UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Modelling and characterization of Steller sea lion haulouts and rookeries using oceanographic and shoreline.. 2005

You don't seem to have a PDF reader installed, try download the pdf

Item Metadata

Download

Media
ubc_2005-0157.pdf [ 9.65MB ]
Metadata
JSON: 1.0074885.json
JSON-LD: 1.0074885+ld.json
RDF/XML (Pretty): 1.0074885.xml
RDF/JSON: 1.0074885+rdf.json
Turtle: 1.0074885+rdf-turtle.txt
N-Triples: 1.0074885+rdf-ntriples.txt
Citation
1.0074885.ris

Full Text

Mode l l i ng and charac te r i za t ion of S te l le r s e a l ion hau lou ts and rooke r i es u s i n g o c e a n o g r a p h i c and sho re l i ne type data by STEPHEN S. BAN B.Sc. McGill University, 1999 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE in THE FACULTY OF GRADUATE STUDIES • .... . Zoology April 2005 UNIVERSITY OF BRITISH COLUMBIA © Stephen S. Ban, 2005 Abstract Steller sea lions range across the Pacific rim from Southern California in the east to northern Japan in the west, where they have continuously occupied terrestrial resting sites (haulouts) and breeding sites (rookeries) for hundreds of years, if not longer. Why they choose (and stay) at these locations, and what their preferred habitat is, remains unknown. Thus, two aspects of the Steller sea lion's habitat usage were examined—the oceanographic and the terrestrial. For the oceanographic aspect, spatial models were constructed to determine which oceanographic factors are associated with haulouts and rookeries, and how conditions near sites might differ from conditions elsewhere. The two modelling techniques employed (logistic regression and supervised classification) were evaluated using the kappa statistic (K n o), and receiver-operating characteristic (ROC) plots. Supervised classification was found to produce better-fitting models than logistic regression. In general, Steller sea lions showed preferences for sites associated with waters that were relatively shallow, well-mixed, had higher average tidal speeds and less-steep bottom slopes. Conditions within 1 nautical mile of land were better predictors of haulout and rookery locations than were conditions within 10, 20, and 50 nautical miles. No consistent differences were found in the physical characteristics of waters surrounding sites in the eastern and western populations of Steller sea lions, or between haulouts and rookeries. Regarding the terrestrial aspect of their habitat, anecdotal accounts describe Steller sea lions as predominantly occupying exposed, rocky shorelines, but this habitat preference has never been quantified. Locations of haulouts and rookeries were compared against a coastline type database to identify the shoreline preferences of Steller sea lions and to look for other spatial trends in site characteristics. Haulouts and rookeries were preferentially located on exposed rocky shorelines and wave-cut platforms. No relationship was found between either latitude or longitude of a site and its average non-pup count. The results indicate that there are differences in both the oceanographic and terrestrial characteristics of sites used by Steller sea lions versus areas of coastline where they are not found. The models could be used to predict changes in habitat use given changing physical conditions, and could be applied to any central-place forager. - iii - Table of Contents Abstract ii Table of Contents iv List of Figures v List of Tables vi List of Plates viii Acknowledgements ix Chapter 1 - Introductory Chapter 1 Background 1 Biogeography of haulouts and rookeries 3 Methodology 5 Assumptions and limitations 7 Hypotheses 9 References 10 Chapter 2 - Physical oceanographic characteristics associated with Steller sea lion haulouts and rookeries throughout their range in the North Pacific 15 Introduction 15 Methods 16 Results 25 Discussion 29 Summary 40 References •. 42 Chapter 3 - Quantification of Terrestrial Habitat Preferences of Steller Sea Lions 48 Introduction :....48 Methods 50 Results 55 Discussion 57 Summary 59 References 60 Chapter 4 - Conclusions 64 References 69 Appendix I - ROC values and K n o values for models at varying probability thresholds 80 Prediction of Rookery Locations - Mainland & Offshore 80 Prediction of Haulout Locations - Mainland & Offshore 82 Prediction of Rookery Locations - Offshore Only 84 Prediction of Haulout Locations - Offshore Only 86 Appendix II - Environmental Sensitivity Index (ESI) shoretypes 89 - iv - List of Figures Figure 2.1 Four physical variables used to characterize the waters surrounding Steller sea lion haulouts and rookeries 18 Figure 2.2 Example of training area mask, showing application of a 20 nautical mile buffer around Steller sea lion rookeries in the Gulf of Alaska and Eastern Aleutian Islands..! 20 Figure 2.3. Area under curve (AUC) values for models 27 Figure 2.4 K n o values for classification models, comparing models trained^ using eastern population sites vs. western population sites 28 Figure 2.5 Means and standard errors of habitat and non-habitat areas 31 Figure 3.1 Usage of shoreline types by haulouts and rookeries versus available shoreline across entire range 56 Figure 3.2 Latitude and longitude vs. average non-pup count of haulouts and rookeries 57 - v - List of Tables Table 3.1. NOAA ESI Shoreline types 52 Table 3.2. British Columbia Coastal Response and Oil Spill Atlas 53 Table 3.3. Conversion between BC Shoreline classification and NOAA ESI 54 Table A1.1. Kn 0 (Kappa) values for predicting rookery locations using rookeries as training areas for classification model 81 Table A1.2. ROC values predicting rookeries using rookeries as training areas.81 Table A1.3. Logistic model performance "(Kno) predicting rookeries using rookeries as training areas at varying probability thresholds 81 Table A1.4 K n o (Kappa) values for predicting haulout locations using rookeries as training areas 83 Table A1.5 K n o (Kappa) values for predicting haulout locations using haulouts as training areas 83 Table A1.6 ROC values predicting haulouts using rookeries as training areas...83 Table A1.7 ROC values predicting haulouts using haulouts as training areas. ...83 Table A1.8. Logistic model performance (Kno) predicting haulouts using rookeries as training areas 84 Table A1.9. Logistic model performance (Kno) predicting haulouts using haulouts as training areas 84 Table A1.10 K n o (Kappa) values for predicting rookery locations using rookeries as training areas 85 Table A1.11 ROC values predicting rookeries using rookeries as training areas.85 Table A1.12. Logistic model performance (K n o) predicting rookeries using rookeries as training areas 85 Table A1.13. Classification model performance (K n o) predicting haulouts using rookeries as training areas 86 Table A1.14. Classification model performance (Kno) predicting haulouts using haulouts as training areas 87 Table A1.15. ROC values predicting haulouts using rookeries as training areas.87 - vi - Table A1.16. R O C values predicting haulouts using haulouts as training areas. 87 Table A1.17. Logistic model performance (K n o ) predicting haulouts using rookeries as training areas 88 - vii - List of Plates Plate 1. ESI shoreline type 1 - exposed rocky shores 89 Plate 2. ESI shoreline type 2 - exposed rocky platforms 90 Plate 3. ESI shoreline type 3 - Fine-grained sand beaches 90 Plate 4. ESI shoreline type 4 - coarse-grained sand beaches 91 Plate 5. ESI shoreline type 5 - mixed sand and gravel beaches 91 Plate 6. ESI shoreline type 6a - gravel beaches 92 Plate 7. ESI shoreline type 6b - riprap structures 92 Plate 8. ESI shoreline type 7 - exposed tidal flats 93 Plate 9. ESI shoreline type 8a-sheltered rocky shores 93 Plate 10. ESI shoreline type 8b - Sheltered artificial structures 94 Plate 11. ESI shoreline type 9 - sheltered tidal flats 94 - viii - Acknowledgements I would like to thank my supervisor Andrew Trites (and his funders, the National Oceanographic and Atmospheric Administration and the North Pacific Universities Marine Mammal Research Consortium ) for generously providing a stipend for the duration of this research. I would also like to thank the members of my supervisory committee, Daniel Pauly and Mike Foreman, for all of their useful advice and input. I am also grateful for all of the members of the Marine Mammal Research Unit, but especial ly Arl iss Winship and Ruth Joy for all their help with the