**1. Introduction**

Climate change represents a globally pervasive stress on natural ecosystems. Temperature and precipitation regimes drive ecosystem productivity and natural dynamics, such as the rate of plant growth, the frequency of natural wildfire, and seasonal streamflow [1]. Paleoecological research has shown that past episodes of climate change triggered transformation of natural communities at regional and local scales with varying speed and magnitude [2,3]. As the rate of climate change increases, substantial shifts in key ecological processes are likely to cascade through natural communities, resulting in altered productivity, change in species composition, local extinctions, and many instances of ecological degradation or collapse [4].

Conservation practitioners often lack a su fficient understanding of the many linkages between changing climate and key ecological processes. Nor do they fully understand the many interactions of climate-induced stress with other ecological stressors, such as those tied to land use, which may have already reduced the resiliency of many natural communities [5]. However, because of the controlling link between climate and many ecological processes, and the individualistic responses of component

species, natural communities could transform in unprecedented ways [6,7]. Therefore, in any given place, a transparent assessment of climate change vulnerability for natural communities is needed to help quantify risk of ecological degradation or collapse.

A repeatable and transparent index of climate change vulnerability designed for natural communities helps to determine those types that, in all or part of their distribution, are most at risk of climate change impacts. It can provide an early warning of elevated risk for associated species, and a baseline for developing scientifically grounded, ecosystem-based strategies for climate change adaptation. This Habitat Climate Change Vulnerability Index (HCCVI) presented in this paper integrates variables from other assessments and results in products that support practical decision making and communication.

#### *Vulnerability Assessment at Di*ff*erent Levels of Ecological Organization*

Climate change vulnerability assessments may address di fferent levels of ecological organization, such as species, communities, or landscapes. The species level is the most common focus for vulnerability assessment and consequently has received extensive attention in the literature [8–11]. Based in autecology, trait-based approaches examine projected climate change where the species occurs, aspects of the genetic variation, natural history, physiology, and landscape context to assess sensitivity and adaptive capacity [12].

Assessments of landscapes often produce spatially explicit results for interpretation at regional scales. Evaluation of exposure may result in maps showing where climate stress is indicated to be greatest, whereas examination of the potential climate-change e ffects on disturbance regimes or invasive species can address aspects of sensitivity [13–15]. Adaptive capacity can be measured through examination of the heterogeneity of topography, moisture gradients, or microclimates under the assumption that more diverse landscapes provide more opportunities for organisms to find climate refugia than homogeneous ones [16].

Assessing the vulnerability of natural community types can provide a useful complement to both landscape and species assessments. Whereas landscape assessments indicate a high potential for climate-change impacts in certain regions, analysis of component communities is based on synecology and aims to more directly measure how climate change will impact species assemblages, ecological processes, structure and function, and is a next logical step to identify practical adaptation strategies where local managemen<sup>t</sup> of vegetation is a common form of resource managemen<sup>t</sup> [7].

Few examples exist for assessing vulnerability at the community level of organization. In two examples, they assessed vulnerability of European forests [17], and coastal communities [18], but both took conceptual approaches based on these socio-ecological systems at continental scales. They were concerned mainly about ecosystem outputs of goods and services and aiming to inform societal responses in forestry and fisheries sectors.

In one recent example focusing more squarely at natural community types themselves [19], 31 major vegetation types in California were assessed considering climate projections by the year 2100, sensitivity and adaptive capacity estimates of dominant species for each type, and spatial disruption that might inhibit species movement over time. They used a basin characterization model [20] to express climate change exposure. Working at a global scale, assessment for terrestrial ecosystems [21] emphasized the need to address all three factors of exposure, sensitivity, and adaptive capacity.

Here, we demonstrate a practical framework for climate change vulnerability assessment, focusing on natural community types themselves. This current assessment builds upon prior e fforts [22] to address major vegetation types, such as sagebrush shrubland and pinyon-juniper woodlands [23], with an initial focus on types that dominate lands managed by the Bureau of Land Management (BLM) in the conterminous Western US. Several of these types also have extensive distributions in neighboring Mexico or Canada. Below we explain our methods with the outputs of each step illustrated with one common sagebrush shrubland type, and then summarize findings and discuss implications for all 52 types that extend over 3.2 million km<sup>2</sup> of Western North America.

We integrate measures of climate exposure with a series of measures for ecosystem resilience. Our intent is to complete analysis of terrestrial ecosystems or natural community types conceptualized at relatively local scales. Our approach facilitates the accumulation of results for many ecosystem types occurring across countries and continents. Results from this form of ecologically based analysis can then be applied to climate change adaptation in a variety of socio-ecological contexts across the range of distribution of the natural community type.

