. .1... I 3: . ., xivuyiaufivz. 310...?» . 1.1. .5»: rahafiue 3:...) .Ci .2 .. s . . q A u“! inn-3 dubmith‘ . $4 . \Ert v. :1}. 3’ I. 2.53%. L 2 in... ....mm.fi%w It a . . i . 19.... .. a x ffgihvfl‘r. 1.1.. i. .1 : . m3- : 9.1.3:.“ Lil) 538.33”. .1 {ailfki v$ , \ A ‘ Ii‘...‘ {3 313.59.,- glwi‘l nil..- \. It! A 6.1.314... .3 hvflfkiéxrwfi. (if... Ofinvati . I . I» . wag. , ”x33." 2 in. .3... 3. 3431......3 2.3... .1. . avafiaefim .3. . Us . . 7 .31.... . , .33.... 1.. i . a 3? m LIBRQ‘ISESNIIVERSITY MICHIGAN STA EAST LANSING, MICH 48824-1048 This is to certify that the dissertation entitled INCORPORATING SATELLITE IMAGERY INTO ANALYSES OF AVIAN DISTRIBUTION PATTERNS ACROSS FORESTED LANDSCAPES presented by_ EDWARD J. LAURENT has been accepted towards fulfillment of the requirements for the Doctoral degree in Fisheries and Wildlife %%/\ Major Professor’s Signature April 5, 2005 Date MSU is an Affirmative Action/Equal Opportunity Institution PLACE IN RETURN Box to remove this checkout from your record. TO AVOID FINES return on or before date due. MAY BE RECALLED with earlier due date if requested. DATE DUE | DATE DUE DATE DUE wag?” 1'JJAN122009 CE§5118920'078 .07 2 b 138 \ <3 2/05 CICIRCIDateDqundd-pJS —______.-.. _. m..- INCORPORATING SATELLITE IMAGERY INTO ANALYSES OF AVIAN DISTRIBUTION PATTERNS ACROSS F ORESTED LANDSCAPES By Edward J. Laurent A DISSERTATION Submitted to Michigan State University in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY Department of Fisheries and Wildlife 2005 ABSTRACT INCORPORATIN G SATELLITE IMAGERY INTO ANALYSES OF AVIAN DISTRIBUTION PATTERNS ACROSS F ORESTED LANDSCAPES By Edward J. Laurent While numerous studies have documented the relationships among bird communities and gradients of vegetation structure and composition, there is still little information regarding specific fine scale habitat associations of bird species over large areas of managed forests. Furthermore, the utility of satellite imagery for identifying influential factors on bird species occurrences and richness is rarely considered during field data collection. As part of a multidisciplinary partnership, I investigated landscape patterns of bird species occurrences and richness over a ~400,000 ha forested region of Michigan’s Upper Peninsula. Small sampling plots comparable in size to pixels of Landsat 7 ETM+ imagery (30 meter radius, n= 433) were surveyed for birds, vegetation and land cover. Bird data were also collected using 50m, 100m and unlimited distance thresholds. Landsat 7 ETM+ imagery was used to investigate spectral relationships among the data. Several vegetation variables describing northern hardwood stands had significant independent contributions to the occurrences of 4 bird species. Some variables had both positive and negative relationships, indicating that horizontal and vertical diversity within northern hardwood stands need to be an important consideration during forest management activities. In contrast, bird species richness across the study region was highest when small areas contained large proportions of the same land cover types that dominated their surroundings. This relationship was detected through a spatially variable association between bird species richness, the Normalized Difference Vegetation Index (NDVI) and land cover types. Avian species richness estimates varied spatially in relation to NDVI and the proportion of non-deciduous land cover pixels surrounding each plot influenced the relationship. However, NDVI values were positively dependent on the proportion of deciduous forest within them. Species richness was therefore highest in deciduous forests within regions dominated by deciduous forest and the relationship was reversed in regions dominated by non-deciduous forest. I also investigated the potential of using unclassified spectral information for predicting the distribution of three bird species. Accuracy statistics for each species were affected in different ways by the detection thresholds of point count surveys used to stratify plots into presence and absence classes and window sizes used in spectral signature development. Comparisons with rule-based maps created using the approach of Gap Analysis showed that spectral information predicted the occurrences of the investigated species better than could be done using known land cover associations. Combined, the research presented in this dissertation increases the general understanding of bird-habitat relationships across forested landscapes. Furthermore, I provide several novel methods for quantifying patterns and understanding the processes causing the patterns of species distributions and gradients of species richness over large landscapes. Important processes were found to be ecologically relevant, geometrically constrained, and methodologically dependent. This knowledge has strong potential for future wildlife managment efforts. However, plot size and study extent have a strong effect on quantified relationships and need to be considered when interpreting results. ACKNOWLEDGEMENTS I would like to thank my wife Sherrie Emerine, my parents Ed and Nancy Laurent, my advisor Dr. Jack Liu and my committee members Dr. Catherine Lindell, Dr. David Lusch, and Dr. Brian Maurer for their support. Joseph LeBouton and Dr. Mike Walters were great collaborators for collecting data and editing manuscripts. The PHASEI program could not have been developed without the programming expertise of Dr. Demetrios Gatziolis and Dr. Haijin Shi. In addition, R. Doepker, M. Donovan, K. Hall, F. Lupi, and L. Raceveskis helped with project development. Bird crews extraordinaire included N. Brown, C. Caux, M. Covell, A. Keaveney, A. Levine, E. Morrisette, and M. Straus. Vegetation data were dependent on the great help of Megan Daniels. Other vegetation team members included T. Balk, R. Bernard, K. Campbell, A. Chediack, K. Daugherty, L. Gilner, S. Kissman, P. Nagelkirk, R. Nagelkirk and H. Todd. International Paper and Mead Corporation allowed me to access their properties for this study. Financial support was provided by the Michigan Department of Natural Resources, the NASA Earth System Science Fellowship program, the Budweiser Conservation Scholarship program sponsored by the National Fish and Wildlife Foundation, the George and Martha Wallace Research Award, the USDA National Research Initiative, USFS McIntire-Stennis grants, and the Department of Fisheries and Wildlife at Michigan State University. Fuzz therapy was provided by Max, Pinto Solo, Nub, Fruitcake and Remey T. Bonzai. iv TABLE OF CONTENTS List of Tables ................................................................................... vii List of Figures ................................................................................... viii CHAPTER 1 BACKGROUND AND RESEARCH SUMMARY ........................................ 1 Background .................................................................................... 1 Research Summary ........................................................................... 5 CHAPTER 2 HABITAT ASSOCIATIONS OF BIRDS WITHIN INDUSTRIAL NORTHERN HARDWOOD FORESTS ACROSS MICHIGAN ’8 CENTRAL UPPER PENINSULA ........................... 13 Introduction .................................................................................... 13 Methods ....................................................................................... 19 Sample Plots ..................................................................... 19 Land Cover ....................................................................... 20 Bird Surveys ..................................................................... 21 Vegetation ........................................................................ 22 Patch Context .................................................................... 26 Landscape Context .............................................................. 27 Hierarchical Partitioning ....................................................... 27 Logistic Regression ............................................................ 29 Accuracy Assessment ........................................................... 30 Results .......................................................................................... 3 1 Discussion ...................................................................................... 40 Management Implications .................................................................... 46 CHAPTER 3 PLOT SIZE, STUDY EXTENT AND SPATIAL HETEROGENEITY AFFECT THE RELATIONSHIP BETWEEN NDVI AND SPECIES RICHNESS ................................................................... 48 Introduction .................................................................................... 48 Methods ......................................................................................... 52 Study Region ..................................................................... 52 Survey Plots ..................................................................... 55 NDVI .............................................................................. 56 Bird Surveys ..................................................................... 58 Species Richness — NDVI Analysis .......................................... 59 NDVI — Land Cover Analysis ................................................. 61 Results .......................................................................................... 66 Species Richness — NDVI Analysis .......................................... 66 NDVI — Land Cover ............................................................ 67 Variation of Local Slopes ...................................................... 69 Discussion ..................................................................................... 73 CHAPTER 4 USING THE SPECTRAL AND SPATIAL PRECISION OF SATELLITE IMAGERY TO PREDICT WILDLIFE OCCURRENCE PATTERNS ................................................................. 77 Introduction .................................................................................... 77 Methods ........................................................................................ 83 Study Region ..................................................................... 83 Study Species ..................................................................... 86 The GRAIN Protocol for Field Sampling .................................... 87 Image Processing ................................................................ 92 Validation Data .................................................................. 95 Signature Creation for Multiple Window Sizes .............................. 96 Image Classification ............................................................. 96 Cross-validation .................................................................. 97 Majority Map Validation ........................................................ 98 Gap Analysis ..................................................................... 98 Results ........................................................................................ 100 Information Content of Imagery ............................................. 100 Majority Map Validation ..................................................... 106 Gap Maps ....................................................................... 107 Discussion .................................................................................... 1 10 CHAPTER 5 SUMMARY, IMPLICATIONS AND CONCLUSIONS ............................... 115 Summary ..................................................................................... 115 Implications for Future Research ......................................................... 118 Implications for Future Management .................................................... 120 Other Data Uses ............................................................................. 121 Conclusions .................................................................................. 122 LITERATURE CITED ....................................................................... 125 vi LIST OF TABLES CHAPTER 2: HABITAT ASSOCIATIONS OF BIRDS WITHIN INDUSTRIAL NORTHERN HARDWOOD FORESTS ACROSS MICHIGAN ’8 CENTRAL UPPER PENINSULA Table 2.1 Descriptions of variables used in hierarchical partitioning .................... 25 Table 2.2. Results of hierarchical partitioning of models describing relationships among vegetation measurements and bird species occurrences sampled within 30m radius plots of northern hardwood forests in Michigan’s Upper Peninsula. Variables with significant independent explanatory power (90% confidence level) are highlighted in bold. . .. 33 Table 2.3. Model results for logistic regression analyses of species’ occurrences. . ....35 Table 2.4. Statistics variable strength and direction in the logistic regression models ............................................................................................. 40 CHAPTER 3: PLOT SIZE, STUDY EXTENT AND SPATIAL HETEROGENEITY AFFECT THE RELATIONSHIP BETWEEN NDVI AND SPECIES RICHNESS Table 3.1. Correlation coefficients of variables used in analysis of the relationships among NDVI, bird species richness, and land cover ....................................... 65 Table 3.2. Results of linear regressions among NDVI, species richness and proportional land cover. See footnotes of Table 3.1 for descriptions of variables ..................... 67 Table 3.3. Results of geographical weighted regression among NDVI, species richness and land cover. See footnotes of Table 3.1 for descriptions of variables ................ 68 CHAPTER 4. USING THE SPECTRAL AND SPATIAL PRECISION OF SATELLITE IMAGERY TO PREDICT WILDLIFE OCCURRENCE PATTERNS Table 4.1. The number of plots where species were observed within each detection distance. All 433 plots were surveyed using 30-m, 50—m, and 180—m detection distances and 321 plots were surveyed using the 100-m detection distance ........................ 92 Table 4.2. Land cover classes used to classify species presence in Gap analysis ...... 99 vii LIST OF FIGURES CHAPTER 1: BACKGROUND AND RESEARCH SUMMARY Figure 1.1 The study region covers portions of five counties in Michigan’s Upper Peninsula, USA. The star indicates the location of Dickinson county (see text for details) ............................................................................................... 2 Figure 1.2. Relationship between harvest volume and wintertime deer density (courtesy of Robert Doepker, MDNR) ..................................................................... 6 Figure 1.3. Conceptual framework of interactions among forest attributes and their economic values in Michigan’s Upper Peninsula as defined by the multidisciplinary partnership (see text for details) .................................................................. 8 Figure 1.4. Scaling differences among commonly collected bird, vegetation and image data. Data are often collected and summarized in different ways along a vegetation gradient (side view of forest showing different tree species as different shapes) ....... 11 CHAPTER 2: HABITAT ASSOCIATIONS OF BIRDS WITHIN INDUSTRIAL NORTHERN HARDWOOD F ORESTS ACROSS MICHIGAN ’8 CENTRAL UPPER PENINSULA Figure 2.1. The study region contains parts of five counties in Michigan’s Upper Peninsula, USA (A). Within the study region (B), randomly selected landscape units (C; ~1000 ha area with different land cover patches) define areas for the selection of specific plots for bird occurrence surveys. Plots (D; overview showing different tree species) were sampled within a 30m radius from plot centers (D; large circle), so that a single Landsat 5 or 7 pixel (D; square) falling within this area could be precisely characterized. . . . . 15 Figure 2.2. Predictive performance of logistic regression models for 4 species of forest birds detected within 30m northern hardwood plots. The top row depicts discrimination ability using an ROC curve, where values above the dotted line indicate the model performs better than chance. The middle row depicts model discrimination ability using a discrimination histogram. Dark columns are proportions of absence samples and light columns are proportions of presence samples. The bottom row depicts discrimination ability for pseudovalidation samples using a discrimination histogram. Because few presence samples were available for pseudovalidation, counts of presence probabilities are plotted (open circles) with proportions of absence probabilities (dark columns)...36 viii CHAPTER 3: PLOT SIZE, STUDY EXTENT AND SPATIAL HETEROGENEITY AFFECT THE RELATIONSHIP BETWEEN NDVI AND SPECIES RICHNESS Figure 3.1. The study region lies within a single Landsat 7 ETM+ scene and contains parts of five counties in Michigan’s Upper Peninsula, USA (A). Within the study region (B), randomly selected landscape units (C; ~1000 ha area with gradients of NDVI values) define areas for the selection of specific plots for bird occurrence surveys. Plots (D; overview showing different tree species) were sampled within a 30m radius from plot centers (D; large circle), so that a single Landsat 7 ETM+ pixel (D; square) falling within this area could be precisely characterized .................................................... 54 Figure 3.2. Plot of the relationships between proportional cover of (A) deciduous forest within the global kernel bandwidth (8258m radius) of the GWR model of species richness and NDVI plotted against local slopes derived from the same model. Outlier plots (open circles) were visually identified as not occurring within the trend of the relationships with deciduous cover (dotted ovals). A linear models of the fit for outlier plots (dashed line) is provided for comparison with a model describing the fit of the rest of the data (solid line). A map of outlier locations (D) shows outliers occur along the northern ecoregional border of the study region ............................................. 62 Figure 3.3. Mean NDVI values (: 1 SD) for pixels dominated (>50% cover) by different tree species. Solid bars indicate deciduous species and hatch bars are presented for non- deciduous species. Data are provided for a subset of plots surveyed for land cover within the Upper Peninsula of Michigan ............................................................. 69 Figure 3.4. Geographic distribution of species richness and local slopes (Beta) and fits (R-squared) of the GWR model of the association between NDV1130 and avian species richness in relationship to the proportion of deciduous forest (pDeck) and non-deciduous forest (pNonDeck) within the GWR kernel bandwidth surrounding sampled plots. Gray scales indicate interpolated values from a spline model using nearest neighbors. Black lines are ecoregion boundaries. See text for details ........................................ 71 CHAPTER 4. USING THE SPECTRAL AND SPATIAL PRECISION OF SATELLITE IMAGERY TO PREDICT WILDLIFE OCCURRENCE PATTERNS Figure 4.1. The study region lies within a single Landsat 7 ETM+ scene and contains parts of five counties in Michigan’s Upper Peninsula (A). Within the study region (B), randomly selected landscape units (C; township sections or USGS QQQs with coarse land cover classes within landscape units shown as shades of gray) define areas for the selection of specific plots (D; overview showing different tree species) for bird occurrence surveys. Plots were sampled within a 30-m radius from plot centers (D; large circle), so that a single Landsat 7 ETM+ pixel (D; square) falling within this area could be precisely characterized ..................................................................... 84 ix Figure 4.2. Factors affecting the ground instantaneous field of view captured by pixels in Landsat imagery. Gray circles represent 30-m radius field plots. Each grid represents the ground information contained within a nine-pixel window. The bashed center cell represents the ground area used to describe the sampling plot. Factors affecting the location and area covered by the center pixel include A) unknown pixel location, B) non- overlapping pixels of multiple images, C) effects of Earth rotation, and D) inclusion of increasing ground coverage within pixels increasingly farther from nadir ............... 89 Figure 4.3. Relationships between the proportion of plots where species were detected and accuracy measures of proportion correctly classified (PCC; open symbols) and Kappa (filled symbols) values for (A) cross-validation, (B) validation, and (C) Gap analysis. Each symbol represents the results of a single analysis for a given species (Black-throated Green Warbler = triangle; Nashville Warbler = square; Ovenbird = circle), detection distances and window size. F irst-order polynomial model fit is provided only as a relative measure of dependence as each point represents accuracy statistics from random subsets of the same pool of plots classified in different ways .................. 101 Figure 4.4. Proportion correctly classified (gray squares) and Kappa values (black diamonds) for maps predicting the occurrence of A) Black-throated Green Warbler, B) Nashville Warbler, and C) Ovenbird using different detection distances and window sizes. Filled symbols represent mean values of accuracy statistics for 11 cross-validation iterations with error bars indicating minimum and maximum values. Open symbols represent accuracy statistics of a majority map summarizing the cross-validation runs and tested with a subset of data reserved for validation ........................................ 104 Figure 4.5. Distribution of accuracy statistics (proportion correctly classified and Kappa) for majority maps created from the 11 cross-validation runs and tested with a subset of data reserved for validation (filled diamonds) and for Gap analysis maps (open squares) tested with data from all 433 plots. Statistics are shown for A) Black-throated Green Warbler, B) Nashville Warbler, and C) Ovenbird. Larger symbols represent larger detection distances. Symbol location indicates mean values for all window sizes used to classify maps for each detection distance, with error bars indicating minimum and maximum values ................................................................................ 108 CHAPTER 1 BACKGROUND AND RESEARCH SUMMARY Background Much of the philosophy of natural resource management in the United States is founded upon a utilitarian view of nature. A few influential factors directing this perspective include Aldo Leopold’s Game Management (Leopold 1933), Guilford Pinchot’s mandate for the US. Forest Service (Callicott 1990), and the Izaak Walton League’s influence upon federal legislation (Perich 1997). Although there has been a shifting influence over the last few decades towards an emphasis on the preservation of aesthetics and biodiversity, the existing infrastructure and economy associated with the extraction of natural resources still drive many decisions in the United States. In Michigan, this utilitarian framework is no exception. A large part of the central portion of the Upper Peninsula (UP) of Michigan (Figure 1.1) is managed for commercial timber production with the forest industry playing a major role in the regional economy. Stakeholders sharing a large interest in timber harvesting across the area include two large industrial companies (International Paper (IP) and MeadWestvaco (Mead)), the Michigan Department of Natural Resources (MDNR) and several small mills. Forest products from the area are diverse and include dimensional products (for construction, furniture and flooring), poles, posts, veneer, and pulp. While much of the landscape is selectively logged on a 10-15 year rotation (Michigan Department of Natural Resources (MDNR) and International Paper (IP), unpublished forest inventories), stands can be characterized by a wide range of ages (or time since last harvest) with some being cut as long ago as the late 19‘h century. Spatial patterns of forest harvesting on the landscape include areas of high concentration over the last ten years as well as areas of lower intensity, where the harvesting is more dispersed. Figure 1.1 The study region covers portions of five counties in Michigan’s Upper Peninsula, USA. The star indicates the location of Dickinson county (see text for details). Besides timber harvesting, a major recreational activity within the region is deer hunting. The economic activity associated with deer hunting has a significant effect on the local economy. Even though it is far from the state’s major population centers, the western Upper Peninsula of Michigan has averaged about 100,000 deer hunters a year (Frawley 1999, 2000). These hunters spend just over 1 million days of deer hunting effort per year (Frawley 1999, 2000). The MDNR estimates that in 1998, about 45,000 deer were harvested in the western Upper Peninsula of Michigan, while in 1999, about 60,000 deer were taken in the area (F rawley 2000). A considerable portion of the MDNR’s operating costs is raised through deer hunting licenses and matching funds. In return, a considerable portion of the MDNR’s budget and staff resources are allocated towards deer management and deer related efforts. While management efforts focusing on the separate production of timber and deer are important to the regional economy, the independent actions taken to enhance each of these resources have had adverse and interrelated effects. By creating ample supplies of accessible forage (tree regeneration, shrubs and herbs) for white-tailed deer, timber harvesting improves habitat quality for this keystone species (DeCalesta 1994). These actions have likely contributed to the overabundance of whitetail deer in many parts of the United States (McShea and Rappole 2000) including Michigan (Xie et al. 1999). The improved food availability increases deer productivity (Verme 1965) leading to population dynamics that are increasingly difficult to predict and manage (McCullough 1997, Bob Doepker (MDNR) personal communication). The resulting increase in deer numbers has had a severe impact on tree regeneration in many areas, especially in lowland conifer (Van Deelen et a1. 1996) and northern hardwood cover types (IP, unpublished forest inventory). High deer numbers have led to the near elimination of forest regeneration in many areas of the United States (Stoeckler et a1. 1957, Anderson and Loucks 1979, Tilighman 1989, Tilinghan 1989, DeCalesta 1997, Frelich and Reich 1999, Rooney 2001, Liang and Seagle 2002, Rooney and Waller 2003), thus threatening the sustainability of timber production in those areas. F urtherrnore, high deer densities can strongly affect a number of other ecosystem attributes including: (1) decreasing the compositional and structural diversity of vegetation (Stromayer and Warren 1997), (2) decreasing the occcurrence frequency of browse sensitive plant species (Balgooyen and Waller 1995, Augustine and Frelich 1998, Comett et al. 2000, Rooney 2001, Rooney and Waller 2003 ), and (3) negatively impacting habitat for other wildlife via direct and higher-order effects (see Werner 1992) on vegetation composition and structure (DeCalesta 1994, 1997, Brooks 1999, McShea and Rappole 2000). Thus, timber and deer management activities affect a multitude of ecosystem attributes. Changes to vegetation composition and structure through timber harvest activities and deer browsing indirectly affect forest birds. Some bird species, especially those adapted to interior and mature forest stands, have experienced dramatic population declines (Thompson et al. 1993). In fact, the Midwest has provided some of the most robust data demonstrating threats from anthropogenic habitat fragmentation (see Robinson 1995). This reduction in numbers is often recognized as a response to increased disturbance from timber harvesting (Rotenberry et al. 1993) and the resulting younger (Holt and Martin 1997), smaller (Ambuel and Temple 1983, Niemi and Hanowski 1984), and relatively more homogeneous stands (Bunnell and Kremsater 1990, Thompson et a1. 1995). Bird population declines may accelerate in the future, as habitat loss can decrease population resilience after stochastic events (Freemark and Merriam 1986). To reduce the loss of bird species and populations in managed forests, it has become increasingly important to understand how bird species are distributed across forested landscapes and the processes that affect these distributions. Unfortunately, investigations into these pattern-process relationships are rarely communicated in a management context (Amett and Sallabanks 1998, Hejl and Granillo 1998) or are too simple to be useful (Mike Young, IP, personal communication). Furthermore, most “landscape” studies are single patch oriented or are conducted over a few adjacent patches, covering too small an area to provide a general understanding of broad scale relationships (cf. F lather and Sauer 1996). Studies that did consider interactions over large areas are sufficiently coarse (10’s to 100’s of hectares per sample plot) so as to limit their utility to broad generalizations (DeSante and Rosenburg 1998), many of which are contentious (James 1998). What is needed, therefore, are site-specific recommendations for augmenting bird habitat availability and quality, which are also relevant over large geographic regions. Research Summary The research described in this dissertation was conducted as part of a multidisciplinary partnership to understand the ecological and economic consequences of forest and wildlife management in the UP. Much of this region exemplifies the effects of unexpected negative consequences of independent timber and deer management activities. Some areas are characterized by high deer densities (>30/mile2) and relatively little recruitment of seedlings to sapling-sized trees in over 20 years (Mike Young, IP, personal communication). There is also strong evidence that deer have responded positively to timber harvest levels in this area (Robert Doepker, MDNR, unpublished pellet count data) as indicated by the positive relationship between timber harvest volume and deer pellet counts (Figure 1.2) and by an increase in deer populations in northern Dickinson county (Figure 1.1) following concentrated high volume of harvests there in the early 1990’s. These changes to vegetation composition and structure are likely influencing bird communities in the region. However, there is still little knowledge concerning how interactions among management activities affect bird species occurrences in the region. Average pellet counts 5.0x109 10.0x109 15.0X109 Cords harvested, Western Upper Peninsula Figure 1.2. Relationship between harvest volume and wintertime deer density (courtesy of Robert Doepker, MDNR). Identifying and quantifying the ecological consequences of management activities requires the integration of research from many diverse fields over multiple grains (sample unit size) and extents (regions sampled) to test competing hypotheses within an iterative framework (sensu Murphy and Noon 1992). A strong inference approach to understanding broad-scale ecological relationships may be facilitated using remotely sensed, spatially and temporally explicit, quantitative descriptions of landscapes. These data types, such as satellite imagery, are commonly used to extrapolate ground-sampled information via similarities in spectral reflectance with unsampled locations (Franklin 2001 , Kerr and Ostrovsky 2003). Several attempts have been made to link remotely sensed data with field data in the interest of ecological function (e. g., O'Connor et a1. 