UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Spatial models of plant species richness for British Columbia's Garry oak meadow ecosystem Boag, Angela Elaine 2014

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2014_september_boag_angela.pdf [ 5.86MB ]
Metadata
JSON: 24-1.0167479.json
JSON-LD: 24-1.0167479-ld.json
RDF/XML (Pretty): 24-1.0167479-rdf.xml
RDF/JSON: 24-1.0167479-rdf.json
Turtle: 24-1.0167479-turtle.txt
N-Triples: 24-1.0167479-rdf-ntriples.txt
Original Record: 24-1.0167479-source.json
Full Text
24-1.0167479-fulltext.txt
Citation
24-1.0167479.ris

Full Text

    SPATIAL MODELS OF PLANT SPECIES RICHNESS FOR BRITISH COLUMBIA’S GARRY OAK MEADOW ECOSYSTEM  by  ANGELA ELAINE BOAG B.Sc., Queen’s University, 2010  A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF  MASTER OF SCIENCE in THE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES (Forestry)  THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)  May 2014 © Angela Elaine Boag, 2014     ii Abstract  Garry oak meadow ecosystems in British Columbia are fragmented, increasingly degraded, and have been prioritized for conservation. While distribution maps of remnant meadow patches have been developed, the ecological integrity of plant communities in many of these remnants remains unknown. Modeling and mapping ecological integrity could inform conservation prioritization exercises in the region.  The primary goal of this thesis was to develop distribution models of native and exotic plant species richness in Garry oak meadow remnants. Secondly, multiple independent datasets were used to analyze the effects of sample size and sampling bias on the accuracy and reliability of resulting predictive maps, which is an active area of research in species distribution modeling. Finally, I investigated whether Terrestrial Ecosystem Mapping (TEM) – a publicly available GIS layer of plant community associations – provided a valid geographical extent over which to map predictions.  In Chapter 2, different datasets were found to produce different models of species richness. However, different native richness models produced similar distribution maps, while exotic richness maps based on different datasets were less similar. The incorporation of spatial variables into models did not improve model fit, however significant residual spatial autocorrelation at a broad scale was detected in some cases, suggesting that an important environmental covariate is missing from these models. Examples of potential missing covariates include deer density and disturbance history. Overall, this research demonstrates that multiple independent datasets are very important  iii for validating species distribution models, especially in heterogeneous landscapes. Additionally, large sample sizes and sampling broadly across of the area of prediction result in more robust models.  The results presented in Chapter 3 suggest that mapping predictions exclusively over Garry oak ecosystem-classified TEM polygons is potentially overly conservative, as species richness of native meadow species was found to be high in other TEM classifications as well. This suggests that Garry oak meadow plant communities do not exist solely in discreet meadow patches, and that they are dispersed throughout other habitat types including Douglas-fir – salal forests.               iv Preface This dissertation is an original body of work written solely by the author, A. E. Boag. The datasets used for analyses include those collected previously by Dr. Joseph Bennett and Dr. Emily Gonzales, as well as a new dataset collected by the author and her field assistant, Cora Skaien. The models in Chapters 2 were constructed with input from lab member Richard Schuster and project supervisor Dr. Peter Arcese, but were primarily developed by the author.                  v Table of Contents Abstract ..............................................................................................................................ii Preface ...............................................................................................................................iv Table of Contents...............................................................................................................v List of Tables ....................................................................................................................vii List of Figures ...................................................................................................................ix Glossary .............................................................................................................................xi Acknowledgements .........................................................................................................xiii Dedication........................................................................................................................xiv Chapter 1: Introduction....................................................................................................1 Ecology and conservation of Garry oak meadows ..................................................1 Variables affecting species distribution....................................................................3 Species distribution models (SDMs) .........................................................................5 Effects of sampling data quality and bias on SDMs................................................9 Ecological integrity and conservation value...........................................................11 Indicators of conservation value .............................................................................13 Chapter 2: Spatial models of plant species richness derived from multiple independent datasets .......................................................................................................15 Introduction ..............................................................................................................15 Methods .....................................................................................................................18 Study area ..............................................................................................................18 Available datasets..................................................................................................18 Environmental variables........................................................................................22 Repeatability of survey results...............................................................................24 Macroecological models........................................................................................24 Model validation and map comparison .................................................................28 Results .......................................................................................................................28 Repeatability of survey results...............................................................................28 Macroecological models........................................................................................29 Model validation and map comparison .................................................................33 Discussion..................................................................................................................36 Measures of ecological integrity ...........................................................................36 Effects of extent differences, bias, and quality of training data ............................37 Sources of error and limitations ............................................................................40 Conclusions ...........................................................................................................42 Chapter 3: Assessing the indicator value of Terrestrial Ecosystem Mapping classifications for Garry oak meadow ecosystems .......................................................44 Introduction ..............................................................................................................44 Indicator species associations ...............................................................................45 Methods .....................................................................................................................46 Results .......................................................................................................................48 Discussion..................................................................................................................52 Chapter 4: Conclusions ...................................................................................................54 Key findings ..............................................................................................................54 Conservation implications .......................................................................................55  vi Future research ........................................................................................................56 Bibliography.....................................................................................................................58 Appendices .......................................................................................................................70 Appendix 1............................................................................................................70 Appendix 2............................................................................................................80                       vii List of Tables Table 2.1: Datasets used in model development ...............................................................20  Table 2.2: Candidate variables (fixed effects) for SDM development (24 total). All buffers were circular with radius = 500 m around quadrat point locations .......................23  Table 2.3. Coefficient estimates (β) ± standard errors of selected variables in the final non-spatial models and single selected spatial model (bold) predicting native SR for each dataset and the combined dataset.......................................................................................31  Table 2.4. Coefficient estimates (β) ± standard errors of selected variables in final models predicting exotic SR for each dataset and the combined dataset.......................................32  Table 2.5. Pearson correlation coefficients and R2  values (*p < 0.05, **0.01, ***0.001) of observed species richness counts vs. those predicted by final models derived from each dataset. Internal validations (grey) were performed using a split-sample bootstrap procedure with 100 repetitions ..........................................................................................34  Table 2.6. Pearson correlation coefficients and R2  values (*p < 0.05, **0.01, ***0.001)  of observed species richness counts in Southern Gulf Islands quadrats only vs. those predicted by final models derived from Dataset A, whose extent was limited to the Southern Gulf Islands ........................................................................................................35  Table 2.7. Results of Fuzzy numerical comparison. Value indicates no (0) to 100% (1) map similarity ....................................................................................................................36  Table 3.1. ‘GOE’-classified and ‘Other’-classified TEM species associations.................47  Table A1.1. Spearman rank correlations of variables permitted in model selection (r < 0.60) for Dataset A ............................................................................................................70  Table A1.2. Spearman rank correlations of variables permitted in model selection (r < 0.60) for Dataset B.............................................................................................................70  Table A1.3. Spearman rank correlations of variables permitted in model selection (r < 0.60) for Dataset C.............................................................................................................71  Table A1.4. Spearman rank correlations of variables permitted in model selection (r < 0.60) for Dataset ABC .......................................................................................................71  Table A2.1. Mean GOE indicator species count for primary TEM class of quadrats; n denotes number of quadrats (849 total) .............................................................................80   viii Table  A2.2. Significant results (α < 0.05) for Tukey HSD multiple comparison tests between richness and percent cover means. Comparisons not shown were all not significant (NS) ..................................................................................................................81                         ix List of Figures Figure 2.1: Sampling locations in British Columbia’s Coastal Douglas Fir (CDF) zone of 1-m2 quadrats from Dataset A (open circles; n = 184), Dataset B (grey circles; n = 484) and Dataset C (black circles; n = 183). Light grey shading in patches across the study area shows the distribution of potential meadow habitat ..........................................................21  Figure 2.2: Methods flowchart outlining model development, which was carried out using these steps for Datasets A, B and C individually, as well as all datasets combined (Dataset ABC)..................................................................................................................................25  Figure 2.3. Repeatability survey results for native and exotic species richness (SR) and percent cover in 1-m2 quadrats (n = 27). Preliminary quadrat surveys were done in 2003-2006 (Dataset A; open circles) and 2006-2008 (Dataset B; closed circles), and the second survey (Dataset C) was completed in 2013. Results show high repeatability for native richness surveys and moderate repeatability for exotic species richness (* p < 0.05, *** p < 0.0001) ....................................................................................................................................... 30  Figure 3.1. Boxplot showing mean GOE indicator species count for TEM classes classified by experts as potential GOE (nquadrats = 741) is not significantly greater than other classes in which quadrats were located (nquadrats = 107; t = -0.77, p = 0.44)........49  Figure 3.2. Boxplot showing mean GOE indicator species cover for TEM classes classified by experts as GOE (nquadrats = 741) is not significantly greater than other classes in which quadrats were located (nquadrats = 107; t = 1.07, p = 0.28)...................................50  Figure 3.3. Boxplots showing Garry oak meadow indicator species count and percent cover by vegetation structural stage (0-2 = open; 3 = open shrub; 4 = closed shrub; 5 = young forest; 6-7 = mature forest) for dominant (Structural stage 1), secondary (Structural stage 2), and tertiary (Structural stage 3) TEM classifications of sampled quadrats ........51  Figure A1.1. Predicted native species richness in Garry oak meadow habitat based on the final model selected using Dataset A as training data ............................................................ 72  Figure A1.2. Predicted native species richness in Garry oak meadow habitat based on the final model selected using Dataset B as training data ............................................................ 73  Figure A1.3. Predicted native species richness in Garry oak meadow habitat based on the final model selected using Dataset C as training data ............................................................ 74  Figure A1.4. Predicted native species richness in Garry oak meadow habitat based on the final model selected using Dataset ABC as training data ...................................................... 75  Figure A1.5. Predicted exotic species richness in Garry oak meadow habitat based on the final model selected using Dataset A as training data ............................................................ 