#### **2. Materials and Methods**

#### *2.1. Analytical Framework for Vulnerability Assessment*

This index approach to vulnerability assessment aims to organize a series of sub-analyses in a coherent structure that will shed light on distinct components of vulnerability, so that each can be evaluated individually, or in combination. Our approach parallels related indexing of climate change vulnerabilities for species [24]. The components of climate change vulnerability are organized into primary categories of Exposure and Resilience. Resilience is further subdivided into subcategories of Sensitivity and Adaptive Capacity (Figure 1). For the HCCVI, these terms are defined as follows:

**Figure 1.** Analytical framework for the habitat climate change vulnerability index.

Exposure refers to the rate, magnitude, and nature of climate-induced stress on the community. Exposure encompasses trends in climate, such as changes in temperature and precipitation regimes, and any predicted e ffects on ecosystem-specific processes. Analyses of exposure consider change in climate variables themselves, or if possible, their resulting e ffects that cause increasing ecosystem stress, changing dynamic processes such as wildfire or hydrological regimes, or coastal dynamics. This definition of exposure aligns closely with other common applications to climate change vulnerability assessments where changing climate is viewed as an extrinsic driver or pressure on the system of interest [12].

Resilience encompasses intrinsic factors that are likely to a ffect ecological responses of the natural community to changing climate. These include factors commonly described for ecological resilience, as defined by Holling [25] and Gunderson [26]. Walker et al. [27] defined ecological resilience as "the capacity of a system to absorb disturbance and reorganize while undergoing change and while still retaining essentially the same function, structure, identity, and feedbacks." Under the broader measure of resilience, we include subcategories of Sensitivity and Adaptive Capacity.

• *Sensitivity* focuses on ecosystem stressors that are likely to a ffect ecological responses of the natural community to climate change. These emphasize human alterations to characteristic patterns and process that can be readily measured across the range of the community type, such as landscape

fragmentation, e ffects of invasive species, or human alterations to other dynamic processes. These alterations are considered independent of climate change, but once identified, are likely to interact with changing climate.

• *Adaptive Capacity* includes natural characteristics that a ffect the potential for a natural community to cope with climate change. Analyses might consider the natural geophysical variability in climate for the type's distribution. They might also consider functional roles that characteristic species play, such as the relative vulnerabilities to climate change of individual species that provide "keystone" functions, and relative taxonomic diversity within key functional groups (e.g., C3- vs. C4-dominated communities) that characterize the type.

These definitions di ffer in part from those provided by the IPCC because here we emphasize synecology. Whereas sensitivity in species vulnerability is based more strongly on life history characteristics, natural communities encompass multiple species, each with di ffering tolerances. Here our sensitivity definition aims to better account for human alterations that a ffect the community composition and response to changing climate. Our adaptive capacity definition emphasizes natural characteristics of the community type—both biotic (species composition) and abiotic (geophysical), that may (or may not) contribute to resilience in the face of changing climate.

Drawing inspiration from Magness et al. [28] among others in structuring analyses with a logic model, the index scores combine information in two stages, with the first analyses gauging relative climate change resilience by combining scores from intrinsic factors of sensitivity and adaptive capacity. Climate change exposure and resilience are then considered together (combining extrinsic and intrinsic factors) to arrive at an overall gauge of climate change vulnerability (Figure 1).

The HCCVI uses component indicator values to ultimately arrive at a four-level series of scores, i.e., Very High, High, Moderate, and Low vulnerability (Figure 1). These can be derived from relative measures of both Resilience and Exposure. When using quantitative data for measurement, numerical scores are normalized to a 0.0 to 1 scale, with 0.0 indicating ecologically "least favorable" conditions, and 1 indicating "most favorable" conditions. Quartiles of each continuous measure may be used as a starting point to determine the range falling into each of the Very High–Low categories (e.g., ≥0.75 = Low, 0.5–0.75 = Moderate, 0.25–0.50 = High, and ≤0.25 = Very High overall vulnerability). In this application of the framework, all indicators are weighted equally, and we used an arithmetic mean for their combination. Di fferent circumstances and datasets for component measures, or di fferent needs for reporting, may sugges<sup>t</sup> alternative weightings or thresholds to delineate these categories. See Tonmoy et al. [29] for a review of issues associated with indicator treatment in vulnerability assessments. For this framework, we emphasize the need to standardize indicator data in some form to enable combination with other component analysis results.

