Using remote sensing to quantify ecosystem site potential community structure and deviation in the Great Basin, United States

Ecological Indicators 96 (2019) 516–531

Contents lists available at ScienceDirect

Ecological Indicators journal homepage: www.elsevier.com/locate/ecolind

Original Articles

Using remote sensing to quantify ecosystem site potential community structure and deviation in the Great Basin, United States

T

⁎

Matthew Riggea, , Collin Homerb, Bruce Wylieb, Yingxin Gua, Hua Shia, George Xianb, Debra K. Meyerc, Brett Bundec a

InuTeq/U.S. Geological Survey (USGS) Earth Resources Observation and Science Center, Sioux Falls, SD 57198, USA U.S. Geological Survey (USGS) Earth Resources Observation and Science Center, Sioux Falls, SD 57198, USA c SGT/USGS Earth Resources Observation and Science Center, Sioux Falls, SD 57198, USA b

A R T I C LE I N FO

A B S T R A C T

Keywords: Remote sensing Landsat Site potential Great Basin Land health Fractional cover

The semi-arid Great Basin region in the Northwest U.S. is impacted by a suite of change agents including fire, grazing, and climate variability to which native vegetation can have low resilience and resistance. Assessing ecosystem condition in relation to these change agents is difficult due to a lack of a consistent and objective Site Potential (SP) information of the conditions biophysically possible at each site. Our objectives were to assess and quantify patterns in ecosystem condition, based on actual fractional component cover and a SP map and to evaluate drivers of change. We used long-term 90th percentile Landsat NDVI (Normalized Difference Vegetation Index) and biophysical variables to produce a map of SP. Ecosystem condition was assessed using two methods, first we integrated fractional components into an index which was regressed against SP. Regression confidence intervals were used to segment the study area into normal, over-, and under-performing relative to SP. Next, the relationships between SP and fractional component cover produced SP expected component cover, from which we mapped the actual cover deviation. Much of the study area is within the range of conditions expected by the SP model, but degraded conditions are more common than those above SP expectations. We found that shrub cover deviation is more positive at higher elevation, while herbaceous cover deviation has the opposite pattern, supporting the hypothesis that more resistant and resilient sites are less likely to change from the shrub dominated legacy. Another key finding was that regions with significant annual herbaceous invasions tend to have lower than expected bare ground and shrub cover.

1. Introduction Land degradation in arid and semiarid environments is often related to historical legacies, environmental variables, edaphic properties, and physical perturbations (Peters et al., 2006). Sagebrush (Artemisia tridentata) dominated ecosystems are widespread across semiarid portions of western North America (Schlaepfer et al., 2014). Many of these ecosystems have been degraded through complex perturbations including invasive species, fire dynamics, woodland expansion, and land use practices (Wisdom et al., 2005; Chambers et al., 2014a; Stringham et al., 2016), with only 10% estimated to be in a condition unaltered by human disturbance (West, 1999). Some sites are near ecosystem thresholds, and are vulnerable to transitioning to a new steady state (Wisdom et al., 2005). Spatial understanding of which areas are near ecosystem thresholds is critical to avoid shifts to an undesirable state at which point restoration practices have been only modestly successful

⁎

(Schlaepfer et al., 2014). Identifying healthy plant communities on the other hand, allows better prioritization of management resources and facilitates understanding of which practices and environmental conditions lead to that state. Good understanding of ecosystem condition over large areas is invaluable for managing and understanding these resources. However, monitoring rangelands has often depended on personal judgement, and collecting objective data over vast landscapes has historically been expensive and difficult (Booth and Tueller, 2003). Condition interpretation requires information on the spatial patterns of reference condition (Bestelmeyer et al., 2009; Maestas et al., 2016). Further, large temporal and spatial variations in weather conditions can obscure the underlying effects of management and long-term climate change (Wylie et al., 2012). State and Transition Models (STM) are a useful concept to describe potential plant communities within an ecological site, based on soils and climate that possess a similar capacity to respond to

Corresponding author. E-mail address: [email protected] (M. Rigge).

https://doi.org/10.1016/j.ecolind.2018.09.037 Received 31 May 2018; Received in revised form 22 August 2018; Accepted 17 September 2018 1470-160X/ Published by Elsevier Ltd.

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

Fig. 1. Overview of study area location and features. A) Level III ecoregions superimposed and Site Potential (SP) model training extent on a summer 2017 Landsat 8 image composite, B) elevation, C) land management; abbreviations: BIA; Bureau of Indian Affairs, BLM; Bureau of Land Management, DOD; Department of Defense, USFWS; United States Fish and Wildlife Service, NGO; Non-Governmental Organization, USFS; United States Forest Service.

comparing actual measured conditions using remote sensing to those expected by a modelled site potential condition. A key component of this research determines the long-term average ecosystem site potential, which represents the ecosystem potential biomass production in an average year in a non-degraded or disturbed state (Wylie et al., 2012; Rigge et al., 2013b). Growing Season averaged NDVI (Normalized Difference Vegetation Index), or GSN, served as a proxy of biomass production. Site Potential (hereafter; SP) was calculated using a Cubist regression tree model trained on intact sites, using multiple biophysical variables as inputs. The goal of SP is to approximate the historical climax through a GSN map produced using a combination of biophysical data and a long-term archive of satellite imagery developed at 30 m resolution. Remote sensing offers new approaches to quantifying and

management and disturbance. Composition, cover, and production for communities of varying states within an ecological site are provided in a STM. STM describe the disturbance or management pathways among states and documents shifts that are difficult to reverse (Bestelmeyer et al., 2009; Bestelmeyer et al., 2011). Modelling variation in potential vegetation is difficult, however, so the reference state for STMs are often based on extrapolations from similar systems of current well managed areas. STM have not yet been completed for much of the western U.S., and their spatial scale sometimes prevents adoption by federal land managers (Stringham et al., 2016). Wylie et al. (2008, 2012) developed an approach that separates yearly weather influence on ecosystem condition from that of management, disturbance, or long-term climate changes. This approach evaluates ecosystem performance anomalies in biomass productivity by 517

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

SP Model Independent Variables Landsat growing season 90th percenƟle NDVI from 1985-2016 Compound topographic index, steep slopes, POLARIS soils database organic maƩer and available water capacity

Resample data from 30 m to 250 m

SP Model Dependent Variable

FracƟonal cover maps

MODIS growing season integrated NDVI median value from 2000-2015

(shrub, sagebrush, herbaceous, annual herbaceous)

SP model, developed at 250 m resoluƟon, using regression tree

Final SP, apply the model to 30 m independent data

IdenƟfy SP model training locaƟons

ASI, actual score index based on amount and type of fracƟonal cover

Legend Data Regress ASI on SP

Regress each fracƟonal component on SP

Site potenƟal deviaƟon, based on ASI deviaƟon from regression

Component deviaƟon, based on component deviaƟon from regression

250 m Process

30 m Process Data Flow

Fig. 2. Data and processing flow diagram.

term weather patterns, resulting in differing carbon fluxes and phenology (Rigge et al., 2013b). For this research, our objectives are to: 1) convert fractional cover component maps to ASI and produce a SP map, 2) assess and quantify patterns in ecosystem condition across the Great Basin based on pixel deviation from the SP to ASI relationship, segmenting deviation based on regression confidence into normal, over-, and under-performing, 3) use the SP relationships with individual fractional components to map deviation in cover from expected conditions, and 4) evaluate the contribution of change drivers, including fire, climate, and vegetation treatments which are all capable of shifting conditions from SP. Our approach offers an objective and synoptic view of ecosystem condition that is only available using remote sensing in readily understood units of fractional cover. Our data can be used to aid in the identification of degraded states, areas near transition, and those in good condition at a regional scale.