1996, Oindo 2002, van Rensburg et al. 2002, Chown et al. 2003, Kerr and Ostrovsky 2003). However, these efforts are typically limited to large grain datasets. Furthermore, detailed wildlife and vegetation data collected in the field are rarely spatially compatible with satellite imagery. I was interested in exploiting the information contained in satellite imagery to investigate general ecological relationships over large landscapes and to extrapolate detailed ecological relationships quantified over small areas. As the wildlife contingent of the aforementioned multidisciplinary partnership, my doctoral advisor, Dr. Jack Liu, and I worked in close association with faculty and students in the Departments of Forestry (Dr. Mike Walters, Joseph LeBouton) and Agriculture Economics (Dr. Frank Lupi, Laila Racevskis) at Michigan State University. I also sought information and advice from members of public (MDNR, Michigan Natural Features Inventory, US. Forest Service) and private (IP, Mead, The Nature Conservancy) stakeholder groups. Our conceptual framework is shown in Figure 1.3. In summary, forest management activities and deer browsing (as a function of deer density) act to alter vegetation composition and structure. Changes to vegetation characteristics will result in changes to bird communities via the combined responses of multiple individuals from multiple species reacting to spatial and temporal differences in the availability of life history requirements (e. g., food, safety and nest locations). Therefore, the occurrences of particular species and bird species diversity are the direct result of vegetation composition and structure, and an indirect result of forest management practices and deer density. Forest Ecological and Economic Relationships ECOLOGICAL INTERACTIONS Timber harvesting 0.00.06.60.000000000: 6 s . Vegetation . mm: .222. : . structure and composition , z 3 ECC):NOMIC VALUES“ . - ¢ . . . - . 3 : diversity Forest bird Visual aesthetics (Wood Productg ‘ Plant diversity Non-market values ~Market value - Figure 1.3. Conceptual framework of interactions among forest attributes and their economic values in Michigan’s Upper Peninsula as defined by the multidisciplinary partnership (see text for details). To understand the pattern-process relationships of birds and vegetation in the UP, several types of data were needed. First, it was important to have a detailed inventory of bird species. While historical accounts of bird species nesting locations in the region were available (Brewer et a1. 1991), the spatial precision of these data (township section) were relatively coarse and therefore not useful for developing knowledge of fine scale habitat associations. However, these historical accounts were very helpful in determining which species to expect when developing training materials for survey crews. Another important need was a contemporary land cover map of the region. Land cover data can be very useful for determining landscape patterns of bird distributions and their causes (e. g., McGarigal and McComb 1995, Brooker et al. 1999, Saab 1999, Golet et al. 2001, Lawler and Edwards 2002, Crozier and Niemi 2003, Lepczyk et al. 2003). Several historical land cover maps existed, including 1) a map of vegetation prior to European settlement derived from early land surveys, 2) a 1978 land cover map derived from digitized aerial photographs and 3) a 1991 deer habitat map created through the unsupervised classification of Landsat TM imagery (metadata and maps available at http://wwwmcgi.state.mi.us/mgdl/). However, these maps had limited utility because of the dynamics of our study region (i.e., changes caused by timber harvesting) and because of many inaccuracies in the 1991 map (R. Doepker, personal communication). Additionally, land cover maps provide only a general description of vegetation composition, while many bird species respond to structural characteristics of forests (e.g., Collins 1981, Crawford et al. 1981, Maurer and Whitmore 1981, Niemi and Hanowski 1984, Robichaud and Villard 1999, Smyth et al. 2002). Therefore, detailed vegetation information was needed to quantify specific vegetation associations of birds in the UP. Many protocols for sampling birds, vegetation and land cover are not spatially compatible. After reviewing the literature of protocols for these disparate data types, I was concerned that differences in sample plot sizes among data collection protocols (Figure 1.4) could add substantial error or otherwise influence the results of analyses using a combination of variables (see Levin 1992, Peterson and Parker 1998a, b, Hamer and Hill 2000, Hill and Hamer 2004). For example, there will be some measure of vertical and horizontal heterogeneity along any transect placed within a forest landscape. A forest manager may assess these differences and delineate what she/he perceives to be a homogeneous unit for the management of commercially important tree species. Perceived homogeneous units may also be delineated from remotely sensed imagery and then surveyed to provide references for classifying land cover (e. g., Space Imaging, Inc. unpublished protocol). In these cases, spectral reflectance data measured within individual pixels (e.g., 30-m x 30-m) and summarized over multiple pixels are used to differentiate boundaries. However, bird data are often collected using 50-m, 100-m or unlimited distance point counts (Ralph et al. 1993, Ralph et al. 1995), which can sample multiple management units and land cover patches. When smaller areas are sampled, such as around bird nest locations (Martin et al. 1997), the plot characteristics are usually measured over areas much smaller (e.g., 5-m or 11.3-m) than managed for timber or measured by satellites. A standard scale of data collection was therefore needed to integrate these disparate data types within a spatial framework. In collaboration with Joseph LeBouton, a forestry student involved with the larger multidisciplinary project, I collected bird, vegetation and land cover data at the spatial grain (i.e., smallest sampling unit) of Landsat 7 ETM+ imagery. The results of different investigations conducted by project members were consequently spatially compatible. In addition, field data and spectral reflectance values were used to investigate and extrapolate relationships over the extent of the study region. This dissertation describes how these data were collected and ways I integrated these data to describe patterns of bird species’ distributions, patterns of bird species richness and causes of those patterns. 10 Figure 1.4. Scaling differences among commonly collected bird, vegetation and image data. Data are often collected and summarized in different ways along a vegetation gradient (side view of forest showing different tree species as different shapes). The goal of this research was to integrate and develop state-of—the-art approaches for understanding ecological and spectral relationships in a spatial fiamework. The next three chapters were written as individual manuscripts for publications in journals. However, they are all spatially integrated with Landsat 7 ETM+ imagery, which is often used for land cover mapping in North America. My objective was to exploit the spatial and spectral precision of satellite imagery for analyses of forest bird distribution patterns and their causes. In Chapter 2, I describe the fine scale habitat associations of birds occurring within northern hardwood land cover patches over the extent of the study region. The data used in this study were comparable to the ground information contained within pixels of Landsat 7 ETM+ imagery. The third chapter describes the spatially variable relationships among a measure of bird species richness, a spectrally derived vegetation index, and descriptions of landscape composition. Data for this series of analyses were spatially equivalent to individual pixels or averaged over larger areas sampled for birds. The fourth chapter provides the foundation of a novel method for extrapolating bird species’ distribution patterns unclassified satellite imagery. In this chapter, I identify ways to assess the influences of spatially relevant image classification decisions on prediction accuracies. In the final chapter, I summarize the findings of the previous three chapters, describe how the data were used for other purposes and conclude with implications for future research and forest management. 12 CHAPTER 2 HABITAT ASSOCIATIONS OF BIRDS WITHIN INDUSTRIAL NORTHERN HARDWOOD FORESTS ACROSS MICHIGAN ’8 CENTRAL UPPER PENINSULA Introduction Numerous studies have documented the effects of forest management practices on bird species occurrences and communities. Often these studies focused on the effects of harvest treatments such as even vs. uneven-aged management (e. g., Derleth et al. 1989, Lorimer 1989, Morrison 1992, Costello et al. 2000, Gram et al. 2003). Additionally, some form of ordination was often used to describe a large dataset of vegetation composition and structure variables into a smaller number of variables (sensu James and Shugart Jr. 1970, James 1971). However, knowledge is lacking about specific local conditions within management units (stands) that favor bird species occupancy, especially over large areas still dominated by forests. While intensive studies of species responses to conditions within and around stands have been conducted, they typically focus on study regions only moderately larger than the sample plots used to describe them (Holt and Martin 1997 , Saab 1999, Jobes et al. 2004, King and DeGraaf 2004). Extending the conclusions from this research to other areas may not be valid if the results would differ out of context (i.e., landscape context (Donnelly and Marzluff 2004)). Therefore, studies having small grain (plot size) and large extent (study region) are needed to identify regionally important factors associated with local species occurrences I was interested in identifying local aspects of vegetation composition and structure responsible for the occurrences of bird species within managed northern hardwood forests across a ~400,000 ha region of Michigan’s Upper Peninsula (UP, 13 Figure 2.1). This region is of particular interest because land use is primarily managed timber production by large organizations (International Paper, MeadWestvaco, and Michigan Department of Natural Resources (MDNR)) that have structured management plans based on well-established forest management principles. The region also contains relatively little agriculture or human development, making it a good candidate for exploring the effects of present forest management policies without the complicating effects of other land uses. Although an understanding of the relationships between present northern hardwood forest characteristics and bird species occurrences does not directly address historical changes in bird species composition due to forest management, it does provide a foundation for predicting the consequences of future management scenarios. The prevalent land cover in this region of interest is a patch mosaic of lowland coniferous forest (24%) and northern hardwood forest (22%), along with aspen and mixed upland forest associations (12% and 10%, respectively). Northern hardwood is of particular interest because it is extensively managed for hi gh-value timber products. Both even-aged and uneven-aged silviculture systems are used to manage northern hardwoods, but uneven-aged single tree selection are predominant. In a typical cycle for uneven-aged northern hardwood management, stands between 2 and60 ha are selectively logged on a 10-15 year rotation (MDNR and IP unpublished forest inventories). At each entry, approximately 25% of standing basal area is removed by selecting single trees or small groups of trees over a range of size classes distributed throughout the stand. 14 Figure 2.1. The study region contains parts of five counties in Michigan’s Upper Peninsula, USA (A). Within the study region (B), randomly selected landscape units (C; ~1000 ha area with different land cover patches) define areas for the selection of specific plots for bird occurrence surveys. Plots (D; overview showing different tree species) were sampled within a 30-m radius from plot centers (D; large circle), so that a single Landsat 5 or 7 pixel (D; square) falling within this area could be precisely characterized. 15 16 Given the predominance of single tree canopy openings, the numerically large cohorts of smaller trees are expected to be dominated by shade tolerant species, especially sugar maple (Acer saccharum). This harvest pattern results in canopy openings expected to facilitate natural tree seedling establishment and ultimately sapling recruitment to the canopy where they will grow and become part of the future tree crop. Thus, harvesting methods can either maintain or create (over several harvest entries) an uneven-aged diameter distribution characterized by tree numbers decreasing linearly with the logarithm of diameter. Despite decades of efforts in forest planning, the consequences of uneven-aged northern hardwood management activities were not always predictable. Vegetation felled by harvesting and regrowth following harvest resulted in temporarily abundant supplies of accessible forage (tree regeneration, shrubs and herbs). Abundant forage often increases white-tailed deer (Odocoileus virginianus) numbers until they exceed their food supplies and effectively devour all the vegetation within their reach (McCullough 1997). Excessive browse pressure can negatively affect future timber production by nearly eliminating forest regeneration (Stoeckler et al. 1957, Tilighman 1989, DeCalesta 1994), eliminating browse sensitive plant species (Rooney and Waller 2003), and reducing compositional and structural diversity of vegetation (Stromayer and Warren 1997). Thus, myriad forest characteristics including canopy and understory species composition, diameter and height distributions of trees, canopy openness, and basal area are strongly affected by harvest activities and deer browsing. While many stands are still classified and managed as northern hardwoods due to the prevalence of sugar maple and other 17 characteristic species within them, the vertical and horizontal diversity within these stands vary as a function of management, browsing pressure and regeneration. I hypothesized that canopy gaps resulting from selective removal of individual trees would have a strong effect on the occurrence of some northern hardwood dwelling bird species in our region of interest. I also hypothesized that deer browsing pressure would have an indirect on bird species occurrences by removing low branches and palatable understory species. The goal of this research was therefore to understand the extent to which management and deer browsing directly and indirectly affect bird species’ occurrences via changes in vegetation characteristics within northern hardwood stands. I sought to provide recommendations for harvesting strategies within northern hardwood stands through knowledge of specific vegetation characteristics associated with bird species’ occurrences. Three important factors should be considered when conducting local species- habitat relationships over large areas: 1) scale, 2) landscape context and 3) multicollinearity of independent variables. Scaling issues arise when plots of different sizes are used to ask the same question (see Peterson and Parker 1998b, Hamer and Hill 2000, Hill and Hamer 2004). Larger plots contain greater structural and compositional complexity compared to smaller plots within the same type of vegetation community, thereby influencing variance estimates used in analyses. Landscape context describes the influence of factors outside the sampling area on factors within the sampling area. The term landscape implies a spatially heterogeneous system, typically divided into areas (patches or stands) with similar qualities (Turner 1989, Wiens 1992). Landscape context therefore refers to the influences of surrounding patches on the sampled area. 18 Multicollinearity is a problem of intercorrelation among predictor variables. Causal variables can be lost fi'om predictive models during stepwise selections of model subsets in statistical analyses such as multiple regression. In such cases, noncausal variables are retained in models at the expense of correlated causal variables (see Mac Nally 2000, Mac Nally 2002). My objective therefore was to determine aspects of northern hardwood composition and structure associated with bird species occurrences while also limiting the effects of scale, landscape context and multicollinearity on my results. Methods Sample Plots To address issues of scale that could arise from sampling different size areas, I used vegetation and bird data collected at the same sites and over the same area. The data were collected as part of a multidisciplinary project investigating the ecological and economic effects of timber harvesting and deer browsing pressure on forest resources. Within the study region (Figure 2.1A) a total of 433 survey plots were established over the summers of 2001 to 2003. They were located within 96 randomly selected landscape sampling units of approximately 1000 ha (Figure 2.1B, LUs). Within each of the 96 LUs, between 2 and 8 plots were selected (Figure 2.1C). A 30—m radius circle was used to define survey plots (Figure 2.1D). This plot size was selected to roughly match the spatial grain of Landsat 5 & 7 imagery used to classify land cover maps for wildlife and forest management in Michigan (Space Imaging, Inc. 2001). Hence, results of this study are spatially compatible with individual pixels of maps used to make management decisions. 19 Plots were selected where field crews perceived that a hypothetical 30-m x 30-m square could be placed anywhere within the circle and expected to have the same vegetation structure and composition as a similar square placed anywhere else within the circle (Figure 2.1D). The choice of 30-m radius circles over 30-m x 30-m squares (grid cell size of Landsat 7 ETM+ imagery) was made for several technical reasons including image distortion due to the rotation of both the Earth and Landsat 7 satellite, georectification error, sensor mechanics (see Chapter 4 for details). In addition, it is easier to estimate bird occurrence within equidistant circles than squares where comers are farther from observers than sides. Land Cover Land cover characteristics were described for each plot using the protocol established by Space Imaging, Inc. (2001). In summary, percent cover was estimated for each plant species by envisioning what would be seen from the satellite’s perspective (looking down from space). Total cover over all species therefore equals 100%, as any overlapped cover is not included in the estimate. Plots with at least 60% combined coverage of maple (Acer spp), beech (Fagus grandifolia), basswood (T ilia americana), white ash (F raxinus americana), cherry (Prunus spp.) and yellow birch (Betula alleghaniensis) were classified as northern hardwoods (Space Imaging Inc. 2001). Surveyed plots were overlaid with a recent (2001) land cover map (available at http://www.mcgi.state.mi.us/mgdl/). Only those plots meeting the conditions of the northern hardwood classification rule and classified as northern hardwood in the land 20 cover map were used in these analyses. In total, there were 124 northern hardwood plots surveyed in 67 randomly selected LUs. Bird Surveys Bird species’ occurrences were detected using 3 surveys conducted within 5 hours of sunrise between June 10 and July 3 of either 2001, 2002 or 2003. Each survey lasted 10 minutes followed by a 1 minute walk around the plot to flush elusive ground nesters (see Chapter 4 for details). Species detected within 30-m during any of the surveys were counted as present. Otherwise, the species were considered absent. Bird species I selected for analyses were listed by Partners in Flight (Matteson et al. in prep) as species of regional concern. Identification of forest conditions favorable for these species is of particular interest for conservation and timber planning efforts. I conducted analyses for species detected within northern hardwood forests in at least 10 of the 67 LUs. This condition was set to reduce the likelihood of retaining non- causal variables in statistical analyses, which can occur with small fiequencies of presence data. I conducted analyses for 6 species: Black-throated Green Warbler (Dendroica virens), Eastern Wood-Pewee (Contopus virens), Least Flycatcher (Empidonax minimus), Ovenbird (Seiurus aurocapillus), Rose-breasted Grosbeak (Pheucticus ludovicianus), and Yellow-bellied Sapsucker (Sphyrapicus varius). Analysis of these species thus provided quantitative descriptions of species-habitat relationships for 4 families (Cardinalidae, Parulidae, Picidae, Tyrannidae). 21 Vegetation Aspects of vegetation structure were measured in three different size categories expected to influence the presence of bird species. These size categories were selected to sample vegetation within vertical layers expected to have firnctional relevance to both birds and deer, as deer browsing pressure in the region has a strong effect on understory composition and structure (LeBouton et al. unpublished data). For seedlings and small saplings (lowstory category) field crews conducted n-tree distance sampling (Lessard et al. 1994, Lessard et al. 2002) from subplot centers spaced 16m apart on a 3x3 grid. At each point on the grid they measured the distance from a subplot center to the base of each of the 5 closest live tree or shrub individuals 0.25 to 1.5-m tall using a 50-m fiberglass reel tape. Maximum search distance for n-tree sampling was 30—m or the edge of the vegetation plot. Consequently, if 5 individuals were not found within 30—m of any particular subplot center, the missing individuals were marked as “missing” and stem density calculations were adjusted by the actual number of measured individuals. Surveyors noted the species of each tree or shrub sampled, measured diameter to the nearest 0.5-cm at 10-cm above the ground using a metal dbh tape, and measured bottom and top of canopy using a 2-m hypsometry pole graduated at l-cm intervals. They defined bottom of canopy as beginning with the first leaves if the canopy was a compact shape. Long-tapering or disjunct canopies presented a more complicated case, and they defined the canopy in these cases as beginning with the bottom of 90% of canopy mass (after Avery and Burkhart 2002). This method introduced a minimum amount of subjectivity to measurements, but resulted in consistent and repeatable measurements. The top of tree canopies were measured at the uppermost leaves. 22 Larger saplings (midstory category) >1 .5-m tall and <10-cm dbh were measured in the same way as smaller vegetation at each of the 9 subplots. However, sapling height was measured to the nearest 0.5-m using either a hypsometry pole or a laser hypsometer and sapling diameter was measured at breast height (1.35-m above the ground). For trees > lO-cm dbh (overstory category), field crews point sampled stems with a cruising prism (BAF = 10-ft2/acre (Avery and Burkhart 2002)).Within each 30-m radius vegetation plot they surveyed 3 subplots placed at the vertices of an equilateral triangle around the plot center. This design was employed with three factors in mind: to minimize double—sampling of large diameter trees, to minimize the potential for sampling out-of-plot (i.e., >30-m from plot center) trees, and to maximize within-plot sampling coverage. They established the first overstory subplot 15-m from the plot center at a random azimuth. The second and third subplot centers were 25-m from the first subplot center, 25-m from each other, and 15-m from the plot center. For each tree, they recorded species, dbh, and height of the bottom and top of the canopy. Stern density for overstory trees is the sum of the basal area of sampled individual tress where: BA/area of each tree = (mz/ha) / (m2 cross-sectional area per tree). Cross-sectional area per tree was = (dbh(m)/2)2 * pi). Heights for overstory trees were measured with either an optical clinometer, an Opti-Logic® 400LH Laser Hypsometer or an Impulse® 200 Laser Rangefinder. Vertical precision for all of these measurements was assumed to be 1-m. Field crews measured canopy openness at each of the 3 overstory tree subplots while facing in 4 directions (N ,S, E, and W) around the center of the subplot using a concave spherical densiometer (Comeau et al. 1998). The mean value of these 4 measurements was calculated to describe canopy openness for the subplot. 23 I summarized vegetation data for each category (lowstory, midstory and overstory) using structural and compositional characteristics of interest to forest management and expected to influence bird species’ occurrences in the study area (Table 2.1). Each category was described by the total number of live stems per hectare and the total basal area (BA) of those stems per hectare. I also calculated the mean height of the bottom and top of the canopy for the midstory and overstory. Canopy height averages were calculated over sampled individuals rather than adjusting for the number of stems they represent. For the overstory, I calculated averages for individuals as a way of weighting for large horizontal canopies associated with trees having large diameter stems (i.e., more likely to be identified for sampling using the cruising prism). For the midstory, I calculated averages for individuals to maintain the structurally representative information in our data as measured using the n-tree sampling design. Several other variables were calculated from only the midstory or overstory. From the midstory data I calculated importance values of three common woody species differentially selected by deer (Acer saccharum, Dirca palustrus and Ostrya virginiana; LeBouton et al. unpublished data), because deer browsing pressure selects against more palatable species (Crete et al. 2001, Liang and Seagle 2002, Rooney and Waller 2003) that could provide essential resources for some birds. Importance values (IV) were calculated using the following equation: IV = (proportion of stems + proportion of basal area) / 2 24 Table 2.1 Descriptions of variables used in hierarchical partitioning Variable Description Understory StemsL Total number of stems/ha in the lowstory" StemsM Total number of stems/ha in the midstorya BAL Total basal area/ha (m2) of live trees in the lowstoryb , BAM Total basal area/ha (m2) of live trees in the midstoryb BotCanM Mean bottom of canopy height of trees in the midstory TopCanM Mean top of canopy height of trees in the midstory ACESiv Importance value of Acer saccharum in the midstoryc DIRPiv Importance value of Dirca palustrus in the midstory a OSTViv Importance value of 0sttya virginiana in the midstory d Overstory Stemso Total number of stems/ha of live trees in the overstory BA 0 Total basal area/ha (m2) of live trees in the overstory BotCano Mean bottom of canOpy height (m) of trees in the overstory TopCano Mean top of canopy height (m) of trees in the overstory TalTree Height of the tallest tree (m)a DdStem Number of dead stems/hae AvOpen Average of canopy openness fiom 3 subsamplesb StDvOpen Standard deviation of canopy openness from 3 subsamplesa Patch Context EdgeDist Distance to the nearest edge of the sampled northern hardwood patchd Area Area of the sampled northern hardwood patcha AT eaP er Area to perimeter ratio of the sampled northern hardwood patchd Core Core area of the sampled northern hardwood patch (area >30-m from edge)d m- Transforrnations used in analyses: (a) ln(x + 1), (b) x033, (c) x2, ((1) V55, (e) x025 25 where the proportions were derived from midstory measurements of the three species. From the overstory data, I included the height of the tallest tree sampled within the plot and the total number of dead stems (snags) per hectare. In addition, I included the average and standard deviation of overstory canopy openness estimates derived from subplot average spherical densiometer readings. Patch Context In addition to vegetation variables within the sampled plots, I also quantified characteristics of patch context for the surveyed plots. Patches used in these analyses were defined by contiguous northern hardwood forest pixels surrounding my sampled plots as described by the aforementioned land cover map. I define patch context here as the location of the plot within patch and the size and shape of the patch. Patch context therefore differs from landscape context, which includes multiple patches of different classes. Patch context variables included in these analyses were the distance between the center of each plot and the nearest edge of its patch, area of the sampled patches, area to perimeter ratio of the sampled patches, and the core area of the sampled patches. I loosely defined core area as the area of the sampled patches greater than 30-m (one pixel) from any edge. This definition was applied as a means of removing edge pixels that are more likely to be misclassified and thus influential to patch area descriptions, rather than accounting for any possible effects of microclimate gradients (e. g., Harrison 1997, Menzel et al. 1999) or predation (e.g., Bollinger and Peak 1995, Donovan et al. 1997) on birds. 