*Very High climate change vulnerability* results from combining high exposure with low resilience (i.e., both trending toward "least favorable" scores). These are circumstances where climate change stress and its e ffects are expected to be most severe, and relative resilience is lowest. Ecosystem transformation is most likely to occur in these types.

*High climate change vulnerability* results from combining either high or moderate exposure with low or medium resilience. Under either combination, climate change stress is anticipated to have considerable impact.

*Moderate climate change vulnerability* results from a variety of combinations for exposure and resilience; initially with circumstances where both are scored as moderate. However, this also results where resilience is scored high, if combined with either high or medium exposure. Where both resilience and exposure are low, some degree of climate change vulnerability remains.

*Low climate change vulnerability* results from combining low exposure with high resilience (i.e., both trending toward "most favorable" scores). These are circumstances where climate change stress and its e ffects are expected to be least severe or absent, and relative resilience is highest.

#### *2.2. Spatial and Temporal Dimensions for Documenting Vulnerability*

Climate change vulnerability assessments need to be placed within explicit spatial bounds. For this effort, we summarized component measurements by 100 km<sup>2</sup> hexagon for the distribution of each vegetation type, and then further summarized results within Level III of the Commission for Economic Cooperation (CEC) ecoregions [30] for each vegetation type (Figure 2). These ecoregions provide an appropriate and consistent spatial structure to systematically document climate change vulnerability at national or regional scales.

**Figure 2.** Mapped distributions of 52 ecological system types (too numerous to list here) in Western North America, centered on the USA, with boundaries of the Commission for Economic Cooperation CEC ecoregions [30] used for reporting on vulnerability.

Similarly, climate change vulnerability assessments require a temporal dimension because the magnitude of component measures could vary over time. For this effort, we based initial vulnerability measures on emerging or "current" climate trends using observed climate data (as of 2014). Given both uncertainties associated with climate projections, and the need to apply vulnerability assessments

to decisions affecting resources in upcoming decades, climate projections over the upcoming 50-year timeframe (e.g., between 2020 and 2070) provide subsequent realistic timeframes where climate trends can be estimated within acceptable bounds of uncertainty. Therefore, we also assess vulnerability with the mid-21st century timeframe (2040–2069).

#### *2.3. Ecological Classification and Distribution*

For this project, we used NatureServe's terrestrial ecological systems classification to define types [31]. The advantage of using this classification system is that it represents an established classification of several hundred upland and wetland types that have been extensively described and mapped by US federal and state resource managers [32,33] and extended into adjacent Canada, Mexico, and across Latin America and the Caribbean [34]. The expected pre-Columbian, or "potential", distribution of each type, mapped at 90 m pixel resolution, was used as the base distribution for assessment (Figure 2). Descriptions of each type can be found at http://explorer.natureserve.org/. We completed additional literature review to further document each type with current knowledge of key ecological processes and stressors, and potential measures for assessing climate change resilience. We completed literature searches using individual communities/systems, common stressors, and functional species groups (e.g., "cool season") as key words using Google Scholar and Colorado State University Library Prospector.

Below, we discuss measures for climate change exposure and resilience applied to each type for this project. Appendix A includes a more detailed explanation of component measures used for ecosystem resilience. For purposes of illustrating our methodology, we will use one example—Intermountain Basins Big Sagebrush Shrubland to depict component steps of the HCCVI framework (Figures 3–7).

**Figure 3.** Change in Climate Suitability estimate for 2040–2070 timeframe for Intermountain Basins Big Sagebrush Shrubland.

**Figure 4.** Climate Exposure estimate for 2040–2070 timeframe as index scores (i.e., low–high climate stress) summarized by 100 km<sup>2</sup> hexagon for Intermountain Basins Big Sagebrush Shrubland.

**Figure 5.** Sensitivity measure of landscape condition and index scores summarized by 100 km<sup>2</sup> hexagon for Intermountain Basins Big Sagebrush Shrubland.

**Figure 6.** Overall Resilience measure summarized by 100 km<sup>2</sup> hexagon for Intermountain Basins Big Sagebrush Shrubland.

**Figure 7.** Overall Climate Change Vulnerability estimate for 2040–2070 summarized by 100 km<sup>2</sup> hexagon for Intermountain Basins Big Sagebrush Shrubland.

