Contributed Paper
Improving effectiveness of systematic conservation planning with density data Samuel Veloz,∗ § Leonardo Salas,∗ Bob Altman,† John Alexander,‡ Dennis Jongsomjit,∗ Nathan Elliott,∗ and Grant Ballard∗ ∗
Point Blue Conservation Science, 3820 Cypress Drive #11 Petaluma, CA 94954, U.S.A., email [email protected] †American Bird Conservancy 4249 Loudon Avenue, The Plains, VA 20198, U.S.A. ‡Klamath Bird Observatory P.O. Box 758, Ashland OR, 97520, U.S.A.
Abstract: Systematic conservation planning aims to design networks of protected areas that meet conservation goals across large landscapes. The optimal design of these conservation networks is most frequently based on the modeled habitat suitability or probability of occurrence of species, despite evidence that model predictions may not be highly correlated with species density. We hypothesized that conservation networks designed using species density distributions more efficiently conserve populations of all species considered than networks designed using probability of occurrence models. To test this hypothesis, we used the Zonation conservation prioritization algorithm to evaluate conservation network designs based on probability of occurrence versus density models for 26 land bird species in the U.S. Pacific Northwest. We assessed the efficacy of each conservation network based on predicted species densities and predicted species diversity. High-density model Zonation rankings protected more individuals per species when networks protected the highest priority 10-40% of the landscape. Compared with density-based models, the occurrence-based models protected more individuals in the lowest 50% priority areas of the landscape. The 2 approaches conserved species diversity in similar ways: predicted diversity was higher in higher priority locations in both conservation networks. We conclude that both density and probability of occurrence models can be useful for setting conservation priorities but that density-based models are best suited for identifying the highest priority areas. Developing methods to aggregate species count data from unrelated monitoring efforts and making these data widely available through ecoinformatics portals such as the Avian Knowledge Network will enable species count data to be more widely incorporated into systematic conservation planning efforts. Keywords: conservation prioritization, species distribution models, species diversity, systematic conservation planning, zonation Mejor´ıa de la Efectividad de la Planeaci´ on Sistem´atica de la Conservaci´ on con Datos de Densidad
Resumen: La planeaci´on sistem´atica de la conservaci´on tiene como meta dise˜nar redes de a´ reas protegidas que cumplan con objetivos de conservaci´ on a lo largo de grandes paisajes. El dise˜ no o ´ ptimo de estas redes de conservaci´ on se basa con mayor frecuencia en modelos de idoneidad de h´ abitat o probabilidad de occurrencia de especies, a pesar de la evidencia existente de que las predicciones de esos modelos pueden no estar fuertemente correlacionadas con la densidad de especies. Hipotetizamos que las redes de conservaci´ on dise˜ nadas con las distribuciones de la densidad de especies conservan con mayor eficiencia a las poblaciones de todas las especies consideradas que las redes dise˜ nadas con modelos de probabilidad de occurencia. Para probar esta hip´ otesis usamos el algoritmo Zonation de planeaci´ on de la conservaci´ on para evaluar los dise˜ nos de redes de conservaci´ on basados en la probabilidad de ocurrencia versus los modelos de densidad para 26 especies de aves terrestres en el noroeste del Pac´ıfico en los Estados Unidos. Evaluamos la efectividad de cada red de conservaci´ on con base en las densidades pronosticadas de cada especie y la diversidad de especies pronosticada. Las clasificaciones de Zonation de los modelos de alta densidad protegieron a m´ as individuos por especie cuando las redes protegieron el 10-40% de la m´ as alta prioridad del paisaje. Comparado con los modelos basados en la densidad, los modelos basados en la ocurrencia protegieron a m´ as individuos en
§email [email protected] Paper submitted July 3, 2014; revised manuscript accepted November 11, 2014.
1 Conservation Biology, Volume 00, No. 0, 1–11 C 2015, Society for Conservation Biology DOI: 10.1111/cobi.12499
2
Improving Systematic Conservation Planning
el 50% m´ as bajo de las a ´ reas prioritarias de los paisajes. Las dos estrategias conservaron la diversidad de especies de formas similares: la diversidad pronosticada fue m´ as alta en las localidades de alta prioridad en ambas redes de conservaci´ on. Concluimos que tanto los modelos de densidad como los de probabilidad de ocurrencia pueden ser u on, pero que los modelos basados en ´ tiles para establecer prioridades de conservaci´ la densidad son m´ as adecuados para identificar las a as alta prioridad. Desarrollar m´etodos para ´ reas de m´ agregar datos de conteos de especies a partir de esfuerzos de monitoreo no relacionados y hacer que estos datos est´en disponibles en portales eco-inform´ aticos como la Avian Knowledge Network permitir´ a que los datos de conteos de especies se incorporen m´ as ampliamente en esfuerzos de planeaci´ on sistem´ atica de la conservaci´ on.