26 Landscape Context For all analyses, I used only 1 plot per LU to ensure the reliability of significance tests. The large extent of my study region and my random sampling design allowed me to reduce the possible incursion of pseudoreplication in significance tests that could be caused by sampling multiple plots within the same landscape context. Any significant relationships observed between vegetation variables and species occurrence were therefore the result of resource selection by multiple individual birds across the large study region in spite of differences in the landscape context among plots. Because absence plots generally outnumbered presence plots for each species, a presence plot was first selected at random from within each LU if it existed. If it did not, an absence plot was randomly selected instead. Data used in any given analysis therefore consisted of 67 northern hardwood plots and patches sampled across 67 randomly selected LUs. Hierarchical Partitioning I used hierarchical partitioning to identify vegetation variables having significant independent associations with species’ occurrences within 30-m radius northern hardwood plots (Mac Nally 2000, Mac Nally 2002). Hierarchical partitioning was implemented using the ‘hier.part package’ (Mac Nally and Walsh 2004), as part of the R open source statistical package (R 201). The use of this new statistical method allowed me to consider multicollinearity in the vegetation data and identify variables making significant independent contributions to model fit. This method also identified variables that may have little independent effect but still have a high correlation with the dependent variable resulting from joint correlation with other variables. The independent and joint 27 contribution of each variable to models of species probabilities of occupancy were determined using hierarchical partitioning of negative log-likelihoods (Chevan and Sutherland 1991, Walsh et al. 2004). Hierarchical partitioning jointly considered multiple regression models using all possible combinations of variables to identify the most likely causal factors. The increase in model fit generated by each variable was estimated by averaging its influence over all models in which it appeared. Variables with significant independent contributions were identified using a randomization approach (‘rand.hp’). Results of hierarchical partitioning for each variable were then expressed as Z-scores ([observed —mean{randomizations}] /SD {randomizations}), with the statistical significance based on upper confidence limits. I evaluated the significance of each variable with an upper confidence limit of 90% (Z 2 1.28). These methods depend upon monotonic relationships between the dependent and independent variables. Consequently, all skewed predictor variables were transformed as appropriate (see footnotes of Table 2.1). Because the degree of collinearity among predictor variables is dependent on the data used in calculations, I divided the vegetation and patch variables into three different functional groups (Table 2.1) and applied hierarchical partitioning to each group separately. I grouped data into understory (lowstory and midstory variables), overstory, and patch context descriptions. This approach permitted comparisons of variables likely to be most correlated and similarly influential on bird species’ occurrences, while also staying within the 12 variable limitation of the ‘hier.part package’. Following Smyth et al. (2002), I also used influential variables identified from each grouped analysis in a 28 combined model. Variables were included in the combined model if the mean Z-score > 0 over all bird species or was significant for any individual bird species. Logistic Regression The relative strength and direction of relationships between species’ probabilities of occurrence and any significant independent variables from the combined hierarchical partitioning model were evaluated using logistic regression (SPSS 12.0, SPSS Inc, Chicago, IL). All significance tests using logistic regression were made with a = 0.05, except outlier detection, where significance was considered at a = 0.1. A preliminary analysis was conducted to detect outliers and evaluate the significant of the model constant. Outliers with standardized residuals >2.58 were removed and the analysis was repeated. If the constant was not significantly different from 0, the analysis was repeated without a constant. Significance of the overall model was tested using the log-likelihood test. Because relatively few plots were occupied by species, I was concerned that the interpretability of Wald statistics for independent variables would be unreliable (Menard 2002). Therefore, the direction of association between species’ occurrences and each independent variable was evaluated using its odds ratio. Odds ratios >1 indicate positive association while smaller values indicate negative association. Confidence intervals were generated for the odds ratios to assess their significance (i.e., not equal to 1, a = 0.05) as useful predictors. In cases of multiple independent variables, scale differences precluded me from comparing the strength of these relationships using odds ratios. Therefore, the relative effects of covariates on the dependent variable’s logged odds were compared by 29 multiplying the unsigned (absolute) values of the unstandardized logit coefficients (i.e., B) of each independent variable by its standard deviation (see Menard 2002). Accuracy Assessment The strength of association described by the models was quantified using Nagelkerke's R-Square (Nagelkerke 1991) and by calculating the area under the curve (AUC) from plots of the receiver operating characteristic (ROC) (Pearce and F errier 2000).Both methods are available in SPSS 12.0 (SPSS Inc., Chicago IL). Nagelkerke's R- Square ranges from 0 to 1 (poor fit to perfect fit, respectively) and is similar but not equivalent to the goodness of fit offered by ordinary least squared regression, because the variance of the dependent variable in logistic regression is dependent on its frequency of distribution. AUC on the other hand measures the discrimination capacity of a model in terms of the area under a ROC curve relating the proportion of plots correctly classified as occupied (correct positive) to the proportion incorrectly classified as occupied (false positive) over a continuous range of probability threshold levels. An AUC of 0.5 indicates a model performs no better than chance while values closer to 1 indicate better discrimination. Values between 0.7 and 0.9 are considered reasonable, while values between 0.5 and 0.7 indicate poor to marginal discrimination (Pearce et al. 2002). Additionally, I plotted discrimination histograms (10 evenly spaced classes) of the predicted probabilities associated with occupied and unoccupied survey plots (Pearce and Ferrier 2000, Pearce et al. 2002). The discrimination histograms allowed me to visualize the overlap in predictions for interpretation. 30 I also used data of species occurrences not included in the analyses to test the predictive accuracy of the logistic regression models. Because these data were collected in the same LUs as the model plots and relatively few plots where species were detected were not included in model calibration, the use of these plots may not allow an independent test of model performance (i.e., pseudovalidation). However, these data do provide additional information to evaluate the results in terms of how well data collected among LUs can be used to predict species occurrences within LUs. To evaluate logistic model predictive performance, I calculated occupancy probabilities for the pseudovalidation plots and then graphed these probabilities as discrimination plots to compare predicted probabilities with observed occupancies. Pseudovalidation outliers with standardized residuals >2.58 were not included. Rem Several forest variables independently contributed to explain species’ occurrences (Table 2.2). Eastern Wood-Pewee was strongly associated with the importance values of Acer saccharum in the midstory, while Ovenbird was associated with the midstory’s canopy bottom height. No other understory variable had a significant independent contribution to any of the six species’ probabilities of occurrence. Five of the eight overstory variables made significant contributions to probabilities of bird species’ occurrences (Table 2.2). In particular, the bottom height of the overstory canopy was associated with the occurrence of three species. One of these species, Black-throated Green Warbler, was also associated with the number of large stems and the variability in canopy openness. Variation in canopy openness was a 31 significant predictor of another species affected by the mean height of the canopy bottom, Rose-breasted Grosbeak. The third species, Ovenbird, was only associated with the canopy bottom in the overstory analysis. Of the other three species, only Eastern Wood- Pewee was associated with overstory variables: the mean height of the top of the canopy and the mean canopy openness. In addition, all overstory variables except the bottom of the canopy height and the height of the tallest tree had relatively large (>1) joint contributions to Eastern Wood-Pewee occurrence. A combination of many factors associated with the overstory therefore contributes to the occurrence of this species. Patch context descriptions had significant explanatory power for only one species. Edge distance and area to perimeter ratio of the sampled patches were associated with the occurrence of Rose-breasted Grosbeak. No vegetation or patch context variables had significant independent explanatory power or large joint contributions to the probability of Least Flycatcher or Yellow-bellied Sapsucker occurrences. Several of the variables making significant independent contributions to model fits in the grouped analyses had less of an effect on model fits in the combined model. For example, the number of overstory stems and mean height of the overstory canopy bottom did not make a significant contribution to predictions of Black-throated Green Warbler occurrence in the combined model, even though they did in the overstory analysis. Similarly, patch context variables did not significantly contribute to predictions of Rose-breasted Grosbeak occurrence in the combined models. The opposite situation occurred for Eastern Wood-Pewee. Variability in canopy openness had a large joint contribution, but less than significant independent contribution to Eastern Wood-Pewee occurrence in the overstory model. Removal of one or more variables fi'om the 32 Nod. $d. vod vod. mm; med mmd Ndd and mod. and. end Nd; dwd d_._ de. wad cod 58.695 vod dmd- NNd end mmd- end and- 2; Q..— ved wed Nd; mm.~ an." wwd wwd- mod 3d 5830‘ 2d. cvd Rd ddd dwd. mod mod. dmd. N—d mod. mud- mod :4 ddd 3d :d mod and Edam—SQ Ed- dmd- 2d mod Rd. mmd mod mm; cm; de and mad mnd wwd Ed 3d- mvd- NNd ouch—ah. vod dmd. dmd dad nod- find dvd de wed de and. .md 9m.— mvé mm; 2.7 mod 3.“ 980qu w~d dvd- 3d mod and 3N mmd Nd." w—.N 5d wed- 2d med. 5d- cod cod. mm; CA 0508mm odd. cmd- de mod 3d. 2d mmd. 8d 3d 2d. cod- 2d mod de nwd 2.7 we; mm; £83m 3d- med. dad 8d end- add mmd- wdd- mmd dvd- ed; 3d on; 3d mud vmd- mud- med o_>._.mO wmd- 3d vwd nod- and- Cd Rd. dvd Nwd mdd. mud. mod Ed nmd. dmd Ed. Ed. 2d Emma wmd. Nd and add. add. 2d ~Nd. Nvd. Sud mod cod. mod mmd. nué dmd wed. ovd. and meO< 2d- find. 2d odd. dvd- 3d 3d Ed. 3d de Ned Ed Ed mmd- emd 3d- Ed em; 250mg. Ndd- med. cod vod- th- vod 2d. mm; 2..— vod Ned. Rd end vmd- dmd mvd. cod and 2508mm and Cd mod dmd _~d- dmd mmd- dmd- mmd edd. mod. odd 2d 3d vwd 3d- 3d- 3d :55 3d- mod. 3d vod- 3d. odd odd de. 7nd Sud Rd. dmd add end- 3d mod- mod. 2d 2 $5 ooeomdfiooeomoi 82m 532:3 _m\£ 3km 723 N33 em}: Sim 85338535 flea—odo—z coxozmamm xaodmocmv 25:30 cogodaobm 684 830m 63;? 580 353.32.“; commenémom -895 5053 dogmobdioflm .Eon E 32%:sz 2m :26— 3:2650 £63 338 @233on deoceoncwfi Eucamcwmm £5 33253 .Ezmficom 3&3 NemeomE 5 888m dock—c8: Panto: do 303 $.68 Edm 5539 BEES 80:95.80 88on EB can $5883.88 eozflowg mecca 3230328 wemntomod 2258 do wamcoEr—am 335885 .3 838% .Nd 033. 33 .moEatg docs—2: 550 05 FEB oBatg do 630a Deanna—axe ~52. 3. .anE 8% do meocmNEodSfi Sod voiced Boom-N $8.52 Bade—Sm “N dint? do .533 bounce—98 «codcoaovfi 3 dmd- mdd- dmd mod mwd end vcd- dmd- 2d Cd vvd- odd vod- dmd- odd Ndd dvd Sud humane .dd vod- 5d 9,: SA mwd 2d- 3d- w_d N~d $d- mod vod mmd- mod dod wcd- Nod Enowwm wod cod- ddd mud 3d dvd 3d- wnd- vod omd :d- mmd mNd mnd- ddd Ed- 3d- Cd 32 3d- ovd- 8d Nwd an; bmd dwd odd. find NNd und- Zd dmd EV; mud a; Nam-N mm; conga—um mod dmd- mmd Sud wdd- omd dwd do; mmd mud mod and 0:” had no.“ mmd- 3nd emd coao>< 3d- vvd- :d nod mmd Sud mm; 2; mod mwd dud mvd £4 mmd wmd vmd- ~Nd- mNd cock—mu. mdd nmd- ad 2: nmd Zed wed owd dvd de wmd- :d mad 3..— Nod de- mmd and 055mg. add mmd- NNd mm; mad mad on; we; Nwd mod dud. odd 8d de- 2d Nod- hwd de 0508mm ddd 3d- 3d vod end- fld vod hmd- 3d mod- mud- odd cNN M2; and :d- NNd dmd 035% 2d- wvd- Cd :d- :d- :d dwd mvé :d 2d dmd- :d and mmd- 3d 3d- vmd- mmd Eamodom wdd- 3d- 2d mod- dvd- :d mod- Z-d- add wed wwd- vod d_.~ mfin cwd vod- nod dmd 2mmu< 3:588 and. 8d. one So 85 e3 .85 N3- :3 NS 2.? Ed :3. £5. mg 93. 3o- 33 88 one- 85. and :3 :2 m2 :3 $5- 86 m3 3.? 3o moo- o3- Rd mod- 85 :5 £82 :1- 86 v: as Mad m3 :3. 5o- 85 3o :3 3o 86 Be. 85 N3- and Ed 82 :3. mg. 86 $6 «5 of 2:. 3e- 85 25 Ed. cod 2:. NS- 85 :3- OS. :3 Enema 38:80 53am 4 N d 4 N <_ 4 N 4 4 N d 4 N d 4 N : coo—033nm c.8380 25:30 coaodmobm ammo-— 330m 83.33 580 coEop-Bo=o> dammed-owed 68>? Eodmmm deficit—02m .5253 .3. 29¢ 34 overstory model thus increased the independent explanatory power and significance of this variable in the combined model. Significant log-likelihood tests of logistic regression indicated that the models of Black-throated Green Warbler, Eastern Wood-Pewee, Ovenbird and Rose-Breasted Grosbeak occurrence fit the data well (Table 2.3). Nagelkerke’s R2 values were generally low and the AUC estimates indicated the models provided marginal to reasonable levels of discrimination (Table 2.3, Figure 2.2). The AUC was significantly greater than 0.5 for models of all four species. The independent variables therefore provided some utility for discriminating between presence and absence of these species among plots. Table 2.3. Model results for logistic regression analyses of species’ occurrences. 95% CI 2 . Species x df Sigmodel RZN AUC SENP SlgAuc Lower Upper Black-throated Green Warbler 20.38 1 <0.001 0.36 0.75 0.065 0.002 0.62 0.87 Eastern Wood-Pewee 16.9 4 0.002 0.37 0.85 0.047 <0.001 0.76 0.95 Ovenbird 8.55 2 0.014 0.16 0.65 0.068 0.035 0.52 0.78 Rose-breasted Grosbeak 8.53 2 0.014 0.17 0.7 0.068 0.008 0.56 0.83 x 2: Log-likelihood test chi-square value Sigmodelz Significance of the log-likelihood test RZN: Nagelkerke's R-Square AUC: Area under the curve SENP: Standard error of the AUC using the nonparametric method for unequal frequencies SigAUC: Significance test under the null hypothesis that the true AUC = 0.5 95% CI: Asymptotic 95% confidence interval of the AUC The shape of ROC plots for the model predictions show that while all models performed better than chance, some probability estimates were more reliable than others (Figure 2.2). For example, the estimated probability of detecting Black-throated Green Warbler occurrence was a little better than chance at any probability threshold. The 35 . '_.’ Figure 2.2. Predictive performance of logistic regression models for 4 species of forest birds detected within 30am northern hardwood plots. The top row depicts discrimination ability using an ROC curve, where values above the dotted line indicate the model performs better than chance. The middle row depicts model discrimination ability using a discrimination histogram. Dark columns are proportions of absence samples and light columns are proportions of presence samples. The bottom row depicts discrimination ability for pseudovalidation samples using a discrimination histogram. Because few presence samples were available for pseudovalidation, counts of presence probabilities are plotted (open circles) with proportions of absence probabilities (dark columns). 36 Number Positive '0. t {'0 N ‘- md md vd Nd dd «'2 d ad ad vd Nd ad ad Nd .vd .wd .md O... ad ad vd Nd dd T. . . _ . °.c fl N... a z _ fill-.8 ed 1%.. -3 x3320 08395-301 8:838 .o 3:332“. 83.8... dd dd Vd Nd dd dd dd vd Nd dd . 2 . . o . 2 l O O O O N 1 to . to A a 1 . a .od 1 o o . ad a . . ad 3:92:30 .0 3:332“. 3359:. ad a... to N... a... «.9 ad to «d a... . ad ad . Nd Nd 2 .3 ed . dd dd r ad ad 323m 3.3. 3 ad 2. to no 90 3 «to no to ad ad ..-.- -.-..J 3. 3.. W3 .3 - We... {a mu- 1. W..,.m.o j -3. fl a... 8... Eficgo 326n— -uoo>> 583m V ad 0d ed Nd od dd 0d ed Nd od d.—. ad ad vd Nd o m F P b b -- p pr— = Y 9., - m K. ..__.....'..‘“ dd N. d wosqv uoguodmd od Nd d ed a... so a... a... 33.55 .320 names-guns dé iuasqv 10 inseam uoguoamd angusod 109.1109 37 discrimination histogram shows that logistic regression accurately modeled slightly higher probabilities of species presence for plots where species were detected. In comparison, Eastern Wood-Pewee commission errors (false positive) were generally low at large probability thresholds, with species presence accurately predicted at almost any probability threshold. The bowed curve for Ovenbird and Rose-breasted Grosbeak indicate those models performed poorly at intermediate probability thresholds. Classification accuracy for these models would therefore be highest if the threshold to classify plots as presence or absence was set very high or very low. The discrimination histograms for these two species confirm there was little overlap in presence and absence observations at high and low probability levels. Model performance for pseudovalidation plots was marginal (Figure 2.2). Absence predictions roughly corresponded with calibration predictions. However, occupied plots in the pseudovalidation data generally had lower probability predictions than plots in the calibration data. This observation may be an artifact of sample size as the frequencies of probability of occurrence predictions for Ovenbird and Rose-breasted Grosbeak, the species with more than 3 pseudovalidation occurrence plots, were distributed roughly the same as the calibration predictions. Variation in canopy openness within 30-m plots was negatively associated with the occurrence of Black-throated Green Warbler (Table 2.4), indicating this species may prefer northern hardwood stands without small canopy gaps. Variation in canopy openness did not significantly contribute to models of Eastern Wood-Pewee and Rose- breasted Grosbeak occurrences afier controlling for other variables (odds ratios bounded 0). However, the wide confidence interval of the odds ratios for this variable indicates the 38 lack of significance could be due to an inadequate sample size (i.e., the data contained large variance relative to the number of plots sampled). Average canopy openness showed a similar wide ranging odds ratio confidence interval for Eastern Wood-Pewee, which was not surprising as average openness was strongly positively correlated with variation in openness (Pearson correlation = 0.76, p<0.001). Only the importance value of midstory sugar maple saplings had a significant contribution to the model of Eastern-Wood Pewee occurrence. While the upper confidence interval of the odds ratio for this negatively associated variable was 1.0 indicating a lack of significance, the confidence range was extremely small and did not bound 1.0. Furthermore, its effect size was large compared to the other variables in the model as indicated by the rescaled unstandardized logit coefficients (i.e., abs(B*0)). The negative relationship between this variable and Eastern Wood-Pewee occurrence indicates that the probability of observing this bird species in a northern hardwood plot increases as the amount of sugar maple regeneration decreases. Substitution of sugar maple with Dirca or Ostrjya does not contribute to the probability of observing Eastern Wood-Pewee (Table 2.2). Therefore, this species was commonly detected in plots having an open understory where sugar maple was not the dominant midstory species. The only variable making a significant contribution to the model of Rose-breasted Grosbeak occupancy was the average height of the bottom of large tree canopies (Table 2.4). If the bottom of the canopy was high there was a greater probability of detecting Rose-breasted Grosbeak in a plot than when the canopy bottom was low. However, the wide confidence interval of the odds ratio indicated that a significant positive association with standard deviation of canopy openness might have been detected with a larger 39 sample size. While a higher bottom of canopy height favors the probability of observing Rose-breasted Grosbeak the opposite relationship was found for Ovenbird. Bottom of midstory canopy height was also negatively associated with the probability of Ovenbird occurrence, but the relationship was not significantly different from random. Table 2.4. Statistics variable strength and direction in the logistic regression models. 95% CI for Exp(B) Species Variable B SEB Exp(B) Lower Upper abs(B*a) Black-throated Green warbler StDevOpen -1.07 0.28 0.34 0.20 0.60 Eastern Wood-Pewee ACESiv -0.0003 0.0001 0.9997 0.9995 1.0000 0.91 TopCano 0.11 0.11 1.12 0.90 1.38 0.40 Angpen 1.49 1.24 4.46 0.39 50.37 0.87 StDevOpen 0.32 0.82 1.38 0.28 6.82 0.22 Constant -6.06 2.59 0.002 Ovenbird BotCanM -0.60 0.34 0.55 0.28 1.06 0.50 BotCanO -0.29 0.13 0.75 0.58 0.97 0.64 Constant 4.89 1.87 132.50 Rose-breasted Grosbeak BotCano 0.30 0.14 1.36 1.04 1.77 0.67 StDevOpen 0.67 0.40 l .95 0.89 4.24 0.46 Constant -4.92 1.71 0.01 B: Unstandardized logit coefficient SEB: Standard error of the unstandardized logit coefficient Exp(B): Odds ratio abs(B*0): Unstandardized logit coefficient multipled by the variable's standard deviation Discussion My results indicated that specific descriptions of vegetation composition and structure measured at the grain of Landsat imagery within northern hardwood stands have significant independent explanatory power in models predicting the probability of Black- 40 throated Green Warbler, Eastern Wood-Pewee, Ovenbird and Rose-breasted Grosbeak occurrences. Neither the vegetation nor patch context variables were useful in predicting the occurrences of Least Flycatcher or Yellow-bellied Sapsucker. In addition, patch context variables did not make significant contributions to models predicting the occurrence of any of these species if vegetation variables were also considered. I found that increased variability in canopy openness within 30—m plots had a negative effect on Black-throated Green Warbler. Freedman et a1. (Freedman et al. 1981) also found this species to disappear after the thinning or strip cutting of mixed deciduous forest. One possible explanation is that this species is reluctant to cross through canopy gaps (Rail et al. 1997). The perforation of a forest by selectively removing equally space trees during uneven-aged management might therefore be perceived as habitat loss by Black-throated Green Warblers. In contrast to Black-throated Green Warbler, Rose-breasted Grosbeak appears to select for plots with canopy openings and overstory trees having limbs relatively high above the ground. These conditions would promote the growth of understory shrubs used as nest sites by this species (Ehrlich et al. 1988). However, I did not detect an association between Rose-breasted Grosbeak and the important value of Dirca palustrus, a shrub species relatively insensitive to deer browsing pressure in the region (LeBouton et al., unpublished data). The effect of shrub density on the probability of Rose-breasted Grosbeak occurrence might have been identified by lumping data for all shrub species into a single variable (e. g., shrub stems/ha).Therefore, canopy openings and higher tree limbs could be proxies for shub density. Similarly, previous studies have shown a preference by Rose-breasted Grosbeak for forest edges and young forests (e. g., Probst et 41 al. 1992). I also detected a significant association between Rose-breasted Grosbeak occurrence and both edge distance and perimeter to area ratio in my patch context analysis. However, these variables were not significant in the combined analysis, likely because of their correlation with canopy openness variability and bottom of overstory height. Ovenbird probability of occurrence was also significantly associated with bottom of canopy height. Lower limbs in the overstory and possibly in the midstory increased the probability of detecting this species. Although Ovenbird may be the most studied bird in my analyses, I am aware of no other research indicating an association between Ovenbird occurrence and the height of the bottom of canopies. However, there is ample evidence that this species selects breeding locations based on local conditions. Most reports indicate a consistent preference for areas with mature trees 16-22 m tall and canopy openness of 10-40% (Van Horne and Donovan 1994). Ovenbird territories are sometimes located in areas with larger trees, less open canopies and less ground cover than other locations within the surrounding forest (Smith and Shugart 1987). Contrary to a growing body of literature on the area sensitivity of this species (Temple 1986, Robbins et al. 1989, Pomeluzi et al. 1993, Flaspohler et al. 2001), I detected no significant relationship between Ovenbird occurrence and any patch context variables. This difference could be due to the relative paucity of agriculture in my study region, which can increase the chances of brood parasitism along edges by cowbirds (Donovan et al. 1997). Only two cowbirds were detected in 1299 point count surveys over the study region (Laurent et al. unpublished data). 42 Eastern Wood-Pewee was associated with more variables than any other species I studied. While this result could be due to my small sample size of presence plots for this species, other studies also found Eastern Wood-Pewee associated with several aspects of forest composition and structure. I detected Eastern Wood-Pewee in plots having a relatively open understory as indicated by the negative association with sugar maple sapling importance values. Similarly, Crawford et a1. (1981) found this species to be most abundant in Virginia forests with little understory vegetation. However, deer-induced changes to intemrediate canopy structure may also have a large effect on its occurrence (DeCalesta 1994). My research supports previous findings that Eastem-Wood Pewee occurrence is not sensitive to forest area or edges (e.g., Blake and Karr 1987, Robbins et al. 1989) but is associated with canopy openings within forests (Hespenheide 1971). I failed to detect a significant association between Least Flycatcher or Yellow- bellied Sapsucker occurrence and any understory, overstory or patch context variables. Least Flycatcher form colonial aggregations, which can leave apparently suitable habitat unoccupied in adjacent areas (Davis 1959, DellaSala and Rabe 1987). This behavior adds additional error into analyses, which likely decreased my ability to detect a significant association. My small sample size of presence plots would compound the problem if individual birds selected for proximity to colonies over local conditions. Yellow-bellied Sapsucker, on the other hand, has very specific requirements for young forest (Walters et al. 2002). Detections of these species within 30-m radius plots were therefore likely a consequence of proximity to young forest, rather than the proximity to the stand’s edge or vegetation composition or structure within the stand as measured in this study. 43 My inability to model Least Flycatcher and Yellow-bellied sapsucker occurrences along with the weak to moderate predictions of species associated with 30-m radius vegetation descriptions indicated that unmeasured relationships are very influential in determining whether all six species will occupy a small plot within a stand. Although my sampling design limited any directional effect of landscape context on the results, landscape context could still have increased model error. There is growing evidence that landscape context influences species presence within areas having similar local conditions (Pearson 1993, Jokimaki and Huhta 1996, Saab 1999, Donnelly and Marzluff 2004). For example, Askins and Philbrick (1987) noticed that Black—throated Green Warblers disappeared from several small northeastern forests (<100 ha) possibly due to the removal of surrounding forests, even though the local conditions did not appear to change. A species may therefore occur in one area and be absent in a very similar area due to the surrounding environment. The small effect size of some variables in the predictive models could also be an artifact of my study region’s extent. My region of interest covers multiple Level III ecoregions (Albert 1995), which by definition differ in terms of climate and geology. These lower level processes affect vegetation composition, structure and productivity directly and thereby indirectly influence wildlife species’ occurrences. For example, ecoregional differences in the relationship between proportional deciduous cover and species richness have been observed over my study region (see Chapter 3 for details). Similarly, it is possible that spatial heterogeneity, or “the complexity and variability of a system property in space” (Li and Reynolds 1994) was present in the habitat associations of individual species. Further work is needed to investigate whether spatial heterogeneity 44 played an influential role in my results. I am aware of no existing statistical methods for investigating the influence of spatial heterogeneity on presence/absence data, although several methods are available for continuous variables (see Fotheringham et al. 2002, Shi et al. in press). While landscape context likely played an influential role in my analyses, it is the exploitable local conditions that provide necessary resources for individual growth and reproduction. Therefore, investigations into local conditions consistently associated with species occurrences over large areas are an important step in determining the mechanisms behind species distribution patterns. My results indicated that local conditions play a significant role in determining species occupancy within small northern hardwood plots across my region of interest. The inclusion of more plots or a larger sampling area may have improved my results as the number the of detections would likely increase as a function of the species-area curve relationship (see Rosenzweig 1995). This is because the probability of detecting a species will increase as a function of increasing the sampled area and also because an increased presence frequency will affect the accuracy of logistic regression models (Cumming 2000). However, a larger sample plot could also increase error in the analyses by including areas not used by the species. Similarly, additional error could be incorporated as a function of spatial autocorrelation, where within sample plot variation will increase as the area contained within a plot gets larger. The generally poor performance of the models for predicting occupancy in the pseudovalidation plots could also be a product of the sampling design. All pseudovalidation presence plots were within the same township sections as calibration presence plots. Therefore, the calibration plots could have contained conditions favorable 45 for species, while species occurred in the pseudovalidation plots because of the proximity of these conditions. It has been shown that abundant availability of a critical resource in a small area may compensate for its paucity in other portions of a bird’s territory (Steele 1992). However, the presence plots used for model calibration were chosen at random. I was therefore unaware which plots contained conditions more favorable for occupancy. In general, the pseudovalidation results indicate that patch and/or landscape context likely plays a strong role in whether species are detected within 30-m northern hardwood plots. Local conditions are also important, however, as the fit of my models indicated many factors within stands contribute to species’ occurrences across the study region, regardless of landscape context. The frequencies of pseudovalidation absence predictions roughly matched that of the calibration predictions indicating that the models may have useful explanatory power. F urtherrnore, the predicted probabilities of the few presence pseudovalidation plots provide only weak evidence to the contrary. Maggement Impligations Altering the timing or method of harvest among 30-m radius or 30-m x 30-m blocks within uneven-aged northern hardwood stands will increase the probability of detecting multiple species within the entire stand. For example, selective harvest within one block will increase the variability of canopy openness and therefore could increase the probability of detecting Eastern Wood-Pewee and Rose-breasted Grosbeak. Other unharvested blocks within the stand will have less variability in canopy openness and therefore higher probabilities of Black-throated Green Warbler occupancy. While management for higher limbs on overstory trees in one block will increase the probability 46 of detecting Rose-breasted Grosbeak, management for lower limbs in another will increase the probability of detecting Ovenbird. Although uneven-aged management is typically focused on increasing canopy openness over the entire stand through dispersed selective harvesting, this activity also indirectly affects forest vertical structure by providing more light to shrubs and lower limbs. Dispersed harvesting also evenly distributes the availability of browse for deer, which can have additional indirect consequences on forest composition and structure. Hence, I anticipate that increasing the spatial precision of management activities using clumped rather than dispersed selective harvesting within stands will increase avian alpha (local) diversity across the landscape. While such activities may not directly augment gamma (regional) diversity, they may enhance population stability by providing suitable habitat within parts of stands during any point in time, thus promoting species persistence across the landscape. 