#### *2.4. Climate Change Exposure*

We characterized the baseline climate niche for each vegetation type using historical climate data for the mid-20th century and the potential historical distribution of the type (detail below). This provides a baseline of suitable conditions from which to compare climate trends from subsequent time periods to clarify the significance of measurable change. The entire potential distribution for each community type was used to ensure that the entire range of possible suitable climate conditions was represented for each type. Using distributions that have been affected by land conversion could skew how we define climate suitability for a vegetation type and hence, how we measure climate exposure. Kling et al. [35] provide additional detail on component model design and performance.

For every grid cell of each vegetation type we calculated a composite index of climate change exposure as the sum of two distinct exposure measures: *suitability change*, which quantifies departure from the historical range of spatial climate variability across the geographic range of that vegetation type, and *typicality*, which quantifies departure from the historical range of year-to-year climate variability at a given pixel location. Each component index ranges from 0.0 to 0.5; these are summed to derive the final combined exposure index ranging from 0 (low climate exposure) to 1 (large climate exposure).

Exposure measures were calculated based on changes in 19 bioclimatic variables derived from monthly temperature and precipitation variables [36]. Using a historical baseline period representing typical 20th century climate (1948–1980), we estimated exposure for both recent observed climate change (1981–2014) and projected future change (2040–2069, RCP 4.5). While related regional analyses investigated differences between RCP 4.5 and 8.5 [37], and our methodology could accommodate either here we utilized just RCP 4.5. These calculations were initiated in 2015, and so at that time 2014 was the most recent climate data available. To maximize data quality across the full spatiotemporal extent of our analysis, we combined gridded climate datasets from several sources. Within the contiguous US, for the baseline and recent periods, temperature data were sourced from TopoWx [38] whereas precipitation data were sourced from PRISM version LT71 [39], both at 810 m resolution. Outside the US, and for the future within the US, all variables were based on 1 km<sup>2</sup> data from ClimateNA [40].

The *suitability change* metric uses niche modeling to estimate changes in a site's climatic suitability for a given vegetation type for a given timeframe. For each vegetation type we fit a random forest [41] model using baseline means of the six most important (of 19) bioclimatic variables and 1,000 presence and absence locations sampled from within the rectangular bounding box of the type's geographic extent. The top six di ffer among types, but six variables consistently explained a high proportion of model variability. Our choice of RF over alternative Generalized Linear Models (GLM), Generalized Additive Models (GAM), and MaxEnt algorithms, as well as our selection of variables for each type, was based on extensive model performance testing using spatial block cross-validation [42] to control the overfitting that comes from spatial non-independence. Variables were selected using recursive feature elimination [43], with six judged as the best empirical tradeo ff between performance and parsimony. The fitted models were used to predict suitability for baseline, current, and future periods. Suitability change was then calculated by subtracting suitability measures across these time periods, and then rescaling these di fferences into a 0–0.5 index with 0 representing increasing or unchanged suitability, and 0.5 representing a large decrease in suitability.

The *typicality* metric compares a site's current or future mean climate to the mean and yearly variation during a baseline period. For a given climate variable, typicality is calculated as the proportion of baseline years whose absolute deviation from the baseline mean (e.g., in degrees C) is larger than the absolute deviation of the recent mean from the baseline mean. A value of 0 indicates no change, while a value of 1 indicates that the recent or future mean climate is more extreme, in either direction, than any individual year in the baseline period. For each vegetation type we calculated typicality for each of the same six climate variables used in the RF models, and then averaged them to derive a final typicality index. Final typicality values were divided by two to rescale them to the 0.0 to 0.5 range.

Typicality (0.0 to 0.5) and Change in Climate Suitability (0.0 to 0.5) measures were then added together to produce an overall exposure measure, resulting in per pixel values within a 0.0–1.0 range, with 1.0 indicating high climate exposure and 0.0 indicating no measurable climate exposure. These per-pixel values were subsequently inverted and averaged for one measure per 100 km<sup>2</sup> hexagon. Figure 3 illustrates one example of mid-21st century estimates of increasing or decreasing climate suitability for one major big sagebrush vegetation type. One can see where most substantial decreases in climate suitability for this type are concentrated in such places as Southeastern Washington, the Snake River Plain of Southern Idaho, Northwestern Nevada, the Big Horn Basin of Wyoming, and the Uinta Basin of Utah.

Figure 4 depicts, for the same type, the result of combining typicality and change in suitability estimates, and then summarizing those per-pixel values to 100 km<sup>2</sup> hexagons. Darker blue areas are forecasted to be least stressed (closer to 1) and yellow areas most stressed (closer to 0.0). This image indicates where a substantial proportion of the range-wide extent scores into the high–very high range of climate exposure.