vagaries of S - P L U S and statistics in general, as well as Edward Gregr and Karin Bodtker for all of their help. Thanks also to my wife Natalie for wielding the carrot and stick equally effectively (if perhaps taking too much pleasure in wielding the stick). And this one's for you, Papa . - ix - Chapter 1 - Introductory Chapter Background Steller sea lions {Eumetopias jubatus, also known as northern sea lions) are the largest members of the Otariid family, and inhabit the northern Pacific Ocean from central California in the east, to northern Japan in the west (Loughlin et al. 1984). Since the late 1970s, there has been a decrease in their overall numbers by over 85% (Loughlin et al. 1992, Trites & Larkin 1996, Calkins et al. 1999). Following these declines, the National Marine Fisheries Service (NMFS) designated the Steller sea lion as a threatened species under the Endangered Species Act in 1990 (55 Fed Reg 12645, Apr 5 & 55 FR 49204, Nov 26). In 1997, two separate populations separated at 144° W longitude were recognized. The western population (Aleutian Islands and Gulf of Alaska), which had suffered a greater decline than the eastern population (SE Alaska, Washington, Oregon, California), was reclassified as endangered (62 FR 24345, May 5). In accordance with the requirements of the Endangered Species Act, NMFS established protective regulations and began the population recovery process. NMFS appointed a Steller Sea Lion Recovery Team in 1990 and published a Final Recovery Plan in 1992 that recommended management and research actions to aid the species' recovery. The recovery team was reconvened in 2000, and has been charged with drafting a new recovery plan. An important feature of the Endangered Species Act listing is the designation of the species' critical habitat. Critical habitat is defined as "areas...the loss of which would appreciably decrease the likelihood of the survival and recovery of a listed species or a distinct segment of its population..."(Littell 1992). NMFS designated Steller sea lion critical habitat as all major Steller sea lion rookeries and haulouts in Alaska, as well as terrestrial, air, and aquatic zones surrounding these sites. The terrestrial zone extends 3,000 feet (0.9 km) landward, the air zone extends 3,000 feet (0.9 km) above the terrestrial zone and the aquatic zone extends 20 - 1 - Introductory Chapter nm (37 km) seaward in State and Federally managed waters west of 144° W longitude (50 CFR § 226.202). "Major" rookeries and haulouts were defined as those with two hundred or more animals. In designating critical habitat, NMFS regarded conservation and management of prey resources and foraging areas as essential to the recovery of the Steller sea lion populations (58 Fed Reg 45269, Aug 27, 1993). In proposing the 20 nm aquatic zone, NMFS claimed that aquatic areas surrounding major rookeries and haulouts provided foraging habitats, prey resources and refuges that are not only essential to lactating females and juveniles, but also encompass areas for non- breeding animals year-round and for reproductively active animals during the non-breeding season. Although NMFS admitted that specific foraging sites and their constancy over time have not been well defined, it decided to rely on their ongoing studies using satellite telemetry and their findings of summertime foraging range of postpartum females occurring mainly in relative shallow waters within 20 nm of the rookeries (58 Fed Reg 17181, Apr 1, 1993). Such findings are consistent with the information provided in the Final Recovery Plan, which cited unpublished data of the National Marine Mammal Laboratory (NMFS 1992). The reasons for Steller sea lion population declines have not been determined (DeMaster & Atkinson 2002, Trites & Donnelly 2003) and population trends show considerable regional variation. Hunter & Trites (2001) identified twelve hypotheses for the Steller sea lion decline: competition with fisheries, juvenile mortality, nutritional stress (junk food and starvation), increased predation, intentional and accidental human kills, migration to other populations, regime shift, disease, pollution, trophic cascade (causing changes in prey availability), and adult mortality. Considerable research effort has focused on the possible effects on Steller sea lions of anthropogenic disruptions such as fishing (Springer 1992, Aydin 2002), incidental and direct mortality (Bigg 1985, Pascual & Adkison 1994, Perlov 1996, Shima et al. 2000), harassment, and disruption of rookery sites. Examining the spatial influence of oceanographic and other environmental factors is a relatively new area of research. To date, relatively little attention has Introductory Chapter been given to the relationship between physical and abiotic factors and Steller sea lion distributions, even though this approach is common with other species, both terrestrial and marine. There is a need to clarify what "critical habitat" is with respect to the Steller sea lion, and perhaps to go beyond protective measures based simply on fixed-radius buffer zones around rookeries. Boundaries of protected areas could instead be based on natural habitat boundaries such as bathymetric contours or other physical or biological oceanographic features. Biogeography of haulouts and rookeries Locations of rookeries and haulouts have been very stable historically, with some sites documented to have been in use for more than four centuries (Lyman 1988, Walker et al. 2000)The establishment of truly "new" rookeries (as opposed to recolonizations) appears to be relatively rare (Raum-Suryan 2002). The stability of site locations presumably indicates that they possess certain favourable characteristics—such as proximity to reliable foraging areas, protection from predators, or some other combination of factors. This in turn implies that the factors that have shaped the selection of haulout and rookery sites must also have been relatively stable over time, and resilient to interannual (or longer) shifts in environmental parameters. Steller sea lions tend to return to their rookeries of birth as adults, but the return rate is not 100% (Milette 1998, Raum-Suryan 2002). This propensity for natal site fidelity is undoubtedly another factor influencing the choice of rookeries, but it may also simply be the product of favourable conditions that have shaped the successful selection of sites, given that haulouts also appear to be as stable over time as rookeries. Furthermore, sea lions have returned to sites that had been previously abandoned or where they were extirpated many years beforehand (Bigg 1985, 1988). Thus, neither natal site fidelity nor conspecific attraction (Stamps 1988) can completely explain the distribution and stability of Steller sea lion rookeries and haulouts. Introductory Chapter Although Steller sea lions have been the subject of numerous (and extensive) directed kills and culls, very few, if any, sites have been abandoned, although some have reverted from rookeries to haulouts, and others have changed from haulouts to rookeries (P. Olesiuk, Fisheries and Oceans Canada, Nanaimo, pers. comm, L. Fritz, NMFS, Seattle, pers. comm, K. Pitcher, NMFS, pers. comm). The exceptions are at the extreme edges of the Steller's range, such as the Channel Islands (California) that have been abandoned (K. Ono, University of New England, unpubl. data), and in Japan, where there appears to have been a slight range contraction (V. Burkanov, NMFS, unpubl. data). Thus, while individual rookery and haulout sites are only occupied relatively briefly from a geological and evolutionary standpoint of the species, the usage and distribution of currently occupied sites—at least in the Bering Sea and Gulf of Alaska region—have been relatively stable and static for a substantial portion of the species' history in this area. Lyman (1989) quotes several anecdotal accounts of the habitat preferences of Steller sea lions, noting that they breed almost exclusively on rocky areas of offshore islands and that few mainland rookeries or haulouts are known. Steller sea lions are also noted to breed only on offshore islets and rocks, and do not habitually enter bays, estuaries, or river mouths—showing a preference for outer reefs and large offshore rocks. They are rarely found in inland waters and are considered a near-shore species. Bigg (1985) noted that year-round haulouts are