47 CHAPTER 3 PLOT SIZE, STUDY EXTENT AND SPATIAL HETEROGENEITY AFFECT THE RELATIONSHIP BETWEEN NDVI AND SPECIES RICHNESS Introduction The Normalized Difference Vegetation Index (NDVI) derived from remotely sensed imagery has recently been employed in several different studies to investigate relationships between species richness and energy availability (e.g., O'Connor et al. 1996, Fraser 1998, Bawa et al. 2002, Oindo 2002, van Rensburg et a1. 2002, Chown et a1. 2003, Kerr and Ostrovsky 2003, Hawkins 2004). TThese efforts were typically limited to large grain (sample plot) and large extent (study region) datasets (but see Gould 2000, Bailey et al. 2004, Seto et al. 2004). NDVI, which is the normalized ratio between reflected red and near-infrared light, is relatively easy to calculate and has been shown to be a consistent measure of biomass and net primary productivity (Benning and Seastedt 1995, Hobbs 1995, Paruelo et al. 1997, D'Arrigo et al. 2000). According to species-energy theory (e.g., Connell and Orias 1964, Wright 1983, Wright et al. 1993), greater amounts of available energy (as measured in terms of net primary productivity and other proxies) should result in more individuals and more species. Therefore, changes in NDVI values over space should track well with changes in species richness. Although NDVI as a proxy for energy availability has a consistent correlation with key processes characterizing vegetation activity such as net primary productivity and actual evapotranspiration, these processes vary by species (Malmstrom et al. 1997) as well as by individuals of species (D'Arrigo et al. 2000). Pixels containing multiple individuals of multiple species will therefore contain an aggregate signal of reflected 48 light. Furthermore, when groups of pixels are summarized during resampling, which is often done for large extent studies, NDVI values are averaged or “smoothed” by some other method of statistical summarization. Smoothing therefore results in a summary of what is already an aggregate value. If a pixel is assumed to represent a vegetation community, then resampling can be expected to aggregate multiple communities having increasingly disparate qualities as the resampling window gets larger (relative to the grain of sampling). Interpretation of the biophysical mechanisms causing the relationship between NDVI and species richness of any taxa will therefore be dependent on patterns of plant species composition and how the data were collected and aggregated. Because vegetation composition differs over space as a consequence of many factors such as climate, geology and history, it is not surprising that habitat heterogeneity, loosely described as variation in vegetation communities over space, was often cited as having a scale-dependent influence on the relationship between NDVI and species richness. For example, Hurlbert and Haskell (2003) resampled NDVI values fi'om Advanced Very High Resolution Radiometer imagery (1 km2 grain) to investigate the relationship between avian species richness and NDVI at different grains. They found a positive relationship at a large grain (>200,000 ha) and extent (North America) between NDVI and migratory avian species richness estimates from Breeding Bird Surveys (Bystrak 1981). They concluded that productivity affects the number of species and migratory guild composition of breeding communities. However, at larger grains (e.g., 2,000,000 ha) the influence of biome heterogeneity was found to decrease the fit of their results. Hurlbert and Haskell (2003) suggested that biome heterogeneity affected the larger grain relationships due to sampling effects. As the grain of analyses increased, 49 their sampling areas contained a greater variety of vegetation gradients and therefore likely overlapped multiple pools of bird species. Greater precision in the relationship between productivity and species richness was suggested to occur at smaller spatial grains. Any mechanisms behind these scaling relationships are likely tied to the biophysical factors that influence the intensity of reflected red and near-infiared light at different scales. NDVI values are influenced by several biophysical factors. Chlorophyll pigments of healthy plants efficiently absorb radiant energy in the red portion of the spectrum (Jensen 1983, 2000). Higher densities of photosynthetically active plants will absorb more red light and reflect less, thereby increasing NDVI values. Near-infrared wavelengths are reflected at the cell wall-air interfaces of leaves (Gausman et al. 1969, Peterson and Running 1989). The amount of spongy meSOphyll within leaves (where the exchange of oxygen and carbon dioxide take place for photosynthesis and respiration) has a large effect on the amount of reflected near-infrared light (Jensen 2000). Plants with a large amount of spongy mesophyll, such as broadleaf deciduous trees can be expected to have higher NDVI values than coniferous trees, with less spongy mesophyll at equal amounts of leaf area (e.g., Garnon et al. 1995). These and other factors combine to explain NDVI’s usefulness as a measure of energy availability, as photosynthetically active radiation and foliar standing crop are positively related to primary productivity (Webb et al. 1983, Law and Waring 1994). However, spatial variation in the proportion of deciduous vs. coniferous cover may also have large effects on NDVI values depending on the grain of data collection and resampling method (e. g., Gemmell 1995, McDonald et 50 a1. 1998). Other factors that may also affect NDVI values include soil moisture (Todd and Hoffer 1998, Stoms and Hargrove 2000) and forest age (Niemann 1995). Because the proportions of deciduous and coniferous cover vary over biomes and ecoregions within biomes, a weak to moderate relationship between NDVI and avian species richness should be expected as a consequence of ecoregional processes contributing to productivity (e.g., Hawkins 2004). Within ecoregions, a weak to moderate relationship should also be expected because of local differences in productivity and proportional coverage of plant species. Thus, spatial variation in NDVI may not necessarily accompany covariation in productivity. For example, areas of relatively low NDVI values in regions dominated by coniferous forests could be more productive and contain more bird species than deciduous forest with higher NDVI values. Consequently, the interpretation of NDVI as a predictor of species richness should not hinge on NDVI’s relationship with productivity. Rather, investigations into the scale-dependent influences of plant species composition on NDVI could provide more comprehensive information concerning which biophysical factors influence the relationship between NDVI and species richness. Along with the availability of multiple satellites to monitor NDVI remotely at different resolutions, this knowledge could greatly help elucidate causes of spatial variation in species richness over time. The relationship between NDVI and species richness is expected to vary over space as a consequence of multiple scale-dependent processes. Therefore, the influence of spatial heterogeneity, or “the complexity and variability of a system property in space” (Li and Reynolds 1994), should be considered during analyses, especially when investigating relationships over large areas relative to the sampling grain. If the impact 51 of spatial heterogeneity is not considered in regression models, it will result in biased parameter estimates, misleading significance tests, and suboptimal prediction (Anselin and Griffith 1988). As a case study on the relationship between avian species richness and NDVI, I was interested in whether the pervasive positive relationship between migratory bird species richness and NDVI held over a region of managed forests in the Upper Great Lakes, USA. Of particular interest were the local and/or regional effects of land cover affecting the relationship at a grain useful for site-specific forest management decisions. If spectrally detectable environmental factors affecting avian species richness could be identified, knowledge of these factors would be useful as part of monitoring programs or forest planning efforts. My objectives therefore were to: 1) calculate the slope and strength of the relationship between species richness and Landsat derived NDVI at the grain of individual bird point count surveys, 2) identify land cover components affecting NDVI values at the grain of pixels (30-m x 30-m) and point count surveys (~10-ha), and 3) evaluate the influence of spatial heterogeneity on model slope and fit in questions 1 and 2. . By analyzing spatially explicit relationships between avian species richness, NDVI and causes of variation in NDVI, I sought to infer causation of fine grain species richness patterns detectable through NDVI values. Methods Study Region The study region is located in Michigan’s Upper Peninsula, USA and includes parts of five counties (Baraga, Dickinson, Iron, Menominee, and Marquette; Figure 3.1A) 52 and three subsection level ecoregions (Albert 1995). This region covers approximately 400,000 ha. A majority of the land is owned by state government and industry with management primarily focused on wood production. The study region is an ecologically diverse landscape characterized by a spatial mosaic of forest stands that include upland hardwoods (sugar maple (Acer saccharum), quaking aspen (Populus tremuloides), yellow birch (Betula alleghaniensis), basswood (T ilia americana)), lowland hardwoods (black ash (F raxinus nigra), red maple (Acer rubrum)), and lowland (northern white cedar (Thuja occidentalis), black and white spruce (Picea mariana and P. glauca), tamarack (Larix laricina)) and upland conifers (European larch (Larix deciduosa), eastern hemlock ( T suga canadensis), red and jack pine (Pinus resinosa and P. banksiana». Besides the influence of post-glacial topography and other edaphic factors, the composition and structure of canopy and understory vegetation varies spatially due to effects of forest management practices and gradients of deer browsing pressure (Albert 1995, Van Deelen et al. 1996). Data were collected as part of a multidisciplinary project investigating the ecological and economic effects of timber harvesting and deer browsing on forest resources. Due to multidisciplinary project goals, my study region boundaries were delineated qualitatively based on several factors. First and foremost, I was interested in a region where ecological dynamics were largely influenced by disturbance from forest harvesting practices and deer browsing. I therefore selected an area that was predominantly forested and contained only a few small towns and limited agriculture. To the east and south, the study area was bounded by rural housing and farms along interstate highways US-2 and US-4l and state highway MI-35 connecting the cities of 53 Figure 3.1. The study region lies within a single Landsat 7 ETM+ scene and contains parts of five counties in Michigan’s Upper Peninsula, USA (A). Within the study region (B), randomly selected landscape units (C; ~1000 ha area with gradients of NDVI values) define areas for the selection of specific plots for bird occurrence surveys. Plots (D; overview showing different tree species) were sampled within a 30-m radius from plot centers (D; large circle), so that a single Landsat 7 ETM+ pixel (D; square) falling within this area could be precisely characterized. 54 Gwinn, Escanaba and Iron Mountain. I also excluded Delta County to the southeast because of a sharp increase in developed land use (suburban and agriculture) close to the county line. A subsection level ecoregion boundary (Albert 1995) served as the northern border. No linear feature was available to border the west. For this reason, a western boundary was loosely delineated based on access constraints. Survey Plots 433 survey plots were established over the summers of 2001 to 2003. They were located within 96 randomly selected landscape sampling units of approximately 1000 ha (LUs, Figure 3.1B, see Chapter 4 for details). Within each of the 96 LUs, between 2 and 8 plots were selected (Figure 3.1C). These plots were a minimum of 90-m apart with the mean minimum distance of 240-m (:194-m SD; i.e., strong right skew towards greater distances) from the next closest plot. Plot selection in general was dependent on two criteria: 1) permission to access the property and 2) maximizing the biotic and abiotic differences among plots to sample the spatial and spectral heterogeneity of the study region. A 30-m radius circle was used to define survey plots characterizing individual pixels (Figure 3.1D). Plots were selected where field crews perceived that a hypothetical 30-m x 30—m square (the resolution of Landsat 7 ETM+ imagery) could be placed anywhere within the circle and expected to have the same vegetation structure and composition as a similar square placed anywhere else within the circle. In this way, a single pixel of Landsat 7 ETM+ imagery sensed over plot centers could be precisely characterized by land cover data collected within the circle. The choice of 30-m radius 55 circles over 30-m x 30-m squares was made for several technical reasons including image distortion due to the rotation of both the Earth and Landsat 7 satellite, georectification error, and sensor mechanics (see Chapter 4 for details). Survey plots were placed in a wide variety of land cover classes (Space Imaging, Inc. 2001) including aspen associations (n = 85), herbaceous openland (n = 10), lowland coniferous forest (n = 35), lowland deciduous forest (n = 14), lowland shrub (n = 2), mixed upland conifers (n = 4), mixed upland deciduous (n = 6), northern hardwood associations (n = 190), pines (n = 23), other upland conifers (n = 15), upland mixed forest (n = 36), and upland shrub / low-density trees (n = 10). Preference was given to at least one northern hardwood plot if it existed in each LU to satisfy the aforementioned collaborative research priorities. Survey plots were located at varying distances from hard and sofi vegetation edges as long as specific plot selection criteria were met. Locations of plot centers were flagged and georeferenced using a minimum of 80 global positioning system (GPS) point locations collected at 5-second intervals with a Trimble GeoExplorer 3 GPS receiver (Trimble Navigation Ltd.) and later differentially corrected to a precision of : S-m using Coast Guard base station data. Accuracy of plot center locations was determined to be within the precision of Garrnin III+ GPS receivers (~ lS-m; Garmin Corporation) used to navigate to these locations during repeated point count surveys. ND VI All image processing was conducted using Imagine 8.7 software (Leica Geosystems GIS & Mapping LLC). A May 29, 2001 Landsat 7 ETM+ 1G image of Path 24 Row 28 was obtained from the USGS in HDF format. Deciduous plant species have 56 already begun producing leaves during late May in Michigan’s Upper Peninsula, with forest canopies approaching, but not completely representative of full-biomass (personal observation). Late May is also the time when most migratory passerines in the UP are finalizing their nesting territories (Brewer at al. 1991; R. Adams, personal communication). Use of a late May image therefore provides a snapshot of productivity during a time that is critical to migratory birds. I converted the image to at-sensor reflectance using the import utility of lrnagine 8.7. At-sensor reflectance was converted to surface reflectance through a dark object subtraction (McDonald et a1. 1998) of the minimum reflectance value measured for random pixels overlapping the centers of four deep lakes and reservoirs located in the study region. The map of surface reflectance values was then georectified using a road map (root mean square error < 8m). The NDVI index was calculated from surface reflectance values using the normalized ratio of reflected intensities of light in the red (band 3) and near-infrared (band 4) wavelengths (see Jensen 2000). NDVI pixels overlapping plot centers were used to describe plots. However, bird point count surveys cover a larger area. For this reason I averaged the NDVI values within a 6 pixel radius (180-m, see Bird Surveys) from plot centers using the Neighborhood Statistics menu available in the Spatial Analyst 2.0 extension of ArcView 3.2 (Environmental Systems Research Institute Inc., Redland, CA). The mean NDVI values within this larger neighborhood (NDVIjgo within 180—m or 6 pixel radius) were used to describe the area surveyed during point counts. Bird Surveys 57 Plots were surveyed for birds during the summers of 2001 (n = 112), 2002 (n = 96) and 2003 (n = 235). Each plot was visited 3 times between June 4 and July 3 in either 2001, 2002, or 2003. Most species have concluded migrating through the study region by the first week of June, and the timing of my data collection overlaps the breeding period of most species in the study region (Brewer et a1. 1991, R. Adams, personal communication).A total of 8 observers collected data over the 3 years. The surveyed order of LU routes, plots among LUs, and survey timing was randomized, as was the selection of observer for each plot. At least two observers surveyed each plot to account for possible differences in observer’s physical abilities to detect the study species (Ramsey and Scott 1981). Observers were trained to detect the species in the region by song and sight using tapes, field guides and practice before data collection. Bird species were counted as present if detected by song, call or sight within any of the three visits using 10 minute unlimited distance point count surveys. Although bird species composition may have differed within plots over each of the three years of this study (see Morrison 1992), it was assumed that species richness did not vary due to any changes in vegetation composition or structure during this time (see Hansen and Urban 1992, Holmes and Sherry 2001). I assumed all birds were detected within ~180—m (see Wolf et a1. 1995). Presence of flushed individuals detected < 180—m from plot centers by surveyors walking in or out of the plots (as determined by Garmin GPS III+ receivers; Garmin Corporation) was also recorded. However, I did not include detections of species flying over plots without landing, because it was not known if these species used the locations as habitat. 58 My species richness estimates included all migratory passerine species seen, heard or flushed within 180-m during any of the three surveys. I included only migratory species of the suborder Passeres in these analyses because of possible confounding effects caused by differences in habitat use among other taxa (e.g., colonial aggregations of Least F lycatchers (Empidonax minimus; Davis 1959)). Many of the passerine species included in this study, which included neotropical migratory warblers, are of conservation concern in the Midwest due to declining numbers (Robinson 1995). The reduction in numbers is often recognized as a response to increased disturbance fi'om timber harvesting (Rotenberry et al. 1993) and the resulting younger (Holt and Martin 1997), smaller (Ambuel and Temple 1983, Niemi and Hanowski 1984), and relatively more homogeneous forest stands (Bunnell and Kremsater 1990, Thompson et al. 1995). Species Richness - ND VI Analysis The computer software program GWR 2.0.3 (F otheringham et a1. 2002, http://www.ncl.ac.uk/geps/research/geography/gwrl) was used to model the relationships between avian species richness and NDVIlgo. Geographically weighted regression (GWR) is an extension of the traditional regression framework (Zhang and Shi 2004). It explicitly incorporates the spatial locations of data and can therefore be used to investigate the influence of spatial heterogeneity on model fit. The local estimation of model parameters was derived by weighting all neighboring observations using a decreasing function of distance. In this way, the impacts of the neighbors nearby were stronger than those farther away. A threshold, called the kernel bandwidth was also identified to indicate the distance beyond which neighbors no longer have influence on 59 the focal observation (Brunsdon et al. 1996, Fotheringham et al. 2000, 2002). The kernel bandwidth was obtained by minimizing Akaike’s Information Criterion for the GWR model fit. An F -test available in the GWR 2.0 software allowed me to test whether GWR provides a significant improvement in model fit compared to linear regression (see Leung et al. 2000). Prior to analyses, species richness estimates were square root transformed to reduce a slight right skew. There was a strong lefi skew to NDVIjgo. I therefore transformed these data using the loglO of their reflective values ((max + 0.01) — x; (Quinn and Keough 2002)). Transformed NDVI values were then multiplied by —l to flip the variable back to its original orientation for analysis and interpretation. To guard against the influence of spatial autocorrelation on significance tests (see Legendre 1993), the degrees of freedom for all significance tests were adjusted to the number of LUs, my random sampling unit, instead of the number of sampled plots. Significance was tested at a=005 Output from GWR includes parameter estimates and model fit for each point as well as a kernel bandwidth specifying the cut-off distance for incorporating the influence of surrounding neighbors. I used the kernel bandwidth to evaluate whether differences in local slopes of the relationship between NDVI and species richness were due to differences in the forest cover surrounding sampled plots (i.e., landscape context). Circles with radii equal to the kernel bandwidth of the GWR model for the relationship between NDV1130 and species richness were generated around each plot using the buffer feature in ArcView 3.2. Proportional cover of deciduous forest, non-deciduous forest, and non- forest were tabulated within the circles. The non-forest category was removed because of 60 its weak correlation with local slopes (r = 0.14), as was the redundant variable of deciduous cover. The local slope parameters from the GWR model of the relationship between NDVI and species richness were fit to a linear model containing proportional cover of non-deciduous forest within the GWR kernel bandwidth. However, I visually identified several outliers in a preliminary graph of the relationship between local slopes and deciduous cover (Figure 3.2A). A similar graph using non-deciduous cover (Figure 3.23) provided additional evidence that these plots were outliers and could represent a functional difference in the mechanism affecting local slopes in the relationship between NDVI and species richness. After plotting their locations, all outlier plots were found to occur along the northern ecoregional boundary of my study area (Figure 3.2C). The combined evidence supports my assumption that the transition zone fi'om the northern ecoregion was influencing my analyses. These transition zone plots were therefore considered outliers, and additional quantitative outlier identification was deemed unnecessary. I removed the outlier plots and repeated all the analyses with the remaining data (11 = 404, LUs = 87). ND VI - Land Cover Analysis Land cover descriptions were used in GWR 2.0.3 to investigate how the amount of deciduous forest, non-deciduous forest or non-forest cover affected NDVI values and whether the relationships showed evidence of spatial heterogeneity. Proportion cover of different land cover classes was summarized in three ways: 1) within plots, 2) within a 180-m radius from plot centers containing the area covered by point count surveys, and3) 61 Figure 3.2. Plot of the relationships between proportional cover of (A) deciduous forest within the global kernel bandwidth (8258m radius) of the GWR model of species richness and NDVI plotted against local slopes derived from the same model. Outlier plots (open circles) were visually identified as not occurring within the trend of the relationships with deciduous cover (dotted ovals). A linear models of the fit for outlier plots (dashed line) is provided for comparison with a model describing the fit of the rest of the data (solid line). A map of outlier locations (D) shows outliers occur along the northern ecoregional border of the study region. 62 1 . y I 4.69x + 0.18 R’ - 0.48 y I 4.74x - 3.44 R’ - 0.31 o s 1 :1 i ‘\ in .-"f 1" /’ Proportion Deciduous Forest B4 63 1 _ y=-6.94x+1.36 y=7.26x-3.74 R’ - 0.61 R‘ = 0.34 o s «1 0.8 1 .1 T .2 _ 4 - c dProportion Non-deciduous Forest 0 Outliers 0 Sample plots D Econgions Study region 10 0 10 Km within the kernel bandwidth specified by the richness—NDVIlgo GWR model described above. Because of a strong left skew in NDVI130 and NDVI values, I transformed both sets of data using the reciprocal of the reflective loglO transformation mentioned above. Within most plots (n = 427), I visually estimated from the ground the proportional canopy cover by species as I expected it to be seen if looking down upon the canopy (i.e., satellite’s perspective). Six plots were missed during sampling and therefore not included in these analyses. Proportional cover for each plot was summarized by percent cover of broadleaf deciduous trees, non-deciduous trees and non-forest. Deciduous cover estimates were skewed to the left. These data were converted to percentages and then transformed using the reciprocal of the square root of the reflective (-sqrt(101- x)). Non- forest cover was square root transformed. Non-deciduous cover was excluded fi'om analysis because assumptions of normality could not be met due to a large number of zeros. As is to be expected from proportional estimates, a strong negative correlation occurred between deciduous and non-forest cover (Table 3.1). The redundant variable (having lower correlation with NDVI) was therefore removed, leaving deciduous tree cover as the predictor variable used to investigate causes of NDVI values for individual pixels. Both global and local models were created to model the fit of NDVI on deciduous cover. The significance of spatial heterogeneity on model fit was investigated using an F- test available in GWR 2.0.3. NDVIng values describing areas covered by point count surveys were modeled against land cover class proportions using a land cover map (Space Imaging, Inc. 2001) in grid format with 30-m cells. First, land cover was aggregated into three cover classes including: broadleaf deciduous forest (northern hardwood, aspen, oak, mixed upland or Table 3.1. Correlation coefficients of variables used in analysis of the relationships among NDVI, bird species richness, and land cover. NDVI pDec NDVIlgo pDCClgo pNonDeclgo bb pDeck pNonDeck Species Richness -0.42 -0.37 pDec 0.72 pNonFor -0.61 -0.76 pDec1 80 0.77 pNonDec 1 80 -0.54 -0. 84 pNonForlgo ~0.74 ~0.69 0.35 pDeck 0.61 pNonDeck -0.78 -0.82 pNonFork 0.14 -0.48 -0.1 b b : Slope estimate of the relationship between NDVI-richness local slope estimates from NDVI: NDVI 1 80: pDec: pNonF or: PD60 1803 pNonDec 180: pNonFor] 80: pDeck: pNonDeck: pNonFork: the GWR model and proportional land cover within kernel bandwidth surrounding survey plots. Normalized Difference Vegetation Index calculated for focal pixels Normalized Difference Vegetation Index calculated for pixels within a 180m window surrounding focal pixel Proportion of deciduous forest cover within focal pixels Proportion of non-forest cover within focal pixels Proportion of land cover pixels classified as deciduous forest within a 180m window surrounding focal pixel Proportion of land cover pixels classified as non-deciduous forest within a 180m window surrounding focal pixel Proportion of land cover pixels not classified as forest within a 180m window surounding focal pixel Proportion of deciduous forest cover within a window around focal pixels with radius of the GWR kernel bandwidth Proportion of non-deciduous forest cover within a window around focal pixels with radius of the GWR kernel bandwidth. Proportion of non-forest cover within a window around focal pixels with radius of the GWR kernel bandwidth. 