Palabras Clave: diversidad de especies, modelos de distribuci´on de especies, planeaci´on sistem´atica de la conservaci´ on, priorizaci´ on de la conservaci´ on, zonaci´ on
Introduction Systematic conservation planning has arisen as an effective approach for designing conservation networks and setting spatial conservation priorities across large landscapes (Margules & Pressey 2000). The approach provides a transparent methodology for using existing biodiversity and environmental data to identify locations that are of high conservation priority based on biodiversity or other stakeholder priorities (Kukkala & Moilanen 2013). An important step in the systematic conservation planning process is to collect data on biodiversity features such as species and their habitats that occur within the planning area. Unfortunately, comprehensive biodiversity surveys are rarely conducted at large spatial scales (e.g., ecoregional). To cope with this problem, systematic conservation planning typically relies on models to infer the distributions of species of interest in unsurveyed portions of the landscape (Ferrier et al. 2002). This approach provides a robust way to use existing and often eclectic collections of species data, such as from museum collections or citizen science programs, to predict distributions across large landscapes (Elith & Leathwick 2009). Although there are known limitations to the approach (Elith et al. 2010; Franklin 2010; Veloz et al. 2012), many studies show the high predictive accuracy of species distribution models (Elith et al. 2006; Dobrowski et al. 2009; Elith & Graham 2009). Relatively few studies use count data to model species density distributions. Although some models based on presence-absence data can identify the upper limit on species abundance (VanDerWal et al. 2009), other models show weak correlations between species probability of occurrence and species abundance, suggesting that substantially different processes may determine the distribution and abundance of species (Nielsen et al. 2005; Stewart-Koster et al. 2013). In particular, information loss from presence-absence data increases as pixel size increases, but this is not necessarily the case with count data (Aarts et al. 2012). Using areas estimated to have high suitability to predict areas of high abundance can result in overestimates of abundance Conservation Biology Volume 00, No. 0, 2015
(Van Couwenberghe et al. 2013). Thus, reserve networks designed using models of habitat suitability or probability of occurrence may not optimally protect areas with the greatest number of individuals per species. A lack of coordinated surveys that count individuals of species across large landscapes is probably the main limitation to development of species abundance or density models. Large-scale monitoring with standard-effort research techniques requires relatively large numbers of person hours of work. The alternative, an integration of multiple uncoordinated monitoring efforts, is hindered by the lack of proper curation of species monitoring data sets. Presence-absence modelers have been more successful at integrating data collected from different sources because more data are available thanks to large-scale citizen science projects (Greenwood 2007; Silvertown 2009; Hochachka et al. 2012) and because differences in data collection protocol matter less in determining presence or absence than in determining abundance (e.g., Thomas et al. 2001). Data from count surveys can be transformed to presence-absence data, but the converse is only possible under limited circumstances (e.g., Royle & Nichols 2003), resulting in greater availability of presence-absence data sets. Still, numerous smaller-scale surveys of species counts exist that may be aggregated into density estimates for large landscape models. As these data sets become increasingly available through the improvement of ecoinformatics databases and infrastructure (Michener & Jones 2012), we expect species density models will become more available for use in systematic conservation planning. Thus, we tested whether conservation networks designed based on count data are equivalent to networks designed based on presence-absence data to evaluate whether large investments in count surveys results in improved conservation outcomes. We hypothesized that conservation networks designed using models constructed from count data provide more efficient protection of species populations than networks constructed using models of species probability of occurrence. Predictions from models of species probability of occurrence or suitability are likely to be positively correlated with predictions from models of species
Veloz et al.
density, but it is also possible that areas of high suitability may only support a few or no individuals of the target species because the realizations of those probabilities of occurrence may also include nonoccurrence. Thus, a conservation network designed using a suitability model may place undue weight on sites that support fewer numbers of individuals than if estimates of density were used to prioritize the network. We compared how well conservation networks designed using models of species probability of occurrence and using models of species density protect the largest proportions of the landscapewide populations of 26 species of land birds in the U.S. Pacific Northwest. We used data from multiple, uncoordinated monitoring data sets aggregated and curated by the Avian Knowledge Network (AKN). We also evaluated the sensitivity of comparisons between the models as we altered the size of the conservation network and considered the differences in information content in occurrence versus density data and the consequences of using each type of data for conservation planning.
