Hooded Warbler
Ecoregional Scale Conservation Planning
spacer
 
   


Made possible through a partnership with the National Wetlands Research Center

Methods

Avian Models

We developed HSI models for 40 priority bird species in the Central Hardwoods and the West Gulf Coastal Plain/Ouachita Mountains BCRs. We identified priority bird species as silvicolous species with a total Partners in Flight regional combined score ≥ 20 or species designated as a Bird of Conservation Concern by the U.S. Fish and Wildlife Service in either BCR. To develop HSI models, we first performed a thorough literature review to identify site- and landscape-scale habitat factors that affected the occupancy, density, and/or productivity of each species. Empirical data derived from these sources formed the basis for individual suitability functions. We combined individual suitability functions in biologically meaningful ways to produce overall habitat suitability estimates for density and productivity. Once initial models were developed, we solicited reviews from 2-5 experts for each species and revised models based upon reviewer comments.

GIS Data
Landscape Variables

We constrained potential model variables to those available via nationally-consistent geodatasets to maintain a uniform classification system across state boundaries within BCRs and to assure our methodology was easily transferable to other forested biomes. We selected 5 nationally-available geodatasets to define landscape conditions: ecological subsections, the National Land Cover Dataset (NLCD), the National Elevation Dataset (NED), the National Hydrography Dataset (NHD), and the State Soil Geographic Database (STATSGO)

Our map of ecological subsections was based on the National Ecological Unit Hierarchy (Keys and Carpenter 1995), which depicts relatively homogenous regions of topography, geology, climate, and potential natural communities. Therefore, we assumed subsection boundaries would capture a large amount of the variation in the broad-scale abiotic features that affect the composition and structure of the avian community within a BCR.

We used the NLCD 1992 to define the spatial location of forests and categorize forestlands into broad classes. NLCD 1992 delineates 21 landcover classes at 30-m resolution; 7 of these classes represent wooded landcover types that we used to define specific avian cover types: transitional, deciduous, evergreen, mixed, shrubland, orchard/vineyard, and woody wetlands (Vogelmann et al. 2001). Additionally, we included low-density residential as a forested landcover to capture the suburban shade tree habitats that are used by some priority species (e.g., orchard oriole [Icterus spurius]).

Landforms (e.g., ridges, valleys, etc.) are local topographic features that can have a profound effect on both the flora and fauna of a forest community. Because no nationally consistent dataset exists for this feature, we created our own classification from the nationally-available NED, which maps elevation in meters at 30-m resolution (Gesch et al. 2002). We generated a landform geodataset from 5 NED-derived variables: relief, slope, aspect, local topographic position index (TPI), and landscape TPI. We separated areas of high and low relief by examining the standard deviation (SD) of elevation values within a 500-m radius moving window. We considered areas with a SD ≥ 2 to be low relief and areas with a SD = 2 to be high relief. We used a 5 percent threshold to separate high slope and low slope locations. We defined high-exposure (drier) slopes as those with aspects between 157.5 and 292.5 degrees (i.e., SSE to WNW) and all other aspects as low-exposure (moister) slopes. Areas lacking aspect were placed in a third category (flat). Derivation of TPI was based on a protocol developed by Jenness Enterprises (Jenness 2006), where the elevation at a pixel is compared to the mean elevation within a user-defined neighborhood. We calculated 2 separate TPI functions to highlight both local (500-m neighborhood) and landscape (1500-m neighborhood) effects and categorized the resulting spatial products into 3 classes: >1 SD above the mean, >1 SD below the mean, and within 1 SD of the mean. We defined 6 landform classes (floodplains, valleys, mesic slopes, terraces, xeric slopes, and ridges) based on the 108 unique combinations of values from the above 5 variables.

We used medium-resolution NHD (U.S. Geological Survey 1999) to define the location of streams and other small water bodies that were not adequately captured by the NLCD but were important habitat cues for many priority species (e.g., Louisiana waterthrush [Seiurus motacilla]). Similarly, STATSGO data(Natural Resources Conservative Service 1994) were used to identify suitable soil texture and moisture classes for the soil-probing American woodcock (Scolopx minor).

Site variables

While, the geodatasets described above allow us to characterize landscape composition and structure, we relied on FIA data to provide information about site-level forest structure. Staff from FIA’s Spatial Data Services (SDS) center in St. Paul, MN, queried plot data to obtain unique plot numbers and location coordinates for 20522 plots located within the 2 BCRs. These plots were sampled between 1986 and 1995, the years associated with the periodic inventories closest in date to the NLCD 1992 dataset. Although true plot location coordinates were not made available to us, we were able download publicly available PLOT, COND, and TREE tables for each state intersecting the BCRs (http://www.ncrs.fs.fed.us/4801/tools-data/data/). The 3 FIA tables for each state were imported into an Access (Microsoft, Redmond, WA) relational database, then combined and queried to generate tables containing plot-level summaries of the variables needed for our habitat models.