understanding rangeland condition (e.g. Homer et al., 2012; Homer et al., 2013). Fractional cover components capable of characterizing rangeland condition have been produced for the entire floristic Great Basin (Xian et al., 2015). More comprehensive association with rangeland conditions such as degradation and performance can be challenging, since component cover naturally varies based on site characteristics of annual precipitation, topography, and edaphic factors. For example, the amount of bare ground is a commonly used indicator of degradation in rangelands (e.g. Augustine et al., 2012; Weber et al., 2010). However, bare ground varies irrespective of anthropogenic influence, confounding the interpretation of degradation. For our analysis of rangeland condition we selected four components; annual herbaceous cover, perennial herbaceous cover, sagebrush cover, and non-sagebrush shrub cover (Xian et al., 2015). These components were synthesized into a single Actual Score Index (hereafter; ASI), which reflects current community composition/structure. The four selected components are critical to interpreting the health of sagebrush steppe and are responsive to grazing intensity, fire history, climate change, and other perturbations (Augustine et al., 2012; Chambers et al., 2014a; Hanna and Fulgham, 2015). Integration of the ASI in a SP framework provides a spatially explicit understanding of deviations from expected community structure. Pixels with a significant departure from conditions expected by the SP provide increased evidence that on-the-ground conditions are anomalous (Wylie et al., 2012). Deviation from SP is usually related to differences in grazing intensity, other management practices, disturbance, and long-

2. Study area We analyzed portions of the floristic Great Basin with completed USGS fractional shrubland mapping products (Xian et al., 2015) available at: https://www.mrlc.gov/gbs_shrub.php. The study area totals 468,797 km2, including the majority of the Central Basin and Range, Northern Basin and Range, Snake River Plain EPA Level III ecoregions and portions of the Blue Mountains, Eastern Cascades Slopes and Foothills, Middle Rockies, Idaho Batholith, and Mojave Basin and Range (Fig. 1A). Elevation widely varies throughout the study area, 518

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

ranging from −78 m to 4338 m, with a mean of 1638 m (Fig. 1B). Annual precipitation ranges from 1.4 to 140 cm, averaging 29.5 cm based on 1981–2015 Daymet climate data (Thornton et al., 2014). Annual mean temperatures vary from −1.8 to 23.7 C, with a mean of 8.4 C. Much of the land within the study area is managed by the Bureau of Land Management (BLM) and United States Forest Service (USFS) (Fig. 1C). Vegetation is typified by sagebrush steppe in the northern portion of the study area, with an overstory of Wyoming big sagebrush (Artemisia tridentata Nutt. Ssp. wyomingensis Beetle & Young), and an understory including Sandberg bluegrass (Poa secunda J. Presl), basin wildrye (Leymus cinereus Scribn. & Merr.), and crested wheatgrass (Agropyron cristatum L.). Cheatgrass (Bromus tectorum L.) is also abundant, especially in burned areas and in the Snake River Plain. Further south and at lower elevations the graminoid understory becomes sparser and rabbitbrush (Chrysothamnus spp.) and greasewood (Sarcobatus spp. Nees) more abundant. The southern limit of the study area is the ecotone of the Great Basin and Mojave Deserts, with creosote (Larrea tridentata (DC.) Coville) becoming more common and sagebrush limited to the wettest sites. The study area also includes the footslopes of the Sierra Nevada, Wasatch Mountains, and various sub-ranges of the Rocky Mountains. Mesic shrubs such as manzanita (Arctostaphylos spp. Adans.), mountain mahogany (Cercocarpus spp. Kunth), and Gambel oak (Quercus gambelii Nutt.) occur in these transition zones, with lesser amounts of sagebrush.

given high weights, and annual herbaceous cover a low weight. Weight values also considered the histograms of component cover, specifically sagebrush and non-sagebrush shrub tend to have a lower mean than perennial herbaceous, thus a 1% difference in sagebrush cover tends to be more ecologically influential than in herbaceous cover. Annual herbaceous dominant pixels, or those with low overall vegetation cover are given the lowest weight as they are associated with a degraded status. The intermediate weighting of perennial herbaceous and nonsagebrush shrub allows for other desired states such as sub-alpine grasslands and creosote shrublands in the Mojave ecotone to achieve moderate to high ASI. Weight values were tested and refined based on our expert judgment to give pixels with abundant sagebrush cover with an understory of perennial grasses or dense perennial grasslands the highest value, both suggestive of a lack of disturbance in this ecosystem. It is important to note that ASI scores themselves are not a primary focus, rather the deviation in ASI from that expected by SP, specifically we assess how the ASI of each site differs from other sites with the same SP. Final ASI scores were multiplied by a factor of ten to expand the range of values in order to detect more subtle differences. 3.2. Site potential 3.2.1. Model inputs 3.2.1.1. 90th percentile Landsat NDVI. We developed a map of 90th percentile NDVI from August and September in the Landsat TM, ETM+, and OLI archive from 1985 to 2016 to serve as an independent variable in our SP model. Using the 90th percentile was key to removing the influence of most fires and other disturbances, thus making the product more appropriate for our analyses. Late summer imagery was used for several reasons: 1) this time period is less variable in terms of weather patterns among years than earlier in the growing season, thus introducing less phenological variability, 2) it is less cloudy, 3) precipitation is typically low in this time period, often resulting in the senescence of herbaceous species, providing a clearer representation of shrubs.

3. Methods and data Methods include the development of an Actual Score Index (ASI) based on a series of existing fractional vegetation maps, and the development of site potential from regression tree models integrating biophysical data and a long-term 90th percentile Landsat NDVI (see Fig. 2). Deviation analysis was then performed between the ASI and the SP to identify under and over-performing areas. These findings were then quantified spatially by each fractional vegetation map to understand the impact on component change. Results were then independently validated with field plots, weather and fire data.

3.2.1.2. Biophysical variables. We included several topographical variables in the SP model, including compound topographic index (CTI), steep north facing slopes, and steep south facing slopes. The steep north facing slope was a combined continuum ranking score with a maximum value of 10 for aspects of 350° to 359° and slopes greater than 20%, If the slope was less than 5% and the aspect beyond ± 48° from north, a value of zero was assigned. The same rank scoring was done relative to the southerly aspect for the steep south index. Organic matter and available water capacity layers in the 0–30 cm horizon were calculated using POLARIS soils data (Chaney et al., 2016), available at 30 m resolution.