65 lowland deciduous), non-deciduous forest (pines, lowland coniferous forest, other and mixed upland conifers, lowland coniferous and mixed forest) and non-forest (upland and lowland shrubs, herbaceous, agriculture, urban, etc.). Circles of 180-m radius surrounding plot centers were then generated using the buffer feature within ArcView 3.2 (Environmental Systems Research Institute Inc., Redlands, CA). The proportional coverage of each class within circles of 180—m radius surrounding plot centers was summarized using the Tabulate Areas feature of ArcView 3.2 as percentages. Deciduous cover was transformed using the reciprocal of the square root of the reflective. Both non- deciduous and non-forest cover were Log10 + 1 transformed. All three cover variables were highly correlated (Table 3.1) so the redundant variables were removed, leaving only deciduous cover as the predictor in the analyses. Both global and local models were created to model the fit of NDVI130 against the proportion of deciduous land cover cells within the same size area. The significance of spatial heterogeneity on model fit was investigated using an F -test available in GWR 2.0.3. $511M Species Richness — ND VI Linear regression indicated that species richness of migratory passerines had a weak, but significant negative relationship with NDVI (Table 3.2). The GWR model improved the fit over linear regression, and an F-test for the influence of spatial heterogeneity was significant (Table 3.3). The slope of the local relationships between species richness and NDVI ranged from negative to positive, although most points had a negative slope. 66 Table 3.2. Results of linear regressions among NDVI, species richness and proportional land cover. See footnotes of Table 3.1 for descriptions of variables. Dependent Predictor b n dfadj t F R1 Species Richness NDVI 1 80 -l .15 404 87-2 -9.2 85.3 0.18 NDVI pDec 0.15 398 87-2 17.4 304 0.43 NDVI] so pDec 1 80 0.44 404 87-2 23.2 537.3 0.57 dfadj: Degrees of freedom adjusted to the number of random landscape sampling units 11: Number of bird survey plots ND VI — Land Cover NDVI values were positively associated with deciduous cover (Table 3.2). Use of GWR did not make a significant improvement to model fit (Table 3.3). A comparison of pixels dominated (>50% cover) by individual tree species shows that larger NDVI values were obtained for deciduous species compared to coniferous species (Figure 3.3). In addition, NDVI values for some species show high variance (e. g., quaking aspen), while others did not (e.g., sugar maple). The relatively large variation in NDVI values for plots dominated by aspen could be associated with differences in leaf-out timing across the study area captured within the late May image. Regardless, the proportion of deciduous cover both within pixels and within the area of point counts explained a majority of the variance of NDVI values (Table 3.2). The relationship within pixels did not vary spatially but the relationship within point count areas did (Table 3.3). 67 Assam: a 385 =88 a =03 a Assay: gene 98 Axmemmv Engage AEENMV 82558 :82 05 .8.“ 8235 2a 338m .2258 M303 3c :33 ” NM £22: 5 .25.: 550 no 52333 650M H x beeowobwo: .3QO do 85:55 Le aEmcomHEB 30:: do 8535ch mczmod <>OZ< Eat o=_a>-m ” m EoEoSEEM mam E 8208c we 82on ” $36.6 £5258 WHO 05 E Eowooa .«o woodman “ 30% .82? Q2. 3 83::me use are 3 SE58 Ass 3 EEEEE :23 £258 M30 Bob mofifizmo one? .83 ”a 82 .3 a; 8.2 N 36 $5 med 3 33 So- 33% 2:594 $3 _ owN md 3 .M: N vmd mud end Nd m _ .o cod coma SQZ 88 a? mg 3.3 N Ed So was ad mm. _- N2- 3.592 $2208 seam v— ..m Ever—LN“ mBOmmv Bobu SEEN.“ BEN.“ EEN.NH me Q BE Q EE Q wowomfiem «cuficoaoa ave—amt? do mcofitomoc co.“ _.m 038. do 86:88 com .830 98— 28 32:6: 86on A>QZ mecca commocmou BEER,» 18393wa No mzaom .m.m 2an 68 1 . = 56 n = 129 0.95 - B 0.9- 2 0.85 - 0.8 "' I l Acer Populus Thuja Pinus Pinus saccharum tremuloides accidents/is resinosa banksiana Figure 3.3. Mean NDVI values (i 1 SD) for pixels dominated (>50% cover) by different tree species. Solid bars indicate deciduous species and hatch bars are presented for non- deciduous species. Data are provided for a subset of plots surveyed for land cover within the Upper Peninsula of Michigan. Variation of Local Slopes Proportional cover of non-deciduous forest, and to a lesser extent deciduous forest, within the kernel bandwidth of the GWR model affected the local slope of the relationship between NDVIlso and migratory passerine species richness (Figure 3.2). The slope decreased from positive values close to zero to negative values as non-deciduous cover within the kernel bandwidth distance around each plot increased from ~ 20% to 69 ~55% (Figure 3.28). This relationship was significant when degrees of freedom were adjusted to the number of landscape units (b = -6.9, t = -25.1, R2 = 0.61, df = 85). Plots of species richness, local model slopes, local model fits and the proportion of deciduous and non-deciduous forest within kernel bandwidth sized areas surrounding sampled plots were visually compared (Figure 3.4). Local model fit was highest along a transition zone to >40% non—deciduous cover in the middle of the study region. This transition zone of local model fit tracked loosely with an ecoregional boundary. In the eastern ecoregion, the slope of the relationship between avian species richness and NDV1130 was very negative with the most negative local slopes occurring in the center of the study region where the proportions of deciduous and non-deciduous forest were most similar. Because NDVIlgo values were positively dependent on their proportional deciduous cover, plots in the east (where there is more non-deciduous forest) contained more bird species in point count areas with more non-deciduous forest. In the west (where there is more deciduous forest), areas with more deciduous forest contained more bird species. Thus, it appears my study region contains two separate domains, with mid- domain peaks in the relationship between species richness and land cover proportions occurring near the northwestern and southeastern corners of the study region. This condition causes an ecoregional edge effect that could be more dependent on the geometric constraints of the data than land cover or productivity relationships. 70 Figure 3.4. Geographic distribution of species richness and local slopes (Beta) and fits OK-squared) of the GWR model of the association between NDVIlgo and avian species richness in relationship to the proportion of deciduous forest (pDeck) and non-deciduous forest (pNonDeck) within the GWR kernel bandwidth surrounding sampled plots. Gray scales indicate interpolated values from a spline model using nearest neighbors. Black lines are ecoregion boundaries. See text for details. 71 .\ 10 o 10 20 Km Species Richness \ @1548 \ @12-14 \ 69 10-11 a) 8-9 0 4-7 R-squared pNonDeck 72 Discussion Using data at a relatively small grain, I detected both negative and positive relationships between NDVI and migratory passerine species richness. These results indicate that the slope of the relationship between NDVI and species richness at the grain of bird point count surveys was affected by the spatial context of sampling locations. NDVI values were positively influenced by the amount of deciduous forest cover within pixels and when averaged over the areas of the bird point count surveys. Both the strengths and slopes of the relationship between species richness and NDVI and the relationship between NDVI and deciduous forest cover varied across space at a lO-ha grain (180-m radius circle). Species richness was highest along transition zones of non- deciduous forest cover within ~20,000-ha (kernel bandwidth) areas surrounding the point count locations. The spatial locations of outlier plots and a shift from very negative to increasingly less negative and even positive local slopes in the regression of species richness on NDVI indicate that broader scale processes were affecting the finer scale slopes. Furthermore, the influential processes shift roughly along an ecoregion boundary to the north (outlier plots) and in the middle of the study region. These results provide empirical support for hierarchical levels of ecological factors (O'Neill et al. 1986) affecting species richness gradients across landscapes. Furthermore, it is possible to measure these influential relationships from space via differences in the intensities of reflected red and near infrared light. I propose that several mechanisms influence the slope of the relationship between NDVI and passerine species richness in my study region. If NDVI values are assumed a 73 consequence of productivity, my results indicated that a peaked or humped shaped relationship exists between NDVI and species richness in my study area. The relationship between productivity and species richness is often hypothesized to be hump shaped, with species richness first increasing along the productivity gradient and then decreasing as a consequence of increased competition among species (e. g., Grime 1973, Huston 1979, Rosenzweig 1992). However, my results indicate this relationship is detectable along productivity gradients measured at larger grains than the data were collected (~20,000 ha; Figure 3.4). The abrupt change in local slopes of the NDVI-species richness relationship in the center of my study region and the occurrence of outlier plots along my northern ecoregional boundary also suggest that edaphic factors such as climate and soil characteristics play at least an indirect role in the relationship between NDVI and species richness. Similarly, H-Acevedo and Currie (2003) found climate variables described bird diversity better than NDVI, and Hawkins et al. (2003) reported that annual actual evapotranspiration explains a large amount of variance in bird species richness over many geographical regions. These factors affect plant species composition as well as productivity, and therefore directly influence the proportional coverage of deciduous and coniferous forest while indirectly influencing NDVI. Other factors hypothesized to affect the NDVI-richness relationship in my study include small grain impacts of forestry practices over the extent of the study region. The purpose of these activities is to maximize merchantable productivity with a byproduct often being reduced compositional complexity. My study region is very productive and has some of the highest NDVI and species richness values in North America (Hurlbert and Haskell 2003). However, management activities focused on maximizing wood output 74 have converted what was once structurally diverse mixed forest into less diverse, even- aged deciduous stands and conifer plantations in many areas (Albert 1995). The resulting structurally and compositionally simpler landscape, although as highly productive as more complex landscapes, provides fewer nesting and foraging locations for many bird species (see Hobson and Bayne 2000). In addition, the canopies of deciduous leaves absorb much of the red light and reflect much of the infrared light, thereby increasing NDVI values and decreasing the slope of the relationship between NDVI and species richness. Only in the far western portion of my study region, which contains large unbroken stands of deciduous forest, was NDVI positively associated with species richness. Because managed forests are divided into stands of differing management activities and periods of entry, birds with rates of reproductive success sensitive to forest area (Blake and Karr 1984, Villard et al. 1995) or edges (Donovan et a1. 1995, Donovan et al. 1997, Rodewald and Yahner 2001) could be affected by regional proportional coverage of land cover classes (Trzcinski et al. 1999). These landscape relationships are known to vary over geographic regions (Hansen and Urban 1992, Smith et a1. 1998) and could cause the relationship between NDVI and species richness to shift towards positive local slopes within regions with more deciduous forest. Further study is needed to determine if these landscape relationships are responsible. Similar to larger grain/extent studies of birds (e.g., Hurlbert and Haskell 2003, Hawkins 2004, Hurlbert 2004), mammals (Oindo 2002) and plants (Gould 2000) I found a trend between NDVI and species richness. However, the trend I detected varied spatially and ranged from negative to positive. My results, in combination with the 75 current body of research on the NDVI-richness relationship, provide support for the following explanations I propose regarding scaling relationships between NDVI and species richness. At a large grain (e. g., 2,000,000-ha, Hurlbert and Haskell (2003)), studies are likely measuring the influence of coarse scale edaphic factors that limit the distribution of different plant species. Here, geometric constraints of the data (cf. J enz and Rahbek 2001) will affect results when species pools are aggregated among ecoregions and biomes (e. g., Hurlbert and Haskell 2003). At a medium grain (e.g., 200,000-ha, Hurlbert and Haskell (2003)), studies are likely measuring metacommunity dynamics (see Cottenie et al. 2003, Cottenie and De Meester 2004) within ecoregions and biomes where local aggregations of species from the regional pool are limited by available energy and competition. At a small grain (e.g., lO-ha, this study), vegetation structure and the means by which individual wildlife organisms use their environment affect the observed relationship between species richness and productivity. However, processes occurring at each grain do not covary and quantified relationships observed at one grain run the risk of being misinterpreted if taken out of context. Therefore, it is critical that future interpretations of NDVI as a predictor of species richness consider mechanisms other than “productivity” as a generic construct. Satellite imagery shows strong potential for identifying temporally and spatially explicit factors affecting species distribution patterns. Consequently, spectrally detectable and functional measures of landscape composition can be integrated with satellite imagery at different resolutions to investigate broad scale processes affecting patterns of species richness. The results of these analyses could then be used to guide management activities, which in turn could be monitored remotely to evaluate their effectiveness. 76 CHAPTER 4 USING THE SPECTRAL AND SPATIAL PRECISION OF SATELLITE IMAGERY TO PREDICT WILDLIFE OCCURRENCE PATTERNS Introduction One of the greatest impediments to accurate mapping of the Earth’s resources is a paucity of spatially referenced information (Franklin 2001). Remote sensing technologies could improve this situation for objects detectable at the spectral frequencies and grain, or smallest spatial sampling unit, of the sensor. In the case of wildlife habitat studies, the Landsat Thematic Mapper series of sensors are often used due to their relatively low price, short repeat-time (16 days), spatial resolution (approximately 30-m x 30-m), and spectral resolution capable of detecting differences in vegetation (Lillesand and Kiefer 1999). Landsat imagery is often used to map wildlife distribution patterns indirectly. This is accomplished by first classifying images into land cover categories and then reclassifying land cover categories for wildlife occurrence by using known vegetation affinities of each species (e.g., Scott et al. 1993, Morrison et al. 1998). In order to characterize land cover, however, a classification scheme must be used to instruct image processing software how to aggregate pixels of remotely sensed imagery into discrete categories. Land cover classification schemes divide continuous ecological gradients (Whittaker 1956, Austin 1985) into compositionally distinct features to achieve an objective (Zube 1987, Foody 1999, Townsend 2000). While this allows large regions to 77 be divided in ways that humans can understand, wildlife often perceive and respond to landscape heterogeneity in substantially different ways (Johnson et al. 1992, Tang and Gustafson 1997). For example, rather than choosing habitat based on a particular forest type or age class, some bird species select for understory vegetation or forest structure (Probst et al. 1992, Stouffer and Bierregaard 1995, Anders et al. 1998). Thus, most land cover classifications have limited utility for predicting patterns of these species’ occurrences. Similarly, when classification schemes do not include key habitat features under selection by target species, investigations into the dependence of wildlife species’ occurrences on land cover classes are not likely to identify causal mechanisms underlying the observed distributions (Wiens et al. 2002). Thus, the disparity between human and wildlife perceptions of and responses to landscape heterogeneity can add substantial error into statistical analyses and management prescriptions. In attempts to minimize predictive errors potentially fostered by inappropriate or inaccurate land cover maps, a logical approach to extrapolating patterns of wildlife distributions across large areas is to directly classify imagery for an individual species’ occurrence using raw spectral reflectance data. Using this approach, spectral characteristics of locations where species are known to occur are employed when extrapolating their distributions. The underlying assumption is that species’ occurrences can be predicted by spectrally detectable components of their habitat. A benefit to this approach is that distribution maps are classified using all occurrence locations separately yet simultaneously, therefore no global model relating spectral variables to occurrence sites need be assumed. 78 This direct approach to mapping wildlife distributions is growing in popularity (e. g., Hepinstall and Sader 1997, Conner 2002, Laurent et al. 2002, Jenkins et al. 2003a). While such studies indicate the potential for bypassing a land cover map, they have not yet provided a strong methodological test of raw satellite imagery’s capability to map species’ distributions. Some of the choices made by the authors during classification were arbitrary and no framework was established for investigating the influence of these choices, or the range of possible choices on prediction accuracy. Several decisions must be made during image analyses in order to extrapolate the results of presence/absence surveys across a landscape using spectral information. Some options include the type of imagery, the classification scheme, the method of classification, the choice of pixels used to represent classes, and the parameter values (both spatial and spectral) used to classify maps. While these choices could be described as the “art” of image processing, it is also possible to place them within a hypothesis- testing framework to quantify their effects on prediction accuracy. To investigate the influence different choices in classification options have on predicting wildlife occurrence maps, I have created a software program (PHASEl) as the first phase in the development of a Habitat Analysis By Iterative CLASSification procedure (HABICLASS). The purpose of HABICLASS is to minimize anthropocentric classification bias when predicting wildlife distribution patterns (first phase) and to determine causes of the predicted patterns (second phases). PHASE] was created to make the prediction phase methodologically rigorous through the use of statistical models. This is accomplished by creating maps of species’ occurrence patterns, iteratively modifying classification methods used to create the maps, and evaluating divergence in the 79 predictive accuracy of maps created in different ways. The goal therefore is to identify better means of improving prediction accuracy within a strong inference framework. Here I use PHASE] of HABICLASS to predict forest bird occurrence with the spatial precision of Landsat imagery. Although these analyses are conducted using spectral data, PHASE] may be used in combination with any raster dataset. PHASE] is superficially similar to other image classification software. It is used to create a spectral signature for each field plot describing grid values at that location. These signatures are then used to identify spectrally similar pixels and classify them into categories to create a thematic map. PHASE] differs from many other image classification software through its application of cross-validation to stratify and repeatedly sample random subsets of signatures. As a result, a group of maps, rather than a single map, are created from a set of reference plots. The accuracy of the group of maps can then be summarized to describe a statistical distribution of accuracy for comparisons with maps made through other methods. However, reference data for this approach need to be spatially compatible to deny the introduction of scale or aggregation effects (Dutilleul and Legendre 1993, Dutilleul 1998) that could influence class descriptions. Reference data used in signature creation should therefore describe the same size area, or grain, for proper use within the PHASE] classification framework. To ensure grain equivalency within datasets used in PHASE] , I developed a Grain Representative Assessment and IN ventory protocol (GRAIN). GRAIN permits a per-pixel extrapolation of spatially referenced survey data across large areas with the spatial and spectral precision of satellite imagery. For bird point count surveys, GRAIN requires the surveyor’s location be georeferenced and the distance between the surveyor 80 and each bird recorded. The georeferenced position is employed to identify the pixel used for signature development in PHASE]. Species presence and absence at those locations define classes used in supervised image classification. However, proximity thresholds (hereafter referred to as detection distances) can be set to exclude species detected beyond specified distances from observers, effectively converting a presence location into an absence. In addition to pixels overlapping georeferenced locations, signatures may also contain values for pixels surrounding the georeferenced position if they meet specified spatial and spectral criteria. Such decisions made during image classification can be iteratively modified, allowing the analyst to repeatedly tune the classification parameters and create a series of maps for investigating the influence of these choices on map accuracies. The goal of this study was to identify methods of image classification that could predict the regional occurrence of Neotropical migrant warblers with the spectral and spatial precision of Landsat 7 ETM+ imagery and independent from any other system for categorizing the landscape (e. g., land cover maps). Furthermore, I was interested in whether this method could obtain results comparable to Gap Analysis (Scott et a1. 1993, Donovan et a1. 2004). Gap Analysis is used to classify maps of species’ fundamental niches (Hepinstall et al. 2002, Gap maps), or areas of potential but not necessarily occupied habitat, using known species-habitat associations, land cover data and expert knowledge within a geographic information system. Land cover maps provide the foundation for most predictions within the Gap framework. Thus, comparisons of spectrally derived species’ occurrence maps with Gap maps provide a useful and 81 contemporary reference for how well Landsat imagery can be used to directly predict species’ occurrences. My specific questions were 1) do individual pixels of Landsat 7 ETM+ imagery contain some of the information needed to predict the occurrence of forest bird species, especially those that select for understory conditions during the breeding season? If the answer to question (1) is yes, then 2) can I identify the influence of detection distances for categorizing species occurrence and alternative classification options on prediction accuracy, and 3) how does the accuracy of spectrally derived maps compare to those created from land cover maps using the methods of Gap Analysis? I used the GRAIN protocol to collect ground reference data of bird occurrence over a large forested region. PHASE] was employed to automate signature creation describing surveyed plots and classify maps predicting species occurrence. I also examined the influence of varying two spatially explicit classification parameters on prediction accuracy: 1) the window size used to average spectral values in signature creation, and 2) the threshold distance required for bird detections to be counted as present. The accuracy of maps predicting species’ occurrences was validated with ground data not used during classification and compared with the accuracy of recent Gap maps. Accuracy was assessed using two common measures, and the information content of these measures was compared. 82 Methods Study Region The study region is located in Michigan’s Upper Peninsula, USA and includes parts of five counties (Baraga, Dickinson, Iron, Menominee, and Marquette; Figure 4.1A). This region covers ~400,000 ha or approximately 11% of a single Landsat 7 ETM+ scene (Path 24 Row 28). A majority of the land is owned by state government and industry with management primarily focused on wood production. The study region is an ecologically diverse landscape characterized by a spatial mosaic of forest stands that include upland hardwoods (sugar maple (Acer saccharum), quaking aspen (Populus tremuloides), yellow birch (Betula alleghaniensis), basswood (T ilia americana)), lowland hardwoods (black ash (F raxinus nigra), red maple (Acer rubrum)), lowland conifers (northern white cedar (Thuja occidentalis), black and white spruce (Picea mariana and P. glauca), tarnarack (Larix laricina» and upland conifers (European larch (Larix deciduosa), eastern hemlock (T suga canadensis), red and jack pine (Pinus resinosa and P. banksiana)). Besides the influence of glacial topography and other edaphic factors, the composition and structure of canopy and understory vegetation has high spatial variability due to differences in forest management practices and gradients of deer browsing pressure (Albert 1995, Van Deelen et al. 1996). 83 Figure 4.1. The study region lies within a single Landsat 7 ETM+ scene and contains parts of five counties in Michigan’s Upper Peninsula (A). Within the study region (B), randomly selected landscape units (C; township sections or USGS QQQs with coarse land cover classes within landscape units shown as shades of gray) define areas for the selection of specific plots (D; overview showing different tree species) for bird occurrence surveys. Plots were sampled within a 30-m radius from plot centers (D; large circle), so that a single Landsat 7 ETM+ pixel (D; square) falling within this area could be precisely characterized. 84 85 Study Species I chose three bird species for analysis (Latin name and American Ornithologists’ Union four-letter code): Black-throated Green Warbler (Dendroica virens; BTNW), Nashville Warbler (Vermivora ruficapilla; NAWA), and Ovenbird (Seiurus aurocapillus; OVEN). All three of my study species commonly breed in deciduous and mixed forests of northern North America and select territories on the basis of subcanopy composition and structure in addition to dominant overstory components (Morse 1976, Collins 1981, Collins et al. 1982, Dunn and Garrett 1997). These birds perceive and respond to landscapes at a small spatial grain for many functional reasons including those related to niche partitioning and territory delineation (MacArthur and Pianka 1966, Schoener 1968, Robinson and Holmes 1982). Predicting the occurrences of these warblers therefore provides a strong test for assessing the information embedded in Landsat 7 ETM+ imagery. Although the study species are included in the same family and have all shown a general affinity for breeding in forested areas, they have markedly different habitat associations. BTNW nest in a wide variety of mature coniferous, deciduous and mixed forests (Morse 197 6, Brewer et al. 1991, Dunn and Garrett 1997). Most foraging and nesting take place in the midlevels of vegetation, therefore requiring a multi-storied layering of vegetation, often with a shrub or coniferous understory component (Collins 1983, Norton 1999). NAWA nest within dense ground cover in a variety of wet and dry open woodlands (Brewer et al. 1991; Dunn and Garrett 1997) and often near ecotones (Williams 1996). In drier areas, they commonly select early successional forests that arise following fires or deforestation. OVEN nest on the ground and often in large stands of 86 mature deciduous and mixed forests (Zach and Falls 1979, Smith and Shugart 1987, Brewer et al. 1991). Dry upland areas are most commonly used but they sometimes are found in lowland forests and swamps. In all OVEN breeding habitat, however, leaf litter is essential for foraging and nest construction (Van Horne and Donovan 1994). Males of all three species are territorial and vociferous during the breeding period and have distinctive songs that make them easy to distinguish fi‘om other species in the study area. In addition, they have characteristic plumages that can be identified by trained observers. The GRAIN Protocol for Field Sampling The GRAIN protocol was used to select and characterize locations for BTNW, NAWA and OVEN occurrence surveys. GRAIN uses the random sampling method described by Lillesand et al. (1998) for the Upper Midwest Gap project but modified to incorporate the field data collection approach of Liu et al. (2001). In 2001, random United States Geological Survey (U SGS) quarter quarter quads (QQQs; n = 28) served as landscape sampling units (LUs). In 2002 and 2003, random township sections (n = 36 and 32, respectively) served as LUs (Figure 4.1B). This switch between roughly equal area QQQs and township sections (1000 ha vs. 1037 ha, respectively) was made in order to allow my data to be more easily integrated with public and private databases such as Michigan’s atlas of breeding birds (Brewer et al. 1991). Within each LU, between 2 and 8 plots (Figure 4.1C) were selected for bird surveys for a total of 433 plots surveyed over the course of the study. These plots were a minimum of 90-m apart with the mean minimum distance of 240-m (:194-m standard deviation, strong right skew) from the next closest plot. Survey plots encompassed a 30- 87 4‘ m radius area so that a single pixel of Landsat 7 ETM+ imagery sensed over plot centers 3 could be precisely characterized by data collected within this area (Figure 4.1D). The specific plot selection criterion was that a hypothetical 30-m x 30-m square could be placed anywhere within the plot and perceived by the field crews as having the same vegetation structure and composition as a similar square placed anywhere else within the plot. Plot selection in general was dependent on two criteria: 1) permission to access the property and 2) maximizing the biotic and abiotic differences among plots to sample the spatial and spectral heterogeneity of the study region. Preference was given to at least one northern hardwood plot in each LU to satisfy collaborative research priorities if northern hardwood plots could be located in the field. Locations of plot centers were georeferenced using a minimum of 80 global positioning system (GPS) point locations collected at 5-second intervals with a Trimble GeoExplorer 3 (Trimble Navigation Ltd.) and later differentially corrected to a precision of : 5-m using Coast Guard base station data. Site centers were flagged, as were 30-m distances in the 4 cardinal directions (sensu Huff et al. 2000). Accuracy of plot center locations was determined to be within the precision of Garmin III+ GPS receivers (15-m; Garmin Corporation) used to navigate to these locations during repeated point count surveys. There were several reasons why I used a 30—m radius circle for selecting survey locations. First, many of my ground surveys were conducted prior to image acquisition so the exact pixel location was unknown (Figure 4.2A). Also, I used multiple season imagery for analysis (see Image processing) and pixels fi'om different images do not always overlap (Figure 4.28). Third, the ground information attributed to a pixel is 88 Nadir Off- Nadir > Figure 4.2. Factors affecting the ground instantaneous field of view captured by pixels in Landsat imagery. Gray circles represent 30-m radius field plots. Each grid represents the ground information contained within a nine-pixel window. The hashed center cell represents the ground area used to describe the sampling plot. Factors affecting the location and area covered by the center pixel include A) unknown pixel location, B) non- overlapping pixels of multiple images, C) effects of Earth rotation, and D) inclusion of increasing ground coverage within pixels increasingly farther from nadir. 