#### *2.5. Resilience—Ecosystem Sensitivity and Adaptive Capacity*

Our measures of Resilience address predisposing conditions—such as extant ecosystem stressors, or natural abiotic or biotic characteristics of the type—that are likely to a ffect ecological responses of the natural community to changing climate. For example, if exposure measures indicate the need for component species to migrate toward other elevations or latitudes, but the natural landscape is fragmented by intensive land uses, the relative vulnerability of community types in that fragmented landscape could increase [44]. Similarly, the introduction of non-native species may displace native species and/or alter key dynamic processes such as wildfire regimes [45], and both could be exacerbated by climate change. These factors would describe relative climate change sensitivity for a natural

community type. This differs in approach from species vulnerability assessments, in that they tend to focus on life-history traits of individual taxa.

Inherent adaptive capacity of natural communities could consider the natural geophysical variability in climate for the type's distribution or the functional roles of species in the community type. Again, climate exposure might indicate a high level of climate stress, natural communities occurring within topographically flat landscapes with potential to retain little variability in microclimates could lack any built-in buffer effects for component species. Likewise, if certain ecosystem functions, such as nitrogen fixation, are limited to one or a few species that characterize the natural community, this could bring additional vulnerability due to the potential for their extirpation over time.

Upon completion of a literature review of each type, and evaluation of available spatial data, we selected four primary indicators suitable for measuring relative sensitivity. To address effects of landscape fragmentation, we used a spatial model for landscape intactness or condition. Since many assessed types are known to be affected by invasive annual grass invasion, a model aiming to measure relative invasion severity was selected. Similarly, since most assessed types have a characteristic natural wildfire regime, a spatial model estimating fire regime departure was used. For forest types, measures of elevated risk from insects or diseases were identified. Three measures of adaptive capacity include both biotic and abiotic factors. Biotics factors included scoring of each type for diversity within identified key functional species groups and relative vulnerability of any identified "keystone" species. One measure of topo-climatic variability was applied for the distribution of each type.

In each of these cases involving spatial models, the model was overlain with the distribution of each vegetation type and scores were transformed to indicate a relative degree of sensitivity or adaptive capacity within a 0.0–1 range, again with 0.0 indicating most severely impacted, or least favorable, conditions while 1 indicating highest integrity, or apparently unaltered, conditions. These scores were each summarized to average values per 100 km<sup>2</sup> hexagon. Figure 5 depicts results from one sensitivity measure of landscape condition for Intermountain Big Sagebrush Shrubland. Darker blue areas indicate apparently least fragmented areas and yellow areas most fragmented. One can see where in substantial portions of this distribution, such as in Eastern Washington, across the Snake River Plain, and south along the Wasatch Front of Utah, this sagebrush type occurs in landscape highly fragmented by intensive land uses.

Again, Appendix A includes a detailed explanation of component measures used for ecosystem resilience.

#### *2.6. Overall Resilience Scores*

Overall resilience scores were derived by averaging results for each measure of Sensitivity and for Adaptive Capacity (Figure 1). This combination of up to four Sensitivity measures and up to three Adaptive Capacity measures were averaged together per 100 km<sup>2</sup> hexagon to establish an overall score for resilience. As an example, Figure 6 depicts this overall measure for Intermountain Basins Big Sagebrush Shrubland. One can again see the general pattern of moderate to low resilience in the regional landscapes—such as the Snake River Plain of Southern Idaho—where most intensive land use and effects of invasive plant species are concentrated.

#### *2.7. Overall Climate Change Vulnerability*

As noted in Figure 1, the combination of climate change exposure with resilience scores results in the relative vulnerability estimate for a given timeframe. Figure 7 depicts this result as projected for the mid-21st century for Intermountain Basins Big Sagebrush Shrubland. Patterns of vulnerability in this type vary across its distribution, as depicted with 100 km<sup>2</sup> hexagons, and are driven strongly by climate change exposure measures, but are also substantially influenced by component measures of resilience (e.g., Figure 4). While per hexagon outputs are summarized along the 0.0–1 continuum, summary statistics for climate change vulnerability may be desirable with categories expressed as "Very High" "High" Moderate" or "Low" (Figure 1). Here, we used default break-points with quartiles

of each continuous measure to determine the range falling into each of the Very High-Low categories (≥0.75 = Low, 0.5–0.75 = Moderate, 0.25–0.50 = High, and ≤0.25 = Very High overall vulnerability). In the case of the Intermountain Basins Big Sagebrush Shrubland, nearly all its distribution is forecasted to fall within the "Moderate" vulnerability as of 2014 and in the "High" range of vulnerability by the mid-21st century.