Methods Avian Distribution and Density Data We downloaded georeferenced point count data on the presence or absence and counts of 26 bird species in the states of California, Oregon, and Washington from the California Avian Data Center, a node of the AKN (http://data.prbo.org/multimap-v3/index.php). For each species, data from outside its breeding range or season were discarded. Breeding range was based on the Digital Distribution Maps of the Birds of the Western Hemisphere (version 3.0) (Ridgley et al. 2007). Breeding season window varied among species and ranged from a start date of 1 April to an end date of 15 August. These dates were based on information taken from the Birds of North America series and our expert opinions. We further filtered the avian data to those detections within 50 or 100 m of survey points. We produced individual estimates of density for both distance categories and multiplied the former by 4 to combine results with those of the latter. This ensured that all data were used, not just data conforming to either a 50- or 100-m cutoff distance. In all, we compiled 252,913 count records (i.e., with one or more individuals detected per count) from 96 different projects in which 38 variations of the basic point count protocol were used (Ralph et al. 1993). These were the result of 143,801 survey events sampling 19,407 points from 1992 to 2012.
Environmental Data Our models of probability of occurrence and density were both based on climate and models of major vegetation types. We acquired contemporary climate data
3
to calibrate our vegetation and bird distribution models from the PRISM Climate Group (http://www.prism. oregonstate.edu/). For models of vegetation (see below), we used bioclimatic variables frequently used in modeling exercises (Stralberg et al. 2009; Elith et al. 2011). We used the biovars function in the dismo package in R (Hijmans et al. 2012) to create 19 bioclimatic variables (http://www.worldclim.org/bioclim). We created a correlation matrix to examine the relationships between the bioclimatic variables. When variables were highly correlated with one another (r ࣙ 0.8), we retained the variable that was least correlated with the remaining layers, leaving us with 9 variables (Table 1). We also included a set of geophysical variables based on soil and topography (Table 1). For our bird models, we selected climate covariates based on possible physiological tolerances. We hypothesized that 8 climate variables could constrain the distribution of birds during months when birds typically breed in the region (April–July). Variables were all summarized across the breeding season. After examining correlations between the predictor variables, we retained 3 variables to use in the bird models (Table 1). All environmental variables were resampled to a common spatial resolution of 0.00833 × 0.00833 decimal degrees (approximately 900 × 900 m within our study extent). Vegetation Modeling We modeled the potential distribution of 78 different vegetation types based on hybrid vegetation maps we created and current maps of climate, topographic, and soil characteristic variables. We created distribution models of vegetation types, rather than using the existing vegetation maps, because we needed to coarsen the spatial resolution of the vegetation map to match climate grids and to allow for multiple vegetation types to have a nonzero probability of occurring in each pixel. Furthermore, bird surveys were not conducted in all vegetation types, which required some adjustment to the vegetation maps to make predictions spatial (see below). We created a hybrid vegetation classification scheme derived from the GAP vegetation classification system (http://gapanalysis.usgs.gov/gaplandcover/). We initially considered all the vegetation classes that occur in Washington, Oregon, and California. When we had an adequate number of avian observation records in the finest GAP class (Ecological System), we modeled that class using current climate and geophysical variables. Otherwise, we used a lower resolution class in the classification hierarchy (lower classes based on the NVCS standards) (http://www.fgdc.gov/standards/projects/FGDCstandards-projects/vegetation/NVCS_V2_FINAL_200802.pdf/download [Supporting Information]). To model the suitability of each vegetation class, we used boosted regression tree (BRT) models (Elith et al. Conservation Biology Volume 00, No. 0, 2015
Improving Systematic Conservation Planning
4 Table 1. Spatial covariate data used to construct vegetation and bird distribution models. Vegetation models Geophysical variables sand fraction (5 cm) silt fraction (5 cm) clay fraction (5 cm) soil porosity (5 cm) soil ph (5 cm) soil permeability (5 cm) available water content (100 cm) distance to perennial streams slope solar radiation Climate variables mean diurnal temperature range temperature seasonality mean temperature of the wettest quarter mean temperature of the driest quarter mean temperature of the coldest quarter precipitation seasonality precipitation of the wettest quarter precipitation of the driest quarter precipitation of the coldest quarter Bird models vegetation (78 classes, individual layers) mean monthly temperature (April–July) temperature range (April–July) total precipitation (April–July)
