Leveraging local habitat suitability models to enhance restoration benefits for species of conservation concern

Our study focused on the six GUSG satellite populations in southwest Colorado: Crawford, Cerro Summit-Cimarron-Sims Mesa (CCS), Piñon Mesa, Poncha Pass, San Miguel, and Dove Creek (Fig. 1). As of 2019, each of these populations were declining or already below conservation targets (GUSG Rangewide Steering Committee [GSGRSC] 2005) with an estimate of < 5000 birds remaining range wide (with 2019’s 3-year estimates for the Gunnison Basin population at 3787 birds, and 554 birds across the satellites; USFWS 2019, 2020a). The 2020 USFWS Final Recovery Plan (FRP; USFWS 2020a) and Recovery and Implementation Strategy (RIS; USFWS 2020b) outline steps needed to improve species viability in the long-term, based on concepts of resiliency, redundancy, and representation (i.e., the 3Rs; USFWS 2016, Smith et al. 2018). While the Gunnison Basin population largely satisfies the resiliency requirement due to its relative size and stability (USFWS 2019), considerable concern surrounds the persistence of the satellite populations (Saher et al. 2022), which are critical to maintaining redundancy and representation, and may provide significant adaptive capacity to future environmental change (USFWS 2019). In addition to highlighting the importance of conserving remaining habitats within these populations, the FRP and RIS also invoke restoration of habitats as critical to improving habitat quantity and quality across the species’ range, underscoring the importance of such actions. We did not include the Gunnison Basin core population or the Monticello satellite population in Utah in our analysis because reference RSF models (see below) were not available for these populations. See the Supplemental for additional details on the study area.

As the spatial scale of habitat selection responses can be influential, we chose a previously developed suite of RSF models (Saher et al. 2022; hereafter “reference models”) that characterize GUSG habitat selection responses at multiple scales based on contemporary vegetation spatial products (Rigge et al. 2020). We used these empirically driven, season-specific models and inputs to assess the habitat improvement potential (or degradation) for specific habitat actions aimed at improving habitat conditions for the six remaining GUSG satellite populations in Colorado. All modifications of spatial data layers and model runs were conducted using ArcGIS® Pro [v.2.8.0, 2021 Esri Inc.]. We restricted our analysis to crucial habitats (Fig. 1) defined by the corresponding “patch-scale” models of Saher et al. (2022; representing 95% of currently utilized habitats) because this best represented the spatial scale at which most land-management actions are applied (e.g., pinyon-juniper removal, invasive weed control, sagebrush seeding/planting, restoration of mesic habitats, etc.).

The reference models generated by Saher et al. (2022) used the exponential form of an RSF (Manly et al. 2002), but their sum of linear responses could not be exponentiated when generating mapped predictions because some locations had incalculably large values, so they simply mapped the linear predictor and scaled between 0–1, to visualize the habitat suitability predictions. The linear formula by itself does not directly reflect a true probability of selection (i.e., compared to a Resource Selection Probability Function [RSPF] estimated using logistic regression) and, thus, is a relative function related to the true probability of selection (Manly et al. 2002). For this reason, we transformed the RSF suitability indices into ranks based on the number of habitat use locations that were observed at or below that RSF value. This was similar to a percentile, but rather than 100 bins the number of bins was determined by the number of use locations (that differed for each satellite population and model). For each RSF map, we extracted the RSF values at habitat use locations used in that model, and calculated the proportion of observations that were less than or equal to the RSF value (Range 0–1). This resulted in each model having its own unique classification, where the RSF values at each use location were assigned 0–1 percentile ranks (see the supplemental for table examples). Each model’s RSF surface was then reclassified using this table as a series of thresholds for binned ranks. Following Saher et al. (2022), we defined “habitat” as those pixels with habitat suitability ranks exceeding the value that captured 95% of use locations in the reference models; those falling below were considered “non-habitat”. The resulting map suitability values (MapRSF) for habitats ranged > 0.00 to 0.95, with smaller values indicating higher ranked habitats (i.e., if there were 100 observations, all mapped RSF values between the 2nd-greatest estimated use location and greatest estimated use location would receive a rank of 0.02). The difference in suitability scores between the reference and modified maps represents the change in value based on binned percentiles for habitat use locations (< 1% increments). Change in habitat suitability could therefore respond non-linearly to change in RSF values, with the steepest increases in suitability occurring where smaller increasing RSF values caused greater increases in the number of use locations captured.