3.1. Actual score index The ASI reflects community composition/structure based on fractional cover component maps developed by Xian et al. (2015), later expanded to a larger region for circa 2014 conditions. The ASI, a single integrated layer considering multiple facets of community composition was necessary to build a robust relationship with SP. Four components were input to the ASI model; annual herbaceous, herbaceous, sagebrush, and shrub cover. Annual herbaceous cover subtracted from herbaceous cover produced perennial herbaceous, while sagebrush cover subtracted from shrub cover produced non-sagebrush shrub cover. Cover values were multiplied by weights (Table 1) and summed to produce ASI. ASI considers both the type and amount of cover, where sagebrush, non-sagebrush shrub, and perennial herbaceous cover are

3.2.2. Site Potential model development The dependent variable of the SP model was long-term (2000–2015) MODIS GSN from mid-April to early-October, developed at 250 m resolution. We determined the 2000–2015 median value of the GSN at each pixel, then averaged the yearly GSN values which were above the median, to serve as a proxy of SP. Since our dependent variable had 250 m resolution, we rescaled all independent variables to 250 m resolution using bilinear interpolation, as needed. We chose to use the GSN as the dependent variable over the Landsat 90th percentile NDVI (i.e. develop the model at 250 m instead of 30 m) as GSN is less influenced by cloud and aerosol contamination and is a closer surrogate to total seasonal biomass production than is NDVI. Another advantage of using GSN values as the dependent variable is that it allows various mixtures of shrub and herbaceous cover to have equal or greater ASI than a sagebrush dominated site, resulting in a more generic model. The SP model was trained on good condition MODIS pixels where fractional component (Xian et al., 2015) sagebrush cover is greater than

Table 1 Fractional component weighting factors and actual score index (ASI) calculation example. Final ASI scores were multiplied by ten, rounded to the nearest integer; 238 in the example below. Example Component

Weight

Cover

Cover Score

Annual Herbaceous Perennial Herbaceous Sagebrush Non-Sagebrush Shrub

0.0625 0.4375 0.625 0.5

10 20 15 10 ASI

0.625 8.75 9.375 5 23.75

519

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

and/or rare to be accurately mapped by the RS model, resulting in pixel deviation.

40% of total vegetation, annual herbaceous cover is less than 10% absolute cover and there were no fires between 1993 and 2014. To be included in model development at least 70% of the 30 m pixels within a 7 by 7 focal window had to meet the above criteria, which included a broad distribution over 10% of the study area (Fig. 1A). Training points (n = 10,700) were randomly selected in equal portions from three sagebrush productivity gradients to force the model to include data from the entire GSN range. Independent variables for the regression tree model included POLARIS (Chaney et al., 2016) organic matter content and available water capacity, both in the top 30 cm, long-term Landsat NDVI 90th percentile, CTI, and steep south and north facing slopes. Dependent and independent data were applied in several Cubist regression tree models (RuleQuest Research, 2008) using an optimization procedure where-in multiple trial runs mitigated differences between test and error. Briefly, we varied the maximum number of rules in the regression tree model from 1 to 100 and the percent of random points as training from 10 to 100 for Cubist model optimization testing (Gu et al., 2016). By selecting models that minimized over- and underfitting based on low training and test errors, we optimized accuracy and reduced overfitting tendencies. The most important independent variables were the 90% percentile NDVI, POLARIS organic matter content and available water capacity, and CTI. Many of these variables were also important in the mapping of ecosystem condition (Wylie et al., 2012). Finally, we applied the model developed at the 250 m scale to 30 m versions of the same variables to generate a 30 m map of SP. The SP model includes some contemporary data, so the model does not solely represent reference conditions, rather, the best approximation of these conditions using contemporary data. The overarching goal of the SP model was to produce a long-term and accurate representation of potential biomass productivity, robust to a range of vegetation communities. The initial SP product was rescaled so its mean was similar to that of the ASI, in an effort to make the relationship between SP and ASI close to 1:1.

3.4. Component deviation We produced a series of maps of deviation in 2014 fractional component cover from SP expected cover. First, we plotted SP against the fractional component cover of shrub, sagebrush, herbaceous, annual herbaceous, bare ground, and litter in 15,000 randomly selected points in unburned (from 1984 to 2014) areas across the study area. From this pool of points we selected those (n = 13,253) that were within the 90% confidence limits of the SP vs. ASI regression (Fig. 5). Excluding the points beyond the 90% confidence limits is likely to remove most of the anomalous pixels which may be impacted by disturbance and management (Wylie et al., 2012). Using these points, we developed regressions between fractional cover and SP (Fig. 7). Litter and annual herbaceous cover were fit with third-order polynomial functions, with linear regressions applied to the remaining fractional components. Using the regression model for the line of best fit for each component we converted SP into expected fractional cover for each component. Expected cover was subtracted from actual cover to produce deviation values, which were mapped for every component (Fig. 8). We summarized the average component deviation in unburned and burned portions of dominant level III ecoregions with our study area (Fig. 9). Component deviation was averaged in unburned areas of the Northern Basin and Range ecoregion by 100 m elevation bins and we conducted a regression analysis between the average deviation and elevation bin. Next, we superimposed the 90% SP residuals on the SP vs. component cover regressions and evaluated their position in relation to the line of best fit developed from the ‘normal’ performing pixels. Specifically, for points above and below the 90% regression confidence limits we calculated 1) the percent of observations above the line of best fit for each component (i.e. higher than expected cover), 2) the average cover deviation from the line of best fit, where positive values indicate higher cover than expected by the model, 3) the average component value, and 4) the correlation coefficient of the SP and ASI to fractional component relationship.

3.3. Site Potential deviation analysis We used a sample of 15,000 random points to regress ASI on SP. From this sample, points occurring on areas burned from 1984 to 2014 were removed. In an effort to reduce the dominance of mid-range values in the regression we partitioned the SP into three bins 1) more than one standard deviation below the mean, 2) within one standard deviation of the mean, and 3) more than one standard deviation above the mean. We randomly thinned the middle bin, within one standard deviation of the mean, insuring equal representation throughout the prediction range. Following these procedures, a total of 7,224 data points remained. We plotted ASI regressed on SP, best fit by a second order polynomial model (R2 = 0.71, p < 0.05, Fig. 5). Next, we developed regression confidence limits around this relationship (Table 2) and mapped the deviation confidence classes spatially, applying the regression developed on the sample pixels to all pixels (Fig. 6). Pixels within a defined range of confidence (i.e. Table 2) about the SP to ASI regression (Fig. 5) are defined as normal performing. Pixels below and above a regression confidence limit, generally 90%, are defined as underperforming, and overperforming, respectively. Underperformance from SP can be thought of as actual community structure degradation (e.g. less shrub and perennial herbaceous cover and/or more annual herbaceous cover) than that expected by SP, the opposite being true for overperforming pixels. Some performance deviation may result from modeling error (in the SP and/or ASI), error in the regression model, or actual deviations in ground conditions. Sites with a biophysicallydriven low ASI due to sparse vegetation cover, low sagebrush cover, and/or low herbaceous cover such as playas and the southern portions of the study area will also have a low SP, resulting in no deviation. Similarly, the wetter portions of the study area will have a high SP and high ASI. In some cases a particular landscape feature may be too small

3.5. Validation 3.5.1. Independent validation Independent validation points (n = 382) were collected in 2014 to evaluate the accuracy of the fractional cover components. We converted the data to ASI, using the weights in Table 1. Next, we compared the ASI and the SP to field measured ASI at these points using a regression analysis, with the field ASI as the dependent variable. 3.5.2. Weather data validation We compared the 2014 water year (September–August) precipitation (PRCP) to the 1981–2015 average, to calculate the 2014 percentage of average on a per pixel basis. Next, we compared the 2014 percentage of average PRCP to the fractional component, SP performance, and component deviation products at 15,000 random points in unburned portions of the study area. The goal of this analysis was to determine if 2014 (mapping year) weather conditions were related to SP performance and component deviation. A strong relationship would indicate that much of the deviation is transient, and related to single year weather patterns, not persistent long-term change. 3.6. Fire case study methods We randomly placed points (n = 25,000) within the perimeters of MTBS (1984–2014) burns occurring in the Snake River Plain, Northern Basin and Range, and Central Basin and Range level III ecoregions. At these points we extracted the values of the component deviations and level III ecological site. Using a space for time concept, we averaged 520

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

Fig. 3. Actual score index (ASI) synthesized from the fractional cover of four rangeland components in circa 2014.

each of the extracted values by time since most recent fire for each of the three ecoregions, excluding years with fewer than 10 observations within each ecoregion. Total redistribution represents the absolute value of deviation in each pixel, summed across shrub cover, bare ground, herbaceous cover, and litter cover. This analysis serves two purposes, 1) it provides regional ecosystem recovery statistics, highlighting differences in rates and end result of succession, and 2) it demonstrates the ability of component deviation to respond to events not accounted for in the SP product.

