disturbance required to emulate the natural disturbance regime would be unacceptable socially, economically, and politically.

Effects of Scale and Context

      The pattern detected in any ecological mosaic is a function of scale, and the ecological concept of scale encompasses both extent and grain (Forman and Godron 1986; Turner et al. 1989; Wiens 1989; Moody and Woodcock 1995). Extent and grain define the upper and lower limits of resolution of a study and any inferences about scale-dependency in a system are constrained by the extent and grain of investigation (Wiens 1989). In the analysis of landscape change, spatial scale is defined by the minimum patch size (grain) and the geographic extent of the landscape; temporal scale is defined by the minimum (grain) and total (extent) period over which landscape change is assessed. We cannot detect patterns or changes in patterns beyond the extent or below the resolution of the grain. This has important implications pertaining to the interpretation of our findings.

      First, we chose to examine landscape structure dynamics at two spatial resolutions: (1) 0.0625-ha (25 m cell size) minimum mapping unit, and (2) 0.5-ha minimum mapping unit. Note, in the coarse-grained representation, the cell size was maintained at 25 m - only the minimum mapping unit was increased. The 0.5-ha resolution was also used to reclassify cover types and stand conditions into a slightly smaller set of aggregated classes in order to highlight habitats of special interest. We expected landscape composition estimates to be insensitive to spatial resolution and indeed this was the case. Surprisingly, the results pertaining to landscape configuration were largely insensitive to spatial resolution as well. In retrospect, it was apparent that small patches had a trivial impact on most configuration metrics and virtually no impact on the metrics selected for interpretation (i.e., area-weighted metrics). This is not to say that the fine-grained patterns of heterogeneity are not important ecologically, only that at the scale of the large landscape extents (10s-100s of thousands of hectares) we examined, the quantitative importance of the fine-grained patterns was dwarfed by the coarse-grained patterns created by the larger patches.

      Second, we designed RMLANDS to operate with a 10-year timestep. Thus, the minimum temporal resolution was fixed and we saw no reason to examine coarser resolutions. In addition, we established the temporal extent of our simulations based on our desire to capture and describe a stable range of variation in landscape structure. Preliminary trials determined that an 800-year simulation, after accounting for a 100-year equilibration period, was necessary to reliably estimate the range of variation in each landscape metric. Again, we deemed it inappropriate for our purposes to examine shorter simulations and it was statistically unnecessary to examine longer simulations. Thus, we did not vary the length of our simulations. However, it is important to recognize (and easy to demonstrate) that the measured range of variation in most metrics is sensitive to simulations shorter than some critical length (Figure-equilibration scale). The critical length varies among metrics and patch types, but it is safe to conclude that a minimum of 100-300 years is needed to capture the full range of variation in most metrics, and twice that long to confirm that it is stable. Thus, a management strategy designed to emulate the natural disturbance regime would take 100-300 years to see the landscape fluctuate through its full range of conditions. This is a humbling thought given that most professional careers last no more than 30 years - a blip on the scale of landscape dynamics - and that most policies are geared toward 10- to 20-year planning horizons.

      Third, when we examined progressively smaller spatial units of the entire simulated landscape (Figure-map), temporal variability increased – as would be expected (Turner et al. 1993), and there was an apparent threshold in the relationship between landscape extent and temporal variability (Figure-scale-land). Specifically, the magnitude of variability in landscape composition increased only modestly as the landscape extent decreased from the forest scale (659,246 ha) to the quadrant scale (average = 164,812 ha), but increased dramatically as the landscape extent decreased to the watershed scale (average = 37,929 ha). A similar relationship was evident for three-toed woodpecker habitat capability, the only species for which we were able to complete the habitat capability analysis at all three scales (Figure-scale-wildlife). In addition, each of the sub-landscapes we examined was somewhat unique in its absolute range of variability in landscape structure. Not surprisingly, uniqueness was greatest for the smaller landscapes extents (watersheds). Interestingly, despite the importance of landscape extent and context on the measured range of variability, the degree of departure of the current landscape from the simulated HRV was relatively invariant to scale and context.

      These results have important management implications. First, they demonstrate that no two landscapes in this mountainous region are identical; each has more or less unique characteristics of topography, vegetation, etc. that affect its dynamical behavior. Consequently, there is no one correct scale for assessing HRV. Second, while no one scale is necessarily more correct than another, our results do suggest that some scales may be more appropriate than others for characterizing HRV. Specifically, our results show that at extents larger than the quadrant scale, the relative variability in landscape structure does not change much, but that at smaller extents, the variability increases dramatically. We interpret this to mean that at the quadrant extent (and larger), the landscape is large enough to fully incorporate the disturbance regime and exhibit stable dynamical behavior. At this scale, our simulated system falls within the ‘stable, high variance’ portion of the state-space model developed by Turner et al. (1993). At increasingly smaller extents, the size of the largest disturbance events approaches and eventually exceeds the size of the landscape, producing major fluctuations in landscape structure. Ultimately, at even smaller extents the range of variability in landscape structure becomes so great as to be meaningless. Thus, we conclude that under the simulated disturbance regime, characterizing HRV is best done at the quadrant or forest scale. Lastly, although the relative degree of dynamism is apparently stable at extents larger than the quadrant scale, the absolute range of variation in landscape structure can vary among quadrants by more than 50%. Thus, each district exhibits a slightly different absolute range of variation in landscape structure. Most of these differences can be attributed to differences in landscape composition (Table-areal coverage). The challenge to managers is in deciding whether to give explicit recognition to the these differences when establishing management direction, or to subsume these difference at the forest level on pragmatic grounds. We believe that it is probably sufficient to characterize HRV at the forest scale for purposes of general communication, but that it would be wise if possible to use the quadrant-specific HRV results when setting management targets. In this manner, the spatial variation in ecological patterns and dynamics across the forest are given explicit consideration.

Concluding Remarks

      In closing, it is important to remember that our simulation study was intended to complement the detailed landscape condition analysis completed for the South Central Highlands Section of southwestern Colorado and northwestern New Mexico (Romme et al. 2003) and extend it to the lower elevation cover types. Our study provides a detailed quantitative analysis of the simulated vegetation dynamics under the historic reference period that complements the detailed, but qualitative, landscape condition assessment of the previous report. Overall, our findings are in complete qualitative agreement with the previous assessment. In addition to enhancing our general understanding of landscape dynamics, our HRV results are of paramount use as a reference or benchmark for comparison with alternative future land management scenarios - the focus of the next phase of this project.

      As with any study, our results and conclusions must be interpreted within the scope and limitations of this study. In particular, our analyses were designed to simulate vegetation dynamics under a specific historic reference period. We chose the period from about 1300 to the late 1800s, representing the period from Anasazi abandonment to EuroAmerican settlement as the reference period (often referred to as the period of indigenous settlement). Thus, our results pertain to landscape conditions during that period. More importantly, our results are based on a simulation model (RMLANDS), and this model, like any model, is an abstract and simplified representation of reality. Given the design limits of this model and the challenges of parameterizing a complex model like this, our results should not be interpreted as “golden”. Rather, they should be used to help identify the most influential factors driving landscape change, identify critical empirical information needs, identify interesting system behavior (e.g., thresholds), identify the limits of our understanding, and help us to explore “what if” scenarios.

Literature Cited