Source USGS (U.S. Geological Survey) State Soil Geographic database (STATSGO) STATSGO STATSGO STATSGO STATSGO STATSGO STATSGO USGS National hydrography data set USGS Global Multiresolution Terrain Elevation Data (2010) USGS Global Multiresolution Terrain Elevation Data (2010) PRISM Climate Group PRISM Climate Group PRISM Climate Group PRISM Climate Group PRISM Climate Group PRISM Climate Group PRISM Climate Group PRISM Climate Group PRISM Climate Group USGS Gap Landcover database PRISM Climate Group PRISM Climate Group PRISM Climate Group
2008) and the climate and geophysical covariates (explained above) as predictors. We selected the optimal number of trees, tree complexity (TC), and learning rate (LR) following an optimization routine described in Elith et al. (2008). We standardized the results of the vegetation models by dividing the suitability for each class by the sum of the suitabilities for all vegetation classes in each pixel. In this way, the values for a given class at a pixel receive lower values if many classes have a high suitability. We did not attempt to determine which vegetation class would dominate a cell; rather, we retained the modeled suitability of all vegetation classes and thus allowed for multiple vegetation classes to give weight to the bird model prediction of each cell. Bird Distribution Modeling Using the bird data for our distribution and density models required that we estimate occupancy (Mackenzie et al. 2002) and determine the true density of bird species at survey points accounting for imperfect detection during field surveys (Supporting Information). Using the corrected presence-absence data from above, we fit an initial probability of occurrence BRT model (5000 trees, TC = 3, LR = 0.01) with a binomial (Bernoulli) link and all climate and vegetation variables. For each species, any variable with a relative influence
Improving effectiveness of systematic conservation planning with density data Samuel Veloz,∗ § Leonardo Salas,∗ Bob Altman,† John Alexander,‡ Dennis Jongsomjit,∗ Nathan Elliott,∗ and Grant Ballard∗ ∗
Point Blue Conservation Science, 3820 Cypress Drive #11 Petaluma, CA 94954, U.S.A., email [email protected] †American Bird Conservancy 4249 Loudon Avenue, The Plains, VA 20198, U.S.A. ‡Klamath Bird Observatory P.O. Box 758, Ashland OR, 97520, U.S.A.
Abstract: Systematic conservation planning aims to design networks of protected areas that meet conservation goals across large landscapes. The optimal design of these conservation networks is most frequently based on the modeled habitat suitability or probability of occurrence of species, despite evidence that model predictions may not be highly correlated with species density. We hypothesized that conservation networks designed using species density distributions more efficiently conserve populations of all species considered than networks designed using probability of occurrence models. To test this hypothesis, we used the Zonation conservation prioritization algorithm to evaluate conservation network designs based on probability of occurrence versus density models for 26 land bird species in the U.S. Pacific Northwest. We assessed the efficacy of each conservation network based on predicted species densities and predicted species diversity. High-density model Zonation rankings protected more individuals per species when networks protected the highest priority 10-40% of the landscape. Compared with density-based models, the occurrence-based models protected more individuals in the lowest 50% priority areas of the landscape. The 2 approaches conserved species diversity in similar ways: predicted diversity was higher in higher priority locations in both conservation networks. We conclude that both density and probability of occurrence models can be useful for setting conservation priorities but that density-based models are best suited for identifying the highest priority areas. Developing methods to aggregate species count data from unrelated monitoring efforts and making these data widely available through ecoinformatics portals such as the Avian Knowledge Network will enable species count data to be more widely incorporated into systematic conservation planning efforts. Keywords: conservation prioritization, species distribution models, species diversity, systematic conservation planning, zonation Mejor´ıa de la Efectividad de la Planeaci´ on Sistem´atica de la Conservaci´ on con Datos de Densidad