4.2. Validation Both the ASI (r = 0.69, p < 0.05) and SP (r = 0.66, p < 0.05) had moderately strong, but significant, relationships with field measured ASI at independent validation sites. This finding indicated the datasets are comparably related to on the ground conditions. The stronger relationship with ASI was anticipated as the SP is not responsive solely to conditions in 2014. While the fractional component inputs are significantly correlated with 2014 PRCP anomaly from long-term average, the SP is not. Since SP performance and component deviation are based in part from the fractional components, they do have significant, but very weak spatial relationships with 2014 PRCP anomalies. The strongest of the relationships is bare ground deviation, (r = 0.22, p < 0.05). Since component deviation is only weakly related to 2014 climate anomalies and SP is not related to 2014 climate anomalies, spatial patterns in component deviation are largely related to more persistent factors such as management, disturbance, and long-term weather conditions (and not 2014 weather conditions). Areas in which the SP model were trained on (Fig. 1A) have slightly higher mean SP and ASI than those that do not meet training criteria (Table 3). Component deviation also varies by agreement with training criteria where shrub, sagebrush, and bare ground cover deviation is highest and annual herbaceous and herbaceous cover deviation is lowest in pixels that meet training criteria.

4. Results 4.1. Actual site score and Site Potential The final SP model had low error and a very high correlation with training data (r = 0.98). We regressed ASI on SP using a second order polynomial model (R2 = 0.71, p < 0.05) (Fig. 4). It is important to note that while a strong relationship between SP and ASI was critical, a coefficient of determination near 1 would be undesirable. While our model is robust, the model error (in our case 29%) is essential, as it represents in part the influence of disturbance and management accounted for in the ASI, but not the SP. The strength of our SP to ASI relationship is similar to that reported by Wylie et al. (2012). The SP and ASI possess similar (in terms of strength and direction) relationships with fractional components. Spatially, the ASI (Fig. 3) and SP (Fig. 4) are strongly related (r = 0.84, p < 0.05). Both show generally lower values in the south and at lower elevations with higher values associated with mountain ranges.

4.3. Site Potential deviation The proportional area of SP performance confidence (Table 2) 521

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

Fig. 4. Site Potential (SP) derived from regression tree models integrating biophysical data and a long-term 90th percentile Landsat NDVI.

4.4. Component deviation

reveals that much of the study area is within the range of normal performance, but tends to have more area in underperformance relative to overperformance. Fig. 6 depicts the deviation confidence classes of the SP to ASI relationship. Wetter sites are generally more resilient and resistant to change and therefore show positive residuals. Recent fires appear as negative residuals. Some small features are not completely captured in the SP model, such as the lava flows at Craters of the Moon National Monument, ID which appears as a negative residual.

The SP values within the 90% confidence limits of the regression between SP and ASI (Fig. 5) was used to build relationships with selected fractional component values (Fig. 7). The relationship between SP and the fractional components were as expected, with bare ground strongly negatively related (r = −0.83, p < 0.05), and shrub and herbaceous cover positively related to SP (r of 0.71, p < 0.05 and 0.66, p < 0.05, respectively). Annual herbaceous and litter cover had nonlinear relationships with SP, achieving a maximum value at the midFig. 5. Relationship between Site Potential (SP) and Actual Site Index (ASI), fit by a second order polynomial regression model (blue line). 90% regression confidence limits are shown in red. Pixel performance is assessed relative to its position on the regression model, not the 1:1 line. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

522

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

residuals (Fig. 5, blue points in Fig. 7) above the best fit line. The higher occurrence of −90% residuals above the line of best fit for bare ground (88%) and the higher occurrence of +90% residuals above the line of best fit for herbaceous (77%), litter (56%), sagebrush (94%) and shrub (81%) support that the bulk of these pixels are at or above fair condition. The exception to this logic is annual herbaceous cover, which is more likely in +90% residuals (54%). The fractional value of components significantly differs between the −90% and +90% residuals (Table 4B). The means of annual herbaceous, herbaceous, litter, sage, and shrub are all significantly higher in +90% anomalies, while bare ground is significantly lower. We applied the relationships between SP and fractional component cover (Fig. 7) to all pixels in the study area (Fig. 8). Doing so translates the SP into expected component cover, allowing us to map deviations of component cover. Again, there were significant differences in the mean component cover deviation between −90% and +90% anomalies (Table 4C). Annual herbaceous, herbaceous, litter, sagebrush, and shrub cover mean deviations were all significantly greater for +90% anomalies than −90% anomalies. Withholding bare ground, the means of the +90% residuals were all above 0, and the means of −90% residuals below 0, indicating higher and lower than expected cover, respectively. With the exception of annual herbaceous, the positive mean deviations of the desirable rangeland components in the +90% pixels indicate fair or above rangeland condition for much of the area.

Table 2 Classification of actual score index (ASI) deviation from the regression model (Fig. 5) based on pixel position relative to regression confidence intervals, and proportion of study area within each class. Class

Confidence level about line of best fit

Study Area (%)

1 2 3 4 5 6 7

> 90% confidence above 75–89% confidence above 50–74% confidence above ± 50% above and below 50–74% below 75–90% below > 90% below

5.8 5.5 11.6 48.8 15.1 7.7 5.5

range. These relationships demonstrate the SP is a robust indicator of landscape component composition and cover. The locations of the −90% and +90% residuals from Fig. 5 superimposed on the SP versus fractional component cover regression space are non-random (Table 4, Fig. 7). For fair or better rangelands, we would expect desirable rangeland components of herbaceous, litter, herbaceous, sagebrush, and shrub cover to have higher frequencies of +90% residuals (+90% residuals from Fig. 5 represented as blue points in Fig. 7) above the line of best fit (black line in Fig. 7). Fair condition rangeland’s undesirable rangeland components (bare ground and annual herbaceous cover) should have higher frequencies of −90%

Fig. 6. Spatially applied deviation confidence classes (Table 2) of the actual site index (ASI) regressed on Site Potential (SP) (shown in Fig. 5). Higher absolute values indicate increasing confidence in pixel performance deviation from SP. 523

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

Fig. 7. Relationships between Site Potential (SP) and fractional component cover. Regression model for each component is based on data within the 90% confidence limits of Fig. 5. 90% confidence residuals of the ASI regressed on SP are superimposed.

and annual herbaceous in the Blue Mountains and Snake River Plain. Across most ecoregions the herbaceous and annual herbaceous deviations are greater in the burned areas than unburned areas, while the shrub and sagebrush deviations tend to be lower. Bare ground deviation in burned areas vary by ecoregion which is partly related to when the fires occurred within each ecoregion.