89 actually a parallelogram that is not square due to the rotation of the Earth as the Landsat 7 satellite travels in a circumpolar orbit (Figure 4.2C). Fourth, the Landsat satellites have sensors that sweep back-and-forth past a central point called nadir. The ground instantaneous field of view attributed to pixels farther from nadir is therefore larger than that of pixels closer to nadir (Figure 4.2D). Although geometric correction during level 1G processing of Landsat 7 imagery makes adjustments for the last two factors, the resulting image values are still dependent on the geometric limitations of the sensor and the influence of reflective features adjacent to the ground area contained in the pixel (Cracknell 1998). Finally, circular plots are easier to inventory for birds than parallelograms. In addition to detections within the 30-m sampling radius, species were also recorded if detected within larger radii circles surrounding plot centers. These radii included 50-m and unlimited distance during all three years. During 2002 and 2003, an additional detection distance of lOO-m was used. These larger spatial thresholds are commonly employed during point count surveys for birds (e. g., Ralph et a1. 1993, Ralph et al. 1995). Bird detectability was assumed to be consistent among plots and all detections were assumed to be made within approximately l80-m from plot centers (see Wolf et al. 1995). LUs were divided into 7 (2001), 6 (2002), and 12 (2003) groups for daily bird surveys to minimize travel time (not necessarily minimizing distances) among LUs surveyed during any day. The surveyed order of LU groups, plots among LUs, and survey timing was randomized, as was the selection of observer for each survey. Beginning at sunrise, observers conducted surveys within a 5-hour period when weather 90 conditions did not preclude birds from singing (Ralph et al. 1993; Ralph et al. 1995). Observers had been trained to detect the three species in this study and other species in the region by song and sight using tapes, field guides and practice prior to data collection. Bird species were counted as present if detected by song, call or sight within each detection distance over an 11 minute period. Timing of surveys began immediately when observers reached the center of the plot. After 10 minutes an additional 1 minute was spent walking around the plot center to flush elusive individuals within 30-m. The distance of flushed individuals detected < 180-m from plot centers by surveyors walking in or out of the plots was also recorded. Each plot was visited three times during the breeding period for these species between June 4 and July 3 in either 200], 2002, or 2003. At least two observers surveyed each site to account for possible differences in observer’s physical abilities to detect the study species (Ramsey and Scott 198]). A total of 8 observers collected data over the 3 years of this study. In all, 433 plots (n = 112, 86 and 235 for years 2001, 2002 and 2003, respectively) were surveyed for bird species’ occurrences 3 times during the breeding period (Table 4.1). Survey plots were placed in a wide variety of land cover classes (see Space Imaging, Inc. 2001 for land cover descriptions) including aspen associations (n = 85), herbaceous openland (n = 10), lowland coniferous forest (n = 35), lowland deciduous forest (n = 14), lowland shrub (n = 2), mixed upland conifers (n = 4), mixed upland deciduous (n = 6), northern hardwood associations (n = 190), pines (n = 23), other upland conifers (n = 15), upland mixed forest (n = 36), and upland shrub / low-density trees (11 = 10). Within these land cover classes, surveys were conducted over a wide variety of size classes (non-forest, sapling, pole, mature), management histories (clear cuts, selective 91 cuts, no cuts) and ranges of vertical structure (homogeneous to diverse). Survey plots were placed at varying distances from hard and soft vegetation edges as long as specific plot selection criteria were met (see Figure 4.1 D). Table 4.1. The number of plots where species were observed within each detection distance. All 433 plots were surveyed using 30-m, SO-m, and 180-m detection distances and 321 plots were surveyed using the lOO-m detection distance. Detection Distance 30-m SO-m 1 OO-m 1 80-m Species n = 433 n = 433 n = 321 n = 433 Black-throated Green Warbler 44 175 196 271 Nashville Warbler 88 208 208 27] Ovenbird l 50 309 297 409 Image Processing I used Normalized Difference Vegetation Index values (NDVI, Jensen 2000) derived fiom bands 3 and 4) and short-wave infrared values (SWIR, band 5) from multiple seasons of Landsat 7 ETM+ imagery to create spectral signatures describing survey plots and to create maps predicting bird species’ occurrences. Multiple season images have been found useful in distinguishing among forest classes with differing phenological leaf canopy trajectories (Wolter et al. 1995, Mickelson Jr. et al. 1998). NDVI is often used for vegetation classification because plants reflect or absorb different 92 amounts of red (band 3) and near-infrared (band 4) light depending on biophysical factors such as chlorophyll content and leaf area (Jensen 2000). The longer SWIR wavelengths are also functional for vegetation discrimination because they are influenced by leaf moisture content and canopy cover (Asner and Lobe112000, Ceccato et al. 2001) as well as coniferous timber volume (Garnmel 1995). Multiple seasons of NDVI and SWIR values can therefore be useful in distinguishing among differences in overstory composition. Leaf-off images can also be helpful in separating deciduous woodlands by non-deciduous components under the canopy (Mickelson Jr. et a1. 1998). Specific subcanopy factors of deciduous and mixed forests that may be influential to the warbler species in this study include leaf litter, coarse woody debris, and the presence of balsam firs, among others. Two Landsat 7 ETM+ level 1G images of Path 24 Row 28 were obtained from the USGS. Because the study area is very cloudy due to the influence of Lake Superior and Lake Michigan, I was limited in my choices for contemporary, cloud-free imagery. An April 27, 2001 image was selected to describe early spring leaf-off conditions just after snowmelt. Photosynthesizing species during this time include conifers, grasses and early spring ephemeral herbs and forbs (personal observation). A May 29, 2001 image was chosen to represent early leaf-on conditions. During late May in Michigan’s Upper Peninsula, most deciduous species have small, young leaves (personal observation) which likely allow some vegetation under the canopy to contribute to reflectance values. All image processing was conducted using Imagine 8.7 software (Leica Geosystems GIS & Mapping LLC). The two images were converted to at-sensor reflectance using the import utility. At-sensor reflectance was converted to surface 93 reflectance through a dark object subtraction (sensu McDonald et al. 1998). Each image was georectified (RMSE < 8m) using 200 road intersections within and around the study area. A nearest neighbor transformation was employed during georectification to maintain the information content within pixels. Road intersections were identified using a digital road map obtained from the state of Michigan’s Center for Geographical Information (http://www.mcgi.state.mi.us/mng). This road map is a level 3b product, originally created from Census Bureau TIGER line files and most recently repositioned using USGS 1:12,000 Digital Ortho Quarter Quad aerial photography. Visual inspection of the resulting images revealed overlap of the two images and 50 ground control road intersections independently georeferenced in the study region, hence indicating accurate georectification. Two to four corners of the ground control road intersections were georeferenced using a minimum of 100 GPS point locations collected at 5-second intervals with a Trimble GeoExplorer 3 (Trimble Navigation Ltd.) and later differentially corrected to a precision of 1 2m using Coast Guard base station data. A lack of canopy at these intersections permitted greater spatial precision than could be obtained at many of the forested plots where bird surveys were conducted. After georectification, the spectral data were relativized (see McCune and Grace 2002). Relativization of spectral values is necessary when using simple distance measures in spectral space during classification. In this way, the covariance structure among the bands is maintained but the values within bands are modified so that equal weight can be given to a similar increment of difference in any band (see Image Classification). For these analyses, I rescaled the image bands to unsigned 8-bit integer values. This conversion type is an artifact of an earlier PHASEl-type analysis conducted with scripts 94 written for proprietary software. The PHASE] software can use long integers and floating point values, however the use of integers substantially reduces computer processing time. The NDVI values were calculated from surface reflectance values of the April and May images. Surface reflectance SWIR values of these images were range shifted to have minimum values of 0. For each image, SWIR values and NDVI values greater than zero were multiplied by 600. The multiplication constant of 600 was chosen because it allowed me to take advantage of the full 256-value range of an 8-bit integer. The 4 grid layers (2 NDVI and 2 SWIR) were exported as ASCH files for use in PHASE]. Validation Data Before each classification procedure, a subset of survey plots was randomly selected to evaluate the accuracy of derived occurrence maps (validation data). Detections of a species within a specified proximity threshold from plot center were stratified into presence and absence (e.g., ovenbird detections within 30-m). One-third of the plots from each of these strata were randomly selected for validation. The remaining plots were used to classify maps of species occurrence within that detection distance using cross-validation (classification data). The detection distance was then iteratively modified, all the occurrence data were re-stratified, and the process was repeated for each species and detection distance. 95 Signature Creation for Multiple Window Sizes Signature creation, as well as image classification (see Image Classification), cross-validation (see Cross-validation) and majority map generation (see Majority map validation) were automated by means of the PHASE] program created with the C++ programming language. Points indicating plot centers were overlaid with the processed imagery to identify focal pixels. A window size was specified around each focal pixel, and within that window, all pixel values were averaged for each grid layer (2 NDVI and 2 SWIR). These 4 values served as a multivariate spectral signature describing each plot. Four sets of signatures were created using 4 different window sizes. They included 0 (per-pixel classification), 1 (3x3 square of 9 pixels), 2 (2 pixel radius containing 13 pixels), and 3 (3 pixel radius containing 29 pixels). Image classification The following classification procedure was followed for each window size, bird species, and detection distance. The set of signatures was stratified by species presence and absence. Two-thirds of the signatures within each species occurrence stratum were randomly selected. The signatures were then used to classify all the pixels in the image based on their minimum spectral Euclidean distance to sampled plots. Spectral Euclidean distance is a relatively simple, non-parametric method of image classification. Images were classified using the following discriminate function (Richards and J ia 1999): min[d(x,sl.)2]=(x—si)’(x—sl.) [1] 96 where 5,, i=1,. . .M are the signature values, x is a 4 x 1 vector of values of the pixel to be classified, and t indicates that the vector is transposed. In other words, each of the 4 grid values for any given pixel was evaluated for its distance to the respective grid values described in the signatures. The squared distances for each grid were summed, and the signature with the smallest square root summed distance to the pixel was used to classify that pixel. In cases where two or more signatures had equal distances from the pixel in question, a random signature from among the alternatives was assigned. Cross-validation PHASE] implements a cross-validation procedure to quantify the statistical distribution of accuracy for the maps it generates in each combination of window size, species and detection distance. I generated eleven maps during each combination. Each map was generated from two-thirds of the signatures randomly selected fiom both occurrence strata. The remaining one third of the classification data were used to test the accuracy of the map. Accuracy statistics of proportion correctly classified (PCC) and Kappa were calculated for each of the 11 maps. The distribution moments from the 11 instances for each combination were then calculated. Like PCC, the Kappa coefficient (Cohen 1960) generally ranges from 1 for perfect agreement to 0 for no agreement (negative values indicating less than chance agreement are dependent on marginal distributions (Rosenfield and Fitzpatrick-Lins 1986). Unlike PCC, Kappa removes chance agreement from consideration in accuracy assessment. Kappa uses all the cells in the error matrix and therefore includes a measure of overall thematic classification accuracy as well as omission and commission error for 97 each class. While other accuracy measures such as sensitivity and specificity (e. g., Hepinstall et a]. 2002) can be calculated from the error matrix, for simplicity I compare only FCC and Kappa. Majority Map Validation A majority classification was implemented in PHASE] to summarize the 11 maps created during cross-validation. The majority classification created a new map whereby each pixel in the majority map was labeled using the most common occurrence class for that pixel from the 11 maps created during the cross-validation procedure. Accuracy statistics for this map were assessed using the 1/3 of plots reserved for validation (see Validation data). Because the validation data were randomly selected for each species and detection distance but not window size used in signature development, differences in classification accuracy within species and detection distances were solely due to differences in window sizes. Gap Analysis The Michigan Gap Analysis Program (MI-GAP) was recently instituted and initial maps predicting the occurrence of bird species across the state were released in 2004 (Donovan et al. 2004). These maps were developed based on known habitat associations (e. g., Brewer et al. 1991) and expert opinion (J. Skillen, personal communication). Using a recently derived land cover map (Space Imaging, Inc. 2001), land cover classes considered to be habitat for each species (Table 4.2) were classified as presence while all others were classified as absence. The spatial precision of Gap maps was presented to the 98 public at a 90-m resolution. The maps, however, represented resampled versions of the original 30-m resolution Landsat imagery-derived maps, which I obtained and used in this study. Table 4.2. Land cover classes used to classify species presence in Gap analysis. Black-throated Green Warbler Nashville Warbler Ovenbird N. Hardwood Assoc. N. Hardwood Assoc. N. Hardwood Assoc. Mixed Upland Deciduous Oak Assoc Oak Assoc Pines Aspen Assoc Aspen Assoc Other Upland Conifers Mixed Upland Deciduous Mixed Upland Deciduous Mixed Upland Conifers Pines Pines Upland Mixed Forest Other Upland Conifers Upland Mixed Forest Mixed Upland Conifers Upland Mixed Forest Lowland Deciduous Forest Lowland Coniferous Forest Lowland Mixed Forest Lowland Shrub The accuracies of Gap maps were assessed in several ways using the entire set of field data. Pixels overlapping plot centers were assessed for their ability to predict species’ occurrences using PCC and Kappa for each detection distance. For example, a 30-m x 30-m pixel overlapping a survey plot and classified by MI-GAP for Ovenbird occurrence (using known land cover associations) was assessed for its accuracy in predicting whether Ovenbirds were present or absent within 30-m, SO-m, lOO-m, and 99 180-m. I also assessed whether pixels in regions surrounding the plot centers accurately predicted species’ occurrences using the other window sizes employed during signature creation (see Image classification). If the Gap maps predicted species’ presence within any pixels contained within each window size then the plot was labeled present, otherwise, it was labeled as absent. The accuracies of MI-GAP predictions were thus compared with field observations over all combinations of detection distances and window sizes. Results Information Content of Imagery Presence and absence were accurately predicted better than chance using unclassified imagery for all species within most detection distances (Figure 4.3A). However, there were large differences between accuracy statistics. Comparisons between PCC and Kappa showed that PCC was greatly influenced by the proportion of plots where the species were detected, independent of which species, detection distance, or window size were used in map classification (Figure 4.3A). For example, in situations where a species was rarely detected (e. g., BTNW within 30-m radius; Table 4.1), classification of the entire landscape as absent for that species would yield high PCC values. Alternatively, in situations where a species was nearly ubiquitous (e. g., OVEN within 180-m; Table 4.1), classification of the entire landscape as presence would yield high PCC values. Such classifications, however, do not yield predictions better than could be made with a random assignment of pixels to classes. The Kappa coefficient accounts for this problem 100 Figure 4.3. Relationships between the proportion of plots where species were detected and accuracy measures of proportion correctly classified (PCC; open symbols) and Kappa (filled symbols) values for (A) cross-validation, (B) validation, and (C) Gap analysis. Each symbol represents the results of a single analysis for a given species (Black-throated Green Warbler = triangle; Nashville Warbler = square; Ovenbird = circle), detection distances and window size. First-order polynomial model fit is provided only as a relative measure of dependence as each point represents accuracy statistics fi'om random subsets of the same pool of plots classified in different ways. 10] Accuracy Accuracy Accuracy 1“ Cross Validation 0.8 ~ 0.6 _ 2 PCC R = 0.88 0.4 -' A . A 9 , g Kappa R‘ = 0.34 i I A 0 . 0.2 «E ‘ 1 :3 a a an 3 A GD 8 no ° 0 A % I l l r o—fi 0 0.2 0 4 0.6 0.8 1 -02 3 1 “g Validation 0.6 . PCC R2= 0.93 Kappa R2 = 0.15 0 0 0.2 0 4 0.6 -0.2 9 1 l GAP 0.8 ~ PCC R2 = 0.94 0.6 < 0.4 Kappa R2 = 0.14 o o 0.2 ~ 2 9 o A o ‘ A 0 i A o r E .3 - J... .. n_o._ a 0 0.2 0.4 0.8 1 -0.2 J Proportion Present 102 by including row and column totals of the error matrix in its calculation and therefore provided a measure of prediction accuracy that was less dependent on the proportion of plots where species were detected (as indicated by lower R2 values in Figure 4.3). The accuracy of maps differed among species. However, both accuracy measures were most affected by the detection distance used to separate sites into species occurrence strata. For BTNW, Kappa increased and FCC decreased with detection distance (Figure 4.4A). Maps classified for NAWA occurrence show a similar, although less distinct relationship between the two accuracy measures (Figure 4.4B). OVEN occurrence maps on the other hand, had a unimodal relationship between the Kappa coefficient and detection distance (Figure 4.4C). PCC values for these maps increased with detection distance. Window sizes used in signature creation had less of an effect than detection distances on Kappa values of BTNW and OVEN maps. For example, the difference between mean Kappa values for BTNW predictions at detection distances of 30—m and 180-m was larger than the range of mean Kappa values for any window size within those detection distances (Figure 4.4A). However, window size appears to have a larger effect on Kappa values within detection distances where Kappa was highest. For example, highest mean Kappa values were obtained for BTNW at a detection distance of lOO-m but also showed a greater range of variability over window sizes within this detection distance (Figure 4.4A). Similar results were obtained for OVEN within a detection distance of SO-m (Figure 4.4C). Kappa values for NAWA were affected primarily by window size (Figure 4.4B). 103 Figure 4.4. Proportion correctly classified (gray squares) and Kappa values (black diamonds) for maps predicting the occurrence of A) Black-throated Green Warbler, B) Nashville Warbler, and C) Ovenbird using different detection distances and window sizes. Filled symbols represent mean values of accuracy statistics for 11 cross-validation iterations with error bars indicating minimum and maximum values. Open symbols represent accuracy statistics of a majority map summarizing the cross-validation runs and tested with a subset of data reserved for validation. 104 BTNW 1‘ 30m 50m 100m 180m ; {I r 0.8} le T >0.0i {Ml fill illl ’0? iii l ll lill ' 00 w it i -0.2 gr“ .1. m 0123 0123 0123 0123 NAWA 30m 50m 100m 180m 1'? 0.81i Ml §°'“ lili- iili fl“ 3 0.4-1 o o o ‘ 11 ill °°l Ill 01 :1 ¥ ° il l i 0123 0123 0123 0123 OVEN 1. 30m 50m 100m 180m {Eff iii! 0.8-- . li . . l l 306 fill { 4 g 0.4l ¥ 0 . < l i l W lll °i lél lo 0 0123 0123 0123 0123 Detection Distance Window Size 105 In summary, differences in Kappa values allowed insight into factors affecting the predictive accuracy of species occurrence maps. While these results had a weak dependence on the proportion of presence or absence plots used as signatures during classification, both the detection distance used to separate plots among occurrence strata and the window size used to create spectral signatures caused most of the variation in Kappa values. However, the relative influence of these factors was species dependent. PCC on the other hand, was almost completely dependent on the proportion of presence or absence plots used as signatures. PCC therefore only summarized the relative number of plots per signature category and provided little information about the influence of detection distance or window size on classification accuracy. Majority Map Validation Majority maps accurately predicted species’ occurrences from validation data better than chance for most detection distances and window sizes (Figure 4.3B). Similar to the maps created via cross-validation, PCC provided little information. In fact, majority classification increased the dependence of PCC and reduced the dependence of Kappa on the distribution of plots among occurrence strata (increased and decreased R2 values, respectively; Figure 4.3B). Majority maps classified validation data better than any of the cross-validation runs for some detection distances and window sizes (Figure 4.4). Like the cross- validation maps, detection distance had a strong effect on Kappa values of majority maps. 106 However, window size had a stronger influence on Kappa than detection distance for NAWA (Figure 4.48). Gap Maps Gap maps yielded a linear relationship between PCC and the proportion of plots where species were detected regardless of species, detection distance or window size (Figure 4.3C). As the proportion of detections increased within the dataset (e. g., via larger detection distances), PCC increased. Kappa values for Gap maps, like those for spectrally derived species’ occurrence maps, were less dependent on the proportion of plots where species were detected and therefore provided a better measure than PCC of how detection distance and window size affected prediction accuracy. Detection distance (Figure 4.5 symbol sizes) had little effect on Gap map Kappa values. Most of the variation in these values was due to the window size used to assess prediction accuracy (Figure 4.5 error bars). Only the SO-m detection data of OVEN (Figure 4.5C) were comparable to maps created using PHASE]. However, window size also had a very strong impact on these values. Spectrally derived maps yielded larger Kappa values than Gap maps in most situations when comparing the ranges of accuracy statistics between methods of predicting species’ occurrences (Figures 4.3 & 4.5). For all three species, highest Kappa values were obtained using spectral associations. In particular, NAWA occurrence was predicted better than chance only by using spectral associations (Figure 4.5B). OVEN, on the other hand, was predicted better from spectral associations only when using moderate detection distances (Figure 4.5C). 107 Figure 4.5. Distribution of accuracy statistics (proportion correctly classified and Kappa) for majority maps created from the 11 cross-validation runs and tested with a subset of data reserved for validation (filled diamonds) and for Gap analysis maps (open squares) tested with data from all 433 plots. Statistics are shown for A) Black-throated Green Warbler, B) Nashville Warbler, and C) Ovenbird. Larger symbols represent larger detection distances. Symbol location indicates mean values for all window sizes used to classify maps for each detection distance, with error bars indicating minimum and maximum values. 108 Kappa 0.4 - 0.3 ~ 0.2 i 0.1 r -0.1 BTNW l i-EEH 0.4 - 0.3 4 0.2 a 0.1 r NAWA -0.1 0.4 - 0.3 - 0.2 ~ 0.1 - -0.1 OVEN 0.2 0.4 PCC 109 0.6 0.8 Discussion Using a relatively simple, non-parametric method of image classification, I predicted the regional occurrences of three warbler species with the spatial and spectral precision of Landsat 7 ETM+ imagery and independent of a land cover map. Maps were validated for all three species with Kappa values >03 and PCC >0.6. Furthermore, spectral information was used to predict the occurrence of these species that use forest subcanopy components, with Kappa values 0.1 to 0.3 higher than achieved by the Michigan Gap program. Bypassing subjective land cover categorization thus avoided land cover classification errors and exploited a large range of information provided in six bands of imagery. For example, all three species are known to use areas of mixed pine species in Michigan (Brewer et al. 1991). Mixed pines, however, are included in the same land cover class as pine plantations. In the study region, none of the study species were found in pine plantations. Thus, Gap maps would err on the side of commission for these species in pine plantations. This explicit intent of Gap analysis to err on the side of commission for purposes of mapping “potential” habitat (Hepinstall et al. 2002) also likely contributed to the linear relationship between PCC and the proportion of plots where species were detected (Figure 4.3C). If subtle spectral differences among different combinations of pine in the study region are detectable within pixels of Landsat 7 ETM+ imagery, then occurrence maps created using spectral associations would not systematically incur this error. Although classification methods differed between Gap maps and direct use of spectral associations, similar factors likely affected their predictions. Misclassification of 110 both types of maps could have occurred via the influence of spectrally undetectable components of habitat on prediction accuracy. These components include vegetation structure (Collins et al. 1982, Collins 1983, Smith and Shugart 1987, Probst et al. 1992) and landscape factors such as edge effects, patch sizes and patch isolation (Kotliar and Wiens 1990, Dijak and Thompson 2000, Sisk and Haddad 2002). Future uses of PHASE] and efforts by MI-GAP need to integrate additional data types into predictive models to account for these habitat associations and improve prediction accuracy. For example, additional data sources such as digital elevation models (O'Neill et al. 1997), radar (Imhoff et al. 1997), and lidar (Lefsky et al. 1999) will likely provide some of the non- spectral information needed to improve species’ occurrence prediction accuracy and land cover map precision. Indices of landscape configuration and context (Pearson 1993, Saab 1999, Conner 2002), land use and other land owner activities (Boren et a1. 