We categorized each covariate in the reference models as having a positive or negative relationship to habitat suitability and determined whether that relationship was linear or quadratic based on marginal effects predictions from Saher et al. (2022). Covariates generally summarized spatial layers representing habitat conditions over moving window buffer radius or distance decay scales ranging 45 m to 570 m (Saher et al. 2022). For most habitat variables, we only assessed breeding season models because summer reference models were unavailable for all satellite populations; however, when available (i.e., for all satellite populations except Dove Creek), we also assessed summer models. These seasonal comparisons of mesic improvements are valuable to consider because of their critical importance in the summer when broods depend on these habitats for food and cover.

We simulated changes in habitat characteristics that were expected to result from habitat restoration actions (e.g., an increase in sagebrush percent cover resulting from seeding or planting sagebrush) by modifying values in the original 30-m × 30-m geospatial input layers (i.e., pre-moving window) used in the reference models, within a 1-km buffer of each population’s spatial extent. We only evaluated management actions that could (1) improve habitat suitability for GUSG in the satellite populations based on these relationships (see Connelly et al. 2011; Wisdom et al. 2011; Aldridge et al. 2012; Knick et al. 2013; Young et al. 2020; Walker et al. 2016), and (2) be reasonably modified using current on-the-ground management actions to improve GUSG habitats (e.g., those employed by the BLM, Table 1). Additional details on covariate modifications can be found in the Supplemental. After generating the modified input layers that simulated management action outcomes across the landscape for each satellite population, we recalculated all moving window and decay function values, substituted them one-at-a-time into the reference model equations (Supplemental Table S1), and reran the models using the ArcPy package in ArcGIS® Pro to generate a comparative RSF layer for each action. For models that included multiple covariates calculated from the same original geospatial input layer (e.g., percent mesic cover [moving window] and distance to mesic [decay function] in Dove Creek), we substituted both covariates simultaneously. To generate our final RSF map, we reclassified each model output from the linear RSF into percentile bins of habitat use locations, using the reclassification table derived from the original RSF model.

Management agencies may apply multiple actions that overlap across space into their management strategies for improving GUSG habitats. We extended our single-covariate breeding season assessments by creating maps representing combined actions to investigate the potential to further improve suitability (e.g., increasing sagebrush cover plus annual herbaceous removal). We used the same methodologies as with single covariates above but instead substituted pairs of modified covariates that spatially coincided with the reference models. We simulated only a subset of possible paired actions to demonstrate how habitat improvement heatmaps can further inform strategic management planning, selecting pairs likely to result in the greatest improvements to habitat suitability based on single action heatmaps.

Prioritizing monitoring efforts for invasive species can lead to more effective management of invasions (Tarbox et al. 2022). To identify where habitats might be most sensitive to increases in invasive plant cover, and therefore where monitoring and rapid response might be most valuable, we created an additional set of heatmaps that simulated new or worsening annual herbaceous and pinyon-juniper invasions. Our annual herbaceous heatmaps increased, rather than decreased, the raw covariate percent cover values by 1 standard deviation (SD) of the values in the buffered crucial habitat extents. To simulate new or worsening pinyon-juniper encroachment (represented by binary presence/absence raw layers of < 5% or 5–10% encroachment), we increased the moving window inputs by 1 SD of the buffered extent. We capped values in the modified layers at 100% cover to remain within logical bounds.

We calculated change in habitat suitability using the ArcPy package in ArcGIS® Pro as:

We created two sets of heatmaps for each habitat action scenario to serve as decision-support tools for habitat restoration planning. We first generated uncategorized maps using values generated by Eq. 1. We then produced final, categorized maps, using nested conditional statements to identify areas where new habitat was created (i.e., that transitioned from non-habitat to habitat) and areas that remained non-habitat despite simulated management interventions. Some model covariates (e.g., percent mesic cover, mean sagebrush height) had a positive but quadratic (decreasing) relationship to suitability, meaning habitat improvement actions would be beneficial, but only to a point. Beyond this threshold, further increases in the variable are expected to be reduce habitat use. For this reason, we visualized resulting negative impacts on habitat suitability in our heatmaps because we expected very high proportions of these habitat characteristics to be detrimental to GUSG.