Spatially, the component value deviations (Fig. 8) are highly interrelated. The strongest positive associations in deviation are that of herbaceous and annual herbaceous cover (r = 0.77, p < 0.05) and shrub and sagebrush cover (r = 0.53, p < 0.05). Shrub, sage, and herb have displacement relationships with bare ground deviation; (r = −0.27, −0.38, and −0.61, respectively all p < 0.05), sites with higher than expected vegetative cover tend to have lower than expected bare ground, and vice versa. Another expected displacement relationship is between shrub and herbaceous cover (r = −0.28, p < 0.05), indicating that a limited amount of competition does take place between the two. The mean component deviation by EPA level III ecoregion underscore the interrelationships among the component value deviations (Fig. 9). The highest component deviations are found in the Blue Mountains and Snake River Plains ecoregions, both having lower than expected bare ground as a result of higher than expected herbaceous cover. All other ecoregions have mean component deviation in unburned areas of less than 5% from expected. The patterns are more extreme for burned areas, with larger positive deviations of herbaceous

4.5. Fire case study Our space-for-time analysis demonstrated that shrub cover had a negative deviation (lower than the cover expected by SP (e.g. Fig. 7)) for 10–15 years after the most recent fire in all ecoregions (Fig. 10). The rate of shrub recovery by times since fire in the Snake River plain ecoregion tended to be slower than the Northern and Central Basin and Range. The other assessed components varied more by ecoregion in their response to fire. Bare ground cover deviation was significantly lower in the Snake River Plain and herbaceous and annual herbaceous cover higher, compared with the Northern and Central Basin and Range. Across all ecoregions, bare ground, herbaceous cover, and 524

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

Fig. 8. Deviation from expected component cover based on the Site Potential (SP) to fractional cover component relationships (Fig. 7).

5. Discussion

annual herbaceous cover deviation tended to decrease with time since most recent fire. Total redistribution of components also tended to decrease with time since most recent fire, with the Snake River Plain showing the slowest overall recovery. Variability exists in the recovery trends, most prominently in the Snake River Plain, resulting from a variety of sources including; 1) variation in burn intensity by year, 2) differing locations of burns within an ecoregion by year, 3) occurrence of fires on previously burned areas impacting the rate of recovery, 4) multi-year drought/wet episodes impacting the rate of recovery, 5) resprouting of non-sagebrush shrubs (e.g. West and Hassan, 1985; Ellsworth and Kauffman, 2017) and survival of shrubs (Chambers et al., 2017), 6) season of burn (Ellsworth and Kauffman, 2017), and 7) dynamic response of cheatgrass to available moisture (Bradley and Mustard, 2005).

5.1. Actual site index and Site Potential The use of the long-term 90th percentile Landsat NDVI was instrumental in producing a SP model which is largely insensitive to modern disturbance and improving our results relative to previous efforts. For example, Rigge et al. (2013a) found a negative relationship (R2 = 0.51) between SP and bare ground cover (from Homer et al., 2012) in Wyoming. Rigge et al. (2013a) found 90% of underperforming pixels were above the regression line of SP plotted against bare ground cover (developed from normal performing pixels) while 21% of over performing pixels were above the regression line. The current study found a stronger relationship (R2 = 0.73) between SP and bare ground cover and 88% of underperforming pixels having higher than expected bare 525

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

Fig. 9. Average component value deviation from Site Potential (SP) cover by EPA Level III ecrogeion in A) unburned areas from 1984 to 2014 and B) burned at least once during the 1984–2014 period. Refer to Fig. 1 for locations of level III ecoregions.

period of comparison. We have strived to remedy this scenario by training the SP model on good condition sites. Indeed, higher mean SP and ASI values in pixels that meet SP training criteria (Table 3) suggest that the SP model was trained on more intact portions of the study area. While relict areas without direct human influence do exist in isolated patches, even they do not completely represent SP as they lack components of historical ecosystem function such as certain animal species and fire cycles (West, 2000).

ground, with only 10% of over performing pixels having a higher than expected bare ground. Moreover, our methods do not necessitate the need for land cover specific SP models as in Rigge et al. (2013), Wylie et al. (2012), and Boyte et al. (2015). These studies strived to also account for annual variations associated with weather. The ASI provides a convenient single value representing ecosystem status and is related to both site biophysical conditions and vegetation growth and succession, damage from insects, climate change, and disturbance across many ecosystems, including gradual and abrupt change both of which can be readily detected by Landsat (Vogelmann et al., 2016). The SP map echoes many of the spatial patterns of the ASI, but is based on long-term datasets and is less influenced by disturbance, plant invasion, land management, modern climate changes, and short-term temporal and spatial weather variations. However, land degradation was widespread prior to the satellite record; e.g. cheatgrass, fire, overgrazing, and erosion in Oregon (e.g. Platt and Jackman, 1946), from which recovery tends to be slow (Welch, 2005), likely resulting in some of these land use and disturbance legacies manifest in the SP data. The SP data are consistent and objective however, and can be thought of; 1) at minimum; the best land condition in the 1985–2016 period, 2) at maximum; historical climax conditions. The more disturbance legacy in a pixel, the lower the SP value will tend to be, and the shorter the

5.2. SP and component deviation There are two scenarios in which a low ASI can occur, first biophysically driven, second, disturbance and management have lowered conditions from the SP. Our method of assessing site deviation relative to SP conditions removes the former scenario to a large extent. Indeed it is clear that climate, fire (Figs. 6, 8, and 9), and land management (Fig. 11) are primary drivers of SP and component deviation. SP and component deviation based on a single year provides a valuable snapshot of landscape condition although a chronic deviation or deviation slope product could allay concerns related to the high degree of interannual variability that exists in herbaceous cover (Chambers et al., 2014a; Maestas et al., 2016) and biomass productivity (Bradley and

Table 3 Average Site Potential (SP), Actual Score Index (ASI), and component deviation by SP model training critera from a sample of 30,000 random points. Averages were calculated for sites that did not meet SP training criteria, for those that did not meet critera and have no burn history from 1984 to 2015, and those that do meet criteria. Meets Training Criteria

No No, and unburned Yes

SP

94.62 89.55 99.78

ASI

90.59 88.42 107.74

Component Deviation Annual Herb. Cover

Bare Ground

Herb. Cover

Litter Cover

Sagebrush Cover

Shrub Cover

2.59 1.83 −2.09

−0.30 0.50 0.71

3.05 1.94 −4.26

0.75 0.20 1.83

−1.52 −1.32 6.55

−1.53 −0.52 3.25

526

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

Table 4 A) Percent of observations above the line of best fit of the Site Potential (SP) vs actual score index relationship shown in Fig. 4, B) Average component value by anomaly class, C) average component value deviation by anomaly class, with positive values indicating higher than expected cover. For A–C −90% anomaly n = 767, +90% anomaly n = 602. For B-C, letters indicate a significant difference (p < 0.05) in means. Annual Herb. Cover

Bare Ground

Herb. Cover

Litter Cover

Sagebrush Cover

Shrub Cover

A) Percent of observations above line of best fit −90% anomaly 35 +90% anomaly 54

88 10

15 77

20 56

0 94

23 81

B) Average component fractional cover (%) −90% anomaly 7.38 b +90% anomaly 10.01 a

55.12 a 27.59 b

13.69 b 30.76 a

14.40 b 19.75 a

3.34 b 22.35 a

17.85 b 29.39 a

C) Average component cover deviation (%) −90% anomaly −0.06 b +90% anomaly 2.22 a

15.44 a −15.01 b

−8.32 b 10.00 a

−4.73 b 0.68 a

−9.84 b 9.87 a

−6.27 b 6.47 a

low resilience and resistance in western North America, often Wyoming big sagebrush ecosystems, are frequently dominated by cheatgrass. (Chambers et al., 2014b), and more sensitive to grazing (Eldridge et al., 2016). It is clear from our deviation maps that the more mesic a site is within our study region, the more likely it is to have a positive SP anomaly (Fig. 6) and positive shrub component anomaly (Fig. 9). Shrub cover deviation in unburned portions of the Northern Basin and Range for example, is positively correlated with elevation bins (r = 0.92, p < 0.05) with the x-intercept (point at which the mean deviation is 0 = 1500 m). Herbaceous cover deviation on the other hand is