usually found in places that are directly exposed to oceanic swells, whereas winter-only haulouts are generally not exposed directly to these swells, and are sheltered to some extent by the surrounding topography. Kastelein & Weltz (1991) studied two sites in Prince William Sound, Alaska, and observed that haulout behaviour was probably influenced by the physical geography of a colony site, particularly regarding variations in the number of animals at a site as tidal height changes. Fiscus (1970) also reported that sea lions preferred rookery beaches composed of sand, clay and small cobblestones or gravel over sections composed of boulders and large rocks. He also believed that they favoured large, Introductory Chapter fairly level rock ledges over boulder beaches. Rookeries may be selected according to slightly different criteria from haulouts, since protection of pups from exposure may be an important consideration. Edie (1977) and Fiscus (1970) reported the death of many pups due to wave action from storms—thus suggesting that sites affording some protection from wave exposure may be favoured for rookeries. Apart from anecdotal descriptions of their habitat preferences, there have been no quantitative studies to date of the factors - biotic or abiotic - that may influence the selection of these sites. My thesis thus has two main objectives: to determine and quantify what terrestrial factors influence Steller sea lion habitat selection, and to determine what physical oceanographic factors might be important in influencing this selection. Methodology Logistic regression has frequently been used to model habitat, but typically for terrestrial animals such as songbirds (Dettmers & Bart 1999), muskrats (Nadeau etal. 1995), and wolves (Mladenoff & Sickley 1998). In the aquatic environment, such models are common for fish (Beauchamp et al. 1992, Yu et al. 1995, Parrish et al. 1997, Diller & Wallace 1999, Knapp & Preisler 1999, Guay et al. 2000, Porter et al. 2000, Broad et al. 2001, Oberdorff et al. 2001, Mattingly & Galat 2002, Morita & Yamamoto 2002, Sato et al. 2002). Otherwise, there have been few attempts until very recently to model the habitats of other organisms in the pelagic environment using this approach (Moses & Finn 1997, Gregr & Trites 2001, Hamazaki 2002). Maintaining the spatial nature of data is central to the concept of a geographic information system (GIS). This allows for methodologies that combine an ecology-based spatial analysis with statistical treatment. A recent review of factors influencing sperm whale distribution emphasized the "absence of consideration of the spatial and temporal scales at which relevant oceanographic Introductory Chapter processes occur" and concluded that "multiscale studies . . . are needed" (p.55, Jaquet 1996). Hence, some of the reasons for using a GIS are: 1) Retaining the geographic characteristics of the data to produce a meaningful output (e.g. maps correlating oceanographic data with rookery site selection); 2) Statistical analysis of a spatial nature is necessary to determine the strength and nature of the relationships between various oceanographic variables such as sea surface (Sydeman & Allen 1999), upwellings (Su & Sheng 1999, Sydeman & Allen 1999), currents, and rookery site selection; 3) Ecologically based analysis ensures that any statistically derived relationships make sense within an ecological context. Oceanographic data such as coastline data and bathymetry have been used in conjunction with observer sightings and satellite data using GIS as a means to develop statistical models for cetaceans (Barber & Chavez 1983, Moses & Finn 1997), but such efforts are often complicated by the three-dimensional (actually, four-dimensional) nature of the marine environment, and by the elusive nature of purely aquatic animals. Due to their life history, pinnipeds present an interesting and useful (from a modeling standpoint) combination of characteristics: their need to breed and rest on land, coupled with their at-sea foraging. Furthermore, observer and telemetry data for Steller sea lions is particularly rich, and much research has gone into trying to understand and model their foraging habits. Consequently, it would be beneficial to both our understanding of Steller sea lions specifically, and of pinnipeds in general, to derive a quantitative method to identify potential habitats (e.g. sea lion rookeries) using the capabilities of a GIS. The approach described in Chapter 2 combines the use of a GIS to maintain spatial information with a statistical technique (logistic regression) and a technique adopted from remote sensing (supervised classification) to calculate a cell-by-cell prediction probability that relates to the likelihood of waters being near (or being used by) Steller sea lions. Introductory Chapter Assumptions and limitations To effectively model the habitat usage of Steller sea lions, there are some simplifying assumptions that need to be made and justified, and some statistical issues that need to be addressed related to these assumptions. The primary assumption being made in the models I developed is that the areas chosen for "presence" and "absence" are in fact reflective of the true distributions of Steller sea lions. For the oceanographic model presented in Chapter 2, this is a key distinction, because typically it is the presence or absence of individual animals that is being modeled—in my model, presence and absence refer not to individual animals, but to breeding and congregation sites that have proven to be stable over time. Therefore, the model assumptions centre around the use or importance of the waters surrounding a rookery or haulout rather than the presence or absence of individual animals within those waters. Secondly, the waters surrounding rookeries and haulouts could either be represented as an equally weighted grid within the desired search distance(s), without regard for proximity (i.e., cells close to shore have as much weight as those further away, as long as both are within the specified foraging distance), or they could be weighted based on their distance from shore. I chose to give all cells equal weight, because insufficient telemetry data exists on how Steller sea lions utilize the waters around rookeries and haulouts. The third assumption I made concerned the size of the rookery or haulout with which waters are associated, and whether this should impart a weight to the areas surrounding each site. Equal weighting was again chosen because the models are intended to reflect long-term habitat decisions that may not be reflected by count data that only includes the past 20-30 years. The primary difficulty with building a model pertaining to Steller sea lion habitat is that the geographic range is quite large, and crosses the jurisdictional boundaries of four countries (United States, Canada, Japan, and Russia). This made it Introductory Chapter difficult to find a consistent data set across the entire study area: quality and scale/resolution may differ depending upon the source of the data. Data that were available for most of the range (excluding Japan and the Commander Islands in Russia) were bathymetry, average tidal speeds, shoreline characterization, and the location of haulout and rookery (breeding) sites. Bathymetry is an important data set, as the continental shelf is thought to be a key factor in Steller sea lion foraging, and the accessibility of benthic prey may be dictated by the bottom depth. Additionally, features such as seamounts, underwater ridges, and other regions of topographic complexity are typically sites of high biological productivity. Slope can be derived from the bathymetric data. At large scales, slope often plays a role in determining where regions of upwelling occur, as currents are forced upwards by bottom topography. At finer scales, slope can be a proxy for rugosity (bottom complexity), which typically provides better quality habitat for benthic organisms (Watling & Norse 1998). Predators often make use of areas of high productivity that are associated with areas of high tidal mixing, fronts, or rips (Schneider et al. 1990, Suryan & Harvey 1998), so both average tidal speed and a calculated stratification parameter (Perry et al. 1983) may be useful in identifying these areas. All of the data I used were initially provided as point data, spaced as close as -500 