Habitat suitability maps are important tools for conservation planning and species investigations, but maps at resolutions that are biologically meaningful and useful for local-level decision-making are lacking for many species that use resources at fine scales. Airborne light detection and ranging (LiDAR) describes 3-dimensional structure of forests and other habitats at high resolution, which can improve habitat models for many species and facilitate investigations otherwise impossible with field data or other remote-sensing techniques. We used an information-theoretic approach and logistic regression to model habitat of red tree voles (Arborimus longicaudus) at multiple spatial scales with predictors derived from airborne LiDAR in southwestern Oregon. We developed a priori models to evaluate habitat associations and identify good models useful for high-resolution mapping of habitat suitability across a large spatial extent. Our best models that compared habitats around nest trees of red tree voles to available sites performed well (i.e., accuracy = 0.83) on an independent test data set. We found fair performance in comparisons of habitat around nest trees to unused locations within areas previously identified as red tree vole habitat (i.e., accuracy = 0.71). Red tree vole nests were often in the largest trees in the stand and away from forest edges. Both analyses suggested that fine- and broad-scale predictors are important for modeling habitat of red tree voles. Habitat suitability for red tree voles should be predicted effectively by LiDAR acquisitions even with low point densities because maps were robust to reductions in point densities as low as 1 point/m2. We found that LiDAR was effective for modeling habitat of red tree voles and expect it to improve performance of models for other species with similar habitat requirements. © 2016 The Wildlife Society.
Conservation planning and species investigations often require detailed habitat information provided by habitat modeling and suitability maps (Franklin 2009). These tools rely on habitat descriptions from field or remotely sensed data that have unique limitations in their applications (McElhinny et al. 2005, Bergen et al. 2009, Ackers et al. 2015). Field data can provide descriptions of microhabitat needed for species that select habitats at fine scales, but extrapolating field data to map habitat suitability across large spatial extents is challenging. Furthermore, multi-scale analyses important to understanding habitat requirements are rarely performed with field data because data collection is laborious (Wiens 1989, Scott et al. 2002, Seavy et al. 2009). Although modeling with remotely sensed data can facilitate multi-scale analyses and straightforward mapping of habitat suitability, the resolution and types of habitat information are sometimes too coarse for biologically meaningful analyses or local-level planning (Ohmann and Gregory 2002, Hagar et al. 2014, Ackers et al. 2015).
Airborne light detection and ranging (LiDAR) has become increasingly available in recent years and provides high-resolution measures of structure for forests and other habitats with high accuracy (Reutebuch et al. 2005, Van Leeuwen and Nieuwenhuis 2010). Airborne LiDAR has been used frequently for forest-wide inventory, and its application to wildlife habitat modeling is growing because habitat characteristics such as canopy cover and tree heights commonly measured by biologists in plots on the ground can be measured by LiDAR with similar or better accuracy (Vierling et al. 2008, Bergen et al. 2009). Airborne LiDAR also facilitates habitat descriptors that are biologically meaningful but difficult to measure from the ground (e.g., foliage height diversity (Clawges et al. 2008), edge density (Wilsey et al. 2012)). Such predictors expand the range of candidate models and may improve understanding of habitat requirements. In addition, LiDAR coverage is spatially continuous and can include entire study areas, which allows for habitat mapping at high resolution and multi-scale habitat assessments that range from tree- to landscape-level attributes.
Spatially explicit information on forest structure at high resolution is especially useful for species associated with late-seral forests in the Pacific Northwest, USA where such habitats are now rare and highly fragmented (U.S. Department of Agriculture and U.S. Department of the Interior [USDA and USDI] 1994). The red tree vole (Arborimus longicaudus) is one such species that has received attention from land managers in recent years because of apparent population declines concurrent with harvest of late-seral forests. However, little information on the ecology of red tree voles and their habitat requirements is available to inform decisions about resource management. The red tree vole is endemic to western Oregon and northwestern California, USA (Verts and Carraway 1998) where its geographic range is entirely within the jurisdiction of the Northwest Forest Plan (USDA and USDI 1994). Red tree voles are among a select group of species that require occupancy surveys prior to timber harvest or other activities that may disturb their habitat (USDA and USDI 1994: C-5). In addition, the northwest coastal population of red tree voles in Oregon has been recognized as a distinct population segment and added to the Federal List of Candidate Species under the Endangered Species Act (U.S. Fish and Wildlife Service 2011). Occupancy surveys are difficult and costly because trees with nests often require climbing to confirm use by red tree voles (Dunk and Hawley 2009, Forsman et al. 2009, Huff et al. 2012). Habitat models for red tree voles have been developed from ground measures of forest structure at forest inventory and analysis (FIA) plots, which cover 1 ha and occur at ≥2.5-km intervals within the range of the red tree vole (Dunk and Hawley 2009). Although these measures of forest structure can be used with satellite imagery to estimate forest structure between plots following gradient nearest-neighbor methods, habitat maps derived from these models are not recommended for local-level planning (Ohmann and Gregory 2002). High-resolution habitat mapping for red tree voles can aid ecological studies, conservation planning, and compliance with environmental regulations.
Our objectives were to model habitat of red tree voles with airborne LiDAR to investigate habitat associations at multiple spatial scales and identify a good model for mapping habitat at high resolution across a large spatial extent in southwestern Oregon. We also evaluated effects of using LiDAR datasets with lower point densities on model predictions and habitat maps. Finally, we modeled characteristics of nest trees based on field measures to assess similarity with LiDAR-based models for nest sites. We expected that LiDAR-based predictors for forest canopy complexity at fine spatial scales and the amounts of old forests at broad scales would effectively identify red tree vole habitat.