FIA does not measure all the key forest attributes for avian habitat selection on all Phase 2 plots; therefore, we fit a regression equation to predict small (<2.54 cm dbh) woody stem density (a derived Phase 3 plot variable) from basal area and tree density (Phase 2 plot variables). Similarly, we estimated overstory canopy cover from tree diameter and pole and sawtimber tree density based on an equation developed by Law et al. (1994). All other forest structure attributes were summarized directly from FIA data. We created a summary table containing all forest structure variables and joined this table to the forest patch attribute table via the FIA plot identification number common to both tables. The sampling intensity of periodic and annual FIA plots is not adequate for spatial interpolation (e.g., kriging) on forest structural attributes because the distance between plots is much greater than the distance over which these attributes are spatially correlated (Coulston et al. 2004). The spatial limitations of FIA’s sampling design, coupled with privacy protections that restrict public access to exact plot locations, necessitated the development of an ecologically meaningful protocol for populating each forest patch in our landscape with FIA plot attributes. To accomplish this, we devised a stratification procedure that first defined our BCRs as patches of unique combinations of variables (i.e., strata), then identified the FIA plots within each of these unique combinations, and finally assigned each patch an FIA plot that had the same strata characteristics.

We stratified each BCR by ecological subsection, NLCD forest class, and landform type because we believed that these variables accounted for the greatest amount of variation in forest structure at the landscape scale. To avoid creating singular strata (i.e., unique combinations of variables associated with only 1 FIA plot) that would prohibit data accessibility, we used a reduced number of strata. Thus, we aggregated NLCD into 6 classes: deciduous, mixed, evergreen, woody wetland, transitional-shrubland, and non-forest. This stratification produced a map that contained 36 unique strata combinations (6 NLCD classes * 6 landform classes) in each ecological subsection.

SDS personnel spatially joined actual FIA plot locations to these strata and returned an attribute table containing plot identification numbers (but not coordinates) and the values for each of the 3 strata. Plot identification numbers allowed us to link each plot and its known strata values to our summary table of FIA and derived forest structure attributes. Due to potential security issues associated with some linkages, strata values for a small proportion of plots were not provided to us. Nonetheless, our stratification scheme allowed us associate approximately 97 percent of FIA plots on private land and 60 percent of FIA plots on public land with our geospatial strata.

An inherent artifact of this approach is the wide range of FIA data plots associated with each strata; common strata combinations contained >200 plots whereas rare combinations contained ≤ 1. To prevent all patches in a single strata combination from being represented by a relatively small number of plots, we established a 6 plot minimum threshold for definition of all strata combinations and developed decision rules to guide aggregation of strata to achieve a minimum of 6 FIA plots. First, we identified all strata combinations within an ecological subsection that contained <6 FIA plots and determined the proportion of the subsection represented by that unique landform-NLCD combination. If a stratum covered <5 percent of the subsection, we considered it a rare strata and combined it with a similar NLCD class within the same landform (e.g., plots from floodplain woody wetlands would be aggregated with plots from floodplain deciduous). If a stratum covered >5 percent of the subsection, we combined strata among similar landforms within the same NLCD classes (e.g., plots from floodplain woody wetlands would be aggregated with plots from valley woody wetlands). Through iterative applications of these rules, we combined strata across similar NLCD and landform classes to achieve the 6-plot threshold. However, in some small and predominately non-forested subsections, we combined strata from different subsections to reach the 6 plot threshold. In these cases, we combined subsections within the same ecological section before combining between different ecological sections. Once all strata were assigned at least 6 FIA plots, we assigned an FIA plot to every forested patch in our study area. We used a modified random number generator to assign an FIA plot identification number to each patch from the corresponding pool of plots associated with each unique combination of strata.

Spatial Assignment of FIA and Derived Forest Structure Variables

To spatially map FIA variables, we created individual geodatasets of each forest structural variable by reclassification on the variable of interest. This produced geodatasets wherein every pixel in a forest patch received the attribute value (e.g., basal area) measured on the plot assigned to that patch. Because all attributes of a plot are assigned together, the covariance structure of the FIA data was maintained and improbable combinations of attributes (e.g., high sawtimber tree density and low basal area) were avoided. We caution that the final product of this procedure is a spatially explicit depiction of forest structure attributes, however, it is not spatially exact (i.e., each pixel has a value, but it is not necessarily the value that would be observed at that location). Even so, because the final model outputs are summarized by subsection and FIA data are representative of forest conditions within subsections, spatial exactness of these attributes within a subsection is not required.

spacer