76   x Figure A1.6. Predicted exotic species richness in Garry oak meadow habitat based on the final model selected using Dataset B as training data ............................................................ 77  Figure A1.7. Predicted exotic species richness in Garry oak meadow habitat based on the final model selected using Dataset C as training data ............................................................ 78  Figure A1.8. Predicted exotic species richness in Garry oak meadow habitat based on the final model selected using Dataset ABC as training data ...................................................... 79                      xi Glossary Akaike Information Criterion (AIC): A measure of quality for a statistical model relative to others, which reflects a trade-off between goodness of fit and model complexity. Smaller AIC values indicate a relatively better model.  Coastal Douglas-Fir zone (CDF): A biogeoclimatic zone that extends throughout the Southern Gulf Islands, Straight of Georgia, Nanaimo Lowland and Georgia Lowland ecosections on the southeastern coast of Vancouver Island.   Ecological integrity (EI): The degree to which the composition, structure, and function of an ecosystem operates within its historic or natural range of variation.   Fuzzy numerical comparison: A map comparison tool in Map Comparison Kit that compares numerical values in the same pixel of two raster maps taking into account neighboring pixel values in order to calculate an overall value of map similarity.  Fortress conservation: A conservation strategy where humans and other potential threats are actively or passively excluded from an ecosystem through physical, logistical, or legal barriers, with the goal of protecting the ecosystem from degradation.   Garry Oak Ecosystem Recovery Team (GOERT): Garry Oak Ecosystem Recovery Team; a non-profit dedicated to the recovery of Garry oak and associated ecosystems in Canada.   Generalized linear mixed model (GLMM): A generalized linear model (allowing for response variables with non-normal error distributions) in which the predictor variables include random effects as well as fixed effects.   Map Comparison Kit (MCK): A computer program used to compare raster maps (© Research Institute for Knowledge Systems).  Macroecological modeling (MEM): An approach to species distribution modeling in which a macroecological measure such as species richness is related to environmental variables using statistical techniques.  Negative binomial distribution: A non-normal error distribution frequently used to model over-dispersed count data.   Principal components analysis (PCA): A statistical analysis that summarizes large numbers of variables by collapsing them into a smaller number of principal component values.  Poisson distribution: A non-normal error distribution frequently used to model count data.    xii Spatial autocorrelation (SAC): The correlation of properties in geographic space.   Species distribution model (SDM): A model that predicts the distribution and/or abundance of one or more species using biotic, abiotic, topographic, and/or spatial variables that align with the species of interest.   Southern Gulf Islands (SGI): An island archipelago in the Georgia Straight between southern Vancouver Island and the B.C./Washington state mainland.  Species richness (SR): Here, simply a count of number of species.   Terrestrial Ecosystem Mapping (TEM): A classification system used by the B.C. Government to classify air photos by biogeoclimatic and ecosystem units using indicator species associations.   Variance inflation factor (VIF): An index of how much the variance of a regression coefficient is increased because of collinearity between predictor variables, allowing for only independent variables to be included in model selection processes.                  xiii Acknowledgements The hard work of Dr. Joseph Bennett and Dr. Emily Gonzales, whose research supplied the majority of data analyzed for this thesis, made this project possible. I must give special thanks to Emily Gonzales who was an excellent field botany instructor, and I would also like to thank the Government of British Columbia, Parks Canada, the Capital Region District, the City of Nanaimo, the City of Victoria and several private landowners for access to survey locations.   Funding was provided by NSERC in the form of a Canada Graduate Scholarship to the author and an NSERC Discovery Accelerator grant to Dr. Peter Arcese. The Nature Trust of British Columbia and the Silverhill Institute of Environmental Research and Conservation provided additional funds for fieldwork.   Richard Schuster provided invaluable advice throughout my program, and Dr. Peter Arcese was an incredibly supportive and fun supervisor. Thanks to committee members Dr. Sarah Gergel and Dr. Judy Myers for their feedback, and to the staff and fellow graduate students in the Faculty of Forestry who provided a rewarding and stimulating work environment. Finally I would like to thank my friends for their encouragement, and Alex for putting up with cross-border living and being both my motivation and my source of comfort.     xiv Dedication This thesis is dedicated to all those who work on conservation programs on the ground. There are too few of you, and we scientists should endeavor to support your work as much as possible, if not join you in it. All those involved with the Garry Oak Ecosystem Recovery Team, and those individuals helping to achieve land conservation successes in the Georgia Basin are the true problem-solvers in this threatened landscape. Thank you for your passion and dedication!                 1 Chapter 1: Introduction Ecology and conservation of Garry oak meadows The Garry oak meadow ecosystem in North America’s Pacific Northwest is comprised of low elevation, herb-dominated plant communities within 3 km of the coastline (Clements et al. 2011). The Garry oak, Quercus garyanna, gives these meadow ecosystems their name, but these plant communities also exist in the absence of oak woodlands in associated coastal bluff, maritime meadow, vernal pool, grassland, and rock outcrop habitats, which are all the focus of considerable conservation work due to their high native floristic diversity (Erickson and Meidinger 2007). The extent of these meadow plant communities has shrunk by 95% since European contact with an estimated 200 ha remaining, and 55 Garry oak meadow species are listed under Canada’s Species At Risk Act (SARA) (Lea 2006, Clements et al. 2011).   While First Nations have altered this landscape for thousands of years – primarily through controlled burning to maintain food-bearing herbaceous plant communities – the arrival of Europeans saw widespread land conversion and the arrival of invasive pasture grasses and other exotic plant species (Lea 2006). Garry oak meadow plant communities are now endangered due to habitat loss, fragmentation, degradation by urban development, invasive species, livestock and ungulate grazing, and the loss of first nations-mediated fire regimes (Lea 2006, Gonzales 2008, Clements et al. 2011, Bennett 2011). While distribution maps of remnant patch locations have been developed, mapping ecological integrity in these meadow patches is a central priority for local conservationists, as it would inform conservation prioritization decisions and facilitate  2 monitoring (Lea 2006, Vellend et al. 2008, Jones et al. 2011, GOERT 2011). It would also contribute to the growing literature on distribution modeling using remotely sensed data in high-human impact systems.   Many ecosystems with high biological diversity like Garry oak meadows exist in increasingly disturbed environments (Rands et al. 2010, Sala et al. 2000). Cape fynbos floristic communities in South Africa are similarly diverse and exist in an increasingly fragmented landscape, typical of many coastal ecosystems in developed regions (Cowling and Bond 1991). Constructing species distribution models that indicate the condition or ecological integrity of endangered plant communities like this is the first step toward applying effective conservation measures (Guisan et al. 2013). These models can provide spatially explicit information on where the most intact ecosystem patches likely remain, and developing such a model for understory plant communities in Garry oak meadows is subsequently the first goal of this thesis. In addition, due to time and funding constraints many ‘training’ datasets used to build SDMs are inadequate due to small sample sizes or sampling bias, and this can compromise their utility (Albert et al. 2010, Zimmermann et al. 2010). Understanding the impact of training dataset quality on resulting species distribution maps has been identified as an important area of research, and the existence of multiple datasets collected from this system allows this to be investigated as the second goal of this thesis (Elith and Graham 2009).     3 Variables affecting species distributions In their projections for 2100, Sala et al. (2000) argue that land use change will become the most important driver of change in terrestrial ecosystems. Therefore, understanding the effects of novel, high human impact landscapes on species and ecosystems is increasingly important for achieving conservation goals (Seiferling et al. 2012). Species distribution models (SDMs) facilitate species and habitat conservation through their contributions to planning and reserve design. They inform biodiversity assessment and habitat restoration, and can be used to predict the effects of environmental change on ecosystems (Franklin 2009, Guisan et al. 2013). For this reason SDMs are increasingly important for environmental research and management, and the methods used to develop them are evolving rapidly (Guisan and Zimmermann 2000, Zimmermann et al. 2010, Franklin 2013).   Before the advent of modern species distribution mapping using geographic information systems (GIS) and statistical modeling techniques, species and habitat conservation decisions were based on principals of island biogeography theory (MacArthur and Wilson 1968, Diamond 1976). This theory predicts that larger habitat patches (originally described using islands) support higher biological diversity (Preston 1962, MacArthur and Wilson 1968). Historically, nature reserves were considered islands in a matrix of anthropogenically-modified land, though scientists soon realized the importance of corridors for connecting habitat and achieving stable metapopulations (Levins 1969). Species-area curves, also adopted from island biogeography, have been used to recommend reserve sizes (May 1975), and together with corridor and connectivity theory,  4 spurred the debate over whether single large or several small reserves yield more effective conservation of biodiversity (Diamond 1976, Simberloff and Abele 1982).   However, contrary to classic theory recent empirical studies find little or no relationship between island or habitat patch size and species diversity in archipelagos of small islands and highly fragmented landscapes (Luna-Jorquera et al. 2011, Bennett 2014). Instead, studies have found that below a certain threshold area, species richness and composition become independent of island or fragment size due to a combination of stochastic processes, habitat heterogeneity, isolation, disturbance and human impacts (Triantis et al. 2006, Burns et al. 2009, Spengler et al. 2011, Bennett et al. 2012). Thus, there has been a move towards decision-making based on an understanding of more context-specific threats posed by the matrix surrounding habitat patches, as opposed to basing conservation plans solely on island or habitat patch size (Luna-Jorquera et al. 2011).  The factors of isolation, disturbance and human impacts, as identified above, have brought the matrix into greater focus as it becomes evident that it harbors threats to even the largest of protected areas (Franklin and Lindenmayer 2009). A meta-analysis of plant and animal occupancy data from six continents found that the type of land cover separating habitat patches most strongly affects the sensitivity of species to patch area and isolation (Prugh et al. 2008). The matrix can harbor many potential threats to biodiversity in so-called “protected” areas ranging from invasive species, drifting pesticides and other pollutants, and human populations that may exploit species of concern for food or trade.  5  Franklin and Lindenmayer (2009) note that “matrix management matters because formal reserve systems will never cover more than a small fraction of the globe; human modified lands – the matrix – overwhelmingly dominate not just forests but all of the world’s ecosystems.” While matrix management is incredibly challenging, it follows that we need to understand how the matrix impacts species within habitat patches in order to design effective conservation plans. This is where SDMs can be applied; anthropogenic disturbance and other threats in the landscape can be tested as predictor variables in models to help understand their role in influencing species distributions (Elith et al. 2014). This type of analysis may suggest priorities for conservation area planning, or indicate that the only successful course of action may be to protect patches isolated from deleterious human influence that could ensure species persistence through ‘fortress conservation’ (Bennett and Arcese 2012). This is a ‘last-resort’ conservation strategy in which ecosystems are maintained using physical, logistical, or legal barriers that prevent degradation by external threats (Brockington 2002).    Species distribution models (SDMs)  Species distribution models (SDMs) (also known as habitat suitability models, HSMs) and subsequent predictive maps can be developed by defining a set of abiotic, biotic and landscape characteristics that, in combination, produce a unique set of mapping areas aligning with the ecosystem or species of interest (Franklin 1995, Austin 1998). The development of SDMs can help to identify knowledge gaps about habitat relationships  6 and spatial data, as well as generate testable hypotheses and guide future research and monitoring efforts (Wu and Smeins 2000).    Species distribution models have historically been produced at relatively course resolutions due to the limited availability of fine-resolution remotely sensed data, but increasingly these types of data are becoming available (Wulder et al. 2004). The types of predictor variables generally used in SDMs are derived from sources such as climate models, digital terrain maps (e.g. digital elevation models, DEMs), soil maps, disturbance maps, as well as known distributions of vegetation and other associated species (Franklin 2009). All of these can be derived from ground surveys, air photo classification (orthophotography), or increasingly by remote sensing tools such as Moderate Resolution Imaging Spectroradiometer (MODIS) or Light Detection and Ranging (LIDAR) (Wulder et al. 2004).  To determine which biogeographic variables should be measured and analyzed as potential predictors of species or habitat presence, well-supported hypotheses must be developed concerning the influence of various factors on the study system (Franklin 2009). A study in an island archipelago similar in scale to British Columbia’s southern Gulf Islands found habitat diversity to be the most important variable explaining species richness of multiple taxa across islands, and level of human impact was the only variable explaining variation in species composition (Luna-Jorquera et al. 2011). Another study found soil type to be the most important attribute for characterizing habitat for rare plant species at a regional scale in Texas (Wu and Smiens 2000). These examples illustrate  7 how the suitability of predictor variables varies with scale as well as the ecosystem and/or species of interest, and stresses the importance of developing robust hypotheses to justify the selection of predictor variables.   