Mustard, 2005) due to variable PRCP (West, 2000). Resistance, the ability to retain existing structure, and resilience, the ability to recover to historical structure in the face of perturbations are key concepts in the assessment of rangeland condition. Mesic sagebrush ecosystems have a greater resource availability (Chambers et al., 2014b) and are more resistant to the invasion of cheatgrass after fire and have a higher overall restoration potential as compared to drier sites (Wisdom et al., 2005; Chambers et al., 2014a; Maestas et al., 2016). The most common sources of lowered resilience and resistance are fire and inappropriate grazing (Chambers et al., 2014b). Areas with

Fig. 10. Component deviation from Site Potential (SP) cover averaged by time since most recent MTBS fire, stratified by ecoregion (n = 25,000). Total redistribution is the summation of the absolute deviation of shrub, bare, herbaceous, and litter (not shown) cover, representing the total departure from expected conditions. 527

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

Fig. 11. Examples of disturbance and land management impact on Site Potential (SP) and component deviation. A) Disturbance related to heavy livestock use near watering/salt/feeding points (piospheres) and impact of a low intensity 1994 burn, B) examples of contrasting deviation across land management units, and C) differing deviation by post-fire management practices, 1999 low to moderate intensity fire indicated by hatching in top panel.

negatively related to elevation (r = −0.91, p < 0.05, x-intercept = 2100 m). In summary, sites below 1500 m tend to have higher than expected herbaceous cover and lower than expected shrub cover, with the opposite true for sites above 2100 m. This finding agrees with the hypothesis that more resistant and resilient sites are less likely to change, specifically in this case, by resistance and resilience to fire and annual herbaceous invasion and thus maintenance and/or recovery of shrub cover.

Regions of high annual herbaceous cover (largely cheatgrass) such as the Snake River Plain ecoregion and northern Nevada (Bradley, 2010) tend to have a negative SP performance (average value of −11.9) (Fig. 6), however this pattern is somewhat tempered by the negative relationship between bare ground deviation and annual herbaceous and herbaceous cover deviation. Specifically, bare ground cover (not deviation) is negatively associated with annual herbaceous cover (r = −0.52, p < 0.05) to a lesser degree than herbaceous cover 528

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

results indicate that fire has a widespread (Fig. 9) and lasting (Figs. 9 and 10) impact on vegetation structure. Other regions have not yet burned but higher than expected herbaceous cover and lower than expected bare ground render them susceptible to burns in the future. Conversely, some fires in cheatgrass dominated regions showed little lasting impact of SP and component deviation due to quick recovery. Our space for time analysis of succession following burns generally agrees with the literature in that depression in shrub cover is long-term (Chambers et al., 2017) and recovery was slow the first 10 years, but generally returned to pre-disturbance levels within thirty years, as reported by Hanna and Fulgham (2015) in a Northeast California mesic sagebrush steppe. Others report sagebrush recovery times after fire of 5–14 years or more (Welch, 2005). In the Snake River Plain ecoregion shrub cover was not fully recovered after 30 years, and sites were still largely herbaceous dominated, due in part to the low resilience in this low and dry ecoregion (Davies et al., 2012). The timing of peak bare ground deviation varied by ecoregion, with the Central Basin and Range decreasing in the second year after fire, similar to West and Hassan (1985) in a good condition sagebrush steppe site in Utah. Lower than expected bare ground in the first year after fire in the Northern Basin and Range and in the Snake River Plain ecoregions is due to the large influx of herbaceous cover. In the Central Great Basin our results again concur with West and Hassan (1985) who found higher herbaceous cover in the second year after fire. Hanna and Fulgham (2015) found annual herb cover peaked in the 3rd year after fire, and declined at a more rapid rate than perennial herbaceous, similar to our results in the Central Basin and Range.

(r = −0.80, p < 0.05) and shrub cover (r = −0.63, p < 0.05). A similar pattern is observed in the component deviation spatial relationships, where bare ground deviation is more weakly related to annual herb deviation (r = −0.30, p < 0.05) than herb cover deviation (r = −0.61, p < 0.05). This implies that shrub and herbaceous cover have a more exclusionary relationship with bare ground than does annual herbaceous cover. The net result is that regions with significant annual herbaceous invasions tend to have lower than expected bare ground cover (Fig. 8) and thus a less negative SP performance. Ecologically, this infers that herbaceous cover, especially annual herbaceous, sometimes occurs in niches devoid of vegetative cover prior to invasion. Though shrubs protect a portion of grasses from grazing under and around its canopy (Welch, 2005), our results show that total herbaceous cover increases as shrubs decline, occupying niches and utilizing resources previously allocated by shrubs. Thus, landscape degradation by annual herbaceous invasion and resulting alteration of the fire cycle results in landscape redistribution of lower than expected shrub, sage, and bare ground, and higher than expected herbaceous and annual herbaceous cover (e.g. Snake River Plain results in Fig. 9). Moreover, the majority of the positive deviation in herbaceous cover is driven by higher than expected annual herbaceous cover (e.g. 71% in the Snake River Plain Fig. 9). This finding concurs with that of Tausch et al. (1995) in NW Nevada, who reported that total bare ground cover in ∼15-year-old burns was lower than in adjacent unburned areas due to increased perennial grass cover and cheatgrass cover in the burned plots. This complex of change tends to occur in sites with low initial resistance, and are made even less resistant following annual grass invasion and corresponding reduction in perennial grass and shrub cover (Chambers et al., 2017). Restoration goals in such ecosystems with disrupted function (Bradley, 2010) that have already crossed a potentially non-reversible threshold (West, 2000) may simply be to restore ecosystem function, with less regard given to structure (Briske et al., 2005). Grazing has a widespread presence in the sagebrush steppe which can significantly impact ecosystem function (Sims and Singh, 1978; Boyd et al., 2014; Eldridge et al., 2016; Chambers et al., 2017), including driving changing states in STM status, potentially lowering resistance to climate change pressures (Beschta et al., 2013), reducing perennial grass cover (Boyd et al., 2014; Munson et al., 2016), increasing cheatgrass cover (Boyd et al., 2014), and increasing bare ground (Augustine et al., 2012). Grazing can impact condition deviation among management units due to differences in intensity, duration, and season of grazing interacting with species composition and disturbance history, and within pastures due to uneven utilization. Higher than expected levels of bare ground and lower than expected perennial herbaceous vegetation present a management concern (Weber et al., 2010; Munson et al., 2016) and are related to soil stability, runoff rates, and watershed health (Booth and Tueller, 2003). Differing SP and component deviations are evident in burns, piospheres, and across management units (Fig. 11), which may indicate excessive grazing levels over a prolonged period (Wylie et al., 2012). Our component and/ or SP performance maps could be integrated in order to evaluate range condition, for example; 1) sites with higher than expected bare ground and shrub cover and lower than expected herbaceous cover are indicative of negative grazing influence (West, 2000), 2) sites with lower than expected bare ground and shrub cover and higher than expected herbaceous cover are symptomatic of cheatgrass invasion and fire complex discussed previously, 3) sites with lower than expected bare ground and higher than expected herbaceous cover suggest good condition range.