We modeled habitat of red tree voles on 4,927 km2 within Josephine, Jackson, and Douglas counties, southwestern Oregon, USA. Public lands administered by Bureau of Land Management (BLM) accounted for nearly half of the study area and were interspersed with private lands following a checkerboard pattern with 259-ha parcels. The study area included portions of the Cascade Range, Siskiyou Range, and lowlands of the Rogue River Valley where elevation ranged from 75 m to 2,038 m. Climate had maritime influences with warm, dry summers and cool, wet winters. Most of the study area was densely forested with a wide range of sere and forest management schemes including clear-cut harvesting and commercial thinning. Common land uses included timber production, grazing, agriculture, and recreation. Vegetation zones within the study area included Mixed-conifer, Mixed-evergreen, True fir, and Interior Valley (Franklin and Dyrness 1988). Many south-facing slopes were dominated by oak (Quercus spp.) woodlands or pine (Pinus spp.)-oak forests, whereas north-facing slopes had fir or mixed-conifer forests. Principle tree species were Douglas-fir (Pseudotsuga menziesii), white-fir (Abies concolor), Garry oak (Quercus garryana), black oak (Quercus kelloggii), and ponderosa pine (Pinus ponderosa). The red tree vole is among 11 species of arvicoline rodents that occur in southwestern Oregon (Verts and Carraway 1998). Red tree voles are commonly eaten by northern spotted owls (Strix occidentalis) in this region (Forsman et al. 2004).
Vole Surveys and Sampling
We developed habitat models from records of nest trees used by red tree voles found on 21,457 ha surveyed within our study area between 1995 and 2012. These records were accumulated by BLM primarily through pre-disturbance surveys required under the Northwest Forest Plan. Red tree voles were surveyed following standardized protocols prior to projects that could significantly diminish red vole habitat, defined as forest stands with quadratic mean diameter >40 cm and complex structural features such as multi-layered canopies typical of mature or older conifer forests of western Oregon (Biswell et al. 2002, Huff et al. 2012). Conifer forests with >60% canopy cover and ≥5 large conifers/ha with structures that can support nests also qualified as suitable areas for red tree voles. Surveyors walked transects spaced 25 m apart through forests and searched tree canopies for nest structures indicative of use by red tree voles (i.e., spherical mass of branchlets and resin ducts of fir needles). Most nest trees were climbed and searched for evidence of red tree vole occupancy. Within the study area, BLM documented 4,094 nest trees of which 3,067 were climbed and confirmed as used by red tree voles. In survey areas without nest detections, several large trees were climbed to search the upper canopy to minimize errors of false absences. Surveys followed BLM protocols for documenting nests of red tree voles to meet environmental regulations (Huff et al. 2012) and met animal welfare guidelines (Sikes et al. 2011).
We ordered the nest trees randomly and selected trees for analysis provided they had been climbed to verify use by red tree voles and no other previously selected nest tree occurred within 800 m. We maintained >800-m spacing among all samples because our largest scale of analysis had an 800-m radius and we found little or no evidence of spatial autocorrelation among samples at this scale and following these criteria (|Moran's I| < 0.07 for all predictors). We used timber harvest records from BLM to screen nests with nearby (<800 m) logging activity that occurred between the time of nest detection and the LiDAR acquisition. Based on these criteria, we found 260 nests suitable for analysis and randomly set aside 25% of these nests for a test data set to evaluate model performance.
We modeled habitat following 2 strategies because of differences in scope of inference associated with each analysis. In the first strategy (used vs. unused), we compared nest sites to a random selection of unused sites within surveyed areas without detection of red tree voles. Because these survey areas were within forests considered habitat for red tree voles based on BLM guidelines, we also used a second strategy (used vs. available) that compared nest sites to random sites selected from the entire study area to ensure the full range of forest types were represented in the analysis. We expected models from the comparison of nests to available sites to be most appropriate for mapping habitat throughout the LiDAR acquisition.
For comparisons of nest sites and unused areas, we used Geospatial Modeling Environment (GME Version 0.7.2.1, http://spatialecology.com/gme, accessed 14 Dec 2014) to randomly select 184 locations within survey polygons where no red tree voles or nests were detected. For this selection, we ensured that no other nest or unused location occurred within 800 m of each sample location. Sample size of unused locations was limited by our selection criteria and the availability of survey polygons without detections of red tree voles. Of these locations, 25% were selected at random for a test data set to evaluate model performance. For comparisons of nest sites to available sites, we randomly selected 670 locations within forested areas of the study area such that each sample was >800 m from known nest locations and other samples. Forested areas were defined by areas with return heights >3 m based on the canopy height model derived from LiDAR. We set aside 25% of these locations for evaluations of model performance.
Airborne LiDAR was acquired from 8 March to 16 August 2012 by Watershed Sciences (Corvallis, OR, USA). Pilots flew over the study area with Cessna Caravan 208B Cessna, Wichita, KS, USA) or Partenavia P.38 (Vulcanair, Naples, Italy) aircrafts equipped with Leica ALS50 or ALS60 sensors (Leica, Wetzlar, Germany). Laser pulses were emitted at 52 kHz or 47 kHz with scan angles of 30° or 28° off nadir from 900 m or 1,300 m above the ground. Average native pulse density was set at 8 points/m2, which yielded an average pulse density of 10.36/m2 with up to 4 returns recorded per pulse. Root mean square error for vertical accuracy was 0.05 m. The LiDAR acquisition covered 5,435 km2 of which we excluded 508 km2 from analyses because it was beyond the geographic range of the red tree vole.
We processed LiDAR data (LAS files) in FUSION (FUSION Version 3.1, http://forsys.cfr.washington.edu/fusion/fusionlatest.html, accessed 14 Dec 2014) to develop a canopy height model at 1-m resolution and to extract predictors based on forest characteristics described by the distribution of LiDAR returns (i.e., point cloud). We subtracted the bare earth layer from the return elevations in the LAS files to find the height above ground for each return. The canopy height model was a raster in which values for each cell represented the highest return in each square meter of the study area. We screened returns >100 m, the tallest tree known in Oregon. We then used the CloudMetrics function in FUSION to extract summaries of the point cloud characteristics around sample points for each scale of analysis.