To gauge how using our habitat improvement heatmaps could improve returns on restoration investments, we simulated habitat improvement actions guided by our heatmaps and compared outcomes to those not guided by these maps. We employed an independent technician, without prior knowledge of our heatmap predictions or the reference models, to develop two sets of hypothetical treatment locations: non-targeted scenarios based on past BLM treatment data and targeted scenarios that used our heatmaps to inform placement of treatments (see the Supplemental for detailed methods) to maximize habitat suitability improvement outcomes for GUSG (Table 2). We assessed three types of actions for which sufficient management records or planning data existed: reduction of non-sagebrush shrub cover, pinyon-juniper removal, and mesic improvements. We limited the simulated reduction of non-sagebrush shrub cover to within San Miguel because this was the only satellite population where models indicated removal could have a beneficial impact on suitability at local sites.

To simulate areas with mesic improvement potential within the satellite populations, we used a stream reach valley bottom polygon layer developed by The Nature Conservancy (Nagel et al. 2014, Data contact Teresa Chapman); see Supplemental for details). We used these pre-defined polygons to generate our set of non-targeted mesic action polygons by selecting the ten largest stream segments that did not overlap areas of substantial existing mesic, water, wetland, or agriculture cover, based on the reference map covariate layers used in the reference models. We then used our heatmaps to inform selection of our targeted polygon set, again with the aim of maximizing habitat suitability outcomes by improving the availability of mesic habitat resources for GUSG.

We used the resulting restoration action polygon sets to simulate changes on the landscape following realistic habitat restoration actions, this time only modifying the covariate values within the treatment polygons (rather than the entire landscape). We mapped improvements and calculated improvement metrics two ways for each set, including: (1) the overall projected improvement in habitat suitability across the patches, and (2) the total projected area of new habitat created, calculating for each the percent gain of the targeted action polygon sets over the corresponding non-targeted sets.

We observed considerable spatial variability in projected change in habitat suitability values across the crucial breeding season habitats for most habitat restoration action scenarios, as well as large differences in the magnitude of change among many action types. Our categorized maps for the Crawford, Piñon Mesa, Poncha Pass, and San Miguel satellite populations [Figs. 2, 3, 4, 5 (also see Supplemental Fig. S1)] highlight areas where 32 simulated management actions had maximal or minimal projected benefits for improving breeding and summer habitat suitability. They illustrate areas where: (1) new habitat could be created, (2) existing habitat could be improved, or (3) areas were likely to remain non-habitat despite management interventions. Results for the CCS and Dove Creek satellite populations involved numerous modeling and interpretation caveats and are reported exclusively in the Supplemental (Figs. S2–S4). Uncategorized RSF percentile change maps show calculated values from Eq. 1 for each 30-m² pixel for each satellite population (Supplemental Figs. S5–S8).

Several notable trends were seen in responses to habitat actions across multiple satellite populations. Increasing sagebrush cover and height, litter cover, and herbaceous cover generally resulted in widespread improvements in habitat suitability (Figs. 2f, 3a, f, and 5e, h [also see Supplemental Figs. S3f, and S4b]). Paired management actions resulted in greater improvements for all satellite populations compared to single actions (Figs. 2g, 3h, and 4g [also see Supplemental Fig. S1a, S2d, and S3g]), and increased the overall area of new habitat that could be created by an average of 30.1% compared to the next best single action.

Removal of annual herbaceous grasses generally resulted in large improvements in habitat suitability but with variable outcomes among satellite populations (Figs. 2a, 4f, and 5a [also see Supplemental Fig. S3a S4a]). New or worsening annual herbaceous invasions (simulated by increasing covariate values within the buffered extents) eliminated nearly all existing habitats (i.e., habitats that transitioned to non-habitat) within crucial habitat extents (Fig. 6a, b), except for San Miguel (Fig. 6c) where cheatgrass (Bromus tectorum, a pervasive and non-native annual herbaceous grass) cover was already relatively substantial in crucial habitats. Pinyon-juniper removal also resulted in large gains in suitability across varying extents [Figs. 2b, 3b (also see Supplemental Fig. S3b)], except for San Miguel where existing pinyon-juniper cover was relatively limited within the crucial habitats (Fig. 5b). New or worsening pinyon-juniper encroachment showed considerable spatial variation in potential impacts (Figs. 7a–c). These patterns appear to be relative to the magnitude of current invasion or encroachment within each satellite population.