Resumen: La planeaci´on sistem´atica de la conservaci´on tiene como meta dise˜nar redes de a´ reas protegidas que cumplan con objetivos de conservaci´ on a lo largo de grandes paisajes. El dise˜ no o ´ ptimo de estas redes de conservaci´ on se basa con mayor frecuencia en modelos de idoneidad de h´ abitat o probabilidad de occurrencia de especies, a pesar de la evidencia existente de que las predicciones de esos modelos pueden no estar fuertemente correlacionadas con la densidad de especies. Hipotetizamos que las redes de conservaci´ on dise˜ nadas con las distribuciones de la densidad de especies conservan con mayor eficiencia a las poblaciones de todas las especies consideradas que las redes dise˜ nadas con modelos de probabilidad de occurencia. Para probar esta hip´ otesis usamos el algoritmo Zonation de planeaci´ on de la conservaci´ on para evaluar los dise˜ nos de redes de conservaci´ on basados en la probabilidad de ocurrencia versus los modelos de densidad para 26 especies de aves terrestres en el noroeste del Pac´ıfico en los Estados Unidos. Evaluamos la efectividad de cada red de conservaci´ on con base en las densidades pronosticadas de cada especie y la diversidad de especies pronosticada. Las clasificaciones de Zonation de los modelos de alta densidad protegieron a m´ as individuos por especie cuando las redes protegieron el 10-40% de la m´ as alta prioridad del paisaje. Comparado con los modelos basados en la densidad, los modelos basados en la ocurrencia protegieron a m´ as individuos en
§email [email protected] Paper submitted July 3, 2014; revised manuscript accepted November 11, 2014.
1 Conservation Biology, Volume 00, No. 0, 1–11 C 2015, Society for Conservation Biology DOI: 10.1111/cobi.12499
2
Improving Systematic Conservation Planning
el 50% m´ as bajo de las a ´ reas prioritarias de los paisajes. Las dos estrategias conservaron la diversidad de especies de formas similares: la diversidad pronosticada fue m´ as alta en las localidades de alta prioridad en ambas redes de conservaci´ on. Concluimos que tanto los modelos de densidad como los de probabilidad de ocurrencia pueden ser u on, pero que los modelos basados en ´ tiles para establecer prioridades de conservaci´ la densidad son m´ as adecuados para identificar las a as alta prioridad. Desarrollar m´etodos para ´ reas de m´ agregar datos de conteos de especies a partir de esfuerzos de monitoreo no relacionados y hacer que estos datos est´en disponibles en portales eco-inform´ aticos como la Avian Knowledge Network permitir´ a que los datos de conteos de especies se incorporen m´ as ampliamente en esfuerzos de planeaci´ on sistem´ atica de la conservaci´ on.
Palabras Clave: diversidad de especies, modelos de distribuci´on de especies, planeaci´on sistem´atica de la conservaci´ on, priorizaci´ on de la conservaci´ on, zonaci´ on
Introduction Systematic conservation planning has arisen as an effective approach for designing conservation networks and setting spatial conservation priorities across large landscapes (Margules & Pressey 2000). The approach provides a transparent methodology for using existing biodiversity and environmental data to identify locations that are of high conservation priority based on biodiversity or other stakeholder priorities (Kukkala & Moilanen 2013). An important step in the systematic conservation planning process is to collect data on biodiversity features such as species and their habitats that occur within the planning area. Unfortunately, comprehensive biodiversity surveys are rarely conducted at large spatial scales (e.g., ecoregional). To cope with this problem, systematic conservation planning typically relies on models to infer the distributions of species of interest in unsurveyed portions of the landscape (Ferrier et al. 2002). This approach provides a robust way to use existing and often eclectic collections of species data, such as from museum collections or citizen science programs, to predict distributions across large landscapes (Elith & Leathwick 2009). Although there are known limitations to the approach (Elith et al. 2010; Franklin 2010; Veloz et al. 2012), many studies show the high predictive accuracy of species distribution models (Elith et al. 2006; Dobrowski et al. 2009; Elith & Graham 2009). Relatively few studies use count data to model species density distributions. Although some models based on presence-absence data can identify the upper limit on species abundance (VanDerWal et al. 2009), other models show weak correlations between species probability of occurrence and species abundance, suggesting that substantially different processes may determine the distribution and abundance of species (Nielsen et al. 2005; Stewart-Koster et al. 2013). In particular, information loss from presence-absence data increases as pixel size increases, but this is not necessarily the case with count data (Aarts et al. 2012). Using areas estimated to have high suitability to predict areas of high abundance can result in overestimates of abundance Conservation Biology Volume 00, No. 0, 2015