m in near-shore areas and -1-2 km apart offshore. These data were interpolated using an inverse-distance weighting algorithm, which, although generally considered to be not as accurate as kriging, tends to be less computationally intensive - an important consideration when dealing with data sets of this size. Thus interpolated, the data then formed a contiguous raster grid fully populated with values estimated from the initial data sets. The problems associated with data interpolation and spatial autocorrelation are acknowledged in the literature (e.g. Gregr & Trites 2001), and remain contentious Introductory Chapter among the spatial modeling community. While they may not be fully resolvable, it is important to acknowledge the limitations and possible faults of a technique that has not yet fully matured. Hypotheses Site selection likely involves either an optimization or compromise of two factors: proximity to favourable at-sea foraging areas, and availability or accessibility of terrain that allows both ingress and egress during variable tidal heights. If Steller sea lions select sites solely on the criterion that they are close to productive at- sea foraging areas, then the current distribution of sites would be expected to correspond with the available distribution of shoreline types. Conversely, if only terrestrial characteristics of sites are important, the characteristics of waters surrounding rookeries and haulouts would not be expected to differ from waters that are not near Steller sea lion sites. The two principal chapters of my thesis address these two aspects of the Steller sea lion's habitat usage—the oceanographic and the terrestrial. In the first chapter, my primary hypothesis is that waters near Steller sea lion haulouts and rookeries are different from other coastal waters. Similarly, I expected the waters near rookeries to differ from those near haulouts. Finally, the utility and efficacy of two different techniques for generating habitat models were evaluated. In the second chapter, I tested two hypotheses concerning the terrestrial habitat needs of Steller sea lions. The first was that the distribution of Steller sea lion haulouts and rookeries is not random with respect to the availability of different shoreline types, and second, that rookeries differ from haulouts in the preferred type of shoreline, presumably because pups and young animals might be less agile than mature animals, and thus less able to access steep or rough sites. Finally, I compared the latitude and longitude, and nature of sites against average non-pup counts to determine whether a relationship existed between either of these factors and the success or popularity of a site. Introductory Chapter References Aydin K (2002) The Eastern Bering Sea. Fisheries Centre Research Reports 10(2), University of British Columbia, Vancouver, B.C. Barber RT, Chavez FP (1983) Biological consequences of El-Nino. Science 222:1203-1210 Beauchamp JJ , Christensen SW, Smith EP (1992) Selection of factors affecting the presence of brook trout (Salvelinus fontinalis) in Adirondack lakes: A case study. Can J Fish Aquat Sci 49:597-608 Bigg MA (1985) Status of the Steller Sea Lion Eumetopias jubatus and California Sea Lion Zalophus californianus in British Columbia Canada. Can Spec Publ Fish Aquat Sci 77:1-20 Bigg MA (1988) Status of the Steller sea lion, Eumetopias jubatus, in Canada. Can Field Nat 102:315-336 Broad TL, Townsend CR, Arbuckle C J , Jellyman DJ (2001) A model to predict the presence of longfin eels in some New Zealand streams, with particular reference to riparian vegetation and elevation. J Fish Biol 58:1098-1112 Calkins DG, McAllister DC, Pitcher KW (1999) Steller sea lion status and trend in Southeast Alaska: 1979-1997. Mar Mamm Sci 15:462-477 DeMaster D, Atkinson S (2002) Steller sea lion decline: Is it food II. AK-SG-02- 02, University of Alaska Sea Grant College Program Report Dettmers R, Bart J (1999) A GIS modeling method applied to predicting forest songbird habitat. Ecol App 9:152-163 Diller LV, Wallace RL (1999) Distribution and habitat of Ascaphus truei in streams on managed, young growth forests in north coastal California. J Herpetol 33:71-79 Edie AG (1977) Distribution and movements of Steller sea lion cows (Eumetopias jubata) on a pupping colony. M.Sc, University of.British Columbia Fiscus CH (1970) Steller sea lions at Ugamak Island, Aleutian Islands, Alaska. Department of the Interior, United States Fish and Wildlife Service, Bureau - 1 0 - Introductory Chapter of Commercial Fisheries, National Marine Mammal Laboratory, Seattle, Washington Gregr EJ , Trites AW (2001) Predictions of critical habitat for five whale species in the waters of coastal British Columbia. Can J Fish Aquat Sci 58:1265- 1285 Guay JC, Boisclair D, Rioux D, Leclerc C, Lapointe M, Legendre P (2000) Development and validation of numerical habitat models for juveniles of Atlantic salmon {Salmo salar). Can J Fish Aquat Sci 57:2065-2075 Hamazaki T (2002) Spatiotemporal prediction models of cetacean habitats in the mid-western North Atlantic Ocean (from Cape Hatteras, North Carolina, U.S.A. to Nova Scotia, Canada). Mar Mamm Sci 18:920-939 Hunter AMJ, Trites AW (2001) An annotated bibliography of scientific literature (1751-2000) pertaining to Steller sea lions (Eumetopias jubatus) in Alaska. Fisheries Centre Research Reports, University of British Columbia, Vancouver, B.C. Jaquet N (1996) How spatial and temporal scales influence understanding of Sperm Whale distribution: A review. Mammal Rev 26:51-65 Kastelein RA, Weltz FC (1991) Distribution and behaviour of Steller sea lions (Eumetopias jubatus) in Prince William Sound, Alaska, June 1989. Aquat Mamm 17:91-97 Knapp RA, Preisler HK (1999) Is it possible to predict habitat use by spawning salmonids? A test using California golden trout pncorhynchus mykiss aguabonita). Can J Fish Aquat Sci 56:1576-1584 Littell R (1992) Endangered and other protected species: Federal law and regulations. The Bureau of National Affairs:p.185 Loughlin TR, Perlov AS, Vladimirov VA (1992) Range-wide survey and estimation of total number of Steller sea lions in 1989. Mar Mamm Sci 8:220-239 Loughlin TR, Rush DJ, Fiscus CH (1984) Northern sea lion distribution and abundance: 1956-80. J Wildl Manage 48:729-740 Introductory Chapter Lyman RL (1988) Zoogeography of Oregon Coast marine mammals: The last 3,000 years. Mar Mamm Sci 4:247-264 Lyman RL (1989) Seal and sea lion hunting: A zooarchaeological study from the Southern Northwest coast of North America. J Anthrop Archaeol 8:68-99 Mattingly HT, Galat DL (2002) Distributional patterns of the threatened Niangua darter, Etheostoma nianguae, at three spatial scales, with implications for species conservation. Copeia 2002:573-585 Milette LL (1998) Reproductive life history, survival, and site fidelity of Steller sea lions (Eumetopias jubatus) in Alaska. Alaska Department of Fish and Game, Wildlife Conservation Division Mladenoff DJ, Sickley TA (1998) Assessing potential gray wolf restoration in the northeastern United States: A spatial prediction of favorable habitat and potential population levels. J Wildl Manage 62:1-10 Morita K, Yamamoto S (2002) Effects of habitat fragmentation by damming on the persistence of stream-dwelling charr populations. Consery Biol 16:1318-1323 Moses E, Finn J (1997) Using geographic information systems to predict North Atlantic right whale (Eubalaena glacialis) habitat. J Northwest Atl Fish Sci 22:37-46 Nadeau S, Decade R, Lambert D, St-Georges M (1995) Nonlinear modeling of muskrat use of habitat. J Wildl Manage 59:110-117 NMFS (1992) Recovery plan for the Steller sea lion (Eumetopias jubatus). NOAA, Silver Spring, MD. Oberdorff T, Pont D, Hugueny B, Chessel D (2001) A probabilistic model characterizing fish assemblages of French rivers: a framework for environmental assessment. Freshw Biol 46:399-415 Parrish FA, DeMartini EE, Ellis DM (1997) Nursery habitat in relation to production of juvenile pink snapper, Pristipomoides filamentosus, in the Hawaiian Archipelago. Fish Bull 95:137-148 - 1 2 - Introductory Chapter Pascual MA, Adkison MD (1994) The decline of the Steller sea lion in the Northeast Pacific: Demography, harvest or environment? Ecol App 4:393- 403 Perlov AS (1996) Catch as one of the main causes of decline in the abundance of Steller's sea lion. Tinro Tsentr, Vladivostok (Russia) Perry Rl, Dilke BR, Parsons TR (1983) Tidal mixing and summer plankton distributions in Hecate Strait, British Columbia. Can J Fish Aquat Sci 40:871-887 Porter MS, Rosenfeld J , Parkinson EA (2000) Predictive models of fish species distribution in the Blackwater drainage, British Columbia. N Am J Fish Manage 20:349-359 Raum-Suryan KL, Kenneth Pitcher, Donald G. Calkins, John L. Sease and Thomas R. Loughlin (2002) Dispersal, rookery fidelity, and metapopulation structure of Steller sea lions (Eumetopias jubatus) in an increasing and a decreasing population in Alaska. Mar Mamm Sci 18:746-764 Sato Y, Okabe T, Takebayashi H (2002) Prediction model of occurrence/non- occurrence of stream fishes in the Katsuura River, Tokushima Prefecture. Jpn J Ichthyol 49:41-52 Schneider DC, Harrison NM, Hunt GL, Jr. (1990) Seabird diet at a front near the Pribilof Islands, Alaska. Stud avian biol 14:61-66 Shima M, Hollowed AB, VanBlaricom GR (2000) Response of pinniped populations to directed harvest, climate variability, and commercial fishery activity: A comparative analysis. Rev Fish Sci 8:89-124 Springer AM (1992) A review: walleye pollock in the North Pacific - how much difference do they really make? Fish Oceanogr 1:80-96 Stamps JA (1988) Conspecific attraction and aggregation in territorial species. Am Nat 131:329-347 Su Y, Sheng Y (1999) Visualizing upwelling at Monterey Bay in an integrated environment of GIS and scientific visualization. Mar Geod 22:93-104 - 1 3 - Introductory Chapter Suryan RM, Harvey JT (1998) Tracking harbor seals (Phoca vitulina richardsi) to determine dive behavior, foraging activity, and haul-out site use. Mar Mamm Sci 14:361-372 Sydeman WJ, Allen S G (1999) Pinniped population dynamics in Central California: Correlations with sea surface temperature and upwelling indices. Mar Mamm Sci 15:446-461 Trites AW, Donnelly LP (2003) The decline of Steller sea lions in Alaska: A review of the nutritional stress hypothesis. Mammal Rev 33:3-28 Trites AW, Larkin PA (1996) Changes in the abundance of Steller sea lions (Eumetopias jubatus) in Alaska from 1956 to 1992: How many were there? Aquat Mamm 22:153-166 Walker PL, Kennett DJ, Jones TL, DeLong R (2000) Archaeological investigations at the Point Bennett pinniped rookery on San Miguel Island. In: Browne D, Mitchell K, Chaney H (eds). US Department of the Interior, Minerals Management Service, Pacific OCS Region 770 Paseo Camarillo Camarillo CA 93010-6064 USA Watling L, Norse EA (1998) Disturbance of the seabed by mobile fishing gear: A comparison to forest clearcutting. Conserv Biol 12:1180-1197 Yu S-L, Peters EJ , Stroup WW (1995) Application of logistic regression to develop habitat suitability criteria for sand shiner, Notropis stramineus. Rivers 5:22-34 Characteristics of Steller haulouts and rookeries Chapter 2 - Physical oceanographic characteristics associated with Steller sea lion haulouts and rookeries throughout their range in the North Pacific Introduction Steller sea lions {Eumetopias jubatus) range across the north Pacific from central California in the eastern Pacific to northern Japan in the west. They rest on land at haulouts, and breed on rookeries. There are 88 known rookeries range-wide, and nearly 600 haulouts. Steller sea lions have been divided into two genetically distinct stocks: the eastern stock, which lies east of 144°W, and the western stock, which lies west of this line. The eastern stock has been stable or increasing in numbers over the past several decades, whereas the western stock has experienced a sharp decline since the mid- to late-1970s, and has been listed as an endangered species in Alaska (Trites & Larkin 1996, NMFS 2001a). Locations of rookeries and haulouts have been very stable historically, with some sites documented to have been in use for more than four centuries (Lyman 1988, Walker etal. 2000). The stability of site locations presumably indicates that they possess certain favourable characteristics—such as proximity to reliable foraging areas, protection from predators, or some other combination of factors. This in turn implies that the factors that have shaped the selection of haulout and rookery sites must also have been relatively stable over time, and resilient to interannual (or longer) shifts in environmental parameters. Steller sea lions tend to return to their rookeries of birth as adults, but the return rate is not 100% (Milette 1998, Raum-Suryan 2002). This propensity for natal site fidelity is undoubtedly another factor influencing the choice of rookeries, but it may also simply be the product of favourable conditions that have shaped the successful selection of sites, given - 1 5 - Characteristics of Steller haulouts and rookeries that haulouts also appear to be as stable over time as rookeries. Furthermore, sea lions have returned to sites that had been previously abandoned or where they were extirpated many years beforehand (Bigg 1985, 1988). Thus, neither natal site fidelity nor conspecific attraction (Stamps 1988) can completely explain the distribution and stability of Steller sea lion rookeries and haulouts. The objective of my study was to determine whether stable site usage by Steller sea lions could be driven by correspondingly stable environmental cues, and whether these cues could be used to predict future changes in habitat usage should these cues change. My approach was to construct spatial models describing the physical characteristics associated with rookeries and haulouts (e.g. bathymetry, tidal speed, etc.). I used two different modelling approaches (logistic regression and supervised classification) to identify locations along the coast that share desireable features associated with sites currently used by Steller sea lions. Comparing selected or preferred sites ("habitat") with non-preferred ("non- habitat") should allow qualitative statements to be made about the habitat preferences of Steller sea lions—from which it might be possible to infer whether these differences represent a life-history strategy in an unpredictable and highly variable environment (Barber & Chavez 1983, Shima et al. 2000). My primary hypothesis was that waters near Steller sea lion haulouts and rookeries were different from other coastal waters. Similarly, I expected the waters near rookeries to differ from those near haulouts. Finally, the utility and efficacy of two different techniques for generating habitat models was evaluated. Methods Models of species-habitat association generally share similar methodologies in terms of input data sets, but differ with regards to how data are used to predict habitat. Most require data about where species have and have not been seen, while some require only presence data (Hirzel et al. 2002, Kery 2002). The most - 1 6 - Characteristics of Steller haulouts and rookeries common type of habitat models use discriminant analysis, general linear models, decision trees, or artificial neural networks. Where remote sensing data is available, techniques specific to these data sets (supervised/unsupervised classifications) are often applied (e.g. Luoto et al. 2002). I applied remote sensing multispectral classification techniques to the waters surrounding Steller sea lion rookeries and haulouts. Multispectral image classification can be divided into two broad categories: supervised and unsupervised. In an unsupervised classification, the number of desired output categories (classes) is typically specified. For example, in a satellite image there are typically a number of different layers (called bands), each contributed by a different sensor that is sensitive to a particular portion of the electromagnetic spectrum. When these sensors correspond to red, green, and blue, the combined result is a "true-colour" image, i.e., what we see in a normal colour photograph. Image classification uses algorithms to analyse each component band of a digital image to identify clusters of similar values within each band. In a supervised classification, areas on the image that have a known class membership (e.g. water or trees) can be selected and used as "training areas". An algorithm would then analyse the remaining pixels in the image to find those that are most similar in value (for each layer or parameter of interest) to the training areas. Both types of analysis (supervised and unsupervised) are aspatial with respect to pixel values (i.e., the spatial location of a pixel does not necessarily