Scale can also play a large role in the development of accurate SDMs (Guisan and Zimmermann 2000, Heinänen et al. 2012). For example, Wu and Smeins (2000) developed and compared regional-, landscape-, and site-scale habitat models for rare plant species in Texas for use in highway construction planning. With regards to their predictor variables, they found almost 80% agreement between landscape-level GIS data and site-level field data, although landform (slope) was only a 35% exact match. They go on to suggest that landscape-scale remotely-sensed models should not be used solely in conservation and development planning, and that field-surveyed environmental variables should be used to produce more accurate habitat maps.  The efficacy of SDMs also depends on whether they are used for making predictions via interpolation or extrapolation. Interpolation applies model predictions to sites within the same range of environmental conditions as the ‘training’ data (the data used to fit the model), within the same general time frame, and generally within the same region (Hirzel and Le Lay 2008, Elith and Leathwick 2009). Extrapolation is when SDMs are used to predict in new areas in different geographic domains, or in different time periods through forecasting or reconstructing historic distributions. Extrapolating should be done with extreme caution, because the species-environment relationships in new geographic areas or time periods may not match those embodied in the model training data (Hirzel and Le  8 Lay 2008, Elith and Leathwick 2009). For example, the authors of a study on seabird nest distributions as they relate to biogeographic variables at a fine microhabitat scale of 10-m2 recognized that the utility of their final model may be restricted to certain seabird species in the same island archipelago (Heinänen et al. 2012).   Finally, one of the most important aspects of SDM development is evaluating model performance (Elith and Leathwick 2009, Zimmermann et al. 2010). Assessing model performance is primarily done in two ways. First, through cross-validation or bootstrap resampling, in which a subset of the original training data is left out during model development, and is used as the ‘testing data’ to evaluate the accuracy of predictions made by the model (Jorgensen 1994, Guisan and Zimmermann 2000). Random samples of training and testing data can be resampled multiple times and the correlations between the predicted and observed values averaged to evaluate model fit. Alternatively, a model can be fit with the full training dataset then used to predict independently collected testing data (Guisan and Zimmermann 2000). This second method is preferable because it evaluates the actual predictive ability of the model at new sites, and effective prediction across space is the goal of many SDMs. However, it is often difficult for investigators to collect independent data for validation due to time, funding, or access constraints (Albert et al. 2010). This is a major issue in SDM research, as many models are never evaluated using independent data, bringing into question their predictive ability (Barry and Elith 2006). Thorough model evaluation is especially important when it is suspected that the training data may be inadequate due to small sample sizes or bias issues (Sardà-Palomera et al. 2012).   9  Effects of sampling data quality and bias on SDMs Spatial or temporal bias in sample data has been recognized as a less-studied problem in species distribution modeling (Boakes et al. 2010, Bean et al. 2012). Data can be biased in space or time, either due to sampling efforts conducted unevenly across geographic space or in different time periods characterized by different environments (Barry and Elith 2006, Bean et al. 2012). Many datasets used in SDMs are biased in space, particularly towards sites that are more accessible (Barry and Elith 2006, Bean et al. 2012). Accessibility can be dictated by terrain, vegetation, land ownership, safety concerns, costs, or simply time, as sites that are far apart require more time to access (Hirzel and Guisan 2002). This introduces a trade-off between sample ‘quality’ and sample size, with high quality data generally being free from bias and reflective of the species’ true spatial distribution (Barry and Elith 2006, Sardà-Palomera et al. 2012).  A simulation study found that grid sampling (i.e. sampling points at the nodes of a pre-defined grid laid over the sampling area in a GIS) did not improve accuracy greatly over random sampling (i.e. all points in the landscape have an equal probability of being selected), but was found to have lower variability over multiple modeling runs (Hirzel and Guisan 2002). This same study also evaluated the effects of sample size by comparing the Pearson correlation coefficients between known values and those predicted by models built using different sample sizes. A sample size of 120 yielded a mean correlation coefficient of 0.58, which rose to 0.63 with 240 samples, and 0.66 with 800 (Hirzel and Guisan 2002).   10  A study of lichen species distributions in Washington state compared models developed using different datasets and also found that larger sample sizes resulted in more precise predictions of species occurrence (Edwards et al. 2006). They also compared models from datasets collected using different methods, and found that models fit from probability-based systematic grid sampling outperformed those based on nonrandom sampling by experts. These experts used their knowledge of lichen life history traits to only survey areas they thought likely to harbor the species of interest. The authors concluded that well-sampled data with few records are better than biased data of any sample size (Edwards et al. 2006).   In contrast, a study on the distribution of Western Capercaillie (Tetrao urogallus), a large European forest bird, found that non-probabilistic, haphazardly-sampled data collected by volunteers over a large study area produced models that outperformed those produced from systematic transect counts conducted over a smaller subarea irrespective of sample size or modeling method (Braunisch and Suchant 2010). They noted that the larger sampling extent of the volunteer-collected data covered a broader range of species-habitat relationships, and therefore achieved better model quality. The authors concluded that data unsystematically collected over a large representative region are preferable to systematically collected data from a restricted region (Braunisch and Suchant 2010). While these findings contrast with those of Edwards et al. (2006) in some ways, both studies support the general consensus that ‘good quality’ data for SDM development is characterized by large sample sizes collected over the full range of environmental  11 gradients experienced by the species of interest (Barry and Elith 2006, Sardà-Palomera et al. 2012).   Overall, SDMs provide critical information for research, planning, and management needs at landscape scales (Lindenmayer and Possingham 1996, Pavlik 1997), and are also important public education tools that can improve communication and facilitate conflict resolution among stakeholders (Wu and Smeins 2000). However, the quality of these models is only as good as the data used to build them, the suitability of the predictor variables and statistical tools that are chosen for the models, as well as how effectively the models are validated. Their utility also depends on whether the response variable they are modeling (e.g. occupancy data for a single species or a macroecological metric like species richness) is most appropriate for answering the questions of interest.   Ecological integrity and conservation value Aldo Leopold said “a thing is right when it tends to preserve the integrity, stability and beauty of the biotic community. It is wrong when it tends otherwise” (1949: 262). This perspective dominates the conservation community and guides those tasked with managing ecosystems. In line with this axiom, prioritization exercises in conservation planning rely on detailed information about the conservation value of land units, as well as broad conceptual frameworks through which this information can be evaluated and compared between sites. Ecosystem integrity (EI) is one such umbrella concept, and has been widely adopted by conservation bodies such as Parks Canada, who define EI as: “...a condition that is determined to be characteristic of [an ecosystem’s] natural region  12 and likely to persist, including abiotic components and the composition and abundance of native species and biological communities, rates of change and supporting processes” (Parks Canada Agency 2008).   While it is evident that the concept of EI is somewhat vague and philosophical, research shows that the associated concepts of ecological complexity and species diversity contribute to the persistence of functioning ecosystems, which conservation managers inherently work to promote. A meta-analysis of 17 grassland biodiversity experiments showed that 84% of 147 grassland plant species promoted ecosystem functions such as biomass production and nutrient uptake, which are associated with ecosystem services such as carbon storage. Different species promoted functions on different temporal and spatial scales, and under different regimes of environmental change, suggesting that complex interactions are necessary to maintain ecosystem services (Isbell et al. 2011). Another study found that plant diversity effects cascade up the food web and modify interactions across all trophic levels, promoting ecosystem stability (Rzanny and Voigt 2012). EI is frequently quantified by some measure of biological diversity, lack of degradation, or level of ecosystem function. And, despite criticisms of its ambiguity, the concept is considered a valid tool for conservation prioritization and is easily conceptualized by stakeholders and the public (Andreasen et al. 2001, Barbour and Paul 2010, Reza and Abdullah 2011). It is used by Parks Canada as noted, who do conservation work on the Garry oak meadows discussed in this thesis.    13 Indicators of conservation value  Defining ecological integrity in a specific context requires that researchers decide on units of measurement. Historically some have argued that biodiversity and subsequently ecological integrity can be defined solely by species richness (e.g. Redford and Sanderson 1992), and indeed a recent review found that species richness was the most frequently calculated diversity index in a literature search of papers whose titles contained the words “biodiversity” and “indicators” (Heink and Kowarik 2010). However, richness does not take the abundance of species into account, and is therefore not synonymous with diversity, according to most modern ecologists (Tuomisto 2010).    LaPaix et al. (2009) warn that because measures of diversity can over-simplify ecological data, segregated calculations of indices using groups of species with a common status (e.g. native or alien) or degree of sensitivity to disturbance may be more appropriate, and this approach is taken in this study. Invasive species show a consistent positive correlation with anthropogenic habitat degradation, and therefore they have been recognized as valuable indicators of a loss of EI (LaPaix et al. 2009).  As part of carrying out this modeling exercise and in order to explore how dataset quality and bias influences SDM predictions, Chapter 2 addresses the question: Which bioclimatic and/or remotely-sensed land use variables most effectively predict on-the-ground ecological integrity in Garry oak meadow ecosystems, and are these variables consistent across models fit using different datasets?   14 Chapter 3 then explores the implications of using a limited geographic prediction extent for final maps. Finally, Chapter 4 summarizes the main conclusions of this research, suggests ways it could be applied to conservation planning in the Coastal Douglas Fir zone, and recommends directions for future research.                      15 Chapter 2: Spatial models of plant species richness derived from multiple independent datasets  Introduction Around the world threatened ecosystems exist in fragmented patches within human-dominated landscapes, such as the Cape Flats Sand Fynbos ecosystem in South Africa and Carolinian forest in Canada (Cowling and Bond 1991). As human disturbances proliferate across landscapes and alter community composition, traditional niche-based SDMs developed using only climate and topography data will no longer be sufficient for conservation planning. With the increasing availability of high resolution remotely sensed data, more detailed SDMs that incorporate complex threat and disturbance information can now be developed to prioritize conservation investments (Guisan et al. 2013).   Due to time, funding and land access constraints, many datasets used for SDMs are biased, either due to incomplete or nonrandom sampling of geographic space (Albert et al. 2010, Zimmermann et al. 2010). It has been noted repeatedly that despite a growing number of landscape-scale distribution models, few account for variation in distribution over time or validate models with independent data (Peltzer et al. 2008, Beale and Lennon 2012).  Several studies have demonstrated the importance of exploring different sampling approaches and their effects on model performance and bias (Edwards et al. 2006, Le Lay et al. 2010). Biases in sampling data can cause the relationships detected using SDMs to reflect patterns at sampled sites rather than those that occur across the entire region of interest, and this can lead to spatial variation in prediction uncertainty  16 (Barry and Elith 2006). This is an important issue in SDM research, because the ability of SDMs to inform conservation decisions largely depends on their successful interpolation across the region of interest, or cautious extrapolation to new areas or time periods. One study using simulated data found that probabilistic sampling strategies (e.g. simple random) resulted in parameter estimates closer to the known truth and with lower bias than non-probabilistic sampling strategies (e.g. sampling along roads) (Albert et al. 2010). Another study compared bird species richness models derived from four different survey types, ranging from systematic breeding bird surveys to citizen science observations, and also found higher accuracy for models fit from the systematically collected data (Sardà-Palomera et al. 2012).   In human-dominated landscapes, it may be more difficult to approximate species-environment relationships and develop useful SDMs because complex relationships between species and different disturbance types may vary across space. To test this idea, I developed models of plant species richness using three datasets collected by different observers in different Garry oak meadow patches. Stakeholders interested in conserving and restoring the meadows are seeking to maximize the diversity of native plants while limiting the number of nonnative species (GOERT 2011) and therefore we compared models for native and exotic understory species and cross-validated predictions between datasets using a map comparison tool (RIKS BV 2010).   The datasets used for modeling were collected over 10 years, and so the first question I asked was: 1) Is species richness or percent cover stable enough over short time scales for  17 comparing models of ecological integrity in Garry oak meadow ecosystems between datasets? This question was addressed because SDMs represent static models based on their training data. If species richness and percent cover have changed over ten years at specific sites, then models based on older data will reflect different species-environment equilibria. I hypothesized that species richness would stay relatively constant over this short time period, while species abundance will have changed due to short-term stochastic factors such as climate (Hernández Plaza et al. 2012).  