We developed predictors at 7 spatial scales in ArcMap 10.1 (Environmental Systems Resource Institute, Redlands, CA, USA) to model fine- and broad-scale habitat associations considered important to red tree voles (Fig. 1). We predicted that probabilities of use would increase with canopy cover, canopy connectivity, canopy height, variability of canopy heights, foliage height diversity, and area of forests with large trees at fine scales. We also predicted that probability of use would increase with amounts of forests, particularly older forests, at broad-scales because habitat amount should correlate with population viability. Furthermore, red tree vole nests tend to occur in clusters, which may reflect their limited dispersal ability and suggests that minimum thresholds of habitat amounts may be needed to support a population, especially within isolated habitat patches. We expected that older forests would support populations of red tree voles and that coverage by forests of any sere >3 m tall in surrounding areas would affect probabilities of use positively by aiding dispersal and connectivity. Finally, we expected nest occurrences to decrease with density of forest edges at all spatial scales of analyses because forest fragmentation has been considered a threat to population viability.
For fine-scale analyses (i.e., 12.5-m, 25-m, 50-m radii), we extracted the mean, standard deviation, and maximum height from the canopy height model around each sample location. We excluded heights <3 m from calculations of these predictors to ensure metrics characterized canopy only. Because red tree voles travel among trees via interconnected branches, predictors that describe canopy connectivity should be useful for modeling habitat for this species. The resolution of structural information on forest canopies provided by LiDAR should be sufficient to describe levels of canopy connectivity important to red tree voles, but no such metric has been tested in previous studies. We developed indices of upper and lower horizontal canopy connectivity based on the areas of canopy above 20 m and 10 m. For each index, we reclassified the canopy height model and converted it to canopy and non-canopy polygons based on these height breaks. We defined core areas as the areas that remained after subtracting the outer 3 m of each canopy polygon. We calculated the horizontal connectivity index as the ratio of the areas for core and canopy polygons after summing their areas within each plot (Smith et al. 2011). We calculated canopy cover as the ratio of first returns above 3 m to all first returns. We calculated foliage height diversity based on percent cover measures within 5-m strata that spanned from 3 m above the ground to the top of the highest canopy in the data set. We estimated cover within each stratum as the ratio of returns within the stratum to all returns within and below the stratum. We divided the cover estimate for each stratum by the sum of the cover estimates to find the cover proportion (pi) for the common index of foliage height diversity (Clawges et al. 2008, Bergen et al. 2009):
For all scales of analysis, we estimated edge density and the areas of forests with heights above 3 m, 40 m, and 50 m. We calculated areas of forest with tree heights above specified breaks through reclassifications of the canopy height model. To calculate edge density, we first reclassified the canopy height model to delineate forest edges at the intersections of forest (heights >3 m) and non-forested (heights <3 m) areas. Next, we extracted forest polygons >900 m2 and erased forest gaps <900 m2 to minimize contributions of small stands and openings to edge density. This criterion corresponded to the home range of the red tree vole (Swingle and Forsman 2009) and effectively eliminated many polygons that represented isolated trees and small forest gaps that do not function as forests or edges. Then, we converted the forest polygons to line features and summed their lengths within each plot to find the edge density (m/ha).
Model Development and Selection
We developed a priori models based on previous studies of red tree voles (Gillesberg and Carey 1991, Meiselman and Doyle 1996, Dunk and Hawley 2009) and followed an information-theoretic approach for model selection (Burnham and Anderson 2002). We used logistic regression to model the relative probability of use by red tree voles as a function of habitat variables extracted from LiDAR (Hosmer and Lemeshow 2000). We performed all analyses with R software (R Version 2.15.1, http://www.R-project.org/, accessed 14 Dec 2014). Prior to formal modeling, we examined boxplots and univariate models to assess the need for transformations and squared terms. We used squared terms for elevation, slope, and aspect in a priori models when quadratic relationships were evident in univariate analyses. Then, we fit models for each scale of analysis and ranked them based on Akaike's Information Criterion (AIC) in a 3-stage analysis (Burnham and Anderson 2002). This analysis strategy allowed us to identify good models from many variables that represented habitat characteristics at multiple spatial scales. It is an extension of a variable selection procedure that evaluates related variables in model subsets (i.e., stages) to identify important variables for subsequent use in inferential or predictive models (Ramsey and Schafer 2002, Johnston and Anthony 2008). Because some variables were correlated within and across spatial scales of analysis, we divided the variables into groups that represented local habitat, broad-scale habitat, and site characteristics independent of spatial scale (e.g., elevation). Each analysis stage evaluated one of these groups, and we used the best models from each stage as base models in the subsequent stage. We identified inferential and predictive models in the third analysis stage where we evaluated broad-scale variables after accounting for local habitat and site characteristics independent of spatial scale.
In the first stage, we fit fine-scale models with predictors extracted from 12.5-m, 25-m, and 50-m radii. These models represented hypotheses about the importance of canopy cover, tree size, canopy complexity, and edge effects. High correlations (|r| > 0.5) among some variables led us to test 13 alternative full models so that all predictors could be considered in the analysis. Each model had 2 or 3 predictors, many of which represented alternative ways of testing similar hypotheses about forest structure. Highly correlated predictors emphasized somewhat different forest characteristics and explained different amounts of variation in the response. For example, maximum canopy height and foliage height diversity discriminated young from mature forests and were highly correlated, but the latter may have greater explanatory power for habitat modeling by conveying additional information on structural complexity of the canopy. This approach allowed evaluation of correlated variables and helped with finding a good predictive model for habitat mapping. In addition, we tested 26 models with 1 or 2 predictors that were subsets of these full models for each spatial scale.