In the Crawford satellite population, the greatest amount of overall improvement in suitability of existing habitats was projected to occur from management actions that increased sagebrush cover and could benefit much of the crucial habitat extent (Fig. 2f). However, removal of annual invasive grasses (Fig. 2a) and pinyon-juniper (Fig. 2b) were also expected to have relatively large impacts in isolated areas. Projected improvements and new habitat created from pinyon-juniper removal generally complemented areas improved by increasing sagebrush cover and annual herbaceous removal. While some single actions were projected to result in the creation of new habitats, the total area with this potential was relatively small (max of 4.89 km² for the combined actions) compared to the total area where existing habitats could be improved (max of 27.92 km²).

In Piñon Mesa, we projected pinyon-juniper removal could create the most habitat and result in the greatest improvements in habitat suitability across crucial habitats of any single management action (Fig. 3b). Increasing litter cover (e.g., by modifying grazing practices) was projected to result in large gains overall that were largely complementary to pinyon-juniper removal, though also reduce suitability in some areas (Fig. 3a). Increasing sagebrush height and decreasing shrub height (e.g., reducing non-sagebrush shrubs) resulted in similar outcomes, though with relatively moderate benefits across much of the crucial habitat extent (Fig. 3f, g). Pairing pinyon-juniper removal with decreasing shrub height led to very strong improvements across the entire extent (Fig. 3h).

In Poncha Pass, removing annual herbaceous cover (Fig. 4f) and increasing non-sagebrush shrub cover (Fig. 4a) were projected to result in the greatest and most widespread habitat suitability improvements and new habitat created. Paired actions resulted in substantial and widespread gains (Fig. 4g). Our simulations indicated that decreasing non-sagebrush shrub cover in this satellite population would result in widespread negative impacts to habitat suitability across most of the crucial habitat extent (Fig. 4b). Mesic improvements in the breeding season were expected to result in habitat suitability gains in many areas (Fig. 4c), but the areas seeing the most benefit in this season differed from those in the summer/brood-rearing season (Fig. 4d). Projected breeding season gains in suitability resulted in negative impacts to suitability in the summer (i.e., late brood-rearing) season.

In San Miguel, increasing herbaceous cover (Fig. 5e) and sagebrush height (Fig. 5h) had the largest and most widespread gains in suitability and new habitat created. Mesic improvements were projected to substantially improve breeding habitat suitability across the western half of crucial habitats and the westernmost of the eastern habitat patches (Fig. 5c), though some of these actions resulted in negative impacts on suitability in summer in those same areas (Fig. 5d). Because of the complex quadratic relationship between non-sagebrush shrub cover and habitat suitability, simulated increases and decreases in shrub cover (a 1-SD directional change) both resulted in negative impacts across much of the extent, with a few isolated areas of relatively large gains in habitat suitability (Fig. 5f, g). Annual herbaceous removal resulted in relatively large benefits in the western half of the crucial habitat extent, where cheatgrass was already well-established (Fig. 5a).

Spatially targeting treatment polygons using our categorized heatmaps resulted in a nearly fourfold increase in habitat suitability values (improvement factor = 3.95) and more than a 15-fold increase in habitat created (improvement factor = 15.77) compared to non-targeted polygons (Table 2). The mean projected change in habitat suitability for each set of simulated population-specific management actions and the total area of new habitat created for targeted and non-targeted actions, as well as the percent increase in these metrics that resulted from targeting, are reported in Table 2. Targeting treatment polygons using our uncategorized heatmaps (i.e., RSF percentile change maps showing the calculated values from Eq. 1) resulted in larger overall improvements in habitat suitability compared to targeting guided by our categorized maps (improvement factor = 4.18), but slightly lower improvements in habitat created (improvement factor = 14.38; see full comparative results in Supplemental Table S2.