reflect which category or class into which it will be placed). Three main considerations limited my choice of predictor variables. First, the North Pacific study area covered a large geographical range. Second, a fairly high spatial resolution was required to distinguish sites in high-density areas (i.e. areas that had large numbers of rookeries and haulouts). Finally, it was important that factors (e.g. bathymetry and tidal speeds) be temporally invariant (i.e. insensitive to seasonal changes and climatic regime shifts). These considerations effectively narrowed the available data to just four sets: bathymetry, slope, mean tidal speed, and a mixing index (Figure 2.1). - 1 7 - Characteristics of Steller haulouts and rookeries Figure 2.1 Four physical variables used to characterize the waters surrounding Steller sea lion haulouts and rookeries. The diagonal line shows the division between the eastern and western populations of Steller sea lions, a) Bathymetry. Shallow areas are in light grey; deeper areas in dark grey, b) Slope: Areas in light grey have low slope; areas in dark grey have high slope, c) Mixing index: Areas in dark grey are well-mixed; areas in lighter grey are poorly-mixed (stratified), d) RMS tidal speed: Areas in light grey have high tidal speeds; areas in dark grey have low tidal speeds. Tidal speeds and bathymetry were obtained from M . G . G . Foreman (Institute of Ocean Sc iences, Sidney, B.C. , pers. comm); slope was derived from bathymetry; and the Simpson-Hunter mixing index (Simpson & Hunter 1974), or stratification parameter was derived from the depth and tidal speed as follows: 5 = log to where S is the stratification parameter, h is depth (in cm), U the instantaneous tidal velocity (cm s" 1), and Cd the bottom drag coefficient, which was set at 0.0025 - 1 8 - Characteristics of Steller haulouts and rookeries (Perry et al. 1983). This index indicates whether waters are well-mixed (S < 1.0), or stratified (S > 2.0). ArcView(ESRI 1992-1999) and ArcGIS (ESRI 1999-2002) were used to perform data input, interpolation, and projection (conversion from degrees of latitude and longitude to metres), while IDRISI32 software (Eastman 2002) was used to perform image analysis and develop the supervised classification models. S P S S (1997) was used to develop the logistic regression models, and to collect summary statistics. Data points were placed on a 600-meter grid and interpolated as necessary using an inverse-distance weighted method. This interpolation was required for three reasons. First, a meaningful slope cannot be calculated from isolated depth data points. Second, the data points from the depth and tidal speed layers do not overlay exactly, which would make calculation of the mixing index (derived from a combination of depth and tidal speed data) impossible unless this interpolation was performed; and third, image classification algorithms require a continuous image, not discrete points, in order to function. Cell size was based on the minimum distance between sample points in the tidal model, with each point also having an associated depth. The classification algorithm also required that all input data sets have the same resolution. Data were projected from latitude-longitude format to a Lambert projection, which ensured that area measurements remained constant across varying latitudes, since the size of a degree of latitude and longitude varies with position. Each data set had to be first converted into an 8-bit (256 value) format for input to the classification algorithm, which normally expects satellite images as input. Hence, data were log-transformed, then reclassified into 256 bins (0-255). Thus, all variables received equal weighting in the supervised classification algorithm. Rookery and haulout locations were obtained from a National Marine Fisheries database (C. Stinchcomb, NMFS, pers. comm.) 52 of 88 rookeries (59%) and 384 of the 594 haulouts (65%) in the database were included for analysis. The - 1 9 - Characteristics of Steller haulouts and rookeries remaining sites in northern Japan, Russia, and the Bering S e a were excluded because they fell outside of the range of available data. Training areas—which are used to inform or "train" the model by selecting areas of known classification—were developed for rookeries in both the western and eastern populations using the accepted division line at 144°W longitude. The same procedure was also applied to haulouts in a separate analysis, and to all "sites" combined (i.e. rookeries and haulouts). Four buffers or search distances of 1, 10, 20, and 50 nautical miles were used to delineate habitat. These varying distances effectively aggregated pixels at different scales, and thus allowed for analysis of scale- dependent effects. The four distances chosen represented typical approximate distances traveled by foraging animals (Merrick & Loughlin 1997), which vary according to sex, season, and age class. Adult males typically travel further than juveniles and adult females, with foraging distances for animals in winter being typically longer than in summer. There are also fishing restrictions within 10 and 20 nmi of certain Steller sites in A laska (NMFS 2001 b) that were implemented by fisheries managers in the belief that they represented important Steller sea lion habitat. Pixels contained within a given buffer distance of an existing site were considered as habitat, while pixels within the same distance of the coastline, but not within range of any sites were considered to be non-habitat. All other pixels, including land and anything further offshore than the search distance, were masked out and not considered in the analysis (Figure 2.2). Figure 2.2 Example of training area mask, showing application of a 20 nautical mile buffer around Steller sea lion rookeries in the Gulf of Alaska and Eastern Aleutian Islands. The dark grey area is a 20 nmi coastline buffer. - 2 0 - Characteristics of Steller haulouts and rookeries Sites were not weighted by population size, since the analysis was intended to represent long-term factors influencing habitat selection over time scales long enough to be insensitive to relatively rapid fluctuations in site-to-site animal numbers. Counts at individual sites often vary greatly from year to year (or from survey to survey), to the extent that even an averaged count for a site over the entire survey period would be unlikely to reflect the long-term "suitability" of a site. Additionally, reliable, comprehensive counts of animals only exist for the past 10-15 years, and thus would have skewed the analysis in favour of sites that have only recently had large numbers of animals, without regard for historical counts. The described training areas were processed by an image classification algorithm in IDRISI called BAYCLASS, or sampled for analysis in a logistic regression model in S P S S . Supervised classification model BAYCLASS is a "soft" classifier, meaning that, for each output pixel, the algorithm expresses a degree of probability that the pixel belongs to a particular class (in this case, "habitat" or "non-habitat"). Specifically, each output pixel contains the posterior probability of belonging to the "habitat" class: 5>(e|fc)*/>(*<) i where: p(h|e) is the probability of the hypothesis being true given the data (posterior probability), p(e|h) is the probability of finding that evidence given the hypothesis being true (this is calculated from the information provided by the training site data), and p(h) is the probability of the hypothesis being true regardless of the data (prior probability). Prior probabilities were set as 50% each for habitat and non-habitat—in effect, indicating no prior knowledge of how pixels would be allocated. A total of 16 -21 - Characteristics of Steller haulouts and rookeries models were generated (four buffer sizes for all sites in both the western and eastern stocks, plus one each for haulouts and rookeries separately). In order to evaluate the overall success of each habitat model, the resulting output probabilities were input to a "hardening" algorithm called MAXBAY, which produces a hard classification from the probability information. This required setting a probability threshold, below which pixels were considered to be "non- habitat". I initially set threshold values at 95% and lowered them in 5% increments (or raised them to 99%) until the