In the second stage, we used competing fine-scale models (ΔAIC < 2) as base models for testing effects of slope, aspect, elevation, and spatial coordinates. We tested all possible combinations of these variables with the base model (26 models). We tested interactions between spatial coordinates in competing models because of their importance reported in previous models for red tree voles (Dunk and Hawley 2009). In the third stage, competing models from the second stage formed base models for broad-scale analyses with predictors for edge density and forest area by height class (i.e., >3 m, >40 m, and >50 m) extracted from 100-m, 200-m, 400-m, and 800-m radii. We tested all possible combinations of these predictors (11 models) for each spatial scale. These predictors tested hypotheses about the importance of habitat quality in surrounding forests beyond the home range of the red tree vole (Swingle and Forsman 2009) after accounting for fine-scale forest structure, topography, and spatial context. We used test data sets from the used versus unused and unused versus available analysis strategies to evaluate performance of the best models based on 1) area under the receiver operating curve (AUC; Fielding and Bell 1997), 2) Cohen's kappa (Cohen 1960, Manel et al. 2001), and 3) classification success rates. We calculated these statistics in R packages ROCR (Version 1.5, https://CRAN.R-project.org/package/web/packages/ROCR/index.html, accessed 14 Dec 2014) and PSYCH (Version 1.4.8, http://CRAN.R-project.org/package=psych, accessed 14 Dec 2014). We expected performance to be higher in evaluations based on the test data set corresponding to the model's analysis strategy compared to that of the alternative strategy (i.e., used vs. available was the alternative strategy to used vs. unused and vice versa).
We averaged, ranked, and constructed 95% confidence intervals for unstandardized coefficients with multi-model inference on their relative association to red tree voles (Burnham and Anderson 2002). We averaged coefficients for predictors at fine scales from all models in the first stage of analysis. We averaged coefficients for topographic and spatial predictors from all models in the second analysis stage that included predictors from the best fine-scale model. We averaged coefficients for broad-scale predictors from all models in the third analysis stage that included predictors from the best model that combined fine-scale, topographic, and spatial variables. We averaged coefficients of predictors and ranked them based on the sums of Akaike weights for models that included the predictor of interest. Coefficient estimates were consistent in sign and magnitude across models, which indicated that any effects of multicollinearity on averages were minimal (Cade 2015). We constructed boxplots of AIC values from each subset of candidate models to evaluate effects of spatial scale in the analyses.
Some of our predictors required analysis of LiDAR point clouds (e.g., foliage height diversity) or substantial manipulation of canopy height models (e.g., edge density) that would require considerable computing power and time to map throughout a large extent. We performed similar model selection and evaluation procedures on a restricted set of a priori models for comparisons of used and available sites to identify a good model with predictors easily extracted and mapped from products commonly delivered by LiDAR vendors (e.g., canopy height models, digital elevation models). We considered only models at spatial scales present in the best or competing models (i.e., 12.5-m, 100-m, and 200-m radii) from the previous analyses. We used the best model from this analysis to map relative probability of use throughout the study area. We repeated this model selection procedure after excluding models with spatial coordinates as predictors to report a good model independent of spatial patterns in red tree vole distribution within the study area. Such a model may be more appropriate for extrapolation beyond the study area.
We assessed sensitivity of model predictions to LiDAR point densities of 8, 4, 2, and 1 pulses/m2 on 976 ha within our study area. First, we used the ThinData function in FUSION to reduce average point densities and then created new digital elevation and canopy height models at 1-m resolution for each level of point density. We mapped values of predictors included in the best models for comparisons of used and available sites and calculated relative probability of use by red tree voles throughout the selected area. We reclassified probabilities as suitable or not suitable for red tree voles based on the threshold probabilities that maximized accuracy in the performance evaluations. For the evaluation, we used kappa, accuracy, and percent overlap in red vole habitat to measure similarity in habitat classification between predictions based on full (10.36/m2) and reduced point densities. We also measured correlations and AUC between predicted values estimated at full and reduced point densities. For AUC calculations, we used the suitable and not suitable classifications defined at the full point density as the binary value predicted by the continuous probabilities of occurrence estimated at reduced densities. Because reduced point densities may bias some predictors, we adjusted threshold probabilities to maximize similarity in predictions of suitable and not suitable areas between full and reduced point densities. We then recalculated kappa, accuracy, and percent overlap in red tree vole habitat based on the adjusted threshold probabilities.
Nest Tree Analysis
We compared field-based measures of nest trees used by red tree voles to nearby trees without nests to assess consistency with results of analyses with LiDAR metrics. These data from 120 nest sites within our study area were collected for the purpose of LiDAR calibration and included measures of diameter at breast height (dbh), tree height, and height of lowest living crown for all trees within 10 m of the nest tree. We also used an index of relative tree height based on the ratio of each tree's height to the median height of trees in the plot. For analysis, we randomly selected 98 nests from the data set to avoid use of overlapping plots. We used conditional logistic regression to compare nest trees to their neighboring trees and ranked models based on AIC (Hosmer and Lemeshow 2000). The analysis was restricted to Douglas-fir trees because they commonly supported nests and were structurally distinct from other species, especially deciduous trees. Because all predictors represented tree size and were highly correlated (r > 0.5), we tested only univariate models. We expected probabilities of use for red tree voles to increase with each of these predictors.
The best model for comparisons of used and available sites (UA1 model) performed well based on evaluation with the test data set (Table 1). The UA1 model was better at classifying available sites than used sites. This model performed similarly to the best model from the restricted model set for comparisons of used and available sites, which excluded predictors that were complicated to extract from LiDAR including foliage height diversity, edge density, and canopy connectivity (UA2 model). Although performances of the UA1 and UA2 models were similar, UA2 was not a competing model (ΔAIC = 9.52). The best model from the restricted model set that excluded complex LiDAR predictors and spatial coordinates as predictors (UA3 model) performed decently but not as good as models that accounted for spatial effects. The performance of the best model for comparisons of used and unused sites (UU model) was fair. This model was better at classifying used sites than unused sites.
|Area under curve||0.871||0.883||0.836||0.644||0.458||0.643||0.656||0.754|
None of the models performed as well when applied to the test data set of its alternative analysis, which suggests parameterization depended on scope of inference. That is, coefficient estimates differed between analyses that considered only mature forests versus those that include all forest sere within the study area. When applied to the test data set for the used versus unused strategy, the UA1 model performance was poor indicating it had little discriminatory power within forests considered suitable for red tree voles by BLM guidelines. Overall, the UU model had fair performance when applied to the test data set for the used versus available strategy. The UU model's specificity was much higher than that of the test data for the used versus unused strategy; however, sensitivity was much lower. This suggests that both models categorize poor habitat effectively across the entire study area but struggle to distinguish use in areas considered habitat for red tree voles by BLM.