I also asked: 2) Do distribution models derived from different datasets suggest that the same environmental predictors are most useful for predicting herbaceous plant distributions in the Coastal Douglas Fir zone? Two previous studies identified urban development as a significant negative predictor of native plant species richness in this ecosystem (Lilley and Vellend 2009, Bennett 2014), and thus I hypothesized that this variable will be retained in models developed using each dataset, indicating that it is a robust predictor in this region. A lack of congruence between models would suggest that heterogeneity exists in species-environment and disturbance relationships across the region of interest. Finally 3), Do models fit using data sampled over different geographic extents and different sampling intensities differ in their ability to predict independent data? The available datasets vary in sample size and survey location. Studies show that larger sample sizes and sampling data collected over a range of environmental gradients produces more accurate models resulting in more robust distribution models (Hirzel and Guisan 2002, Braunisch and Suchant 2010). This suggests that the largest of three  18 available datasets will produce the most accurate predictions, and the dataset collected over a limited area will poorly predict over the entire region of interest.   Methods Study area British Columbia’s Garry oak ecosystem (GOE) is patchily distributed throughout the Coastal Douglas-fir (CDF) zone, which includes southeastern Vancouver Island and the southern Gulf Islands of British Columbia’s Georgia Straight.  Similar ecosystems occur in the adjacent San Juan Island archipelago of Washington state as well as sites in Washington’s Puget Trough and the Willamette Valley in Oregon (Clements et al. 2011). The CDF lies in a rain shadow and experiences moderate drought in the summers and mild, wet winters. Mean annual temperatures range from 9.2 to 10.5°C, and mean annual precipitation varies from 647 to 1263 mm (Wang et al. 2006). Drought-tolerant Garry oaks (Quarcus garyanna) grow in woodlands interspersed with grasslands and herb-dominated meadow plant communities in shallow and deep soils (Clements et al. 2011, Meidinger and Pojar 1991). These meadows contain a high diversity of charismatic herbaceous flowering plants, which are the focus of considerable conservation and restoration work (Erickson and Meidinger 2007). The extent of the meadows has been reduced by 95% since European contact (Lea 2006), and 55 Garry Oak meadow species are listed under Canada’s Species At Risk Act (SARA) (Clements et al. 2011).   Available datasets Datasets A and B were collected in previous surveys conducted between April and mid-June in 2003-2006 and 2006-2008 respectively (Gonzales 2008, Bennett 2011), while  19 Dataset C was collected by the author in the spring of 2013. April to mid-June is the peak flowering period for herbaceous plants in this region and is when they are most easily identified. In order to reduce potential bias due to differences in observer identification skills, species data were summarized for 113 common species (nnative = 67; nexotic = 46) that were present in all datasets and easily identified.   All data were collected in a similar manner. Observers identified Garry oak meadows via satellite imagery, air photos, or local knowledge of the area, and randomly sampled 1-m2 quadrats within sites (sites were defined as discreet meadow patches) at an approximate density of 1 quadrat per 200-m2. Random sampling was conducted by finding the previously-identified centroid of a 1-ha grid cell randomly chosen for each patch using a handheld GPS, sampling the first plot there, then sampling subsequent plots in the patch using a combination of random pace number and direction derived from a random list. All three observers used the same quadrats made of PVC piping. A minimum of three quadrats was sampled per site, and quadrat number was increased with site size up to 15. Each quadrat location was recorded using a handheld GPS centered over the quadrat (GPS60, Garmin Ltd, 94 Kansas, USA), and a photo was taken of each plot from above and all species within were identified using herbarium records. If they could not be identified on-site, a sample was collected for later identification. Percent cover of each species was recorded by estimating the proportion of each cell occupied in a 10 cm x 10 cm string grid stretched across the quadrat.     20 The datasets differed in sampling intensity (Table 2.1); Dataset A was not originally collected for species distribution modeling purposes and sampling intensity was lower than Dataset B. Dataset A was also collected exclusively from sites in the Southern Gulf Islands (SGI), representing a reduced geographic extent, while the other datasets also contain data from Vancouver Island and the northern Gulf Islands (Figure 2.1). Dataset C had the lowest sampling intensity due to a focus on sampling over a wider extent of the CDF, and therefore effort within each site was reduced. As such, these three datasets represent varying levels of data quality, with Dataset B being the highest.   Table 2.1. Datasets used in model development.    nsites Quadrats per site (mean) Total nquadrats Dataset A (2003-2006) 33 3-13 (5.58) 184 Dataset B (2006-2008) 66 4-15 (7.30) 482 Dataset C (2013) 41 3-9 (4.46) 183 Combined (ABC) 123 3-21 (6.90) 849   21  Figure 2.1. Sampling locations in British Columbia’s Coastal Douglas Fir (CDF) zone of 1-m2 quadrats from Dataset A (open circles; n = 184), Dataset B (grey circles; n = 484) and Dataset C (black circles; n = 183). Light grey shading in patches across the study area shows the distribution of potential meadow habitat as estimated by GOE-associated TEM polygons.       !(!( !(!(!(!( !(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!(!( !(!(!(!(!(!(!(!(!(!(!(!(!(!( !(!(!( !(!(!(!(!(!(!(!(!(!(!(!!!!!!!!!!!!!!!!!!!!!!!!!!!! ±0 10 20 30 405kmVancouver Island 22 Environmental variables Remotely sensed predictor variables were selected based on a review of the literature (Table 2.2). Data were collected for each variable at each quadrat point using ArcGIS v. 10.1 (ESRI 2012) and Geospatial Modelling Environment (GME) (Beyer 2010) from the following sources: (i) Terrain Resource Information Management (TRIM, http://archive.ilmb.gov.bc.ca/crgb/pba/trim/specs/specs20.pdf); (ii) airphotos to calculate the island sizes; and (iii) Terrestrial Ecosystem Mapping (TEM) of the CDF Zone (Madrone Environmental Services 2008). For landcover data, both the area within a 500-m radius buffer around each quadrat and distance from each landcover type was calculated. The Spatial Analyst Tools extension was used for digital elevation model (DEM) data collection on elevation, slope and aspect. Aspect (0-360°) was used to calculate two linear variables representing north vs. south and east vs. west aspects (Table 2.2). Climate data was derived from ClimateWNA v. 4.70 (Wang et al. 2012), which provides over 80 climate variables from the 1960-1991 Normal Period. ClimateWNA downscales PRISM and ANUSPLINE 1961-1990 monthly normal data (2.5 x 2.5 arcmin) to scale-free point data. To simplify climate data, a principal components analysis was completed on the climate variables from ClimateWNA yielding PC1 and PC2 values for each quadrat that correlated with winter temperature and spring precipitation, and summer mean temperature and precipitation respectively. Spatial variables included Easting, Northing, Easting2, Northing2, and Easting × Northing.     23 Table 2.2. Candidate variables (fixed effects) for SDM development (24 total). All buffers were circular with radius = 500 m around quadrat point locations.  Parameter Code Justification  Source Island size (m2) isl.sz Small, isolated islands experience less disturbance (Bennett et al. 2012)  Orthophotos  Forest area (m2 in 500 m buffer) Nearest forest (m) for.ar  dist.for   forests may serve as a buffer to reduce invasion (Martin et al. 2008)   Terrestrial Ecosystem Mapping (TEM) (Madrone Environmental Services 2008) Meadow area (m2 in 500 m buffer) Nearest meadow (m) mead.ar  dist.mead proximity of similar ecosystems may affect dispersal (Lilley and Vellend 2009)   Agricultural area (m2 in 500 m buffer) Nearest agriculture (m) ag.ar  dist.ag may be a source of exotic propagules, especially exotic grasses from rangelands (Bennett 2011)   Ocean area (m2 in 500 m buffer) Nearest ocean (m) ocn.ar  dist.ocn dispersal barrier (MacArthur and Wilson 1967) or may increase human access (coastal activity)   Developed (urban or suburban area) (m2 in 500 m buffer) Nearest developed (m) dev.ar   dist.dev Urbanization generally increases nonnative richness and decreases native richness (McKinney 2008)AM   Road length (500 m buffer, m) Nearest road (m) rd.lgth  dist.rd invasive propagule pressure (Lilley and Vellend 2009)   Terrain Resources Information Managment (TRIM) (Government of B.C. 1997) Spatial covariates     utmn, utme, utmn2, utme2 utmn × utme  May be used to account for residual spatial autocorrelation (Borcard et al. 1992)   Climate PCA1 (winter temp/spring precip) Climate PCA2 (summer temp/precip)  clim.pc1   clim.pc2 Effect of temperature and precip. variation may differ between generalist exotic species vs. natives (Riordan and Rundel 2009)AM  ClimateWNA v. 4.70 (Wang et al. 2012)        24 Parameter Code Justification  Source Elevation   Northness (=cos((aspect in degrees * PI)/180)  Eastness (= sin((aspect in degrees * PI)/180)   Slope (degrees)  elev   north   east    slp May affect exposure to weather, access by herbivores  Garry Oak species generally occur on warmer, south and east-facing slopes (Erickson 1993)AM     Garry Oak species generally occur on steeper slopes with higher drainage (Erickson 1993)AM   TRIM Digital Elevation Model (DEM) (Government of B.C. 1997)  Repeatability of survey results In order to determine whether species richness or percent cover would be most appropriate for constructing macroecological models, 27 quadrats from Dataset A (n = 8) and B (n = 19) were resampled in 2013. The author went to the UTM coordinate location of each quadrat using a handheld GPS device (error estimate for Garmin GPS device is ±3 m) and sampled species richness and percent cover as described above.  These data were then plotted against the original survey data and evaluated for repeatability using a Spearman rank correlation and simple linear regression.   Macroecological models All models were developed following the procedure outlined in Figure 2.2. Predictive variables that were heterscedastic were log-transformed, and all predictive variables were standardized using z-scores.   25  Figure 2.2. Methods flowchart outlining model development, which was carried out using these steps for Datasets A, B and C individually, as well as all datasets combined (Dataset ABC).  !Collect baseline data  (process two existing datasets and collect third; combine into master dataset) ID predictor variables by developing literature-supported hypotheses Collect environmental data (ArcGIS, ClimateWNA) Prep data for analysis in R  (standardize using Z scores) Assess collinearity using Spearman ρ and VIFs and remove redundant variables AIC-based model selection and averaging  Assess residual spatial autocorrelation (SAC) using spline correlograms Cross-validation using predicted vs. observed values of each dataset Add spatial covariates to variable suite residual SAC detected Final models projected across extent of Garry oak meadow habitat in the CDF  26 Following the recommendations of Zuur et al. (2009) a correlation matrix (Spearman) was constructed for each dataset, and if two variables had correlations > 0.60, one was not permitted to proceed to the model selection stage. I attempted to remove the fewest variables possible. As a secondary check, the Variance Inflation Factor (VIF) of each variable was assessed and those with VIF > 4 were also removed (Zuur 2009). These collinearity analyses and subsequent modeling were conducted in R v. 2.15.3 (R Core Development Team 2012).    I used a macroecological modeling (MEM) approach by using species richness (SR) of native and exotic plants as the response variables, which has been found to be a complimentary approach to the alternative strategy of stacking predictions from individual presence-absence species distribution models (Dubuis et al. 2011). Previous studies found that spatially explicit models were most accurate for this system (Bennett 2011), and so I opted for a Generalized Linear Mixed Modeling framework (GLMM). While challenging to use due to their relatively recent establishment in ecological modeling, GLMMs are potentially very powerful for non-normal data exhibiting spatial autocorrelation (Zuur et al. 2009). Site identity was included in all models as a random effect to account for similarities among quadrats within sites, and for Dataset ABC, site was nested within dataset to account for inter-observer variation. For each model the richness count data were first modeled using a Poisson distribution with a log link function, and if overdispersion was detected a negative binomial distribution was used instead. Model equations had the following structure:   27 SRij = β′ × Xij + bi,  Where SRij is the species richness (native or exotic) at quadrat j in site i; β′ is a vector of the model coefficients; Xij is a vector of the predictor variables for quadrat j in site i; and bi is the random effect for site i.  For model fitting I used the R package ‘glmmADMB’, then used an information theoretic approach for model selection by completing forward and backward selection of all possible models using package ‘MuMIn’. Models with ΔAIC<2 were averaged, and the standard errors of the coefficients were used to test whether the parameter estimates of the selected variables were significant (α = 0.05). To identify residual spatial autocorrelation in final models I constructed spline correlograms using the R package ‘ncf’, and if it was detected, I repeated model selection with third order spatial polynomial covariates included, simplifying the procedure suggested by Borcard et al. (1992) (Schuster and Arcese 2012). They argue that including higher order polynomials allows not only the linear gradient patterns in the species data to be extracted, but also more complex features like patches, which require quadratic terms to be correctly described (Borcard et al. 1992).   Final models were projected onto maps of the CDF zone. Predictions were constrained to potential GOE habitat as defined by TEM polygon classes judged by expert authors of previous studies to be associated with Garry oak meadows (Schuster and Arcese 2012,  28 Bennett 2014). This was done so that the maps could provide maximum utility to practitioners, and the implications of this decision are addressed further in Chapter 3.   Model validation and map comparison Each model was internally validated using a split-sample bootstrap procedure. Each model was fit using 70% of the original dataset (the ‘training’ subset) then used to predict the remaining 30% (the ‘testing’ subset). Pearson correlation coefficients and R2 values were averaged over 100 repetitions. Each model was then used to predict the response variables (native and exotic richness) of the other datasets, and Pearson correlation coefficients and R2 values were recorded. This was also done for a subset of quadrats from Dataset A and B that were located in the Southern Gulf Islands, which is the limited sampling extent of Dataset A, in order to determine if models from Dataset A had improved fit when used to predict data collected from the same geographic extent. The resulting maps were then compared using a Fuzzy numerical comparison tool from the Map Comparison Kit version 3.2.1, which gives a measure of map similarity by comparing values in raster cells from two maps taking into account the values of neighboring cells (RIKS BV 2010).   