kappa statistic was maximized. Probability thresholds ranged from 60% to 99%. Logistic Regression Model Training areas were encoded as 0 (absence: not within buffer distance of a site) or 1 (presence: within buffer distance of a site), and thus used as the dependent variable for the logistic regression model. Each cell containing a value for an oceanographic variable was then associated with a value of the dependent variable. The logistic models were developed using a forward stepwise likelihood ratio procedure in S P S S , which selected the variables with the most explanatory power first. The parameters for these variables were calculated in S P S S , and the values were passed back into ArcView to calculate predicted probabilities according to the logistic model: Vv(proximityto SSL site) = —-— \ + ex where: x - a + * xt i and a, and b-, represent the regression coefficients calculated for each of the *'s (independent variables) selected during the forward-stepwise selection process. This stepwise procedure was also used to validate the selection of variables -22 - Characteristics of Steller haulouts and rookeries used in the classification model rather than running a new classification model for each additional variable included. Validation The classifications generated by the models were quantitatively evaluated using the IDRISI VALIDATE and ROC functions. VALIDATE generates a series of kappa statistics (Pontius 2000) that assesses model fit. For any kappa statistic, a value of one indicates that a given classification is perfect, with any number above zero indicating that the observed (modelled) classification is better than that expected due to chance. Kappa values less than zero indicate that the model performs worse than a random model. A standard contingency table compares the number of observed versus expected cells or pixels in each category between a true (or ideal) output and a modelled output. However, standard measures of correlation are aspatial, and thus cannot distinguish between two models with exactly the same number of correctly classified cells in completely different spatial locations. Kappa statistics take into account both location and quantity when evaluating model performance. For my study, I used the K n 0 statistic rather than the Kstandard because the proportion of classes (habitat to non-habitat) was low. When this is the case, K s t a ndard tends to underestimate model performance versus chance. K n o is more appropriate in this situation because it rewards accurate estimates of quantity more appropriately than Kstandard (Pontius 2000). The performance of each model was thus evaluated in terms of both location and quantity of correctly classified pixels when compared to a theoretical "perfect" classification that was generated by buffering all rookeries (east and west) to the appropriate distance. The variance for each K n 0 value was calculated according to Couto (2003), and used to generate pair-wise (east versus west) z-scores to determine whether the differences in model performance were statistically significant. In the case of haulout predictions, the performance of models trained using rookeries was also evaluated against models trained using haulouts. - 2 3 - Characteristics of Steller haulouts and rookeries The ROC (Receiver Operating Characteristic) function generates a table with the percentage of true positives (sensitivity) versus false positives (specificity) for each model compared to an ideal reference model. ROC plots were originally used in clinical laboratory settings to determine the diagnostic accuracy and utility of tests and assays (Zweig & Campbell 1993), but have since been found useful for evaluating the success of classification models as well (Fielding & Bell 1997, Pontius & Schneider 2001, Poulin et al. 2002). The sensitivity and specificity were calculated for a number of user-specified thresholds. However, these thresholds were set by ranking pixels according to this percentile threshold (e.g. a 90% threshold means that the pixels with the highest 10% of probability values are selected), rather than by using a set probability value. Thus the thresholds used in VALIDATE and ROC were not directly comparable, although the endpoints (0% and 100% cutoffs) were the same in both cases. Calculating the area underneath the curve of an ROC plot provided an overall metric of model performance—the area under curve (AUC) statistic. A model with perfect discrimination (ability to distinguish positives from negatives) would have a plot that passes through the upper left corner of the two axes. A model that randomly assigns presence or absence to a pixel has an AUC value of 0.5 by definition, and is represented by a diagonal line from the lower-left to the upper- right of the plot. Thus, models with AUC values >0.5 are said to perform better than a random model. The AUC statistic is threshold independent, and can be used to get an overall picture of model performance. However, there are instances in which a trivial rule (one that assigns all pixels to one category or another, rather than at random) will outperform all other models according to the K n 0 statistic, but not on the AUC measure. This typically arises when the proportion of habitat to non-habitat pixels is very skewed in one direction or the other, and thus a high proportion of false- positives (or false negatives) can be outweighed by having 100% true positives. The ROC function in IDRISI also calculates K n o values at several user-specified thresholds; in cases where a trivial rule (equivalent to setting a threshold of 0% or - 2 4 - Characteristics of Steller haulouts and rookeries 100%) outperformed any other threshold, but the AUC value was over 0.5, the next-highest K n 0 value at a non-zero threshold was selected. If the AUC value was less than 0.5, no further analysis was conducted and the model was rejected. The AUC score was thus used as the initial criterion to select models that warranted further examination with the Kappa statistic. Prior to calculating the ROC and Kappa statistics, model outputs were first masked to exclude: (1) landmasses, (2) any distance further from shore than the original training data, and (3) areas that contained data that were extrapolated beyond the boundaries of the original input data. Additionally, a "terrestrial predator exclusion" mask was applied that effectively excluded the mainlands of Alaska and British Columbia, as a proxy for the existence of large terrestrial predators such as grizzly and black bears that may have favoured the selection of sites that were inaccessible to these animals (nearly all of the present sites are offshore). While there are no reports of terrestrial predation on Steller sea lions in the literature, the threat of predation may have been an historical influence on the present location of sites. While some near-shore islands may be within swimming distance of large terrestrial predators, detailed information about presence/absence of bears was not available, so exclusion of the mainland was used as a crude proxy. Results Summary of model results A total of 32 different models were generated (4 buffer distances x 2 populations x 2 site types x 2 methods). Rookery-trained models (both classification and logistic) were evaluated for their ability to predict both rookery locations alone as well as haulouts, while haulout models were only evaluated for their ability to predict haulout locations. All models were also tested both with and without the mainland-exclusion mask. - 2 5 - Characteristics of Steller haulouts and rookeries Classification models outperformed (or were not significantly different from) logistic models in every instance, with only two of the logistic models performing better than chance (according to ROC plots) at predicting rookery locations (50 nautical miles with eastern training sites, and 20 nautical miles with eastern training sites). Neither of these models performed better than the corresponding supervised classification models (p<0.001). Thus, only K n 0 and ROC scores for the classification models are presented here. Both classification and logistic models trained using rookery locations were often as successful at predicting haulout locations as models trained using haulouts. Thus, little additional information was gained by including haulouts in the training set. All of the