(Van Couwenberghe et al. 2013). Thus, reserve networks designed using models of habitat suitability or probability of occurrence may not optimally protect areas with the greatest number of individuals per species. A lack of coordinated surveys that count individuals of species across large landscapes is probably the main limitation to development of species abundance or density models. Large-scale monitoring with standard-effort research techniques requires relatively large numbers of person hours of work. The alternative, an integration of multiple uncoordinated monitoring efforts, is hindered by the lack of proper curation of species monitoring data sets. Presence-absence modelers have been more successful at integrating data collected from different sources because more data are available thanks to large-scale citizen science projects (Greenwood 2007; Silvertown 2009; Hochachka et al. 2012) and because differences in data collection protocol matter less in determining presence or absence than in determining abundance (e.g., Thomas et al. 2001). Data from count surveys can be transformed to presence-absence data, but the converse is only possible under limited circumstances (e.g., Royle & Nichols 2003), resulting in greater availability of presence-absence data sets. Still, numerous smaller-scale surveys of species counts exist that may be aggregated into density estimates for large landscape models. As these data sets become increasingly available through the improvement of ecoinformatics databases and infrastructure (Michener & Jones 2012), we expect species density models will become more available for use in systematic conservation planning. Thus, we tested whether conservation networks designed based on count data are equivalent to networks designed based on presence-absence data to evaluate whether large investments in count surveys results in improved conservation outcomes. We hypothesized that conservation networks designed using models constructed from count data provide more efficient protection of species populations than networks constructed using models of species probability of occurrence. Predictions from models of species probability of occurrence or suitability are likely to be positively correlated with predictions from models of species
Veloz et al.
density, but it is also possible that areas of high suitability may only support a few or no individuals of the target species because the realizations of those probabilities of occurrence may also include nonoccurrence. Thus, a conservation network designed using a suitability model may place undue weight on sites that support fewer numbers of individuals than if estimates of density were used to prioritize the network. We compared how well conservation networks designed using models of species probability of occurrence and using models of species density protect the largest proportions of the landscapewide populations of 26 species of land birds in the U.S. Pacific Northwest. We used data from multiple, uncoordinated monitoring data sets aggregated and curated by the Avian Knowledge Network (AKN). We also evaluated the sensitivity of comparisons between the models as we altered the size of the conservation network and considered the differences in information content in occurrence versus density data and the consequences of using each type of data for conservation planning.
Methods Avian Distribution and Density Data We downloaded georeferenced point count data on the presence or absence and counts of 26 bird species in the states of California, Oregon, and Washington from the California Avian Data Center, a node of the AKN (http://data.prbo.org/multimap-v3/index.php). For each species, data from outside its breeding range or season were discarded. Breeding range was based on the Digital Distribution Maps of the Birds of the Western Hemisphere (version 3.0) (Ridgley et al. 2007). Breeding season window varied among species and ranged from a start date of 1 April to an end date of 15 August. These dates were based on information taken from the Birds of North America series and our expert opinions. We further filtered the avian data to those detections within 50 or 100 m of survey points. We produced individual estimates of density for both distance categories and multiplied the former by 4 to combine results with those of the latter. This ensured that all data were used, not just data conforming to either a 50- or 100-m cutoff distance. In all, we compiled 252,913 count records (i.e., with one or more individuals detected per count) from 96 different projects in which 38 variations of the basic point count protocol were used (Ralph et al. 1993). These were the result of 143,801 survey events sampling 19,407 points from 1992 to 2012.