Best Models and Predictors
The majority of predictors and their relationships with nest occurrences were similar between the UA1 and UU models (Table 2). Both models had foliage height diversity and edge presence as predictors at the finest scale of analysis. Despite their occurrence in relatively few candidate models, these predictors had considerably higher Akaike weights than all others in the fine-scale analysis (Table 3). Notably, foliage height diversity was highly correlated (r > 0.8) with several predictors including mean, maximum, and standard deviation of canopy heights yet had much greater explanatory power. Foliage height diversity acted as composite of forest height and canopy cover that scored older forests with evenly distributed canopies higher than others, which may explain its superior performance (Fig. 2). Red tree vole nests occurred in areas with relatively high foliage height diversity, even within forests considered red tree vole habitat by BLM guidelines. This suggests that red tree voles occupied areas with the largest trees and most complex canopy structures relative to others in a stand.
|Foliage height diversity||12.5||Log/nonea||5.85||0.799||2.46||0.447|
|Mean canopy height||12.5||Log||2.12||0.306||1.94||0.284|
|Edge density (m/ha)||12.5||Presence||−0.703||0.387||−0.690||0.446|
|Northing × Easting (UTM)||3.30e−10||1.71e−10||9.60e−10||2.70e−10|
|Forest area >50 m tall (ha)||200||Presence||1.61||0.391||0.902||0.478|
|Forest area >40 m tall (ha)||100||Presence||2.46||0.611||2.62||0.605|
|Forest area >3 m tall (ha)||200||0.280||0.142|
|Edge density (m/ha)||200||None/log||2.40e−3||9.23e−4||0.247||8.24e−2|
|Mean canopy height (m)||12.5||Log||4.50e−7||2.35||0.267||*||2.16e−3||2.15||0.380||*|
|Maximum canopy height (m)||12.5||Log||1.68e−6||2.67||0.311||*||3.37e−2||2.94||0.525||*|
|SD canopy height (m)||12.5||Log||3.05e−6||1.83||0.225||*||0.200||1.82||0.369||*|
|Foliage height diversity||12.5||Log/none||0.998||6.29||0.703||*||0.709||2.48||0.398||*|
|Lower canopy connectivity||12.5||Logit/arc||3.05e−6||1.36||0.300||*||0.222||0.371||9.48e−2||*|
|Upper canopy connectivity||12.5||Logit/arc||1.08e−7||0.483||5.60e−2||*||3.73e−3||0.453||8.59e−2||*|
|Forest area >40 m (ha)||12.5||Presence||3.83e−11||1.39||0.189||*||2.70e−2||1.18||0.250||*|
|Forest area >3 m (ha)||12.5||5.58e−7||9.09||18.7||1.48e−2||21.9||23.1|
|Forest area >50 m (ha)||12.5||Presence||6.83e−23||0.311||0.219||3.78e−6||0.730||0.330||*|
|Edge density (m/ha)||12.5||Presence||0.602||−0.533||0.327||0.329||−0.324||0.414|
|Canopy cover (%)||12.5||Arc/none||1.02e−8||0.508||0.619||2.73e−3||1.51||0.958|
|Forest area >3 m (ha)||200||0.165||1.15||1.48||0.533||0.279||0.146|
|Forest area >40 m (ha)||200||Log||1.75e−3||2.25||1.06||*||0.160||0.600||0.428|
|Forest area >50 m (ha)||200||Presence||0.434||1.57||0.395||*||0.431||0.929||0.480|
|Edge density (m/ha)||200||None/log||0.398||2.73e−3||1.14e−3||*||0.709||0.230||8.92e−2||*|
|Easting × Northing (UTM)||0.331||2.55e−10||1.63e−10||0.998||9.76e−10||2.72e−10||*|
At the finest scales, the best models for both analyses indicated red tree voles avoided forest edges. As the scale of analysis increased, the edge effect dissipated and then became positive at the broadest scales. The best models included negative associations with edge presence within a 12.5-m radius and positive associations with edge density within 200 m. The positive association with edge density at broad scales appeared to be related to the frequent occurrence of older forests as small fragments within landscapes of industrial timber production.
The UU and UA1 models also shared the predictor for occurrence of trees with heights >50 m within a 200-m radius. This predictor was transformed from a continuous measure of forest area with tree heights >50 m because of strong right skewness. On the original scale, the mean for available sites exceeded that of the used sites (Table 4) because of a few available sites with unusually high values. Following transformation, the frequency of occurrence for trees with heights >50 m within a 200-m radius was much higher at used sites (95%) than available (64%) and provided high explanatory power in the analysis.
|Foliage height diversity||12.5||1.80||2.09e−2||1.53||3.21e−2||1.32||2.52e−2|
|Mean canopy height (m)||12.5||25.6||0.524||20.6||0.301||18.2||0.395|
|Edge density (m/ha)||12.5||41.5||12.7||112||29.5||152||17.0|
|Forest area >50 m tall (ha)||200||0.195||2.34e−2||0.107||1.86e−2||0.210||2.28e−2|
|Forest area >40 m tall (ha)||100||0.280||2.14e−2||0.148||1.99e−2||0.174||1.40e−2|
|Forest area >3 m tall (ha)||200||11.1||8.06e−2||10.9||0.114||10.4||9.48e−2|
|Edge density (m/ha)||200||119||7.77||110||9.26||138||6.72|
Predictors for topography and spatial coordinates were included in each of the best models. Probabilities of use increased with elevation in the UU model and had a quadratic relationship in the UA1 model. The inclusion of the squared term in the UA1 model likely reflected greater sample size and scope of inference because the samples were drawn from a wider range of elevations. The UA1 model had the spatial coordinate for northing as a predictor, whereas the UU model included its interaction with the easting coordinate. Differences in effects of spatial coordinates between analyses were related to the distribution of absences across the study area. Relatively few areas in the northeastern portion of the study area were surveyed to determine absences of red tree voles, whereas the distribution of available sites for the used versus available analysis was fairly uniform throughout the study area. Unlike the UU model, the UA1 model included quadratic relationships with slope, an effect that was likely detected because of greater sample size and scope of inference. Unlike the UA1 model, the UU model included a predictor that indicated nest occurrences increased with forest area (i.e., trees >3 m tall) within a 200-m radius.