5.4. Study limitations and applications Similar to Bradley and Mustard (2005), we do not imply that our deviation maps are inclusive of all change as some deviation may represent ephemeral variation resulting from temporary interruption to normal surface dynamics such as flooding, minor insect damage, moderate grazing, and short-term drought, resulting in a within-state change. Our focus however is to highlight abrupt change (e.g. fires, intensive management) and especially gradual change (e.g. Vogelmann et al., 2016), that change which occurs over a period of years to decades with plant succession or shifts in species composition. Though our ASI map only represents one year (2014) some of the constituent components, namely shrub and sagebrush cover, are slow to change and represent the net degradation over a longer term. Some components such as annual herbaceous are weakly related to SP and the stratification of anomalies is noisier, limiting the utility of the relationship, but could be indicating invasive annual expansion related to periodic droughts and not exclusively disturbance. While the values of ASI in Table 1 are tuned to sagebrush steppe, the Mojave Basin and Range ecoregion, which strongly differs in community composition still has pixels above the 90% confidence limits of SP performance (Fig. 6) and has mean component deviations in line with other ecoregions (Fig. 9). These findings suggest that our SP model is general enough that while ASI tends to be lower in non-sagebrush steppe (Fig. 4), the SP is also low in these regions (Fig. 5), producing unbiased deviation products. Still, the ASI and SP could be more regionally tailored (i.e. different table for each ecological site or ecoregion) to produce a more relevant model at a local scale, potentially resulting in a better regression fit and more accurate deviation products. Doing so would reduce the likelihood of rare features being mischaracterized. Though in some cases a more regional SP model such as ours is desired as it glosses over inconsequential differences and produces seamless products. Our products could serve as a first-order view of land health whose accuracy is partly confirmed with on the ground measurements and further supported by our results that needs on the ground ‘spot checking’ prior to utilization in management decision making. Products can inform the evaluation of which management practices result in

5.3. Fire Historically, large fires were infrequent in sagebrush steppe resulting in a landscape dominated by large patches of sagebrush, interspersed with patches of grassland (Bukowski and Baker, 2013). Our 529

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

desirable and undesirable land outcomes in terms of both SP performance and community structure as indicated by component deviation. While our products are designed to work at a regional scale, the results are also clearly applicable at a local scale (Fig. 11). Moreover, the results themselves advance the understanding of SP state and drivers of change in the Great Basin.

Boyd, C.S., Beck, J.L., Tanaka, J.A., 2014. Livestock grazing and sage-grouse habitat: impacts and opportunities. J. Range Appl. 1, 58–77. Boyte, S.P., Wylie, B.K., Major, D.J., 2015. Mapping and monitoring cheatgrass dieoff in rangeland of the Northern Great Basin, USA. Rangeland Ecol. Manage. 68, 18–28. Bradley, B.A., Mustard, J.F., 2005. Identifying land cover variability distinct from land cover change: cheatgrass in the Great Basin. Remote Sens. Environ. 94, 204–213. Bradley, B.A., 2010. Assessing ecosystem threats from global and regional change: hierarchical modeling of risk to sagebrush ecosystems from climate change, land use and invasive species in Nevada, USA. Ecography 22, 198–208. Briske, D.D., Fuhlendorf, S.D., Smeins, F.E., 2005. State-and-transition models, thresholds, and rangeland health: a synthesis of ecological concepts and perspectives. Rangeland Ecol. Manage. 58, 1–10. Bukowski, B.E., Baker, W.L., 2013. Historical fire regimes, reconstructed from land-survey data, led to complexity and fluctuation in sagebrush landscape. Ecol. Appl. 23, 546–564. Chambers, J.C., Miller, R.F., Board, D.I., Pyke, D.A., Roundy, B.A., Grace, J.B., Schupp, E.W., Tausch, R.J., 2014a. Resilience and resistance of sagebrush ecosystems: implications for state and transition models and management treatments. Rangeland Ecol. Manage. 67, 440–454. Chambers, J.C., Bradley, B.A., Brown, C.S., D’Antonio, C., Germino, M.J., Grace, J.B., Hardegree, S.P., Miller, R.F., Pyke, D.A., 2014b. Resilience to stress and disturbance, and resistance to Bromus tectorum L. invasion in cold desert shrublands of Western North America. Ecosystems 17, 360–375. Chambers, J.C., Board, D.I., Roundy, B.A., Weisberg, P.J., 2017. Removal of perennial herbaceous species affects response of Cold Desert shrublands to fire. J. Veg. Sci. 28, 975–984. Chaney, N.W., Wood, E.F., McBartney, A.B., Hempel, J.W., Nauman, T.W., Brungard, C.W., Odgers, N.P., 2016. POLARIS, a 30-meter probabilistic soil series map of the contiguous United States. Geoderma 274, 54–67. Davies, G.D., Bakker, J.D., Dettweiler-Robinson, E., Dunwiddie, P.W., Hall, S.A., Downs, J., Evans, J., 2012. Trajectories of change in sagebrush steppe vegetation communities in relation to multiple wildfires. Ecol. Appl. 22, 1562–1577. Eldridge, D.J., Poore, A.G.B., Ruiz-Colmenero, M., Lentic, M., Soliveres, S., 2016. Ecosystem structure, function, and composition in rangelands are negatively affected by livestock grazing. Ecol. Appl. 26, 1273–1283. Ellsworth, L.M., Kauffman, J.B., 2017. Plant community response to prescribed fire varies by pre-fire condition and season of burn in mountain big sagebrush ecosystems. J. Arid Environ. 144, 74–80. Gu, Y., Wylie, B.K., Boyte, S.P., Picotte, J., Howard, D.M., Smith, K., Nelson, K.J., 2016. An optimal sample data usage strategy to minimize overfitting and underfitting effects in regression tree models based on remotely-sensed data. Remote Sens. 8, 943. Hanna, S.K., Fulgham, K.O., 2015. Post-fire vegetation dynamics of a sagebrush steppe community change significantly over time. Calif. Agr. 69, 36–42. Homer, C.G., Aldridge, C.L., Meyer, D.K., Schell, S.J., 2012. Multi-scale remote sensing sagebrush characterization with regression trees over Wyoming, USA; laying a foundation for monitoring. Int. J. Appl. Earth Obs. 14, 233–244. Homer, C.G., Meyer, D.K., Aldridge, C.A., Schell, S.J., 2013. Detecting annual and seasonal changes in a sagebrush ecosystem with remote sensing derived continuous fields. J. Appl. Remote Sens. 7 (1) No. 073508. Maestas, J.D., Campbell, S.B., Chambers, J.C., Pellant, M., Miller, R.F., 2016. Tapping soil survey information for rapid assessment of sagebrush ecosystem resilience and resistance. Rangelands 38, 120–128. Munson, S.M., Sankey, T.T., Xian, G., Villarreal, M.L., Homer, C.G., 2016. Decadal shifts in grass and woody plant cover are driven by prolonged drying and modified by topoedaphic properties. Ecol. Appl. 26, 2480–2494. Peters, P.C., Bestelmeyer, B.T., Herrick, J.E., Fredrickson, E.L., Monger, H.C., Havstad, K.M., 2006. Disentangling complex landscapes: new insights into arid and semiarid system dynamics. BioScience 56, 491–501. Platt, K., Jackman, E.R., 1946. The cheatgrass problem in Oregon. Bull. 668. Oregon Agr. Exp. Sta., Corvallis, Ore. Rigge, M., Wylie, B., Gu, Y., Belnap, J., Phuyal, K., Tieszen, L., 2013a. Monitoring the status of forests and rangelands in the Western United States using ecosystem performance anomalies. Int. J. Remote Sens. 34, 4049–4068. Rigge, M., Wylie, B., Zhang, L., Boyte, S.P., 2013b. Influence of management and precipitation on carbon fluxes in Great Plains grasslands. Ecol. Indic. 34, 590–599. RuleQuest Research. 2008. Cubist. Version 2.05. St. Ives, NSW: RuleQuest Pty. Schlaepfer, D.R., Lauenroth, W.K., Bradford, J.B., 2014. Natural regeneration processes in big sagebrush (Artemisia tridentata). Range. Ecol. Manage. 67, 344–357. Sims, P.L., Singh, J.S., 1978. The structure and function of ten western North American grasslands. III. Net primary production, turnover and efficiencies of energy capture and water use. J. Ecol. 66, 573–597. Stringham, T.K., Novak-Echenique, P., Snyder, D.K., Peterson, S., Snyder, K.A., 2016. Disturbance response grouping of ecological sites increases utility of ecological sites and state-and-transition models for landscape scale planning in the Great Basin. Rangelands 38, 361–378. Tausch, R.J., Chambers, J.C., Blank, R.R., Novak, R.S., 1995. Differential establishment of perennial grass and cheatgrass following fire on an ungrazed sagebrush-juniper site, pp. 252–257. In: Roundy, B.A., McArthur, D.E., Haley J.S., Mann, D.K. (comp.) Proc. Wildland shrub and arid land restoration symposium. Gen. Tech. Rep. INT-GTR-315. USDA For. Serv. Int. Res. Sta. Ogden, UT. Thornton, P.E., Thornton, M.M., Mayer, B.W., Wilhelmi, N., Wei, Y., Devarakonda, R., Cook, R.B., 2014. Daymet: Daily Surface Weather Data on a 1-km Grid for North America, Version 2. ORNL DAAC, Oak Ridge, Tennessee, USA. Available at: http:// daac.ornl.gov/ (accessed 20 June 2016). Vogelmann, J.E., Gallant, A.L., Shi, H., Zhu, Z., 2016. Perspectives on monitoring gradual change across the continuity of Landsat sensors using time-series data. Remote Sens.