Results Repeatability of survey results As seen in Figure 2.3, richness counts were highly repeatable for native plants and were moderately repeatable for exotic species, but native and exotic percent cover had very low to zero repeatability. Therefore, species richness was used to develop subsequent  29 macroecological models. With regards to within vs. between site variation in species richness, there was more variation between sites overall (ANOVA, F = 6.12, p < 0.0001) and the average deviation of species richness counts per quadrat from the within-site mean was 1.89 for natives and 1.79 for exotics.   Macroecological models Tables 2.3 and 2.4 contain the parameter estimates for the variables selected in each of the final models for native and exotic species richness respectively. A variety of variables were retained in final models, with all models retaining at least one variable and some up to six.            30    Figure 2.3. Repeatability survey results for native and exotic species richness (SR) and percent cover in 1-m2 quadrats (n = 27). Preliminary quadrat surveys were done in 2003-2006 (Dataset A; open circles) and 2006-2008 (Dataset B; closed circles), and were compared to the secondary survey (Dataset C) which was completed in 2013. Results show high repeatability for native richness surveys and moderate repeatability for exotic species richness (* p < 0.05; *** p < 0.0001).        ●●●●●●●●●●●●●● ●●●●●●●0 5 10 15051015native SR: preliminary surveynative SR: secondary survey R2 = 0.61r = 0.79●●●●●●● ●●●●●●●●●● ●●●●●0 5 10 15051015exotic SR: preliminary surveyexotic SR: secondary surveyR2 = 0.18r = 0.43●●●●●●●●●●●●●●●●●●●●●●●●●●●0.0 0.2 0.4 0.6 0.8 1.00.00.20.40.60.81.0native percent cover: preliminary surveynative cover: secondary survey R2 = 0.18r = 0.27●●●●●●●●●●●●●●●●●●●●●●● ●●●0.0 0.2 0.4 0.6 0.8 1.00.00.20.40.60.81.0exotic percent cover: preliminary surveyexotic percent cover: secondary survey R2 = 0.06r = 0.25***          *  31 Table 2.3. Coefficient estimates (β) ± standard errors of selected variables in the final non-spatial models and single selected spatial model (bold) predicting native SR for each dataset and the combined dataset.   Dataset A Dataset B Dataset C Dataset ABC Intercept 1.474± 0.0915 1.687±0.0417  1.4630±0.0694  1.479±0.0642  1.550±0.0828   isl.sz -0.210±0.0891 - - - for.ar - - - - dist.for - - - -0.0797± 0.0347 ocn.ar - - - - dist.ocn - - - - mead.ar - - - - dist.mead -0.296±0.0871 - - - ag.ar - - 0.130±0.0658 0.148±0.0633 0.0671±0.0367 dist.ag - - - - dev.ar - - - -0.183±0.0412 dist.dev - 0.272±0.0479 - - rd.lgth - - - - dist.rd - - 0.145±0.0697 0.162±0.0635 - climate PC1 - - - - climate PC2 - - - - slope - 0.0728±0.0296 -0.0970±0.0544 - east - - - - nor - - - - utmn - - -0.135±0.0681 - utme - - - - utmn2 - - - - utme2 - - - - utmn x utme - - - -  32 Table 2.4. Beta values ± standard errors of selected variables in the final non-spatial models predicting exotic SR for each dataset and combined dataset.  Dataset A Dataset B Dataset C Dataset ABC Intercept 1.936±0.0539        1.5573±0.0405 0.864±0.169 1.5230±0.2258     isl.sz - - - - for.ar 0.218±0.0575 - - - dist.for - - 0.341±0.0953 - ocn.ar - - - -0.137±0.0484 dist.ocn - - - - mead.ar - 0.226±0.0514 - - dist.mead - - - - ag.ar 0.0831±0.0485 0.102±0.0452 - 0.104±0.0408 dist.ag - - - - dev.ar - - - - dist.dev - - - - rd.lgth - - - - dist.rd - - - - climate PC1 - 0.0806±0.0445 - - climate PC2 - - - -0.125±0.0473 slope - 0.0865±0.033 - - east - -0.140±0.0506 - - nor - -0.0874±0.0503 - - utmn - - - - utme - - - - utmn2 - - - - utme2 - - - - utmn x utme - - - -    33 There were only three instances in which the same variable was retained in multiple datasets: Agricultural area was retained in Datasets C and ABC for native richness, and Datasets A, B and ABC for exotic richness, all with positive coefficients. Slope was retained in native richness models from Datasets B and C, albeit with an opposite direction of influence. In general, native richness was negatively associated with developed area and roads as expected, while these variables were not selected in any of the final models for exotic richness.  Model validation and map comparison As seen in Table 2.5, all datasets exhibited moderate to low accuracy using internal split-sample validation. Correlations between predicted and observed data when the models were used to predict other datasets were generally low, with correlation coefficients ranging between 0 – 0.34. Maps resulting from the selected models are in Appendix 1 for reference (Figures A1.1-A1.8; Note: Colour ramps are reversed for native vs. exotic species richness, with green representing potentially high ecological integrity – high native richness or low exotic richness – and red representing low native richness or high exotic richness, or lower EI). The lag distance of significant spatial autocorrelation was less than 1 km for several datasets, suggesting it is probably accounted for by the random site effect, however it is 5.5 km for Dataset C and ABC’s final native model, which may explain why a spatial variable (utmn) was selected for Dataset C’s spatial model.      34 Table 2.5. Pearson correlation coefficients and R2  values (*p < 0.05, **0.01, ***0.001) of observed species richness counts vs. those predicted by final models derived from each dataset. Internal validations (grey) were performed using a split-sample bootstrap procedure with 100 repetitions.  Dataset A Dataset B Dataset C Dataset ABC Training Testing native exotic native exotic native exotic native exotic Dataset A      r                  R2 0.300.10* 0.55 0.30*** 0.04 0 0.34 0.12*** 0.10 0.01 -0.47 0.22*** 0.06 0 0.40 0.16*** Dataset B      r                                             R2  0.040 0.310.09*** 0.410.15*** 0.42 0.18*** 0.15 0.023*** -0.25 0.061*** 0.44 0.19*** 0.20 0.04*** Dataset C      r                R2 0.100.01 0.04 0 0.06 0 0.08 0 0.18 0.034*** 0.33 0.11*** 0.16 0.03* 0.42 0.18*** Dataset          r ABC           R2 0.04 0 0.23 0.05*** 0.20 0.04*** 0.30 0.09*** 0.14 0.018*** -0.13 0.08*** 0.24 0.06*** 0.18 0.08*** lag distance (m) of residual SAC 705 814 1,913 300 5,547 962 5,746 2,712  Predictions from Dataset C’s model for exotic richness were significantly correlated with all observed values from the other three datasets, however all in the opposite direction (all negatively related). This model only had one significant parameter, distance from forest, which was positive, indicating a positive trend in exotic species richness with increasing distance from forest. The other datasets exhibited the opposite trend with regards to forest. The exotic richness model developed using Dataset ABC was the only model to achieve significant correlations with the observed values in all three other datasets. The native richness model predictions from Dataset A had the poorest fit with observed values from other datasets. This was slightly improved when predictions from Dataset A were tested against only the observed values from Datasets A, B and ABC that were sampled from the limited extent of the Southern Gulf Islands (SGI), as seen in Table 2.6.      35 Table 2.6. Pearson correlation coefficients and R2  values (*p < 0.05, **0.01, ***0.001)  of observed species richness counts in Southern Gulf Islands quadrats only vs. those predicted by final models derived from Dataset A, whose extent was limited to the Southern Gulf Islands.  Dataset A Training Testing native exotic Dataset A           r                     R2  0.30 0.10* 0.55 0.30*** Dataset B           r                     R2 0.01 0 -0.09 0.008 Dataset C           r                     R2 0.27 0.08* 0.35 0.12** Dataset ABC     r                     R2 0.12 0.02 0.19 0.03*  Despite low observed vs. predicted correlation values, the Fuzzy numerical comparison done between maps revealed fairly high levels of similarity between native richness maps (0.72-0.82) while exotic richness maps were less similar to one another (0.37-0.78) (see maps in Appendix 1 for comparison). The combined dataset (ABC) had the highest average similarity value to the other three maps for native richness, and had the same average similarity value for exotic richness as Dataset B, suggesting that the exotic richness map from Dataset B provides similar utility. Dataset A yielded the exotic richness map with the lowest average similarity value (0.57), and it also predicted much higher average exotic species richness values than the other datasets, indicating bias in the training data.       36 Table 2.7. Results of Fuzzy numerical comparison. Value indicates no (0) to 100% (1) map similarity.  Dataset A Dataset B Dataset C Dataset ABC  native exotic native exotic native exotic native exotic Dataset A         Dataset B 0.72 0.68       Dataset C 0.79 0.37 0.72 0.54     Dataset ABC 0.80 0.66 0.81 0.78 0.82 0.56   average Fuzzy numerical comparison value 0.77 0.57 0.75 0.67 0.78 0.49 0.81 0.67  Discussion This modeling exercise produced predictive maps of plant species richness for Garry oak meadows in to Coastal Douglas Fir zone, and demonstrated that low dataset quality, sampling bias and differences in geographic extent reduce the accuracy of distribution models. In particular, my results suggest that good sampling techniques for collecting data for use in SDMs include surveying evenly over the geographic extent of prediction, sampling at a high intensity if possible at each sampling site, and ensuring to sample regions with suspected differences in species-environment relationships, as with small islets, large islands and Vancouver Island in our study system. Below, I examine in detail the evidence for these recommendations.   Measures of ecological integrity Species richness counts for native species were reasonably consistent at very fine scales (1-m2 sampling units) even after up to 10 years between sampling visits, suggesting that this is a consistent metric for monitoring ecological integrity in Garry oak meadows. Exotic species richness was also repeatable at a moderate level while percent cover measures had almost no repeatability. Given that the GPS error was at least ±3 m, it is  37 unlikely that I resampled the exact same quadrats, however this result demonstrates that species richness is relatively consistent within a patch. This also aligns with theory, as stochastic factors such as rainfall, temperature, and dispersal can cause large fluctuations in species abundance between years (Hernández Plaza et al. 2012), but the consistent measures of species richness suggest species-environment relationships at these sites have not changed dramatically over short time scales.  Effects of extent differences, bias, and quality of training data Dataset A, with its sampling extent limited to the Southern Gulf Islands, had the lowest ability to predict richness values from the other datasets, which was improved in some cases when the other datasets were reduced to just those data collected in the same geographic area. This demonstrates the importance of only predicting values for areas within your sampling area extent (interpolation) and the risks posed by extrapolation, which has been recognized by SDM experts (Jiménez-Valverde et al. 2008, Elith and Graham 2009, Braunisch and Suchant 2010). The lack of correlation between native richness values predicted by the model from Dataset A and observed in SGI-only plots in Dataset B is likely because most of the 62 plots from the latter dataset were located on small islets that are classified as meadow, and therefore the “distance to meadow” variable for these plots is 0 and their predicted value would be solely dictated by the intercept.  Similarly, the fit between Dataset A predictions and observed Dataset B values for exotic richness in the SGI extent is likely low because forest area within a 500-m buffer, one of the two variables in the model from Dataset A, would be zero for quadrats on most small islets.   38  Overall, the bias inherent in Dataset A is twofold. Firstly, because only small islets and larger Gulf Islands were sampled, the model results reflect dynamics in these two landscape types. Previous studies have shown that small islets free of herbivores like deer and protected by ocean from exotic species invasions have high native richness and more rare native species, while herbivory and the dispersal of exotic species by deer and humans on larger Gulf Islands reduces ecological integrity (Gonzales 2008, Bennett 2011). Island size was one of the two variables selected by the native richness model from Dataset A, reflecting these processes. The impacts of this on the resulting predictive map for the CDF are dramatic (see Figure A1.1) – Vancouver Island sites are all predicted to have lower native richness, purely due to the large ‘island size’ of the CDF on Vancouver Island. Secondly, Dataset A was not originally collected for SDM purposes, and while the observer selected sampling locations randomly, their study focus was on exotic-native species interactions (Gonzales 2008). Therefore, they may have unintentionally sampled sites with higher exotic plant diversity, or this is simply an artifact of locating more sampling sites on large Gulf Islands, which are known to be highly invaded by exotic plants as discussed above. Either way, it resulted in higher predictions of exotic species richness than observed in the other two datasets. This is reflected in the low similarity of maps and highlights the effects of sampling bias on model results.  Despite the predictions of Dataset A for native richness being poorly correlated with observed values from Datasets B and C, the map similarity scores are reasonably high  39 (0.72 and 0.79 respectively). The other results are generally confusing in this way, and it begs an important question: How can models with different predictor variables that poorly predict each others’ data result in similar maps?  Decomposing the relationships between each dataset’s selected models provides some explanation. The final model for native richness produced by Dataset A contains two predictor variables: island size and distance of each quadrat to meadow, both with negative coefficients. This means that for quadrats from Dataset A, species richness is higher on smaller islands and closer to (or in) TEM-classified ‘meadow’ polygons. This produces a map where small islands and TEM-classified meadow polygons are predicted to have high species richness. Dataset B resulted in two selected variables in its final native richness model: distance to urban development, and slope (both positive), indicating that quadrats from this dataset have higher native richness further from urban and suburban areas, and in steeper terrain. Dataset A would poorly predict Dataset B because many of Dataset B’s quadrats are located on Vancouver Island which has a very large ‘island size’ as noted above, but Dataset B will also produce a map where small islands are predicted to have high native richness because these islands are far from developed areas. Hence, their resulting maps will be reasonably similar.  Many of the other seemingly diametric results between predicted vs. observed values and map similarity can be explained using similar logic, but it requires a detailed evaluation of each selected predictor variable and the location of the original quadrats in each dataset. This highlights how the characteristics of sampling data, model selection idiosyncrasies, and broader landscape patterns can interact to produce predictions of  40 greater or lesser similarity. Along with the much lower observed vs. predicted correlations between datasets as opposed to within-dataset bootstrap validation, this emphasizes the high importance of validating SDMs with independent data. The composite maps developed using the combined dataset appear to be most robust, which is likely due to the large combined sample size, which increases accuracy (Elith and Leathwick 2007).   