The UA2 model indicated similar habitat associations as those of the UA1 and UU models (Table 2). Mean height of canopy was the only fine-scale predictor, and presence of trees >40 m tall within the 100-m radii was the only broad-scale predictor. Like the UA1 model, the UA2 model included spatial coordinates, slope, and elevation. The UA3 model was similar to UA2 but excluded Universal Transverse Mercator (UTM) coordinates as predictors. Aside from the intercept, parameter estimates for the UA3 model were similar to those of UA2.
Most predictors were useful for discriminating used and unused or available sites (Table 3). Coefficient estimates were consistent across models within each analysis, and effect sizes were similar between the model averages and the top models. At the finest scale, all predictors related to tree size had model-averaged coefficients larger than 0 based on 95% confidence intervals. The odds of nest occurrence increased 1.9 times with each 0.1-unit increase in the natural log of foliage height diversity within a 12.5-m radius (95% CI = 1.6, 2.2) based on the model-averaged coefficient from the used versus available analysis. Predictors for upper and lower canopy connectivity also had confidence intervals above 0. Canopy cover was highly correlated (r > 0.7) with lower canopy connectivity, but its confidence interval overlapped 0, likely because young forests may have high canopy cover and low connectivity above the height breaks used for connectivity estimates. Although the confidence interval for edge presence within a 12.5-m radius overlapped 0, this predictor was included in the best models for both analyses and had high Akaike weights, which indicated the importance of local edge effects for modeling red tree vole habitat. The odds of a nest occurring within 12.5 m of a forest edge were 0.59 times those of areas without edges (95% CI = 0.31, 1.1) based on the model-averaged coefficient from the used versus available analysis.
At broader scales, model-averaged coefficients for area or presence of tall trees within 200 m were larger than 0 based on 95% confidence intervals in the analysis of used and available sites only. This analysis indicated that occurrence of trees >50 m tall within 200 m increased the odds of nest occurrence 4.8 times (95% CI = 2.2, 10). Although the confidence intervals for all forest area predictors overlapped 0 in the used versus unused analysis, their inclusion in the best model indicated their importance to red tree voles. Confidence intervals for effects of edge density within 200 m were also above 0, consistent with their inclusion in the best models. The odds of nest occurrence increased 1.3 times with each 100-unit increase in edge density within 200 m (95% CI = 1.1, 1.6) based on the used versus available analysis. Predictors for elevation, slope, and spatial coordinates that appeared in the best models were the only topographical and spatial predictors with confidence intervals that did not overlap 0.
For fine-scale analyses, there were large differences in AIC between the best models for the finest scale versus those with 25-m and 50-m radii (Fig. 3). Explanatory power of the models appeared to increase linearly with the resolution of analysis in this subset of models. For broad-scale analyses, predictors with a 200-m radius were more effective than those at other scales. However, predictors with 100-m radii were present in some competing models in the analysis of used and available sites (Table 5). Cross-scale correlations (r > 0.8) among predictors at similar scales were evident and likely explain the competitive performance of predictors with 100-m radii. For example, edge densities extracted from 100-m and 200-m radii were highly correlated. We did not find cross-scale correlations between the fine- and broad-scale predictors used in the broad-scale analyses.
Good predictive power at the finest scale of analysis permitted habitat mapping at high resolution with the UA2 model (Fig. 4). Large differences in predicted use corresponded to sharp transitions in structure from young to old forests in the canopy height model. Variations in predicted use also followed subtle differences in canopy structure and forest gaps within old forests. Young forests within 100 m of old forests had higher probabilities of use relative to young forests elsewhere because of the broad-scale predictor for presence of trees >40 m tall within a 100-m radius.
We mapped habitat suitability on 492,675 ha and defined 59,338 ha as red tree vole habitat based on probabilities above the threshold value that maximized accuracy in the performance evaluation. Lands managed by BLM accounted for 48% of the study area but supported 81% of the habitat for red tree voles. Private lands accounted for 50% of the study area and supported 15% of red tree vole habitat. Remaining habitat for red tree voles was on lands administered by the State of Oregon or United States Forest Service.
Habitat maps based on the UA1 and UA2 models were robust to reductions in LiDAR point densities (Fig. 5). Measures of similarity between maps derived from the full versus reduced point densities yielded values for correlation, kappa, AUC, accuracy, and overlap of red tree vole habitat ≥0.95 in nearly all cases. Prior to adjustment of the threshold probabilities, some minor departures were evident with maps based on the UA2 model at 2 points/m2 and 1 point/m2; their kappa values were 0.85 and 0.73, respectively. Although these values indicate high similarity, they were relatively lower than the kappa values from higher point densities, which suggested some sensitivity of predictions to reduced point densities. However, all measures of similarity were high (≥0.93) following adjustment of threshold probabilities, which indicated maps were nearly identical in their depiction of relative probabilities of use. Mean height of canopy and the area of forests >40 m tall increased with point density, which likely explains the improvements following adjustment of threshold probabilities. With fewer LiDAR points, high branches were less likely to be detected, resulting in slightly lower values for these predictors and probabilities of use in the map based on the UA2 model. In contrast, similarity was high regardless of point density for maps based on the UA1 model because reductions in point density had little effect on foliage height diversity, edge density, elevation, and slope.