6. Conclusions The SP data provide a consistent and objective framework for assessing ecosystem condition and are strongly related to fractional component cover. The use of the long-term 90th percentile Landsat NDVI trained over good condition sites was instrumental in producing a SP model which is largely insensitive to modern disturbance and improving our results relative to previous efforts. Conversion of SP into expected component cover, then into component cover deviation based on the actual cover of components was a key step in making our results more readily interpretable to land managers and was a novel effort of the current work. Further, we demonstrate the utility of fractional mapping components (Xian et al., 2015) in rangeland applications. Overall, much of the study area is within the range of normal performance in SP performance, but tends to have more area in underperformance versus over performance due to the impact of disturbance, management, and plant invasions. We found that shrub cover deviation is more positive at higher elevation, while herbaceous cover deviation has the opposite pattern, supporting the hypothesis that more resistant and resilient sites are less likely to change from the shrub dominated legacy. Another key finding was that regions with significant annual herbaceous invasions (i.e. have positive herbaceous and annual herbaceous cover deviations) tend to have lower than expected bare ground and shrub cover. Annual herbaceous invasion is then associated with an ecosystem structure redistribution complex wherein herbaceous cover, especially annual herbaceous, sometimes occurs in niches devoid of vegetative cover prior to invasion. Our case study demonstrates that depression of shrub and sagebrush cover following fire persists for a long period, at minimum 10 years. Further, the rate and end result of recovery vary spatially. In future work we plan on expanding the view of deviation to longer times scales including the entire Landsat record (1984-present), permitting investigation of chronic deviation, deviation trends, and more explicit impacts of land management. Further, we plan on integrating the concepts of deviation presented here with a state and transition model framework and resistance and reliance. Acknowledgments Work by Rigge, Shi, and Gu was under USGS contract G13PC00028, work of Meyer and Bunde was under USGS contract G10PC00044. We acknowledge the efforts of Spencer Schell and Lauren Cleeves in collecting field data for the fractional component mapping. References Augustine, D.J., Booth, D.T., Cox, S.E., Derner, J.D., 2012. Grazing intensity and spatial heterogeneity in bare soil in a grazing-resistant grassland. Rangeland Ecol. Manage. 65, 39–46. Beschta, R.L., Donahue, D.L., DellaSala, D.A., Rhodes, J.J., Karr, J.R., O’Brien, M.H., Fleischner, T.L., Deacon Williams, C., 2013. Adapting to climate change on western public lands: addressing the ecological effects of domestic, wild and feral ungulates. Environ. Manage. 51, 474–491. Bestelmeyer, B.T., Goolsby, D.P., Archer, S.R., 2011. Spatial perspectives in state-andtransition models: a missing link to land management: spatial state-and-transition models. J. Appl. Ecol. 48, 746–757. Bestelmeyer, B.T., Tugel, A.J., Peacock, G.L., Robinett, D.G., Shaver, P.L., Brown, J.R., Herrick, J.E., Sanchez, H., Havstad, K.M., 2009. State-and-transition models for heterogeneous landscapes: a strategy for development and application. Rangeland Ecol. Manage. 62, 1–15. Booth, D.T., Tueller, P.T., 2003. Rangeland monitoring using remote sensing. Arid Land Res. Manage. 17, 455–467.

530

Ecological Indicators 96 (2019) 516–531

M. Rigge et al.

wildfire. J. Range Manage. 38, 131–134. Wisdom, M.J., Rowland, M.M., Tausch, R.J., 2005. Effective management strategies for sage-grouse and sagebrush: a question of triage? Trans. North Am. Wildlife Natural Resour. Conf. 70, 206–227. Wylie, B.K., Zhang, L., Ji, L., Tieszen, L.L., Jolly, M., 2008. Integrating modeling and remote sensing to identify ecosystem performance anomalies in the boreal forest, Yukon River Basin, Alaska. Int. J. Dig. Earth 1, 196–220. Wylie, B.K., Boyte, S.P., Major, D.J., 2012. Ecosystem performance monitoring of rangelands by integrating modeling and remote sensing. Rangeland Ecol. Manage. 65, 241–252. Xian, G., Homer, C., Rigge, M., Shi, H., Meyer, D., 2015. Characterization of shrubland ecosystem components as continuous fields in the northwest United States. Remote Sens. Environ. 168, 286–300.

Environ. 185, 258–270. Weber, K.T., Glenn, N.F., Tibbitts, J., 2010. In: Final Report: Forecasting Rangeland Condition with GIS in Southeastern Idaho, pp. 101–112. Welch, B., 2005. Big sagebrush: A sea fragmented into lakes, ponds, and puddles. Gen. Tech. Rep. RMRS-GTR-144. Fort Collins, CO: U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station. 210 p. West, N.E., 1999. Managing for biodiversity of rangleands. In: Collins, W.W., Qualset, C.O. (Eds.), Biodiversity in Agroecosystems. CRC Press, Boca Raton, Florida, pp. 101–126. West, N.E., 2000. Synecology and disturbance regimes of sagebrush steppe ecosystems. pp. 15–26. In: Entwistle, P.G., Debolt, A.M., Kaltenecker, J.H., Steenhof, K., (compilers) Proc. Sagebrush Steppe Ecosystems Symposium. Bureau of Land Mgt. Publ. No. BLM/LD/PT-00 1 1001 + 1150. West, N.E., Hassan, M.A., 1985. Recovery of Sagebrush-grass vegetation following

531