Sources of error and limitations The low to moderate correlation between predicted and observed values between datasets suggests that several sources of error exist. Models from Dataset B had the highest accuracy as determined by internal validation, suggesting that small sample sizes may be limiting the model performance of those fit from Datasets A and C. Additionally, it is very possible that 1-m2 quadrats are too fine-scale to adequately reflect relationships between environmental variables sampled at broad-scale resolutions and native or exotic species richness in a meadow patch. A larger sampling unit size, of say 25-m transects or a full count of species in a one hectare survey, may reveal much stronger relationships as found by another study in a subset of our study region (Lilley and Vellend 2009). This is a common issue in data collection for SDMs; if previous surveys used certain sampling methods, it is common for subsequent studies to do the same to bolster sample sizes or provide long-term data, even if it is not the most appropriate way to measure the response variable of interest.   41 There are many other sources of error in SDMs. One of the most common is simply missing covariates from the suite of predictive variables, which the high lag distances from the spatial autocorrelation analysis suggest (Barry and Elith 2006). In this study I am very aware of several important missing covariates. Firstly, deer density, which is known to strongly impact native species richness and abundance, has not been mapped across the CDF (Gonzales and Arcese 2008). The variable ‘island size’ incorporates deer slightly, as very small islets are deer-free. Site history is also an important missing covariate. High levels of unexplained variability exist in plant diversity studies because it is nearly impossible to gain a historical record of stochastic events such as disturbance and animal or human-mediated colonization events (Ostendorf 2011). Data on site history could be acquired by collecting records of grazing history for properties in the CDF, and deer densities have been partially modeled in the Gulf Islands, though they are difficult to map on Vancouver Island (Martin et al. 2011). This study also did not take into account soil characteristics, for which fine scale data were not readily available.  As for adequately accounting for residual spatial autocorrelation during model fitting, it is likely that the procedure recommended by Borcard et al. (1992) was not sufficient. Only one model (native SR for Dataset C) retained a spatial covariate (northing), and as mentioned above, several of the final models still exhibited high lag distances indicating significant residual spatial autocorrelation. Alternatively, it may have been more successful to include a more meaningful autocovariate that adds a distance-weighted function of neighboring response values to the model’s predictor variables to reflect processes such as limited dispersal and site history (Dormann et al. 2007). The best  42 neighborhood size could be determined by comparing models with different neighborhood sizes using a modified Gibbs sampler procedure to generate estimated response values in unsurveyed grid cells (Jewell et al. 2007).    With regards to limitations, these maps do not directly identify locations likely to harbor rare species because rare species were deliberately removed from the training data in order to make a common list of species between datasets. However, research has found that the frequency of rare species increases with native richness in this ecosystem, and therefore one could infer that the sites predicted to have high native richness also likely harbor rare plants (Bennett et al. 2012).   Finally, none of these datasets were collected using a systematic grid sampling design due to land access constraints. Systematic grid sampling has been found to produce robust model results, in one example increasing percent correctly classified presence/absence values for lichen species by approximately 10% over models produced using data collected purposively from sites of expected species occurrence (Edwards et al. 2006). While the data modeled here was collected somewhat randomly from across the landscape, land access constraints are a source of sampling bias, and model accuracy may have been improved appreciably using a more systematic design.   Conclusions  Overall, while different variables were selected in models built using the three different datasets, native richness maps were produced that had reasonably good similarity values  43 and therefore provide some utility. They consistently show areas such as small islands and remnants away from development that experience low human disturbance, which correspond to high native plant species richness. I do not have confidence in the exotic species richness models; these recent arrivals have not yet established strong relationships with the local environment. The native richness maps could be used as part of a GAP Analysis to identify potential gaps in the local conservation network (i.e. existing regional, provincial and national parks and ecological reserves) (Rodríguez et al. 2007), however they should be used with caution as correlations between observed and predicted variables were generally low. Despite this, they do provide important information about the implications of dataset quality and bias. SDMs are frequently constructed out of necessity by using single datasets, with results that are meant to inform management actions in the future. Even with internal validation, their utility remains largely unknown unless they are compared with independent datasets. To my knowledge this is the first time SDMs have been developed and compared using different datasets collected using the same sampling protocol by different observers in a high human impact landscape. The next Chapter explores another potential source of map error with regards to the estimated extent of Garry oak habitat in the CDF.        44 Chapter 3: Assessing the indicator value of Terrestrial Ecosystem Mapping classifications for Garry oak meadow ecosystems  Introduction While most sources of error in species distribution modeling have been identified in the literature, they are only occasionally quantified in modeling studies and factored into final map products (Beale and Lennon 2012). Barry and Elith (2006) segregate uncertainty in SDMs into data error, which includes small sample sizes, biased samples, and lack of absence records; error in variables (e.g. incorrectly interpolated climate data etc.); and model specification errors, including incorrect fit (linear vs. quadratic) and missing covariates. In complex human-dominated landscapes, it is likely impossible to identify and measure all sources of uncertainty, but the utility of predictive models in such regions depends on how much uncertainty is controlled or accounted for in predictions.   Several methods have been proposed to partition and account for uncertainty in SDMs, including relating uncertainty sources to response variables and looking at deviance ratios to see which component of model development contributes most, mapping standard deviations and looking for spatial patterns in uncertainty (Buisson et al. 2010), and creating ‘maps of ignorance’ that illustrate the areas where the reliability of mapped distributions is either known or unknown (Rocchini et al. 2011). In Chapter 2 the issues surrounding data bias, quality and missing covariates were addressed. While creating  45 maps that account for all the identified sources of uncertainty is beyond the scope of this study, the validity of the spatial extent of prediction is explored in this Chapter.  Indicator species associations In order to make species or ecosystem distribution maps most useful to practitioners, predictions must be made over suitable habitat in the landscape. The extent of the Garry oak meadow ecosystem has been reduced by 95% since European colonization, and this reduced area was used to project richness predictions in Chapter 2. This limited extent was identified by selecting Terrestrial Ecosystem Mapping (TEM) polygons – a publicly-available map of land cover classifications – that experts deemed likely to contain Garry oak meadow ecosystems (Gonzales 2008, Bennett 2011, Schuster and Arcese 2012). The 1:5,000 scale TEM landcover classification system was developed by the B.C. Government in partnership with a consulting group using air photo classification (Madrone Environmental Services Ltd. 2008). It assigns letter codes to polygons across the landscape denoting ecosystems (determined by indicator species associations) and anthropogenic landcover types, and also contains information on structural stage of vegetation (ranging from open land to mature forest).   In order to gauge the value of expert-elicited TEM classes as indicators of potential Garry oak meadow, I sought to answer the following questions: 1) Does the number and percent cover of Garry oak meadow indicator plant species recorded in plots of expert-elicited ‘potential Garry oak meadow’ TEM polygons exceed those in other TEM polygon classes? I hypothesized that they have a higher mean count and percent cover of Garry  46 oak meadow indicator species, and an alternative result would indicate that the extent of our distribution map is too limited. 2) Does structural stage as classified in TEM serve as a useful indicator of potential Garry oak meadow? Open structural stage classes are expected to have higher GOE indicator species richness and cover than high structure classes, as these are primarily open meadow species.   Methods Datasets A, B and C introduced and described in Chapter 1 were used in this analysis. Using a map of the CDF classified into TEM polygons (Madrone Environmental Services Ltd. 2008), the TEM class of the polygon that each quadrat (total n = 849) resided within was determined using a spatial join tool in ArcGIS 10.1 (ESRI 2014). In order to test the indicator value of the expert-elicited potential GOE TEM classes, the species richness and percent cover data for 20 Garry oak indicator plant species were extracted from the full species richness data used in Chapter 2. These 20 indicator species have been identified by GOERT as typical and common native Garry oak meadow plants (GOERT 2011):  ! Common Camas (Camassia quamash)/Great Camas (Camassia leichtlinii) ! Harvest Brodiaea (Brodiaea coronaria) ! Long-stolon Sedge (Carex inops) ! Miner’s Lettuce (Claytonia perfoliata) ! Maiden Blue-eyed Mary (Collinsia parviflora) ! Desert Deervetch (Lotus micranthus) ! Seablush (Plectritis congesta) ! Nutka Rose (Rosa nutkana) ! Common Snowberry (Symphoricarpos albus) ! Wild Strawberry (Fragaria vesca) ! Field Chickweed (Cerastium arvense) ! Pacific Sanicle (Sanicula crassicaulis) ! Bracken Fern (Pteridium aquilinium) ! Common Yarrow (Achillea millefolium)  47 ! Cleavers (Galium aparine) ! Western Buttercup (Ranunculus occidentalis) ! California Oatgrass (Danthonia californica) ! Red Fescue (Festuca rubra) ! Blue Wildrye (Elymus glaucus) ! Heath Wood-rush (Luzula multiflora)  Species presence and percent cover were summed to yield a total richness count and total percent cover of indicator species in each quadrat. Mean indicator species richness was then calculated for the expert-elicited potential GOE group (Schuster and Arcese 2013, Bennett 2014), and an ‘Other’ group of the remaining TEM classes (Table 3.1). TEM assigns each polygon a dominant, secondary, and tertiary classification. Therefore, a quadrat was assigned to the GOE group if any of these three contained an expert-elicited potential GOE classification, while quadrats assigned to the ‘Other’ group contained none. Species richness was log-transformed and percent cover was arcsin square root-transformed to achieve normal distributions. A one-tailed Student’s t-test was performed to test the hypothesis that more indicator species would be present and have higher percent cover in the GOE group.   Table 3.1. Expert-elicited potential GOE and ‘Other’-classified TEM classes.  GOE Other CL – Cliff DA – Douglas-fir – arbutus FC – Fescue-camas GO – Garry oak-ocean spray OR – Oceanspray-rose QB – Garry oak-brome RO – Rocky outcrop SC - Cladina-Wallace’s selaginella BE -  Beach CF – Cultivated field DO – Douglas-fir - onion grass DS – Douglas-fir – salal GC – Golf course LM –  Dunegrass-beach pea RF – Western redcedar/grand fir-foamflower RK – Douglas-fir/Western red cedar-kindbergia RW – Rural UR – Urban/suburban Ws50 – Hardhack-sitka sedge   48 To evaluate the relationship between TEM-classified structural stage and GOE indicator species richness and percent cover, a Tukey HSD multiple comparison test was completed to test the difference between means for each structural stage (0, open – 7, mature forest) for the dominant, secondary, and tertiary structural stage classifications.   Results As seen in Figures 3.1, indicator species richness in ‘GOE’-classified quadrats was not higher than in ‘Other’-classified quadrats using a 0.05 significance level (t = -0.77, p = 0.44). Indicator species percent cover also showed no difference between ‘GOE’-classified quadrats and ‘Other’-classified quadrats (Figure 3.2, t = 1.07, p = 0.28). Results by TEM class can be viewed in Table A.2.1. As seen in Figure 3.3, there was no strong trend in richness with increasing vegetation structure for dominant, secondary, or tertiary TEM classifications, though percent cover showed a weak trend towards lower cover with increasing structural stage. In several instances low structural stages had significantly greater richness and cover than higher stages, and the results of the multiple comparison tests can be seen in Table A.2.2.    49  Figure 3.1. Boxplot showing mean GOE indicator species count for TEM classes classified by experts as potential GOE (nquadrats = 741) is not significantly greater than other classes in which quadrats were located (nquadrats = 107; t = -0.77, p = 0.44). ●GOE Other0.00.51.01.52.02.5Expert−elicited TEM groupinglog(indicator sp. count + 1) 50   Figure 3.2. Boxplot showing mean GOE indicator species cover for TEM classes classified by experts as GOE (nquadrats = 741) is not significantly greater than other classes in which quadrats were located (nquadrats = 107; t = 1.07, p = 0.28).       ●●●●●●●GOE Other0.00.51.01.5Expert−elicited TEM groupingarcsin sqrt(indicator sp. % cover) 51  Figure 3.3. Boxplots showing Garry oak meadow indicator species count and percent cover by vegetation structural stage (0-2 = open; 3 = open shrub; 4 = closed shrub; 5 = young forest; 6-7 = mature forest) for dominant (Structural stage 1), secondary (Structural stage 2), and tertiary (Structural stage 3) TEM classifications of sampled quadrats.       ● ● ● ●0 1 2 3 4 5 6 70.00.51.01.52.02.5Structural stage 1● ●0 1 2 3 4 5 6 70.00.51.01.52.02.5Structural stage 2log(indicator sp. count + 1)● ● ●●●●0 1 2 3 4 5 6 70.00.51.01.52.02.5Structural stage 3●●●●●●● ●●●●0 1 2 3 4 5 6 70.00.51.01.5Structural stage 1●●●●●●●●0 1 2 3 4 5 6 70.00.51.01.5Structural stage 2arcsin sqrt(indicator sp. % cover)●●●●●●●0 1 2 3 4 5 6 70.00.51.01.5Structural stage 3 52 Discussion This simple study tested the validity of using a publicly available map of indicator species associations, TEM, to bound spatial predictions for Garry oak meadow ecosystems in the Coastal Douglas Fir zone. The results indicate that neither richness nor percent cover of Garry oak indicator species is appreciably higher in expert-elicited potential ‘GOE’ TEM polygons. This suggests that Garry oak meadow-associated plants do not exist solely in discreet meadow patches, but are dispersed throughout other landcover types as well. This potentially contradicts my initial assumption that the distribution of remaining meadow habitat is known and has been well mapped (Lea 2006, Vellend et al. 2008, Jones et al. 2011). However, this analysis has several key limitations.  Firstly, the data used were collected by individuals who visited sites to collect data on Garry oak meadow-associated plants, and chose these sites based on reconnaissance, local knowledge, and air photos showing open areas. Therefore, this sample is heavily biased, because the ‘Other’-classified sites are much more likely to be misclassified than if these sites were randomly sampled. Additionally, the mimumum mapping unit for TEM is 0.04 hectares, or 400 m2, and therefore quadrats could exist in misclassified areas purely because the meadow patch or clearing they are in is too small to be classified. These limitations also may explain why the results of the vegetation structural stage analysis did not strongly match expectations. There was a slight downward trend in percent cover with increasing structural stage as expected, but richness did not show the same trend. Again, these results may be due to inherent sample bias, and are not adequate  53 to make conclusions on the utility of TEM structural stage classes for identifying meadow habitat.   