Nest Tree Analysis
Analyses of field data indicated nest trees were larger than available trees, consistent with LiDAR-based models (Table 6). All coefficients for predictors indicative of tree size were positive and greater than 0 based on 95% confidence intervals. Diameter at breast height had far greater explanatory power than all other variables based on AIC and an Akaike weight >0.999. Tree height was the second best predictor, but ΔAIC was 40.9, which indicated dbh was better at relating tree characteristics important to red tree voles.
|Tree height (m)||3.89||0.453||40.9||1.28e−9|
|Relative tree height||6.15||0.712||42.1||7.36e−10|
|Lowest living crown height (m)||3.28||0.401||88.0||7.92e−20|
Habitat modeling with airborne LiDAR was effective for red tree voles because it provided spatially explicit measures of forest structure at high resolution. Our models indicated that red tree voles were associated with older forests consistent with our expectations and previous habitat studies (Gillesberg and Carey 1991, Meiselman and Doyle 1996, Dunk and Hawley 2009). The ability of LiDAR to accurately measure tree heights, canopy cover, and other forest attributes allowed us to construct predictors that described or correlated with the complex forest structure characteristic of red tree vole habitat. Much of the variation in our data was explained by the distinction of mature from young forests based on the canopy height model. In addition, LiDAR provided high-resolution information on structure within older forests that increased explanatory power of the models. Our exploration of LiDAR-derived predictors yielded novel or unconventional predictors that captured forest characteristics important to red tree voles. In some cases, these predictors outperformed others and were included in our best models.
Foliage height diversity explained more variation in nest occurrences than other predictors in the fine-scale analysis. This predictor was likely most effective because it differentiated structures of mature and older forests better than others. Foliage height diversity increased with the number of height strata that had LiDAR returns, which indicated presence of canopy. This index increased further with evenness of cover estimates among strata, which conveyed the amount and complexity of canopy. Accuracy and precision of sub-canopy cover estimates from airborne LiDAR are probably lower compared to estimates for upper canopy cover because of attenuation of LiDAR returns (Andersen et al. 2003). However, the effects of inconsistencies in cover estimates on foliage height diversity are probably small relative to the contributing number of strata with cover estimates and the range of values expressed on a landscape. Even under dense canopy, airborne LiDAR has been effective at estimating sub-canopy and shrub cover (Goodwin 2006, Hill and Broughton 2009, Martinuzzi et al. 2009, Johnston 2013). Our assessment of maps generated at varying LiDAR point densities suggested that foliage height diversity was robust to low point densities and had sufficient precision for effective modeling of red tree vole habitat.
Our novel predictors for canopy connectivity had good explanatory power but were not included in our best models. These predictors were exploratory and need further evaluation and refinement to ensure their representation of interconnected canopies that facilitate arboreal travel for red tree voles. Canopy connectivity indices based on the number of trees with branches within <1 m of focal trees have been effective for modeling habitat of gray squirrels (Sciurus spp.; Gregory et al. 2010, Johnston 2013). These field-based measures of canopy connectivity had low correlation with canopy cover readings from densiometers, which suggested that connectivity described important subtleties of canopy structure. It is unclear whether our novel predictors captured these subtleties, especially because of high correlations between our canopy connectivity and cover estimates. Validations of canopy connectivity indices derived from LiDAR were not possible with our data set but would be helpful for modeling habitat of arboreal species.
Edge density was estimable from simple manipulations of the canopy height model derived from LiDAR. Such measures would be difficult or impossible from ground measures or other remotely sensed data, yet edge effects were included in our best models. Our finding of avoidance of edges close to the nest was consistent with hypotheses of fragmentation effects on red tree voles. Dissipation of negative edge effects at increasing spatial scales probably reflected the small home range of red tree voles (Swingle and Forsman 2009) because edges >25 m from the nest are unlikely to occur within areas used by the occupying vole. We were surprised to find positive effects for edge density at broad scales given the small home range, limited dispersal abilities, and old forest associations of red tree voles. We suspect that these positive associations with edge density were an artifact of the landscape context of older forests in this region rather than an indicator of biological importance. In our study area, edge density was high in areas of industrial forestry where older forests exist as small, isolated patches. Habitat connectivity modeling could improve understanding of broad-scale effects of habitat fragmentation on red tree voles (McRae et al. 2008).
Our best models included variables from fine and broad scales demonstrating the importance of modeling at multiple spatial scales. As expected, our results suggest that fine-scale habitat features within 12.5 m of the nest are important to red tree voles. This area covers about half of the typical home range for red tree voles and likely includes areas of highest use by the occupying vole (Swingle and Forsman 2009). Because red tree voles have limited dispersal abilities and nests are commonly clustered within stands, they probably require sufficiently large habitat patches to support multiple individuals for population persistence at the local level. Broad-scale predictors in the best models likely reflected these population requirements and suggest that habitat quality within 200 m of the population center is important to red tree voles.
Performances of the UA1 and UA2 models were similar to the best model reported by Dunk and Hawley (2009), which was developed from field measures at 1-ha FIA plots throughout the range of the red tree vole. Predictors and coefficients from this model were also consistent with our findings; they indicated probability of use increased with forest age and tree size based on basal area of large trees, maximum dbh, and standard deviation of conifer dbh. Similar to ours, this model also included slope and spatial coordinates, which indicated that topography and spatial context were important for habitat modeling.
LiDAR Strengths and Weaknesses
Our analysis of nest tree characteristics from field measures revealed that dbh had much greater explanatory power for use by red tree voles than tree height, despite high correlation. Tree height likely distinguishes young from mature trees effectively, but important structural differences between mature and old trees are probably best distinguished by dbh. Broken tops and other structural anomalies common among older trees can diminish tree height but do not affect dbh. Although we suspect our predictors based on canopy height models and point cloud distributions would have high correlation with the dbh-based predictors modeled by Dunk and Hawley (2009), dbh-based predictors may have greater explanatory power than our height- and cover-based predictors. Predictors of basal area or dbh can be estimated by LiDAR, but we did not test any because of their allometric derivation from canopy heights.