1997, Lepczyk et al. 2004), and the connectivity of potential wildlife occurrence areas (F ahrig and Merriam 1985) could provide additional sources of information for increased prediction accuracy and interpretation of results. Alternative parametric measures of spectral dissimilarity such as Mahalanobis distance (Richards and J ia 1999) also show promise for mapping species’ occurrences through spectral associations (see Conner 2002). Comparison of PCC and Kappa in this study emphasizes the importance of selecting a measure of prediction accuracy that can be interpreted in a meaningful way. PCC did not provide a valid measure of prediction accuracy for binary classification of species presence and absence using either spectral associations or the approach of Gap analysis. For the three species in this study, PCC provided little information other than a summary of the distribution of samples among classes. These results were obtained 11] regardless of species, detection distance or window size. In contrast, Kappa values were nearly independent of the distribution of plots among occurrence classes, especially after majority map classification. Thus, Kappa values provided a more complete picture of classification accuracy than PCC. The range of Kappa values, however, varied given the species, detection distance and window size. Kappa values spanned 20% to 30% the scope of the index over any group of 11 cross-validation iterations. Majority classification helped hone in on the “true” predictability of occurrence maps. Furthermore, majority maps often predicted species’ occurrences better than any of the maps created during cross-validation. Cross- validation followed by majority classification thus shows potential for improving the accuracy of any image classification method. However, no single measure, including Kappa, tells the entire story about the quality of classification accuracy. Other accuracy measures such as sensitivity and specificity (Hepinstall et al. 2002) can shed additional light on reasons for omission and commission errors. My results also indicated that the detection distance used to characterize each species’ occurrence affected Kappa values substantially (Figure 4.4). Moderate-to-large detection distances (lOO-m and 180-m) best classified maps of BTNW and NAWA occurrences. However, moderate detection distances, which ignored remote observations, provided the best source of information for classification of OVEN. Unlimited distance point counts for this very common species in the study region may consequently include detections unrepresentative of the vegetation described by pixels used in signature creation. These larger detection distances are also subject to increased bias from differences in observer’s hearing abilities (Ramsey and Scott 1981). Attention should 112 therefore be given to using appropriate detection distances when creating occurrence maps through spectral associations. Besides detection distances, window sizes used in signature creation also influenced accuracy statistics but to a lesser extent. Highest Kappa values of majority maps were typically obtained using moderate window sizes. These window sizes of 9 to 13 pixels (0.8 to 1.2 ha) used in my analyses are best representative of the study species territory sizes (Schoener 1968, Morse 1976). These results emphasize the importance of considering the species’ natural history when choosing among alternative classification parameters. However, habitat features within their territory may be of the greatest importance for some species, while landscape factors may be more influential for others (Pearson 1993; Saab 1999). In my analyses, window size provided an indication of how averaging the spatial variability of pixels surrounding survey plots affected my results. Spatial variability is expected to affect prediction accuracy as a function of the feature being classified, the information content of the imagery, edge effects, and the interspersion of spatial autocorrelation from higher order edaphic gradients and landscape structure (Hurlbert 1984, Collins and Woodcock 1999, Legendre et a1. 2002). In addition to the influence of window size and detection distance investigated in this study, prediction accuracy is likely dependent on several other factors. Kappa values are expected to vary as a function of the extent of the study area, type and timing of imagery, grain of imagery, image processing methods, and image classification methods. The type of occurrence data used in supervised image classification will also affect prediction accuracy. This is because the occurrences of some species are expected to vary spatially as a function of ontogenic changes in life history strategies (Polis 1984, Temple 113 1990, Kolasa and Waltho 1998) and temporal changes in habitat use (Morse 1985). The influence of all these factors on prediction accuracy can be tested within a strong inference framework (sensu Murphy and Noon 1992, Jenkins et al. 2003b) using maps generated with PHASE]. Descriptions of landscape heterogeneity will therefore differ given the process of interest as well as the grain of analysis, the classification system in use and the variability of spectral information employed. For these reasons, a quantitative examination of all possible factors affecting prediction accuracy is warranted. The GRAIN protocol and HABICLASS procedure introduced in this paper provide a general framework for such an examination. By controlling for scale and aggregation effects and permitting a strong inference approach for investigations into causal mechanisms behind the predicted occurrence patterns, GRAIN and HABICLASS provide a compelling complement to regional mapping efforts of wildlife occurrences such as Gap Analysis. 114 CHAPTER 5 SUMMARY, IMPLICATIONS AND CONCLUSIONS Sum While numerous studies have documented the relationships among bird communities and gradients of vegetation structure and composition, there is still little information regarding specific fine scale habitat associations of bird species over large areas of managed forests. To fill this gap, I quantified the fine scale habitat associations of forest birds within managed northern hardwood stands over a ~400,000 ha region of Michigan’s Upper Peninsula. Small sampling plots (30 meter radius, n= 124) were surveyed for birds and vegetation characteristics. I investigated multicollinearity among vegetation descriptions to identify those with significant independent and joint contributions to the occurrence of 6 bird species: Black-throated Green Warbler (Dendroica virens), Eastern Wood-Pewee (Contopus virens), Least Flycatcher (Empidonax minimus), Ovenbird (Seiurus aurocapillus), Rose-breasted Grosbeak (Pheucticus ludovicianus), and Yellow-bellied Sapsucker (Sphyrapicus varius). The strength, direction of influence and relative contribution of vegetation variables to the probability of bird species’ occurrences were then determined using logistic regression. Several vegetation conditions affected the probability of bird species occurrences within northern hardwood stands. Mean height of the canopy bottom for trees > 10cm dbh and the standard deviation of canopy openness among subsamples within 30—m plots were found to have both positive and negative effects on the probability of bird species occupancies. Both horizontal and vertical diversity within northern hardwood stands 115 therefore need to be an important consideration for promoting probabilities of species’ occurrences within northern hardwood stands across my study region. Other influential variables included mean height of sapling canopy bottom, average canopy openness, height of the tallest tree and the importance value of sugar maple saplings. While additional research is needed to determine whether changes to these aspects of forest composition and structure will affect the demography of these birds, my results indicated that limiting the amount of selective harvesting to a subset of 30—m radius plots within uneven-aged northern hardwood stands during each entry will increase avian species diversity over entire stands. In addition to vegetation data, unclassified remotely sensed imagery also has strong potential for mapping and understanding process influencing bird diversity patterns. For example, I used linear regression and geographical weighted regression (GWR) to investigate small grain and extent relationships between NDVI and migratory passerine species richness and to explore the pctential causes of these relationships. The use of GWR allowed me to model spatial heterogeneity to derive global and local model slopes and fits between variables over the extent of the study area. NDVI values derived from Landsat 7 ETM+ imagery and averaged over point count survey areas (180-m, NDVIlgo) had a weak negative relationship with avian species richness in the linear model (R2 = 0.18, p < 0.05). GWR indicated the local slopes and fits varied spatially (a = 0.05, p < 0.05) and ranged from negative to positive (B ranged fi‘om -2.32 to 0.29, R2 ranged from 0.25 to 0.74). The proportion of pixels classified as non-deciduous land cover surrounding each plot within the kernel bandwidth of the GWR model (8036 m) 116 influenced these local slopes. Large shifts in local slopes of this relationship occurred near transition zones among level III ecoregions. I also found a negative relationship between NDV1130 and avian species richness in regions with large amounts of non-deciduous cover and a relatively positive relationship in regions dominated with deciduous cover. NDVI values of 30-m pixels were positively dependent on the proportion of deciduous forest within them, and the slope and strength of this relationship did not vary spatially. However, the relationship did vary spatially for NDVIlgo when compared to the proportion of overlapping pixels classified as deciduous forest in the land cover map (b ranged from 0.01 to 0.09, R2 ranged from 0.48 to 0.99). These results indicate the observed relationship between NDVI and bird species richness and the mechanisms behind the relationship vary with scale. Interpretation of causes for the relationship between NDVI and species richness therefore needs to be made within the spatial context of the study. However, NDVI values can provide scale-specific information regarding spatial trends in avian species richness. Along with the availability of multiple satellites to monitor NDVI remotely at different resolutions, this knowledge could greatly help elucidate causes of spatial variation in species richness over time. Finally, I investigated the potential of using unclassified spectral information for predicting the distribution of three bird species over the study region using Landsat ETM+ imagery and 433 locations sampled for birds through point count surveys. These species, Black-throated Green Warbler, Nashville Warbler, and Ovenbird, were known to be associated with forest understory features during breeding. I examined the influence of varying two spatially explicit classification parameters on prediction accuracy: 1) the 117 window size used to average spectral values in signature creation and 2) the threshold distance required for bird detections to be counted as present. Two accuracy measurements, proportion correctly classified (PCC) and Kappa, of maps predicting species’ occurrences were calculated with ground data not used during classification. Maps were validated for all three species with Kappa values >03 and PCC >0.6. However, PCC provided little information other than a summary of how equally sample plots used to classify occurrence were distributed into presence and absence classes. Comparisons with rule-based maps created using the approach of Gap Analysis showed that spectral information predicted the occurrence of these species that use forest subcanopy components better than could be done using known land cover associations (Kappa values 0.1 to 0.3 higher than Gap Analysis maps). Accuracy statistics for each species were affected in different ways by the detection distance of point count surveys used to stratify plots into presence and absence classes and window sizes used in spectral signature development. Implications for Future Resegg My sampling strategy allowed me to integrate diverse types of data for novel analyses. I highly recommend others take a similar approach, especially if existing land cover data or imagery are not available at the time of field data collection. However, I suggest a slightly different approach if these data are available. My selection of plots within random township sections resulted in a dataset with variables that often did not meet assumptions of normality. Complex transformations were often needed prior to statistical analyses, making interpretations of the results difficult. Sometimes, data could 118 not be transformed to meet normality assumptions, rendering some potentially interesting analyses impossible using parametric statistics. One possible improvement to the sampling design would be to take an approach similar to that proposed by Urban (2002). Landscape units could be randomly selected within nested grids. The selected landscape units could then be overlaid with property ownership maps, land cover maps and satellite imagery to stratify the selection of patches and spectral values for sampling. As a result, sampling efficiency may be maximized by selecting survey plots over a range of spectral values within a single or just a few land cover classes and patch sizes. This approach could also help the researcher refine her/his research questions based on the range of variation in the data. However, future researchers should also keep in mind that land cover categories used in stratification and the classification methods employed to create the land cover maps will also have a strong influence on results. If accurate predictions of species distributions are all that is needed for research priorities, the PHASE] program shows strong potential. However, further work needs to be done to refine the methodology and determine reasons for the predicted wildlife distributions. Quantified spectral associations of vegetation conditions that influence species occupancies, such as those identified in Chapter 2, would enhance the utility of this approach. Preliminary results indicate the potential for a strong relationship between spectral values and pixel area summaries of canopy openness, conifer basal area, and the variability of tree height. Vegetation species composition also affects the relationships. Data should therefore first be stratified by land cover type prior to future investigations. 119 However, the accuracy, precision and methodology used to create land cover maps will affect the analyses. Imiications for Future Management The results of this body of research have several management implications for bird and forest management. Horizontal and vertical diversity within northern hardwood stands can increase bird diversity. While this observation may not be surprising, the homogeneous structure and composition of forests in the study region, especially northern hardwood stands, leads me to believe that horizontal diversity in particular is not considered when individual trees are selected for harvest. Instead, an even-aged approach appears to dominate harvest decisions. Canopy gaps are evenly distributed within a stand to possibly maximize the spatial allocation of sunlight to the canopy floor. While this strategy has some theoretical merits in a closed system, regeneration in many northern hardwood stands across the study region is minimal at best. Although I am not qualified to predict the impact of alternative harvest strategies on regeneration in these stands, any alternative may be worth considering. My results indicate that aggregating the harvest of merchantable timber to a subset of 30-m radius plots within a stand has the potential for increasing bird diversity over the entire stand. Furthermore, initiating this site-specific modification to harvesting strategies is expected to. have a positive influence on bird species richness over the extent of the study region. It may also affect the processes that currently discourage natural regeneration, such as intensive deer browsing, but additional research is needed to determine if that is the case. An adaptive management approach that 120 considers these relationships may help managers realize biodiversity and sustainable harvesting goals. The results of the species richness and NDVI analyses indicated that at least two separate processes are likely affecting the gradients of avian species richness in the study area. In the southeastern region dominated by coniferous and mixed forests, high proportions of coniferous or mixed forest cover within a 180-m radius was correlated with high species richness. In the northwestern region dominated by deciduous forest, high species richness was associated with high proportions of deciduous forest within a 180-m radius. These existing relationships should be kept in mind when making future management decisions as beneficial activities in one area are likely detrimental in other places. For example, in the southeast portion of the study region, conversions of mixed forest to aspen dominated stands as a result of clearcutting may reduce avian species richness. Similarly, transforming the deciduous forest in the northwest part of the study region to pine plantations could also lower bird species richness. Maintaining or increasing coniferous and mixed forest in the southeast and deciduous forest in the northwest portions of the study region could help maintain current bird species occupancies. Other Data Uses Many of the data collected during the course of this research were distributed to other interested parties. For example, the land cover data collected during the summer of 2001 were provided to Pacific-Meridian Resources for use as ground reference data during the development of Michigan’s latest land cover map. The land cover data 121 collected in 2002 and 2003 were provided to M. Donovan of the MDNR to assess the accuracy of the map. This collaborative effort more than doubled the number of sample locations available for my study region and may therefore have helped increase the accuracy of the map. This was a win-win situation as the land cover map was very useful in my analyses. Hopefully firture land cover map users will find the extra effort helpful as well. I encourage others to engage in similar efforts whenever possible. The bird data I collected were summarized by township section and provided to R. Adams of the Kalamazoo Nature Center for use in the next edition of Michigan’s Breeding Bird Atlas. Mr. Adams has also indicated that he will distribute the data to national databases. People similarly interested in making others aware of the availability of bird data can now quickly enter a few details about the data and their contact information into a web-based form I helped develop as part of the Michigan Bird Conservation Initiative steering committee. The Internet address is wwwmichigan.gov/dnmibci . Conclusions Combined, the research presented in this dissertation increases the general understanding of bird-habitat relationships across forested landscapes. Furthermore, I developed several novel methods for quantifying patterns and understanding the processes causing the patterns of species distributions and gradients of species richness over large landscapes. Important processes were found to be 1) ecologically relevant (e.g., variability in canopy openness decreases the probability of Black-throated Green Warbler occurrence), 2) geometrically constrained (e. g., the relationship between NDVI 122 and species richness varies among ecoregions), and 3) methodologically dependent (e.g., decisions made during image classification affect prediction accuracies). The strength of my approach came from the collection and use of spatially compatible data. Vegetation, land cover and satellite image data collected over roughly the same size areas were all shown to be useful for understanding reasons for bird occupancies. For example, by employing a small detection threshold for bird point count surveys and detailed vegetation measurements collected over a similar size plot (i.e., 30- m radius circle) within stands of northern hardwood forests I was able to investigate fine scale habitat associations of bird species over a large region (~400, 000 ha). While the findings are similar to those of previous studies, the scale of analyses shows that the accuracy of models predicting bird species distributions could be enhanced by using data describing specific vegetation conditions within areas comparable in size to pixels of Landsat imagery (i.e., 30-m x 30-m pixels). This knowledge has strong potential for future wildlife management efforts. If important vegetation conditions for birds have a spectral response detectable by the image analyses, additional data layers having a functional relationship with wildlife occupancy can be developed The results of Chapters 3 and 4 supported this conclusion. Bird species richness had a predictable relationship with NDVI that was mapped and can be monitored using the intensities of reflected radiation measured at different wavelengths by satellites orbiting the Earth. In addition, there is information available in unclassified imagery to permit predictions of some migratory bird distributions with greater accuracy than chance and current Gap analysis predictions. Aggregating pixel values affected the relationships between NDVI and land cover. It also affected the accuracies of species occurrence 123 predictions. Therefore, the effects of plot size and study extent should be considered when interpreting output from similar studies conducted using data of different grains or collected over different regions. 124 LITERATURE CITED Albert, D. A. 1995. Regional Landscape Ecosystems of Michigan, Minnesota, and Wisconsin: A Working Map and Classification. General Technical Report NC- 178, USDA Forest Service, St. Paul, MN. Ambuel, B. and S. A. Temple. 1983. Area-dependent changes in the bird communities and vegetation of southern Wisconsin forests. Ecology 64:1057-1068. Anders, A. D., J. F aaborg, and F. R. Thompson 111. 1998. Postfleding dispersal, habitat use, and home-range size of juvenile wood thrushes. The Auk 115:349-358. Anderson, R. C. and O. L. Loucks. 1979. White-tailed deer (Odocoileus virginianus) influence on structure and composition of Tsuga canadensis forests. Journal of Applied Ecology 16:855-861. Anselin, L. and D. A. Griffith. 1988. Do spatial effects really matter in regression analysis. Papers of the Regional Science Association 65:11-34. Amett, E. B. and R. Sallabanks. 1998. Land manager perceptions of avian research and information needs: a case study. Pages 399-414 in J. M. Marzluff and R. Sallabanks, editors. Avian Conservation: Research and Management. Island Press, Washington, DC. Askins, R. A. and M. J. Philbrick. 1987. Effect of changes in regional forest abundance ' on the decline and recovery of a forest bird community. Wilson Bulletin 99:7-21. Asner, G. P. and D. B. Lobell. 2000. A biogeophysical approach for automated SWIR unmixing of soils and vegetation. Remote Sensing of Environment 74:99-112. Augustine, D. J. and L. E. Frelich. 1998. Effects of white-tailed deer populations on an understory forb in fragmented deciduous forests. Conservation Biology 12:995- 1004. Austin, M. P. 1985. Continuum concept, ordination methods, and niche theory. Annual Review of Ecology and Systematics 16:39-61. Avery, T. E. and H. E. Burkhart, . 1967. 2002. Forest Measurements. McGraw Hill, New York, New York. Bailey, S.-A., M. C. Homer-Devine, G. Luck, L. A. Moore, K. M. Carney, S. Anderson, C. Betrus, and E. Fleishman. 2004. Primary productivity and species richness: relationships among functional guilds, residency groups and vagility classes at multiple spatial scales. Ecography 27:207-217. 125 Balgooyen, C. P. and D. M. Waller. 1995. The use of Clintonea borealis and other indicators to gauge impacts of white-tailed deer on plant communitites in northern USA. Natural Areas Journal 15:308-318. Bawa, K., J. Rose, K. N. Ganeshaiah, N. Barve, M. C. Kiran, and R. Umashaanker. 2002. Assessing biodiversity from space: an example from the Western Ghats, India. Conservation Ecology 6:7. [online] URL:http//www.consecol.org/vol6/issZ/art7. Benning, T. L. and T. R. Seastedt. 1995. Landscape-level interactions between topoedaphic features and nitrogen limitation in tallgrass prairie. Landscape Ecology 10:337-348. Blake, J. G. and J. R. Karr. 1984. Species composition of bird communities and the conservation benefits of large versus small forests. Biological Conservation 30: 1 73- 1 87. Blake, J. G. and J. R. Karr. 1987. Breeding birds in isolated woodlands: area and habitat relationships. Ecology 68: l 724- 1 734. Bollinger, E. K. and R. G. Peak. 1995. Depredation of artificial avian nests: a comparison of forest-field and forest-lake edges. American Midland Naturalist 134:200-203. Boren, J. C., D. M. Engle, and R. E. Masters. 1997. Vegetation cover type and avian species changes on landscapes within a wildland-urban interface. Ecological Modelling 103:251-266. Brewer, R., G. A. McPeek, and R. J. Adams. 1991. The Atlas of Breeding Birds of Michigan. Michigan State University Press, East Lansing. Brooker, L., M. Brooker, and P. Cale. 1999. Animal dispersal in fragmented habitat: measuring habitat connectivity, corridor use, and dispersal mortality. Conservation Ecology 3:21. Brooks, R. T. 1999. Residual effects of thinning and high white-tailed deer densities on northern redback salarnanders in southern new england oak forests. Journal of Wildlife Management 63:] 172-1 180. Brunsdon, C. A., A. S. F otheringham, and M. E. Charlton. 1996. Geographically weighted regression: a method for exploring spatial nonstationarity. Geographical Analysis 28:281-298. Bunnell, F. L. and L. L. Kremsater. 1990. Sustaining wildlife in managed forests. Northwest Environmental Journal 6:243-269. 126 Bystrak, D. 1981. The North American breeding bird survey. Pages 34-41 in C. J. Ralph and J. M. Scott, editors. Estimating the Number of Terrestrial Birds: Studies in Avian Biology 6. Callicott, J. B. 1990. Whither conservation ethics? Conservation Biology 4:15-20. Ceccato, P., S. Flasse, S. Tarantola, S. Jacquemoud, and J. M. Gregoire. 2001. Detecting vegetation leaf water content using reflectance in the optical domain. Remote Sensing of Environment 77:22-33. Chevan, A. and M. Sutherland. 1991. Hierarchical partitioning. American Statistician 45:90—96. Chown, S. L., B. J. van Rensburg, K. J. Gaston, A. S. L. Rodrigues, and A. S. van J aarsveld. 2003. Energy, species richness, and human population size: conservation implications at a national scale. Ecological Applications 13: 1233- 1241. Cohen, J. 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement 20:37-46. Collins, J. B. and C. E. Woodcock. 1999. Geostatistical estimation of resolution dependent variance in remotely sensed images. Photograrnmetric Engineering and Remote Sensing 65:41-50. Collins, S. L. 1981. A comparison of nest-site and perch-site vegetation structure for seven species of warblers. Wilson Bulletin 93. Collins, S. L. 1983. Geographic variation in habitat structure for the wood warblers in Maine and Minnesota. Oecologia 59:246-252. Collins, S. L., F. C. James, and P. G. Risser. 1982. Habitat relationships of wood warblers (Parulidae) in northern central Minnesota. Oikos 39:50-58. Comeau, P. G., F. Gendron, and T. Letchford. 1998. A comparison of several methods for estimating light under a paper birch mixedwood stand. Canadian journal of Forest Research 28: 1 843-1 850. Connell, J. H. and E. Orias. 1964. The ecological regulation of species diversity. The American Naturalist 98:399-414. Conner, L. M. 2002. A technique to locate isolated populations using satellite imagery. Wildlife Society Bulletin 30:1044-1049. 127 Comett, M. W., L. E. F relich, and K. J. Puettmann. 2000. Conservation implications of browsing by Odocoileus virginianus in remnant upland Thuja occidentalis forests. Biological Conservation 93:359-369. Costello, C. A., M. P. Yarnasaki, J. Pekins, and W. B. Leak. 2000. Songbird response to group selection harvests and clearcuts in New Hampshire northern hardwood forest. Forest Ecology and Management 127:41-54. Cottenie, K. and L. De Meester. 2004. Metacommunity structure: Synergy of biotic interactions as selective agents and dispersal as fuel. Ecology 85:114-119. Cottenie, K., E. Michels, N. Nuytten, and L. De Meester. 2003. Zooplankton metacommunity structure: Regional vs. local processes in highly interconnected ponds. Ecology 84:991-1000. Cracknell, A. P. 1998. Synergy in remote sensing - what's in a pixel? International Journal of Remote Sensing 19:2025-2047. Crawford, H. S., R. G. Hooper, and R. W. Titterington. 1981. Songbird population response to silvicultural practices in central Appalachian hardwoods. Journal of Wildlife Management 45:680-692. Crete, M., J. P. Ouellet, and L. Lesage. 2001. Comparative effects on plants of caribou/reindeer, moose and white-tailed deer herbivory. Arctic 54:407-417. Crozier, G. E. and G. J. Niemi. 2003. Using patch and landscape variables to model bird abundance in a naturally heterogeneous landscape. Canadian Journal of Zoology 81:441-452. Cumming, G. S. 2000. Using between-model comparisons to fine-tune linear models of species ranges. Journal of Biogeography 27:441-455. D'Arrigo, R. D., C. M. Malmstrom, G. C. J acoby, S. O. Los, and D. E. Bunker. 2000. Correlation between maximum latewood density of annual tree rings and NDVI based estimates of forest productivity. International Journal of Remote Sensing 21:2329-2336. Davis, D. E. 1959. Observations on territorial behavior of Least Flycatchers. Wilson Bulletin 71:73-85. DeCalesta, D. S. 1994. Effect of white-tailed deer on songbirds within managed forests in Pennsylvania. Journal of Wildlife Management 58:711-718. DeCalesta, D. S. 1997. Deer, Ecosystem Damage, and Sustaining Forest Resources. in Online Proceeding of Deer Management in Maryland Conference, Maryland. 128 DellaSala, D. A. and D. L. Rabe. 1987. Response of Least Flycatchers Empidonax minimus to forest disturbances. Biological Conservation 41:291-299. Derleth, E. L., D. G. McAuley, and T. J. Dwyer. 1989. Avian community response to small-scale habitat disturbance in Maine. Canadian Journal of Zoology 67 :385- 390. DeSante, D. F. and D. K. Rosenburg. 1998. What do we need to monitor in order to manage landbirds? Pages 93-106 in J. M. Marzluff and R. Sallabanks, editors. Avian Conservation: Research and Conservation. Island Press, Washington, DC. Dijak, W. D. and F. R. Thompson. 2000. Landscape and edge effects on the distribution of mammalian predators in Missouri. Journal of Wildlife Management 64:209- 216. Donnelly, R. and J. M. Marzluff. 2004. Importance of reserve size and landscape context to urban bird conservation. Conservation Biology 18:733-745. Donovan, M. L., G. M. Nesslage, J. J. Skillen, and B. A. Maurer. 2004. The Michigan Gap Analysis Project Final Report. Michigan Department of Natural Resources, Lansing, MI. Donovan, T. M., P. W. Jones, E. M. Armand, and F. R. Thompson. 1997. Variation in local-scale edge effects: mechanisms and landscape context. Ecology 78:2064- 2075. Donovan, T. M., R. H. Larnberson, A. Kimber, F. R. 1. Thompson, and J. F aaborg. 1995. Modeling the effects of habitat fragmentation on source sink demography of neotropical migrant birds. Conservation Biology 9: 1396-1407. Dunn, J. and K. Garrett. 1997. A Field Guide to Warblers of North America. Houghton Mifflin Co., New York. Dutilleul, P. 1998. Incorporating scale in ecological experiments: study design. Pages 369-386 in D. L. Peterson and V. T. Parker, editors. Ecological Scale: Theory and Applications. Columbia University Press, New York, New York. Dutilleul, P. and P. Legendre. 1993. Spatial heterogeneity against heteroscedasticity: an ecological paradigm versus a statistical concept. Oikos 66:152-170. Ehrlich, P. R., D. S. Dobkin, and D. Wheye. 1988. The Birder's Handbook: A Field Guide to the Natural History of North American Birds. Simon and Schuster, New York, NY. Fahrig, L. and G. Merriam. 1985. Habitat Patch Connectivity and Population Survival. Ecology 66 1762-1768. 129 F laspohler, D. J ., S. A. Temple, and R. N. Rosenfleld. 2001. Effects of forest edges on Ovenbird demography in a managed forest landscape. Conservation Biology 15: 1 73—1 83. Flather, C. H. and J. R. Sauer. 1996. Using landscape ecology to test hypotheses about large-scale abundance patterns in migratory birds. Ecology 77:28-35. Foody, G. M. 1999. The continuum of classification fuzziness in thematic mapping. Photograrnmetric Engineering and Remote Sensing 65:443-451. F otheringham, A. S., C. Brunsdon, and M. Charlton. 