Another study done on Vancouver Island mapped rare coniferous forest ecosystems using high resolution satellite data, and validated their maps using TEM polygons (Thompson and Gergel 2008). They found that rare associations were mapped with low accuracy, and argued that this may be due in part to the fact that less common associations routinely occur as subdominant classes within a more common association patch, and occupy areas smaller than TEM’s minimum mapping unit. Finally, given that TEM only has a minimum accuracy requirement of 65% - 75%, a large portion of this result could also be due to classification error.  Overall, these results cannot be used to determine whether mapping GOE plant richness predictions by ‘GOE’-classified TEM polygons is valid. As in many cases, this analysis suffers from data deficiency, because the only data available is highly biased. Alternatively, a highly conservative extrapolation of predictions could be done using a map of Garry oak tree occurrence developed for the CDF using LiDAR (Jones et al. 2011). This would be conservative because my own field observations indicate that Garry oak meadow plant communities often occur in the absence of actual oak trees.       54 Chapter 4: Conclusion Key findings This research outlines how field collected and remotely sensed data can be used to construct predictive ecological models at a regional scale. Chapter 2 outlined a process for model and map development and methods for comparing the accuracy and reliability of resulting maps. Chapter 3 investigated whether a commonly used, publicly available plant community dataset was a valid tool for mapping predictions from Chapter 2.   In Chapter 2, the largest of three datasets (in terms of number of quadrats sampled) was found to produce the most robust plant species richness predictions, and a composite model combining all three datasets produced the most reliable maps of plant species richness. The dataset with the smallest geographic sampling extent poorly predicted richness outside of this extent, and the number of quadrats located on small islets, large islands or Vancouver Island had a large effect on the resulting models due to differences in species-environment relationships across these three landscape types. The incorporation of spatial variables into models did not improve model fit, however significant residual spatial autocorrelation at a large scale was detected for Dataset B’s native richness models, suggesting that an important covariate is missing from models. Examples of potential missing covariates include deer density or site use history, which have been found to have large impacts on plant species distributions in other studies (Gonzales 2008).   55 Results from Chapter 3 indicate that expert-elicited ‘potential GOE’ Terrestrial Ecosystem Mapping (TEM) polygon classes may not be a useful extent over which to project predictions for Garry Oak meadow communities, but the bias inherent in the data used in this analysis precludes robust conclusions. However, it also suggests that Garry oak meadow plant communities do not exist solely in discreet meadow patches, but are dispersed throughout other habitat types.  Conservation implications The direct conservation implications of this research are that human development reduces native plant richness in the CDF, and the most intact meadow patches – and therefore possibly the best candidates for conservation – are located on small islets or on rocky outcrops deep in undisturbed coastal forests. These results corroborate the work of others, solidifying local understanding of the factors affecting native Garry oak meadow plant communities (Gonzales 2008, Lilley and Vellend 2009, Bennett 2011).   For those working towards conservation of other ecosystems in high human impact landscapes, a key point from my work is that good models cannot be developed without good data. Therefore, investing in systematic sampling of endangered ecosystems is very important, and a specific emphasis should be placed on creating relationships with private landowners to enable access to patches of habitat across entire regions. Well-designed systematic sampling is what produces robust SDMs in areas with even accessibility, such as large protected areas, and this same goal needs to be championed in highly fragmented landscapes.  56  In general, the complexity of ecosystem dynamics both spatially and temporally suggests that conservation decisions may be most effectively studied on a case-by-case basis, and tools like SDMs can help inform these processes. Systematic conservation planning, which arose in the 1990s, is a process whereby the best conservation outcomes are selected for the least cost (either financial or loss of another desired value) (Margules and Pressey 2000). A systematic conservation planning exercise is currently being orchestrated for the CDF (Schuster and Arcese, unpublished data), and the results of this thesis may become part of this larger project.   Future research This study suggests several different avenues for future research. There are still many opportunities to improve Garry oak meadow ecosystem mapping in the CDF, and these improvements should be sought so that these predictive maps can be most useful for conservation planners. Firstly, how can we develop a better picture of deer density in the CDF, and will this improve the accuracy of plant community maps? Also, would collecting plant data using larger scale sampling methods, such as over 1-ha sampling units or transect surveys, improve models of ecological integrity? Creating maps of site use history including livestock grazing pastures and former areas of native plant cultivation by First Nations may also improve distribution maps.   In general, there also needs to be an increased focus on maps of uncertainty in the SDM literature (Guisan and Zimmermann 2000, Rocchini et al. 2011). While beyond the scope  57 of this thesis, “dissimilarity maps” could be created by overlaying the maps compared in Chapter 2 in order to identify whether there are consistent locations in the landscape where prediction similarity is low, which may help identify model inadequacies.   Overall, with several studies being completed on Garry oak meadow plant communities in the last decade, we now have a good picture of the factors protecting and degrading ecological integrity in this ecosystem. However, the low accuracy of the models presented in this paper suggests that SDMs are challenging to develop in complex, high human-impact landscapes. A key piece of research that would be very useful to the SDM and conservation planning community would be a meta-analysis determining whether SDMs constructed in human modified landscapes are less or more robust than those constructed primarily through niche modeling in “natural” landscapes. I suspect that species distributions will become harder to map using remotely sensed variables as human impacts on landscapes increase, and that conservation managers will become more reliant on ground surveys in these situations. Ultimately, developing robust spatial information products for species distributions facilitates species and habitat conservation, and the tools available to do so are improving rapidly.            58 Bibliography Albert, C. H., N. G. Yoccoz, T. C. Edwards, C. H. Graham, N. E. Zimmermann, and W. Thuiller. 2010. Sampling in ecology and evolution - bridging the gap between theory and practice. Ecography 33:1028–1037. Andreasen, J. K., R. V. O. Neill, R. Noss, and N. C. Slosser. 2001. Considerations for the development of a terrestrial index of ecological integrity. Ecological Indicators 1:21–35. Austin. 1998. An ecological perspective on biodiversity investigations: examples from Australian eucalypt forests. Annals of the Missouri Botanical Garden 85:2-17. Barbour, M., and M. Paul. 2010. Adding value to water resource management through biological assessment of rivers. Hydrobiologia 651:17–24. Barry, S., and J. Elith. 2006. Error and uncertainty in habitat models. Journal of Applied Ecology 43:413–423. Beale, C. M., and J. J. Lennon. 2012. Incorporating uncertainty in predictive species distribution modelling. Philosophical Transactions of the Royal Society of London. Series B, Biological sciences 367:247–58. Bennett, J. R. 2011. Multi-scale comparison of native and exotic communities in the Garry oak ecosystem of British Columbia. PhD Thesis, University of British Columbia, Vancouver, British Columbia, Canada. Bennett, J. R. 2014. Comparison of native and exotic distribution and richness models across scales reveals essential conservation lessons. Ecography 37:120–129.  59 Bennett, J. R., M. Vellend, P. L. Lilley, W. K. Cornwell, and P. Arcese. 2012. Abundance, rarity and invasion debt among exotic species in a patchy ecosystem. Biological Invasions 15:707–716. Beyer, H.L. 2014. Geospatial Modelling Environment version 0.7.2.1, http://www.spatialecology.com/gme/. Brocard, D., P. Legendre, and P. Drapeau. 1992. Partialling out the spatial component of ecological variation. Ecology 73:1045–1055. Braunisch, V., and R. Suchant. 2010. Predicting species distributions based on incomplete survey data: the trade-off between precision and scale. Ecography 33:826–840. Brockington, D. 2002. Fortress Conservation: The Preservation of the Mkomazi Game Reserve, Tanzania. International African Institute, Oxford, UK.  Buisson, L., W. Thuiller, N. Casajus, S. Lek, and G. Grenouillet. 2010. Uncertainty in ensemble forecasting of species distributions. Global Change Biology 16:1145–1157. Burns, K.C., P. McHardy, and S. Pledger. 2009. The small-island effect: fact or artefact? Ecography 32:587-600. Carignan, V., and M.-A. Villard. 2002. Selecting indicator species to monitor ecological integrity: a review. Environmental Monitoring and Assessment 78:45–61. Clements, D., Hooper, T.D., Martell, K., and C. Masson (eds). Restoring British Columbia’s Garry Oak Ecosystems: Principles and Practices © Garry Oak Ecosystems Recovery Team (GOERT) 2011, Victoria, British Columbia, Canada.   60 Clements, D. R., S. Luginbill, D. A. Jordan, R. Van Dragt, and R. K. Pelant. 2011. Techniques to promote Garry Oak seedling growth and survival in areas with high levels of herbivory and competition techniques to promote Garry Oak seedling growth and survival in areas with high levels of herbivory and competition. Northwest Science 85:172–181. Cowling, R. M., and W. J. Bond. 1991. How Small Can Reserves Be$? An Empirical Approach in Cape Fynbos, South Africa. Biological Conservation 58:243–256. Diamond, J.M. 1976. Island biogeography and conservation: strategy and limitations. Science 193:1027–29. Dormann, C., J. M. McPherson, M. B. Araújo, R. Bivand, J. Bolliger, G. Carl, R. G. Davies, A. Hirzel, W. Jetz, W. Daniel Kissling, I. Kühn, R. Ohlemüller, P. R. Peres-Neto, B. Reineking, B. Schröder, F. M. Schurr, and R. Wilson. 2007. Methods to account for spatial autocorrelation in the analysis of species distributional data: a review. Ecography 30:609–628. Dubuis, A., J. Pottier, V. Rion, L. Pellissier, J.P. Theurillat, and A. Guisan. 2011. Predicting spatial patterns of plant species richness: a comparison of direct macroecological and species stacking modelling approaches. Diversity and Distributions 17:1122–1131. Edwards, T. C., D. R. Cutler, N. E. Zimmermann, L. Geiser, and G. G. Moisen. 2006. Effects of sample survey design on the accuracy of classification tree models in species distribution models. Ecological Modelling 199:132–141.  61 Franklin, J. 1995. Predictive vegetation mapping: geographic modeling of biospatial patterns in relation to environmental gradients. Progress in Physical Geography 19:494-519.  Franklin, J.F. and D.B. Lindenmayer. 2012. Importance of matrix habitats in maintaining biological diversity. Proceedings of the National Academy of Sciences of the United States of America 106:349-350. Elith, J., and C. H. Graham. 2009. Do they? How do they? Why do they differ? On finding reasons for differing performances of species distribution models. Ecography 32:66–77. Elith, J., and J. Leathwick. 2007. Predicting species distributions from museum and herbarium records using multiresponse models fitted with multivariate adaptive regression splines. Diversity and Distributions 13:265–275. Erickson, W. 1993. Ecosystems in British Columbia At Risk: Garry Oak Ecosystems. Ministry of Environment, Lands and Parks, Province of British Columbia. Victoria, British Columbia, Canada. Erickson, W. R., and D. Meidinger. 2007. Garry Oak (Quercus garryana) Plant Communities in British Columbia$: A Guide to Identification. B.C. Ministry of Forests and Range Forest Science Program, Victoria, B.C. Franklin, J. 2009. Mapping Species Distributions: Spatial Inference and Prediction. Page 320. Cambridge University Press, Cambridge, UK.  Franklin, J. 2013. Species distribution models in conservation biogeography: developments and challenges. Diversity and Distributions 19:1217–1223.  62 Franklin, J. F., and D. B. Lindenmayer. 2009. Importance of matrix habitats in maintaining biodiversity. Proceedings of the National Academy of Sciences of the United States of America 106:349–350. Garry Oak Ecosystem Recovery Team (GOERT). 2011. Restoring British Columbia’s Garry Oak Ecosystems: Principles and Practices. Victoria, British Columbia, Canada. Gonzales, E. K. 2008. The effects of herbivory, competition, and disturbance on island meadows. PhD Thesis, University of British Columbia, Vancouver, British Columbia, Canada. Gonzales, E. K., and P. Arcese. 2014. Herbivory more limiting than competition on early and established native plants in an invaded meadow. Ecology 89:3282–3289. Government of British Columbia. 1997. Terrain Resource Information Management.  http://archive.ilmb.gov.bc.ca/crgb/pba/trim/. Guisan, A., R. Tingley, J. B. Baumgartner, I. Naujokaitis-Lewis, P. R. Sutcliffe, A. I. T. Tulloch, T. J. Regan, L. Brotons, E. McDonald-Madden, C. Mantyka-Pringle, T. G. Martin, J. R. Rhodes, R. Maggini, S. a Setterfield, J. Elith, M. W. Schwartz, B. a Wintle, O. Broennimann, M. Austin, S. Ferrier, M. R. Kearney, H. P. Possingham, and Y. M. Buckley. 2013. Predicting species distributions for conservation decisions. Ecology Letters 16:1424–1435. Guisan, A., and N. E. Zimmermann. 2000. Predictive habitat distribution models in ecology. Ecological Modelling 135:147–186. Guisan, A., R. Tingley, J. B. Baumgartner, I. Naujokaitis-Lewis, P. R. Sutcliffe, A. I. T. Tulloch, T. J. Regan, L. Brotons, E. McDonald-Madden, C. Mantyka-Pringle, T. G.  63 Martin, J. R. Rhodes, R. Maggini, S. Setterfield, J. Elith, M. W. Schwartz, B. a Wintle, O. Broennimann, M. Austin, S. Ferrier, M. R. Kearney, H. P. Possingham, and Y. M. Buckley. 2013. Predicting species distributions for conservation decisions. Ecology Letters 16:1424–1435. Heinänen, S., J. Erola, and M. Numers. 2012. High resolution species distribution models of two nesting water bird species: a study of transferability and predictive performance. Landscape Ecology 27:545–555. Heink, U., and I. Kowarik. 2010. What criteria should be used to select biodiversity indicators? Biodiversity and Conservation 19:3769–3797. Hernández Plaza, E., L. Navarrete, C. Lacasta, and J. L. González-Andujar. 2012. Fluctuations in plant populations: role of exogenous and endogenous factors. Journal of Vegetation Science 23:640–646. Isbell, F., Calcagno, V., Hector, A., Connolly, J., Harpole, W. S., Reich, P. B., Scherer-Lorenzen, M., Schmid, B., Tilman, D., van Ruijven, J., Weigelt, A., Wilsey, B. J., Zavaleta, E. S., and M. Loreau. 