Habitat models derived from LiDAR provide significant advantages over existing models constructed from ground data or imputation techniques (Hagar et al. 2014, Ackers et al. 2015). Because LiDAR measures are continuous across the areas of acquisition, the LiDAR-derived habitat model can be used directly to estimate suitability throughout the entire area. Although the model from Dunk and Hawley (2009) performed well, mapping habitat suitability is less straightforward because its ground-based measures are only available within 1-ha FIA plots that occur at ≥2.5-km intervals. Habitat suitability for red tree voles has been mapped with forest structure data derived from FIA plots and gradient nearest-neighbor techniques (Ohmann and Gregory 2002, Institute for Natural Resources 2011). This imputation procedure estimates forest structure between FIA plots based on a model that relates structure at FIA plots to 2-dimensional remotely sensed data. Resolution of the habitat map (∼100 m) is coarse for red tree voles because of their small home range sizes (<0.2 ha; Swingle and Forsman 2009) and is not suitable for local-level planning. High-resolution habitat maps derived from LiDAR can match the scale of habitat use by small species and thereby aid conservation planning, population studies, and connectivity modeling at biologically meaningful scales.
Scope of Inference and Model Limitations
Habitat suitability maps are valuable tools for conservation planning, but habitat modeling with existing data requires careful consideration of sampling design, model biases, and scope of inference (Franklin 2009). As in this study, wildlife occurrences are often collected for purposes other than habitat modeling, which can restrict some types of inference and invoke biases in the results. Occurrences of red tree voles in this study were largely collected to meet regulatory requirements for survey of rare species prior to timber harvest. Survey areas were not randomized throughout the study area and were focused on mature and late-seral forests. Consequently, our models predict relative rather than true probability of use (Manly et al. 2002), and the use of young forests by red tree voles may be underestimated. We suggest any biases are likely small because BLM's guidelines for surveying red tree voles were based on results of previous habitat studies, which showed strong associations with late-seral forests, and actual use of earlier sere by red tree voles is likely extremely low. Furthermore, magnitudes of bias likely vary by spatial scale of analysis. Fine-scale analyses were likely the least biased because forest gaps and patches of young forest occurred within survey areas and had good representation in the sample of unused sites. At broader scales, areas of mature and late-seral forests were high in both used and unused samples compared with the available sample because survey areas were in older forests. Substantial field work would be required to obtain a similarly large data set on red tree vole occurrences following procedures that yield unbiased coefficients and true probability of use. Additional field efforts to validate and assess biases in our models are likely more efficient than collecting a new data set.
Performance of our models may be improved by controlling for some potential sources of error. Contamination in analyses of used and available sites can be problematic when absences at available sites are not confirmed in the field. Because red tree voles are rare and have small home ranges, we think contamination in our analysis was low. Contamination may also arise from failures to detect nests during surveys. Although some discrepancy in abundance of red tree vole nests has been found between ground and tree-climbing efforts, contamination has been low for detections of nest presence (Dunk and Hawley 2009, Forsman et al. 2009).
More significant error in our modeling probably comes from the assumption that available sites were actually available to red tree voles. Because of their limited dispersal abilities, red tree voles were probably absent in some suitable areas because of significant isolation from extant populations. We attempted to account for some of this error by modeling at broad scales, but we suspect this factor remained a primary contributor of error, particularly for the fine-scale analysis. Assurance of independence among samples and true availability to extant populations is problematic for modeling broad-scale effects on species with limited dispersal capabilities. Historically, most of the study area probably was available to red tree voles because it falls within the geographic range of the species and significant loss of late-seral forest to timber harvest has been recent. Good performances by our models suggest problems of violating the availability assumption were not serious.
Results of this study supported our concern that scope of inference has important consequences for model parameterization. Although the UU model had fair performance when applied to the test data sampled from the entire study area, the UA1 model performed considerably better despite having a similar set of predictors. Differences in parameterization and performance between the UA1 and UU models probably arise from the lack of samples for young forests in the training data for the UU model. Therefore, we consider the UA1 and UA2 models most appropriate for mapping habitat suitability throughout the study area.
Our models and associated habitat suitability maps are valuable tools for resource managers. The spatially explicit information on the amounts and distribution of habitat for red tree voles provided by these maps can inform efforts that prioritize areas for habitat conservation or assess effects of resource management. This information is also useful for quantifying the amounts of red tree vole habitat across land ownerships and management agencies. Most of the habitat in our study area was located on BLM lands, which underscores their role in maintenance of older forests for conservation of red tree voles. These maps also can aid planning for environmental compliance surveys by identifying sites that may support red tree voles. Furthermore, maps can be used to evaluate habitat suitability and connectivity at local and regional scales or to formulate hypotheses about habitat requirements and population dynamics. As LiDAR acquisitions are repeated over time, changes in habitat distribution and connectivity can be assessed. For example, a fire burned a significant portion of our study area following the LiDAR acquisition used in this study. Models from this study can be applied to the post-fire LiDAR acquisition to determine where and how habitat of red tree voles was affected by the fire. Because of the high spatial resolution provided by LiDAR, habitat changes can be assessed within individual forest stands and across landscapes. Consistent with previous studies, our models indicate strong associations of red tree voles with complex canopy and structures characteristic of old forests in western Oregon. These models also suggest that maintenance of forest cover, particularly older forests, within 200 m of red tree vole nests is important to their persistence. We recommend further validation of our models in their application beyond our study area where LiDAR and similar data for occurrences of red tree voles exist. Good performance of these or other models at similarly high resolution can inform resource managers and scientists about the distribution of habitat for red tree voles at biologically meaningful scales.
We thank G. McFadden of the Oregon State Office of BLM for support of this project. J. Reilly of BLM procured data and shared insight on vole survey protocols. We thank all researchers that contributed to collection of the red tree vole nest locations. A. Kazakova provided guidance on LiDAR post-processing. W. Simpson and his crew collected field data at nest trees. J. Hagar and D. Rosenberg provided helpful reviews of the manuscript. This study was funded by the Bureau of Land Management through the Precision Forestry Cooperative at the University of Washington.
Associate Editor: Jeff Bowman.