Environmental Data Our models of probability of occurrence and density were both based on climate and models of major vegetation types. We acquired contemporary climate data
3
to calibrate our vegetation and bird distribution models from the PRISM Climate Group (http://www.prism. oregonstate.edu/). For models of vegetation (see below), we used bioclimatic variables frequently used in modeling exercises (Stralberg et al. 2009; Elith et al. 2011). We used the biovars function in the dismo package in R (Hijmans et al. 2012) to create 19 bioclimatic variables (http://www.worldclim.org/bioclim). We created a correlation matrix to examine the relationships between the bioclimatic variables. When variables were highly correlated with one another (r ࣙ 0.8), we retained the variable that was least correlated with the remaining layers, leaving us with 9 variables (Table 1). We also included a set of geophysical variables based on soil and topography (Table 1). For our bird models, we selected climate covariates based on possible physiological tolerances. We hypothesized that 8 climate variables could constrain the distribution of birds during months when birds typically breed in the region (April–July). Variables were all summarized across the breeding season. After examining correlations between the predictor variables, we retained 3 variables to use in the bird models (Table 1). All environmental variables were resampled to a common spatial resolution of 0.00833 × 0.00833 decimal degrees (approximately 900 × 900 m within our study extent). Vegetation Modeling We modeled the potential distribution of 78 different vegetation types based on hybrid vegetation maps we created and current maps of climate, topographic, and soil characteristic variables. We created distribution models of vegetation types, rather than using the existing vegetation maps, because we needed to coarsen the spatial resolution of the vegetation map to match climate grids and to allow for multiple vegetation types to have a nonzero probability of occurring in each pixel. Furthermore, bird surveys were not conducted in all vegetation types, which required some adjustment to the vegetation maps to make predictions spatial (see below). We created a hybrid vegetation classification scheme derived from the GAP vegetation classification system (http://gapanalysis.usgs.gov/gaplandcover/). We initially considered all the vegetation classes that occur in Washington, Oregon, and California. When we had an adequate number of avian observation records in the finest GAP class (Ecological System), we modeled that class using current climate and geophysical variables. Otherwise, we used a lower resolution class in the classification hierarchy (lower classes based on the NVCS standards) (http://www.fgdc.gov/standards/projects/FGDCstandards-projects/vegetation/NVCS_V2_FINAL_200802.pdf/download [Supporting Information]). To model the suitability of each vegetation class, we used boosted regression tree (BRT) models (Elith et al. Conservation Biology Volume 00, No. 0, 2015
Improving Systematic Conservation Planning
4 Table 1. Spatial covariate data used to construct vegetation and bird distribution models. Vegetation models Geophysical variables sand fraction (5 cm) silt fraction (5 cm) clay fraction (5 cm) soil porosity (5 cm) soil ph (5 cm) soil permeability (5 cm) available water content (100 cm) distance to perennial streams slope solar radiation Climate variables mean diurnal temperature range temperature seasonality mean temperature of the wettest quarter mean temperature of the driest quarter mean temperature of the coldest quarter precipitation seasonality precipitation of the wettest quarter precipitation of the driest quarter precipitation of the coldest quarter Bird models vegetation (78 classes, individual layers) mean monthly temperature (April–July) temperature range (April–July) total precipitation (April–July)
Source USGS (U.S. Geological Survey) State Soil Geographic database (STATSGO) STATSGO STATSGO STATSGO STATSGO STATSGO STATSGO USGS National hydrography data set USGS Global Multiresolution Terrain Elevation Data (2010) USGS Global Multiresolution Terrain Elevation Data (2010) PRISM Climate Group PRISM Climate Group PRISM Climate Group PRISM Climate Group PRISM Climate Group PRISM Climate Group PRISM Climate Group PRISM Climate Group PRISM Climate Group USGS Gap Landcover database PRISM Climate Group PRISM Climate Group PRISM Climate Group
2008) and the climate and geophysical covariates (explained above) as predictors. We selected the optimal number of trees, tree complexity (TC), and learning rate (LR) following an optimization routine described in Elith et al. (2008). We standardized the results of the vegetation models by dividing the suitability for each class by the sum of the suitabilities for all vegetation classes in each pixel. In this way, the values for a given class at a pixel receive lower values if many classes have a high suitability. We did not attempt to determine which vegetation class would dominate a cell; rather, we retained the modeled suitability of all vegetation classes and thus allowed for multiple vegetation classes to give weight to the bird model prediction of each cell. Bird Distribution Modeling Using the bird data for our distribution and density models required that we estimate occupancy (Mackenzie et al. 2002) and determine the true density of bird species at survey points accounting for imperfect detection during field surveys (Supporting Information). Using the corrected presence-absence data from above, we fit an initial probability of occurrence BRT model (5000 trees, TC = 3, LR = 0.01) with a binomial (Bernoulli) link and all climate and vegetation variables. For each species, any variable with a relative influence