2011. High plant diversity is needed to maintain ecosystem services. Nature 477:199–203. Jiménez-Valverde, A., J. M. Lobo, and J. Hortal. 2008. Not as good as they seem: the importance of concepts in species distribution modelling. Diversity and Distributions 14:885–890. Jones, T. G., Coops, N. C. and Sharma, T. 2011. Exploring the utility of hyperspectral imagery and LiDAR data for predicting Quercus garryana ecosystem distribution and aiding in habitat restoration. Restoration Ecology 19:245–256.  64 Jørgensen, S.E. 1994. Models as instruments for combination of ecological theory and environmental practice. Ecological Modelling 5:75-76 LaPaix, R., B. Freedman, and D. Patriquin. 2009. Ground vegetation as an indicator of ecological integrity. Environmental Reviews 17:249–265. Le Lay, G., R. Engler, E. Franc, and A. Guisan. 2010. Prospective sampling based on model ensembles improves the detection of rare species. Ecography 33:1015–1027. Lea, T. 2006. Historical Garry Oak Ecosystems of Vancouver Island , British Columbia , pre-European Contact to the Present. Davidsonia 17:34–50. Levins, R. 1969. Some demographic and genetic consequences of environmental heterogeneity for biological control. Bulletin of the Entomological Society of America 15:237–240.  Lilley, P. L., and M. Vellend. 2009. Negative native-exotic diversity relationship in oak savannas explained by human influence and climate. Oikos 118:1373–1382. Lindenmayer D.B. and H.P. Possingham. 1996. Modelling the inter-relationships between habitat patchiness, dispersal capability and metapopulation persistence of the endangered species, Leadbeater's possum, in south-eastern Australia. Landscape Ecology 11:79-105  Luna-Jorquera, G., C. E. Fernández, and M. M. Rivadeneira. 2011. Determinants of the diversity of plants, birds and mammals of coastal islands of the Humboldt current systems: implications for conservation. Biodiversity and Conservation 21:13–32. Madrone Environmental Services Ltd. 2008. Terrestrial ecosystem mapping of the coastal Douglas-fir biogeoclimatic zone. Prepared for the Integrated Land Management Bureau, Duncan, B.C.   65 Margules, C.R. and R.L. Pressey. 2000. Systematic conservation planning. Nature 405:242–53. Martin, T. G., P. Arcese, and N. Scheerder. 2011. Browsing down our natural heritage: Deer impacts on vegetation structure and songbird populations across an island archipelago. Biological Conservation 144:459–469. May, R. 1975. Island biogeography and the design of wildlife preserves. Nature 254:177–78. McKinney, M. L. 2008. Effects of urbanization on species richness: A review of plants and animals. Urban Ecosystems 11:161–176. Meidinger, D. 2003. Ecosystem Mapping Accuracy and Timber Supply Applications. B.C. Forest Service.  Meidinger, D. and J. Pojar. 1991. Ecosystems of British Columbia. B.C. Ministry of Forests. Milieu-en Natuurplanbureau (MNP - Netherlands Environmental Assessment Agency). Map Comparison Toolkit v. 3.2.3. © RIKS BV 2014.  Niemi, G. J., and M. E. McDonald. 2004. Application of Ecological Indicators. Annual Review of Ecology, Evolution, and Systematics 35:89–111. Noss, R. F. 1990. Indicators for Monitoring Biodiversity: A Hierarchical Approach. Conservation Biology 4:355–364. Ostendorf, B. 2011. Overview: Spatial information and indicators for sustainable management of natural resources. Ecological Indicators 11:97–102. Parks Canada. 2008. Principles and Guidelines for Ecological Restoration in Canada’s Protected Natural Areas. © Parks Canada and the Canadian Parks Council.    66 Pavlik, B.M. 1997. Perspectives, tools, and institutions for conserving rare plants. Southwestern Naturalist 42:375-383. Peltzer, D. A., S. Ferriss, and R. G. Fitzjohn. 2008. Predicting weed distribution at the landscape scale$: using naturalized Brassica as a model system. Journal of Applied Ecology 45:467–475. Prugh, L. R., K. E. Hodges, A. R. E. Sinclair, and J. S. Brashares. 2008. Effect of habitat area and isolation on fragmented animal populations. Proceedings of the National Academy of Sciences of the United States of America 105:20770–5. R Core Team. 2013. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing. Vienna, Austria. http://www.R-project.org. Reza, M. I. H., and S. A. Abdullah. 2011. Regional Index of Ecological Integrity: A need for sustainable management of natural resources. Ecological Indicators 11:220–229. Rands, M. R. W., W. M. Adams, L. Bennun, S. H. M. Butchart, A. Clements, D. Coomes, A. Entwistle, I. Hodge, V. Kapos, J. P. W. Scharlemann, W. J. Sutherland, and B. Vira. 2010. Biodiversity Conservation: Challenges Beyond 2010. Science 329:1298-1303. Riordan, E. C., and P. W. Rundel. 2009. Modelling the distribution of a threatened habitat: the California sage scrub. Journal of Biogeography 36:2176–2188. Rocchini, D., J. Hortal, S. Lengyel, J. M. Lobo, a. Jimenez-Valverde, C. Ricotta, G. Bacaro, and a. Chiarucci. 2011. Accounting for uncertainty when mapping species distributions: The need for maps of ignorance. Progress in Physical Geography 35:211–226.  67 Rodríguez, J. P., L. Brotons, J. Bustamante, and J. Seoane. 2007. The application of predictive modelling of species distribution to biodiversity conservation. Diversity and Distributions 13:243–251. Rzanny, M., and W. Voigt. 2012. Complexity of multitrophic interactions in a grassland ecosystem depends on plant species diversity. The Journal of Animal Ecology 81:614–27. Sala, O. E., F. Stuart Chapin, III, J. J. Armesto, E. Berlow, J. Bloomfield, R. Dirzo, E. Huber-Sanwald, L. F. Huenneke, R. B. Jackson, A. Kinzig, R. Leemans, D. M. Lodge, H. A. Mooney, M. Oesterheld, N. L. Poff, M. T. Sykes, B. H. Walker, M. Walker, and D. H. Wall. 2000. Global Biodiversity Scenarios for the Year 2100. Science 287:1770–1774. Sardà-Palomera, F., L. Brotons, D. Villero, H. Sierdsema, S. Newson, and F. Jiguet. 2012. Mapping from heterogeneous biodiversity monitoring data sources. Biodiversity and Conservation 21:2927–2948. Schuster, R. and P. Arcese. 2012. Using bird species community occurrence to prioritize forests for old growth restoration. Ecography 35:1–9. Seiferling, I., R. Proulx, P. R. Peres-Neto, L. Fahrig, and C. Messier. 2012. Measuring protected-area isolation and correlations of isolation with land-use intensity and protection status. Conservation Biology 26:610–618. Shannon, C.E. 1948. A mathematical theory of communication. The Bell System Technical Journal 27:623-656. Simpson, E. H. 1949. Measurement of diversity. Nature 163:688.  68 Simberloff, D. S. and L. G. Abele. 1982. Refuge design and island biogeograpic theory - effects of fragmentation. American Naturalist 120:41-56 Spengler, A. 2011. How island size and isolation affect bee and wasp ensembles on small tropical islands: a case study from Kepulauan Seribu, Indonesia. Journal of Biogeography 38:1365-2699. Thompson, S. D., and S. E. Gergel. 2008. Conservation implications of mapping rare ecosystems using high spatial resolution imagery: recommendations for heterogeneous and fragmented landscapes. Landscape Ecology 23:1023–1037. Tierney, G. L., D. Faber-Langendoen, B. R. Mitchell, W. G. Shriver, and J. P. Gibbs. 2009. Monitoring and evaluating the ecological integrity of forest ecosystems. Frontiers in Ecology and the Environment 7:308–316. Triantis, K.A., Vardinoyannis, K., Tsolaki, E.P., Botsaris, I., Lika, K., and M. Mylonas. 2006. Re-approaching the small island effect. Journal of Biogeography 33:914-923. Tuomisto, H. 2010. A consistent terminology for quantifying species diversity? Yes, it does exist. Oecologia 4: 853–860. Tyler, N. J. C., J. M. Turi, M. a. Sundset, K. Strøm Bull, M. N. Sara, E. Reinert, N. Oskal, C. Nellemann, J. J. McCarthy, S. D. Mathiesen, M. L. Martello, O. H. Magga, G. K. Hovelsrud, I. Hanssen-Bauer, N. I. Eira, I. M. G. Eira, and R. W. Corell. 2007. Saami reindeer pastoralism under climate change: Applying a generalized framework for vulnerability studies to a sub-arctic social–ecological system. Global Environmental Change 17:191–206.  69 Underwood, E. C., J. H. Viers, K. R. Klausmeyer, R. L. Cox, and M. R. Shaw. 2009. Threats and biodiversity in the mediterranean biome. Diversity and Distributions 15:188–197. Vellend, M., A. D. Bjorkman, and A. McConchie. 2008. Environmentally biased fragmentation of oak savanna habitat on southeastern Vancouver Island, Canada. Biological Conservation 141:2576–2584. Wang, T., Hamann, A., Spittlehouse, D., and Aitken, S. N. 2006. Development of scale-free climate data for western Canada for use in resource management. International Journal of Climatology 26:383-397. Wang, T., Hamann, A., Spittlehouse, D., and Murdock, T. N. 2012. ClimateWNA – High-resolution spatial climate data for western North America. Journal of Applied Meteorology and Climatology 61:16-29. Wu, X. Ben, and F. E. Smeins. 2000. Multiple-scale habitat modeling approach for rare plant conservation. 51:11-28. Zimmermann, N. E., T. C. Edwards, C. H. Graham, P. B. Pearman, and J.-C. Svenning. 2010. New trends in species distribution modelling. Ecography 33:985–989. Zuur, A. F., E. N. Ieno, N. Walker, A. Saveliev, and G. M. Smith. 2009. Mixed effects models and extensions in ecology with R. Springer, New York, NY, USA.   70 Appendix 1  Table A1.1. Spearman rank correlations of variables permitted in model selection (r < 0.60) for Dataset A.   isl.sz rd.lgth.500 nr.ocn ag.ar.500 dev.ar.500 for.ar.500 nr.mead slp east clim.pc1 isl.sz           rd.lgth.500 0.596          nr.ocn 0.565 0.366         ag.ar.500 0.234 0.450 0.169        dev.ar.500 0.254 0.411 0.251 0.547       for.ar.500 0.513 0.441 0.433 0.184 -0.003      nr.mead 0.174 0.377 0.120 0.275 0.459 0.155     slp 0.498 0.291 0.411 0.048 0.186 0.353 0.112    east -0.207 -0.305 -0.349 -0.198 -0.121 -0.103 -0.140 -0.166   clim.pc1 0.038 0.024 0.039 0.322 0.382 -0.195 -0.025 0.024 0.283  clim.pc2 -0.277 -0.313 0.100 -0.137 -0.373 -0.099 -0.326 0.076 -0.309 -0.411  Table A1.2. Spearman rank correlations of variables permitted in model selection (r < 0.60) for Dataset B.   rd.lgth.500 nr.ocn ag.ar.500 dev.ar.500 for.ar.500 nr.mead slp east clim.pc1 rd.lgth.500          nr.ocn 0.366         ag.ar.500 0.450 0.169        dev.ar.500 0.411 0.251 0.547       for.ar.500 0.441 0.433 0.184 -0.003      nr.mead 0.377 0.120 0.275 0.459 0.155     slp 0.291 0.411 0.048 0.186 0.353 0.112    east -0.305 -0.349 -0.198 -0.121 -0.103 -0.140 -0.166   clim.pc1 0.024 0.039 0.322 0.382 -0.195 -0.025 0.024 0.283  clim.pc2 -0.313 0.100 -0.137 -0.373 -0.099 -0.326 0.076 -0.309 -0.411  71 Table A1.3. Spearman rank correlations of variables permitted in model selection (r < 0.60) for Dataset C.   rd.lgth.500 nr.ocn ag.ar.500 dev.ar.500 for.ar.500 nr.mead slp east clim.pc1 rd.lgth.500          nr.ocn 0.366         ag.ar.500 0.450 0.169        dev.ar.500 0.411 0.251 0.547       for.ar.500 0.441 0.433 0.184 -0.003      nr.mead 0.377 0.120 0.275 0.459 0.155     slp 0.291 0.411 0.048 0.186 0.353 0.112    east -0.305 -0.349 -0.198 -0.121 -0.103 -0.140 -0.166   clim.pc1 0.024 0.039 0.322 0.382 -0.195 -0.025 0.024 0.283  clim.pc2 -0.313 0.100 -0.137 -0.373 -0.099 -0.326 0.076 -0.309 -0.411  Table A1.4. Spearman rank correlations of variables permitted in model selection (r < 0.60) for Dataset ABC.   rd.lgth.500 nr.ocn ag.ar.500 dev.ar.500 for.ar.500 nr.mead slp east clim.pc1 rd.lgth.500          nr.ocn 0.366         ag.ar.500 0.450 0.169        dev.ar.500 0.411 0.251 0.547       for.ar.500 0.441 0.433 0.184 -0.003      nr.mead 0.377 0.120 0.275 0.459 0.155     slp 0.291 0.411 0.048 0.186 0.353 0.112    east -0.305 -0.349 -0.198 -0.121 -0.103 -0.140 -0.166   clim.pc1 0.024 0.039 0.322 0.382 -0.195 -0.025 0.024 0.283  clim.pc2 -0.313 0.100 -0.137 -0.373 -0.099 -0.326 0.076 -0.309 -0.411    72   Figure A1.1. Predicted native species richness values binned by quantiles in Garry oak meadow habitat based on the final model selected using Dataset A as training data.   ±0 10 20 30 405kmVancouverIslandDataset A: native SR0.84 - 4.14.2 - 4.24.3 - 4.84.9 - 5.35.4 - 21 73   Figure A1.2. Predicted native species richness values binned by quantiles in Garry oak meadow habitat based on the final model selected using Dataset B as training data.  ±0 10 20 30 405kmVancouverIslandDataset B: native SR2 - 4.64.7 - 5.45.5 - 6.16.2 - 6.97 - 9.9 74  Figure A1.3. Predicted native species richness values binned by quantiles in Garry oak meadow habitat based on the final model selected using Dataset C as training data.   ±0 10 20 30 405kmVancouverIslandDataset C: native SR2 - 3.63.7 - 4.14.2 - 4.64.7 - 5.25.3 - 9.7 75   Figure A1.4. Predicted native species richness values binned by quantiles in Garry oak meadow habitat based on the final model selected using Dataset ABC as training data.   ±0 10 20 30 405kmVancouverIslandDataset ABC: native SR2.7 - 3.94 - 4.64.7 - 5.35.4 - 5.45.5 - 6.3 76   Figure A1.5. Predicted exotic species richness values binned by quantiles in Garry oak meadow habitat based on the final model selected using Dataset A as training data.   ±0 10 20 30 405kmVancouverIslandDataset A: exotic SR1.7 - 6.66.7 - 77.1 - 7.37.4 - 88.1 - 8.8 77   Figure A1.6. Predicted exotic species richness values binned by quantiles in Garry oak meadow habitat based on the final model selected using Dataset B as training data.   ±0 10 20 30 405 kmVancouverIslandDataset B: exotic SR0.91 - 3.94 - 4.64.7 - 5.45.5 - 6.36.4 - 12 78   Figure A1.7. Predicted exotic species richness values binned by quantiles in Garry oak meadow habitat based on the final model selected using Dataset C as training data.    ±0 10 20 30 405 kmVancouverIslandDataset C: exotic SR1.81.9 - 3.23.3 - 3.63.7 - 4.14.2 - 5.9 79  Figure A1.8. Predicted exotic species richness values binned by quantiles in Garry oak meadow habitat based on the final model selected using Dataset A as training data.        ±0 10 20 30 405kmVancouverIslandDataset ABC: exotic SR0.58 - 4.14.2 - 4.54.6 - 4.74.8 - 5.25.3 - 7.3 80 Appendix 2 Table A2.1. Mean GOE indicator species count of primary TEM classes; n denotes number of quadrats (849 total).   Dataset A Dataset B Dataset C Dataset ABC TEM primary class mean count n mean count n mean count mean count mean count n BE: Beach - - 8.00 3 - - 8.00 3 CF: Cultivated field 1.83 12 - - - - 1.83 12 CL: Cliff 4.11 9 - - 0.50 2 3.45 11 DA: Douglas fir-Arbutus 3.33 15 5.19 32 2.60 25 3.90 72 DO: Douglas fir- Oniongrass 2.21 14 3.73 22 1.40 5 2.88 41 DS: Douglas fir-Salal 2.20 15 5.56 50 1.95 65 3.36 130 FC: Fescue-Camas 4.55 53 5.19 26 4.91 11 4.78 90 GC: Golf course - - 2.00 1 - - 2.00 1 GO: Garry Oak-oceanspray 4.44 9 0.00 0 3.67 6 4.13 15 LM: Dune grass-beach pea 1.75 4 0.00 0 0.00 0 1.75 4 OR: Oceanspray-Rose 4.36 11 4.40 5 4.00 1 4.35 17 QB: Garry Oak-brome 2.29 7 3.62 76 3.40 20 3.49 103 RF: Western red cedar- Foamflower 3.67 3 - - - - 3.67 3 RK: Westerm red cedar- Douglas fir-Kindbergia 2.78 9 - - - - 2.78 9 RO: Rock outcrop 3.57 14 4.41 109 4.17 6 4.31 129 RW: Rural - - 5.26 23 - - 5.26 23 SC: Cladina-Wallace’s selaginella 1.67 3 2.94 120 2.80 41 2.88 164  81  Dataset A Dataset B Dataset C Dataset ABC TEM primary class mean count n mean count n mean count n mean count n UR: Urban/suburban - - 3.38 8 - - 3.38 8 Ws50: Hardhack-Sitka sedge 3.50 6 4.80 5 - - 4.09 11  Table A2.2. Significant results (α < 0.05) for Tukey HSD multiple comparison tests between richness and percent cover means. Comparisons not shown were all not significant (NS).   log(percent cover + 1) arcsin square root (percent cover) Dominant TEM class 0 (open) > 1 (low vegetation)  (p = 0.007) 0 (open) > 5 (young forest)     (p < 0.001) 3 (shrub) > 5 (young forest)    (p = 0.005)  Dominant TEM class 0 (open) > 1 (low vegetation) (p = 0.036) 0 (open) > 6 (mature forest)  (p = 0.015) 2 (low veg.) > 5 (young forest) (p < 0.001) 2 (low veg.) > 6 (young forest) (p = 0.044) 3 (shrub) > 5 (young forest)  (p < 0.001) 3 (shrub) > 6 (young forest)  (p = 0.007)  Secondary TEM class NS   Secondary TEM class 0 (open) > 5 (young forest)  (p < 0.001) Tertiary TEM class 0 (open) > 4 (closed shrub)   (p = 0.001) Tertiary TEM class 0 (open) > 1 (low vegetation)    (p = 0.007)  

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.24.1-0167479/manifest

Comment

Related Items