2000. Quantitative Geography: perspectives on spatial data analysis. SAGE Publications. F otheringharn, A. S., C. Brunsdon, and M. Charlton. 2002. Geographically Weighted Regression: the Analysis of Spatially Varying Relationships. John Wiley & Sons, New York. Franklin, S. E. 2001. Remote sensing for sustainable forest management. Lewis Publishers, Boca Raton, Fla. Fraser, R. H. 1998. Vertebrate species richness at the mesoscale: relative roles of energy and heterogeneity. Global Ecology and Biogeography Letters 72215-220. F rawley, B. J. 1999. Michigan Deer Harvest Survey Report: 1998 Seasons. Michigan Department of Natural Resources, Wildlife Report no. 3303, Lansing, MI. F rawley, B. J. 2000. Michigan Deer Harvest Survey Report: 1999 Seasons. Michigan Department of Natural Resources, Wildlife Report no. 3314, Lansing, MI. Freedman, B., C. Beauchamp, I. A. McLaren, and S. I. Tingley 1981. Forestry management practices and populations of breeding birds in Nova Scotia. Canadian Field-Naturalist 95:307-31 1. Freemark, K. E. and H. G. Meniarn. 1986. Importance of area and habitat heterogeneity to bird assemblages in temperate forest fragments. Biological Conservation 36:1 15-141. Frelich, L. E. and P. B. Reich. 1999. Neighborhood effects: disturbance severity, and community stability in forests. Ecosystems 2:151-166. Garnmel, F. M. 1995. Effects of forest cover, terrain, and scale on timber volume estimation with Thematic Mapper data in a Rocky Mountain site. Remote Sensing of Environment 51:291-305. 130 Garnon, J. A., C. B. Field, M. L. Goulden, K. L. Griffin, A. E. Hartley, G. Joel, J. Penuelas, and R. Valentini. 1995. Relationships between NDVI, canopy structure, and photosynthesis in three Californian vegetation types. Ecological Applications 5:28-41. Gausman, H. W., W. A. Allen, and R. Cardenas. 1969. Reflectance of cotton leaves and their structure. Remote Sensing of Environment 1:] 10-122. Gemmell, F. M. 1995. Effects of forest cover, terrain and scale on timber volume estimation with thematic mapper data in a Rocky Mountain site. Remote Sensing of Environment 51:291-315. Golet, F. C., Y. Wang, J. S. Merrow, and W. R. DeRagon. 2001. Relationship between habitat and landscape features and the avian community of red maple swamps in southern Rhode Island. Wilson Bulletin 113:2]7-227. Gould, W. 2000. Remote sensing of vegetation, plant species richness, and regional biodiversity hotspots. Ecological Applications 10: 1 861-1 870. Gram, W. K., P. A. Pomeluzi, R. L. Clawson, J. Faaborg, and S. C. Richter. 2003. Effects of experimental forest management on density and nesting success of bird species in Missouri Ozark forests. Conservation Biology 17 :1324-1337. Grime, J. P. 1973. Competitive exclusion in herbaceous vegetation. Nature 242:344-347. H-Acevedo, D. and D. J. Currie. 2003. Does climate determine broad-scale patterns of species richness? A test of the causal link by natural experiment. Global Ecology and Biogeography Letters 12:461—473. Hamer, K. C. and J. K. Hill. 2000. Scale dependent effects of habitat disturbance on species richness in tropical forests. Conservation Biology 14:1435-1440. Hansen, A. J. and D. L. Urban. 1992. Avian response to landscape pattern: the role of species life histories. Landscape Ecology 7: 163-1 80. Harrison, S. 1997. How natural habitat patchiness affects the distribution of diversity in Californian serpentine Chaparral. Ecology 78: 1 898-1906. Hawkins, B. A. 2004. Summer vegetation, deglaciation and the anomalous bird diversity gradient in eastern North America. Global Ecology and Biogeography 13:321- 325. Hawkins, B. A., R. Field, H. V. Cornell, D. J. Currie, J .-F. Guegan, D. M. Kaufman, J. T. Kerr, G. G. Mittelbach, T. Oberdorff, E. M. O'Brien, E. E. Porter, and J. R. G. Turner. 2003. Energy, water, and broad-scale geographic patterns of species richness. Ecology 84:3]05-3117. 131 Hejl, S. J. and K. M. Granillo. 1998. What managers really need from avian researchers. Pages 431-438 in J. M. Marzluff and R. Sallabanks, editors. Avian Conservation: Research and Management. Island Press, Washington, DC. Hepinstall, J. A., W. B. Krohn, and J. A. Sader. 2002. Effects of niche width on the performance and agreement of avian habitat models. in J. M. Scott, P. J. Heglund, M. L. Morrison, J. B. Hauffler, M. G. Raphael, W. A. Wall, and F. B. Samson, editors. Predicting Species Occurrence: Issues of Accuracy and Scale. Island Press, Washington, DC. Hepinstall, J. A. and J. A. Sader. 1997. Using bayesian statistics, thematic mapper satellite imagery, and breeding bird survey data to model bird species probability of occurrence in Maine. Photogramrnetric Engineering and Remote Sensing 63:1231-1237. Hespenheide, H. A. 197]. Flycatcher habitat selection in the eastern deciduous forest. Auk 88:61-74. Hill, J. K. and K. C. Hamer. 2004. Determining impacts of habitat modification on diversity of tropical forest fauna: the importance of spatial scale. Journal of Applied Ecology 41:744-754. Hobbs, T. J. 1995. The use of NOAA-AVHRR NDVI data to assess herbage production in the arid rangelands of Central Australia. International Journal of Remote Sensing 16:1289-1302. Hobson, K. A. and E. Bayne. 2000. Breeding bird communities in boreal forest of western Canada: Consequences of "unmixing" the mixedwoods. Condor 102:759- 769. Holmes, R. T. and T. W. Sherry. 2001. Thirty-year bird population trends in an unfiagmented temperate deciduous forest: importance of habitat change. The Auk 118:589-609. ‘ Holt, R. and K. Martin. 1997. Landscape modification and patch selection: the demography of two secondary cavity nesters colonizing clearcuts. The Auk 1 14:443-455 . Huff, M. H., K. A. Bettinger, H. L. Ferguson, M. J. Brown, and B. Altman. 2000. A Habitat-Based Point-Count Protocol for Terrestrial Birds, Emphasizing Washington and Oregon. USDA Forest Service PNW-GTR-SO]. Hurlbert, A. H. 2004. Species-energy relationships and habitat complexity in bird communities. Ecology Letters 7:714-720. 132 Hurlbert, A. H. and J. P. Haskell. 2003. The effect of energy and seasonality on avian species richness and community composition. The American Naturalist 161283- 97. Hurlbert, S. H. 1984. Pseudoreplication and the design of ecological field experiments. Ecological Monographs 54:187-21 1. Huston, M. A. 1979. A general hypothesis of species diversity. American Naturalist 1 13281-1 01 . Imhoff, M. L., T. D. Sisk, A. Milne, G. Morgan, and T. Orr. 1997. Remotely sensed indicators of habitat heterogeneity: Use of synthetic aperature radar in mapping vegetation structure and bird habitat. Remote Sensing of Environment 60:217- 227. James, F. C. 1971. Ordinations of habitat relationships among breeding birds. The Wilson Bulletin 83:215-236. James, F. C. 1998. The growing pains of avian conservation biology. Pages 15-23 in J. M. Marzluff and R. Sallabanks, editors. Avian Conservation: Research and Management. Island Press, Washington, DC. James, F. C. and H. H. Shugart Jr. 1970. A quantitative method of habitat description. American Birds 24:727-736. Jenkins, C. N., R. D. Powell, 0. L. Bass Jr., and S. L. Pimm. 2003a. Demonstrating the destruction of the habitat of the Cape Sable seaside sparrow (Ammodramus maritimus mirabilis). Animal Conservationz29—38. Jenkins, C. N., R. D. Powell, 0. L. Bass Jr., and S. L. Pimm. 2003b. Why sparrow distributions do not match model predictions. Animal Conservation 6:39-46. Jensen, J. R. 1983. Biophysical remote sensing. Annals, Association of American Geographers 73:111-132. Jensen, J. R. 2000. Remote Sensing of Environment: an Earth Resource Perspective. Prentice Hall, NJ. J enz, W. and C. Rahbek. 2001. Geometric constraints explain much of the species richness pattern in African birds. Procedings of the National Academy of Science 98:5661-5666. J obes, A. P., E. No], and D. R. Voigt. 2004. Effects of selection cutting on bird communities in contiguous eastern hardwood forests. Journal of Wildlife Management 68:51-60. 133 Johnson, A. R., J. A. Wiens, B. T. Milne, and T. O. Crist. 1992. Animal movements and population dynamics in heterogenous landscapes. Landscape Ecology 7:63-75. Jokimaki, J. and E. Huhta. 1996. Effects of landscape matrix and habitat structure on a bird community in northern Finland: A multi-scale approach. Omis F ennica 73:97-113. Kerr, J. T. and M. Ostrovsky. 2003. From space to species: ecological applications for remote sensing. Trends in Ecology & Evolution 18:299-305. King, D. I. and R. M. DeGraaf. 2004. Effects of group-selection opening size on the distribution and reproductive success of an early-successional shrubland bird. Forest Ecology and Management 190:179-185. Kolasa, J. and N. Waltho. 1998. A heirarchical view of habitat and its relationship to species abundance. Pages 55-76 in V. T. Parker and D. L. Peterson, editors. Ecological Scale: Theory and Applications. Columbia University Press, New York, New York. Kotliar, N. B. and J. A. Wiens. 1990. Multiple scales of patchiness and patch structure: a hierarchical framework for the study of heterogeneity. Oikos 59:253-260. Laurent, E. J., J. P. LeBouton, M. B. Walters, and J. Liu. 2002. Integrating human, satellite and avian perspectives of the landscape for analysis of forest bird distribution patterns. in D. Chamberlain and A. Wilson, editors. Avian Landscape Ecology: Pure and Applied Issues in the Large-Scale Ecology of Birds. Proceedings of the 11th Annual IALE(UK) Conference. Colin Cross Printers Ltd., Garstang, Great Britain. Law, B. E. and R. H. Waring. 1994. Combining remote sensing and climatic data to estimate net primary production across Oregon. Ecological Applications 4:717- 728. Lawler, J. J. and T. C. Edwards. 2002. Landscape patterns as predictors of nesting habitat: building and testing models for four species of cavity-nesting birds. Landscape Ecology 17 :233-245. Lefsky, M. A., W. B. Cohen, S. A. Acker, G. G. Parker, T. A. Spies, and D. Harding. 1999. Lidar remote sensing of the canopy structure and biophysical properties of douglas-fir western hemlock forests. Remote Sensing of Environment 70:339- 361. Legendre, P. 1993. Spatial autocorrelation: trouble or new paradigm. Ecology 74:1659- 1 673. 134 Legendre, P., M. R. T. Dale, M. F ortin, J. Gurevitch, M. Hohn, and D. Myers. 2002. The consequences of spatial structure for the design and analysis of ecological field surveys. Ecography 25:60] -61 5. Leopold, A. 1933. Game Management. Charles Scribner's Sons, New York. Lepczyk, C. A., A. G. Mertig, and J. Liu. 2004. Assessing landowner activities that influence birds across rural-to-urban landscapes. Environmental Management 33:110-125. Lepczyk, C. A., V. C. Radeloff, C. H. Flather, and J. Liu. 2003. The human influence on birds across landscapes of the Midwest United States. Die Vogelwarte 42:8. Lessard, V. C., T. D. Drummer, and D. D. Reed. 2002. Precision of density estimates from fixed-radius plots compared to N-tree distance sampling. Forest Science 48:1-6. Lessard, V. C., D. D. Reed, and N. Monkevich. 1994. Comparing N-tree distance sampling with point and plot sampling in northern Michigan forest types. Northern Journal of Applied Forestry 11:12-16. Leung, Y., C. Mei, and W. Zhang. 2000. Satistical tests for nonstationarity based on the geographically weighted regression model. Environment and Planning A 3229-32. Levin, S. A. 1992. The problem of pattern and scale in ecology. Ecology 73:1943-1967. Li, H. and J. F. Reynolds. 1994. A simulation experiment to quantify spatial heterogeneity in categorical maps. Ecology 75:2446—2455. Liang, S. Y. and S. W. Seagle. 2002. Browsing and microhabitat effects on riparian forest woody seedling demography. Ecology 83:212-227. Lillesand, T. M., J. Chipman, D. Nagel, H. Reese, M. Bobo, and R. Goldmann. 1998. Upper Midwest Gap Analysis Program Image Processing Protocol. EMTC 98- G001, Report prepared for the US. Geological Survey, Environmental Management Technical Center, Onalaska, Wisconsin. Lillesand, T. M. and R. W. Kiefer. 1999. Remote sensing and image interpretation, 4th edition. John Wiley & Sons, New York. Liu, J ., M. Lindennan, Z. Ouyang, L. An, J. Yang, and H. Zhang. 2001. Ecological degradation in protected areas: the case of Wolong Nature Reserve for giant pandas. Science 292:98-101. Lorimer, C. G. 1989. Relative effects of small and large disturbances on temperate hardwood forest structure. Ecology 70:565-567. 135 Mac Nally, R. 2000. Regression and model-building in conservation biology, biogeography and ecology: The distinction between and reconciliation of 'predictive' and 'explanatory' models. Biodiversity and Conservation 9:655-671. Mac Nally, R. and C. J. Walsh. 2004. Hierarchical partitioning public-domain software. Biodiversity and Conservation 13:659-660. Mac Nally, R. C. 2002. Multiple regression and inference in ecology and conservation biology: further comments on identifying inportant predictor variables. Biodiversity and Conservation 11:1397-1401. MacArthur, R. H. and E. R. Pianka. 1966. On optimal use of a patchy environment. The American Naturalist 100:603-609. Malmstrom, C. M., M. V. Thompson, G. Juday, S. 0. Los, J. T. Randerson, and C. B. Field. 1997. Interannual variation in global-scale net primary production: testing model estimates. Global Biogeochemical Cycles 11:367—392. Martin, T. E., C. R. Paine, C. J. Conway, W. M. Hochachka, P. Allen , and W. Jenkins. 1997. BBIRD Field Protocol. Montana Cooperative Wildlife Research Unit, University of Montana, Missoula, Montana, USA. Matteson, S., A. Paulios, G. Bartelt, G. Butcher, D. Sample, J. Fitzgerald, G. Niemi, J. Hanowski, and R. Howe. in prep. Partners In Flight Bird Conservation Plan For The Boreal Hardwood Transition (Physiographic Area 20 and Bird Conservation Region 12, USA). American Bird Conservancy, The Plains, VA. Maurer, B. A. and R. C. Whitrnore. 1981. Foraging of five bird species in two forests with different vegetation structure. Wilson Bulletin 93:478-490. McCullough, D. R. 1997. Irruptive behavior in ungulates. Pages 69-98 in W. J. McShea, H. B. Underwood, and J. H. Rappole, editors. The Science of Overabundance: Deer Ecology and Population Management. Smithsonian Institution Press, Washington DC. McDonald, A. J ., F. M. Gemmell, and P. E. Lewis. 1998. Investigation of the utility of spectral vegetation indices for determining information of coniferous forests. Remote Sensing of Environment 66:250-272. McGarigal, K. and W. C. McComb. 1995. Relationships between landscape structure and breeding birds in the Oregon Coast Range. Ecological Monographs 65:235-260. McShea, W. J. and J. H. Rappole. 2000. Managing the abundance and diversity of breeding bird populations through manipulation of deer populations. Conservation Biology 14:1161-1170. 136 Menard, S. 2002. Applied logistic regression analysis, 2nd edition. Sage Publications, Thousand Oaks, CA. Menzel, M. A., W. M. Ford, J. Laerm, and D. Krishon. 1999. Forest to wildlife opening: habitat gradient analysis among small mammals in the southern Appalachians. Forest Ecology and Management 114:227-232. Mickelson Jr., J. G., D. L. Civco, and J. A. Silander Jr. 1998. Delineating forest canopy species in the northeastern United States using multi-temporal TM imagery. Photogrammetric Engineering and Remote Sensing 64:891-904. Morrison, M. L. 1992. Bird abundance in forests managed for timber and wildlife resources. Biological Conservation 60. Morrison, M. L., B. G. Marcot, and R. W. Mannan. 1998. Wildlife-Habitat Relationships: Concepts and Applications. The University of Wisconsin Press, Madison. Morse, D. H. 1976. Variables affecting the density and territory size of breeding spruce- woods warblers. Ecology 57:290-301. Morse, D. H. 1985. Habitat selection in North American Parulid warblers. Pages 131-157 in M. L. Cody, editor. Habitat Selection in Birds. Academic Press, Inc., New York. Murphy, D. D. and B. R. Noon. 1992. Integrating scientific methods with habitat conservation planning: reserve design for northern spotted owls. Ecological Applications 2:3-17. Nagelkerke, N. J. D. 1991. A note on a general definition of the coefficient of determination. Biometrika 78:691-692. Niemann, K. O. 1995. Remote sensing of forest stand age using airborne spectrometer data. Photogramrnetric Engineering and Remote Sensing 61:1119-1127. Niemi, G. J. and J. M. Hanowski. 1984. Relationships of breeding birds to habitat characteristics in logged areas. Journal of Wildlife Management 48:438-443. Norton, M. R. 1999. Status of the Black-Throated Green Warbler (Dendroica virens) in Alberta. Wildlife Status Report No. 23, Alberta Environment, Fisheries and Wildlife Management Division, and Alberta Conservation Association, Edmonton, AB. O'Connor, R. J ., M. T. Jones, D. White, C. Hunsaker, T. Loveland, B. Jones, and E. Preston. 1996. Spatial partitioning of environmental correlates of avian biodiversity in the conterrninous United States. Biodiversity Letters 3:97-110. 137 Oindo, B. O. 2002. Predicting mammal species richness and abundance using multi- temporal NDVI. Photogramrnetric Engineering and Remote Sensing 68:623-629. O'Neill, M. P., J. C. Schmidt, J. P. Dobrowski, C. P. Hawkins, and C. M. U. Neale. 1997. Identifying sites for riparian wetland restoration: application of a model to the upper Arkansas river basin. Restoration Ecology 5:85-102. O'Neill, R. V., D. L. DeAngelis, J. B. Waide, and T. F. H. Allen. 1986. A Heirarchical Concept of Ecosystems. Princeton University Press, Princeton, NJ. Paruelo, J. M., H. E. Epstein, W. K. Lauenroth, and I. C. Burke. 1997. ANPP estimates from NDVI for the Central Grassland region of the US. Ecology 78:953-958. Pearce, J. and S. Ferrier. 2000. Evaluating the predictive performance of habitat models developed using logistic regression. Ecological Modelling 133:225-245. Pearce, J. L., L. A. Venier, S. F errier, and D. W. McKenney. 2002. Measuring prediction uncertainty in models of species distribution. in J. M. Scott, P. J. Heglund, M. L. Morrison, J. B. Haufler, M. G. Raphael, W. A. Wall, and F. B. Samson, editors. Predicting Species Occurrences: Issues of Accuracy and Scale. Island Press, Washington, DC. Pearson, S. M. 1993. The spatial extent and relative influence of landscape-level factors on wintering bird populations. Landscape Ecology 8:3-18. Perich, S. 1997. Architects of wildlife management. Outdoor America 62:19-25. Peterson, D. L. and V. T. Parker. 1998a. Ecological Scale: theory and applications. in. Columbia University Press, New York. Peterson, D. L. and V. T. Parker. 1998b. Ecological Scale: Theory and applications. Columbia University Press, New York, New York. Peterson, D. L. and S. W. Running. 1989. Applications in forest science and management. Pages 4210-4473 in Doghouses and G. Asrar, editors. Theory and Applications of Optical Remote Sensing. John Wiley & Sons, New York. Polis, G. A. 1984. Age structure component of niche width and intraspecific resource partitioning: Can age groups firnction as ecological species? American Naturalist 123:541-564. Pomeluzi, P., J. C. Bednarz, L. Goodrich, N. Zawada, and J. Hoover. 1993. Reproductive performance of tenitorial Ovenbirds occupying forest fragments and a contiguous forest in Pennsylvania. Conservation Biology 7 :61 8-622. 138 Probst, J. R., D. S. Rakstad, and D. J. Rugg. 1992. Breeding bird communities in regenerating and mature broadleaf forests in the USA Lake States. Forest Ecology and Management 49:43-60. Quinn, G. P. and M. J. Keough. 2002. Experimental Design and Data Analysis for Biologists. Cambridge University Press, New York, NY. Rail, J. F ., M. Darveau, A. Desrochers, and J. Huot. 1997. Territorial responses of boreal forest birds to habitat gaps. Condor 99:976-980. Ralph, C. J ., G. R. Geupel, P. Pyle, T. E. Martin, and D. F. DeSante. 1993. Handbook of Field Methods for Monitoring Landbirds. General Technical Report PSW-GTR- 144, United States Department of Agriculture, Forest Service, Pacific Southwest Research Station. Albany, CA. Ralph, C. J ., J. R. Sauer, and S. Droege, editors. 1995. Monitoring Bird Populations by Point Counts. USDA Forest Service PSW-GTR-149. Ramsey, F. L. and J. M. Scott. 1981. Tests of hearing ability. Pages 341-345 in J. M. Scott, editor. Estimating Numbers of Terrestrial Birds. Studies in Avian Biology 6. Richards, J. A. and X. Jia. 1999. Remote Sensing Digital Image Analysis: An Introduction. Springer-Verlag, New York. Robbins, C. S., D. K. Dawson, and B. A. Dowel]. 1989. Habitat area and requirements of breeding forest birds of the middle Atlantic states. Wildlife Monographs 103:]- 34. Robichaud, I. and M.-A. Villard. 1999. Do Black-throated Green Warblers prefer conifers? Meso- and microhabitat use in a mixedwood forest. Condor 101:262- 27]. Robinson, S. K. and R. T. Holmes. 1982. Foraging behavior of forest birds: the relationships among search tactics, diet, and habitat structure. Ecological Applications 63:1918-1931. Robinson, S. K. P. i. e., editor. 1995. Threats to breeding neotropical birds in the Midwest. USDA Forest Service. Rodewald, A. D. and R. H. Yahner. 2001. Avian nesting success in forested landscapes: influence of landscape composition, stand and nest-patch microhabitat, and biotic interactions. The Auk 118:1018-1028. Rooney, T. P. 2001. Deer impacts on forest ecosystems: a North American perspective. Forestry 74:201-208. 139 Rooney, T. P. and D. M. Waller. 2003. Direct and indirect effects of white-tailed deer in forest ecosystems. Forest Ecology and Management 181:165-176. Rosenfield, G. H. and K. Fitzpatrick-Lins. 1986. A coefficient of agreement as a measure of thematic classification accuracy. Photogramrnetric Engineering and Remote Sensing 52:223-227. Rosenzweig, M. L. 1992. Species diversity gradients: we know more and less than we thought. Journal of Mammalogy 73:715-730. Rosenzweig, M. L. 1995. Species Diversity in Space and Time. Cambridge University Press, Cambridge. Rotenberry, J. T., R. J. Cooper, J. M. Wunderle, and K. G. Smith. 1993. Incorporating effects of natural disturbances in managed ecosystems. Pages 103-108 in D. M. Finch and P. W. Stangel, editors. Status and Management of Neotropical Migratory Birds. US. Department of Agriculture, Fort Collins. Saab, V. 1999. Importance of spatial scale to habitat use by breeding birds in riparian forests: A hierarchical analysis. Ecological Applications 9: 135-1 5 1. Schoener, T. W. 1968. Sizes of feeding territories among birds. Ecology 49: 123-141. Scott, J. M., F. Davis, B. Csuti, R. Noss, B. Butterfield, C. Groves, H. Anderson, S. Caicco, F. D'Erchia, T. C. Edwards, J. Ulliman, and R. G. Wright. 1993. Gap analysis: A geographic approach to protection of biological diversity. Wildlife Monographs 123:1-41. Seto, K. C., E. Fleishman, J. P. Fay, and C. J. Betrus. 2004. Linking spatial patterns of bird and butterfly species richness with Landsat TM derived NDVI. International Journal of Remote Sensing 25:4309 - 4324. Shi, H., E. J. Laurent, J. P. LeBouton, L. Racevskis, K. R. Hall, M. Donovan, R. V. Doepker, M. B. Walters, F. Lupi, and J. Liu. in press. Local spatial modeling of white-tailed deer distribution. Sisk, T. D. and N. M. Haddad. 2002. Incorporating the effects of habitat edges into landscape models: effective models for cross-boundary management. Pages 208- 240 in J. Liu and W. W. Taylor, editors. Integrating Landscape Ecology into Natural Resources Management. Cambridge University Press, New York, New York. Smith, R., M. Hamas, M. Dalhnan, and D. Ewert. 1998. Spatial variation in foraging of the Black-throated Green Warbler along the shoreline of northern Lake Huron. Condor 1002474-484. 140 Smith, T. M. and H. H. Shugart. 1987. Territory size variation in the ovenbird: the role of habitat structure. Ecology 68:695-704. Smyth, A., R. Mac Nally, and D. Lamb. 2002. Comparative influence of forest management and habitat structural factors on the abundances of hollow-nesting bird species in subtropical Australian eucalypt forest. Environmental Management 30:547-559. Space_lmaging_lnc. 2001. Integrated Forest Monitoring Assessment and Prescription: IF MAP. Pacific Meridian Resources, Ann Arbor, MI. Steele, B. B. 1992. Habitat selection by breeding Black-throated Blue Warblers at 2 spatial scales. Omis Scandinavica 23:33-42. Stoeckler, J., R. Strothman, and L. Krefting. 1957. Effect of deer on browsing on reproduction in the northern hardwood-hemlock type in northeastern Wisconsin. Journal of Wildlife Management 21:75-80. Stoms, D. M. and W. W. Hargrove. 2000. Potential NDVI as a baseline for monitoring ecosystem functioning. International Journal of Remote Sensing 21:401-407. Stouffer, P. C. and R. O. Bierregaard. 1995. Use of Amazonian forest fragments by understory insectivorous birds. Ecology 76:2429-2445. Stromayer, K. A. K. and R. J. Warren. 1997. Are overabundant deer herds in the eastern United States creating alternate stable states in forest-plant communities. Wildlife Society Bulletin 25:227-234. Tang, S. M. and E. J. Gustafson. 1997. Perception of scale in forest management planning: challenges and implications. Landscape and Urban Planning 39: 1 -9. Temple, S. A. 1986. Predicting impacts of habitat fragmentation on forest birds: a comparison of two models. Pages 301-304 in J. Verner, M. L. Morrison, and C. J. Ralph, editors. Wildlife 2000: Modeling Habitat Relationships of Terrestrial Vertebrates. University of Wisconsin Press, Madison. Temple, S. A. 1990. The role of dispersal in the maintenance of bird populations in a fragmented landscape. International Ornithological Congress 20:2298-2305. Thompson, F. R., S. J. Lewis, J. Green, and D. Ewert. 1993. Status of neotropical migrant landbirds in the Midwest: identifying species of management concern. Pages 145- 158 in D. M. Finch and P. W. Stangel, editors. Status and Management of Neotropical Migratory Birds. US. Department of Agriculture, Fort Collins, CO. 141 Thompson, F. R. I., S. K. Robinson, D. R. Whitehead, and J. D. Brawn. 1995. Management of central hardwood landscapes for the conservation of migratory birds. Pages 117-143 in F. R. Thompson 111, editor. Management of Midwestern Landscapes for the Conservation of Neotropical Migratory Birds. USDA Forest Service General Technical Report NC-187. Tilighman, N. G. 1989. Impacts of white-tailed deer on forest regeneration in northwestern Pennsylvania. Journal of Wildlife Management 53:524-532. Tilinghan, N. G. 1989. Impacts of white-tailed deer on forest regeneration in northwestern Pennsylvania. Journal of Wildlife Management 53. Todd, S. W. and R. M. Hoffer. 1998. Responses of spectral indices to variations in vegetation cover and soil background. Photogramrnetric Engineering and Remote Sensing 64:915-921. Townsend, P. A. 2000. A quantitative firzzy approach to assess mapped vegetation classifications for ecological applications. Remote Sensing of Environment 72:253-267. Trzcinski, M. K., L. Fahrig, and G. Merriam. 1999. Independent effects of forest cover and fragmentation on the distribution of forest breeding birds. Ecological Applications 9:586-593. Turner, M. G. 1989. Landscape ecology: The effect of pattern on process. Annual Review of Ecology and Systematics 20 171-197. Urban, D. L. 2002. Tactical monitoring of landscapes. Pages 294-311 in J. Liu and W. W. Taylor, editors. Integrating Landscape Ecology Into Natural Resources Management. Cambridge University Press, New York, New York. Van Deelen, T. R., K. S. Pregitzer, and J. B. Haufler. 1996. A Comparison of Presettlement and Present-Day Forests in Two Northern Michigan Deer Yards. American Midland Naturalist 135: 181-194. Van Horne, M. A. and T. M. Donovan. 1994. Ovenbird (Seirurus aurocapillus). Pages 1- 21 in A. Poole and F. Gill, editors. The Birds of North America. The Academy of Natural Sciences, Philadelphia. van Rensburg, B. J ., S. L. Chown, and K. J. Gaston. 2002. Species richness, environmental correlates, and spatial scale: a test using South Afiican birds. American Naturalist 159:566-577. Venne, L. J. 1965. Reproduction studies on penned white-tailed deer. Journal of Wildlife Management 29:74-79. 142 Villard, M.-A., G. Merriam, and B. A. Maurer. 1995. Dynamics in subdivided populations of neotropical migratory birds in a fragmented temperate forest. Ecology 76:27-40. Walsh, C. J ., P. J. Papas, D. Crowther, P. T. Sim, and J. Yoo. 2004. Storrnwater drainage pipes as a threat to a streamdwelling arnphipod of conservation significance, Austrogammarus australis, in southeastern Australia. Biodiversity and Conservation 13:781-793. Walters, E. L., E. H. Miller, and P. E. Lowther. 2002. Yellow-bellied Sapsucker (Sphyrapicus varius). in A. Poole and F. Gill, editors. The Birds of North America. American Ornithologist's Union, Philadelphia. Webb, W. L., W. K. Lauenroth, S. R. Szarek, and R. S. Kinerson. 1983. Primary production and abiotic controls in forests, grasslands, and desert ecosystems in the United States. Ecology 64:134-151. Werner, E. E. 1992. Individual behavior and higher-order species interactions. The American Naturalist 140:S5-S32. Whittaker, R. H. 1956. Vegetation of the Great Smoky Mountains. Ecological Monographs 26:1-80. Wiens, J. A. 1992. What is landscape ecology, really? Landscape Ecology 7:149-150. Wiens, J. A., B. Van Horne, and B. R. Noon. 2002. Integrating landscape structure and scale into natural resources management. Pages 23-67 in J. Liu and W. W. Taylor, editors. Integrating Landscape Ecology into Natural Resources Management. Cambridge University Press, New York. Williams, J. M. 1996. Nashville warbler. in A. Poole and F. Gill, editors. The Birds of North America. American Ornithologist's Union, Philadelphia. Wolf, A. T., R. W. Howe, and G. J. Davis. 1995. Detectability of forest birds from stationary points in northern Wisconsin. Pages 19-23 in S. Droege and C. J. Ralph, editors. Monitoring Bird Populations by Point Counts. USDA Forest Service General Technical Report PSW-GTR-149. Wolter, P. T., D. J. Mladenoff, G. E. Host, and T. R. Crow. 1995. Improved forest classifications in the northern lake states using multi-temporal Landsat imagery. Photogramrnetric Engineering and Remote Sensing 61:1129-1143. Wright, D. H. 1983. Species-energy theory: an extension of species-area theory. Oikos 4]. 143 Wright, D. H., D. J. Currie, and B. A. Maurer. 1993. Energy supply and patterns of species richness on local and regional scales. Pages 66-74 in R. E. Ricklefs and D. Schluter, editors. Species Diversity in Ecological Communities. University of Chicago Press, Chicago. Xie, J ., H. R. Hill, S. R. Winterstein, H. Campa, HI , R. V. Doepker, T. R. Van Deelen, and J. Liu. 1999. White-tailed deer management options model (DeerMOM): design, quantification, and application. Ecological Modelling 124:121-130. Zach, R. and J. B. Falls. 1979. Foraging and territoriality of male Ovenbirds (Aves: Parulidae) in a heterogeneous habitat. Journal of Animal Ecology 48:33-52. Zhang, L. and H. Shi. 2004. Local modeling of tree growth by geographically weighted regression. Forest Science 50:225-244. Zube, E. H. 1987. Perceived land use patterns and landscape values. Landscape Ecology 1:37-45. 144 lil'ljljjillfljljlliIlljlifll