classification models performed better than chance at predicting site locations, as indicated by AUC scores greater than 0.5 and K n o scores above 0 (Figures 2.3 and 2.4). Exclusion of the mainland to simulate avoidance of terrestrial predators did not generally improve model performance in predicting rookery locations. In six of eight cases (50, 20, and 10 nmi, east and western training areas), models that included the mainland outperformed those that excluded it ^-score, p<0.001, except for 10 nmi west, p<0.01). There was no significant difference (p>0.05) between the mainland and offshore models at the 1 nmi distance. There was no clear trend in model performance between models trained using sites from the western stock versus the eastern stock. In other words, the conditions surrounding sites in the western stock were no more representative of overall conditions than those in the eastern stock. - 2 6 - Characteristics of Steller haulouts and rookeries Mainland included Mainland excluded Wesierm sites Easts* n sites Rookeries as predictors of rookeries Rookeries as predictors of haulouts Haulouts as predictors of haulouts SO nmi 20 nmi 10 nmi I nmi 50 nmi 20ninl 10 nmi 1:nmi Buffer distance Buffer distance Figure 2.3. Area under curve (AUC) values for models. An AUC value of 0.5 (indicated by the dashed line) or below indicates a model performed worse than chance; higher scores represent better model fit. Scores are for single model runs and are presented for visual reference; refer to figure 2.4 to quantitatively compare model performance. -27- Characteristics of Steller haulouts and rookeries Mainland included Mainland excluded n.s. • • Western stock sites i: I Eastern stock sites Rookeries as predictors of rookeries as. 50 nmi 20 nmi 10nmi Irani B u f f e r d i s t a n c e n.s. 50 nmi 20 rani 10 nmi 1 nrnf B u f f e r d i s t a n c e Rookeries as predictors of haulouts Haulouts as predictors of haulouts Figure 2.4 K n o values for classification models, comparing models trained using eastern population sites vs. western population sites. Higher K n o values represent better model fits. Kn0 scores indicated by an asterix were highly significantly different (p<0.001), while those marked n.s. were not significantly different. -28- Characteristics of Steller haulouts and rookeries Oceanographic Conditions Near Rookeries and Haulouts The training area masks were also used to obtain comparative statistics for habitat and non-habitat areas for each of the input variables. In general, waters near sites (both rookeries and haulouts) were shallower, had less steep slopes and were more well-mixed (as indicated by a lower value' of the mixing index). They also had higher average tidal speeds than waters not near sea lion sites. However, there were some exceptions to these overall trends. For example, at the 1 and 10 nmi buffer distances, rookeries in the western stock were surrounded by waters that were slightly deeper than waters elsewhere in the west. Average tidal speeds at 1 nmi around eastern rookeries were also slightly lower than average tidal speeds elsewhere in the east, while the mixing index was the same as non-habitat areas in the east. These results are summarized in Figure 2.5. Discussion My study sought to find a relationship between physical oceanographic variables and the locations of terrestrial resting and breeding sites used by Steller sea lions, and to determine whether conditions around sites in the declining western stock differed from those in the stable eastern stock. I also compared the efficacy of two different modelling methods—supervised classification and logistic regression. There were no consistent differences in the physical conditions associated with rookeries and haulouts used by Steller sea lions in the western and eastern populations. However, waters near sites in both stocks (as defined by the 4 buffer distances) did consistently differ from waters not near sites. Sea lions appear to prefer sites near waters that are more well-mixed, with shallower depths, less steep bottom slopes, and higher average tidal speeds. Conditions within 1 nautical mile of shore also appear to be better predictors of haulout and rookery locations than conditions within 10, 20, or 50 nautical miles. Taken - 2 9 - Characteristics of Steller haulouts and rookeries together, these parameters may be indicative of areas where prey may be reliably obtained, and may indicate that preferred feeding areas are relatively close to shore. The present distribution of Steller sea lions is undoubtedly the product of a number of biological and historical factors, including interspecific competition, predation (and predation risk), and interactions with humans. However, physical oceanographic processes may also have had an influence on their distribution For example, studies of seabirds (Schneider et al. 1990) and tuna (Fiedler & Bernard 1987) have demonstrated the importance of fronts in concentrating prey, Schneider (1990) has also described the existence of a reliable front in the Bering Sea created by bathymetric features. Steller sea lions may well be using similar cues when selecting their terrestrial sites. This difference between potential habitat (available coastline) and actual habitat (individual sites) may be an example of the difference between fundamental and realized niches (Anderson et al. 2002). A fundamental niche constitutes the autecological requirements of a species—that is, the confluence of physical and biological factors that are necessary for the existence and survival of that species. The realized niche is a subset of the fundamental niche that reflects the influence of competition, predation, and other historical factors. Thus, the location of rookeries and haulouts of Steller sea lions probably reflects their realized niche. Based on the foregoing, ephemeral factors such as sea surface temperature and salinity would be expected to be less reliably and consistently associated with haulouts and rookeries compared to factors that vary little (if at all) over time, such as bathymetry, slope, and other characteristics that are relatively unaffected by periodic phenomena such as El Nino events. Physiographic features can concentrate prey through various means, such as by inducing upwelling, and have been shown to influence cetacean feeding behaviour (Baumgartner 1997). - 3 0 - Characteristics of Steller haulouts and rookeries Figure 2.5 Means and standard errors of habitat and non-habitat areas. The overall average includes all areas, both habitat and non-habitat, across both the eastern and western population areas. -31 - Characteristics of Steller haulouts and rookeries It thus seems reasonable to suggest that sea lions might also take advantage of these persistent features when foraging. The variables used in my study were consistent with this reasoning. Both supervised and unsupervised classification have become relatively commonplace for habitat modeling in terrestrial environments (e.g. Andries et al. 1994, Mladenoff & Host 1994, Kitron et al. 1996, Osborne etal . 2001, Luoto et al. 2002), but have only recently been applied to aquatic and marine environments (Stanbury & Starr 1999, Mumby & Edwards 2002, Andrefouet et al. 2003). The apparent lag in applying classification techniques to marine systems may reflect a number of differences between terrestrial and marine processes and. environments. First, marine environments have an extra spatial dimension as well as a temporal dimension that is more dynamic than that found in most terrestrial ecosystems. The abiotic environment (e.g. nutrient availability) in terrestrial ecosystems also remains functionally static for longer periods than in aquatic systems. Another difference is the boundaries of aquatic ecosystems, which are in constant flux, at least at fine scales. At larger scales, spatial patterns emerge among many aspects of marine ecosystems that are associated with prevailing winds, currents, and more permanent features such as seamounts, shelfbreaks, and other topographic features. Thus organisms have presumably evolved life histories and behaviours that operate at a number of different spatial and temporal scales, and with varying degrees of ability (Levin 1992). A number of different spatial and temporal considerations need to be addressed when modeling the habitat usage of Steller sea lions. For example, foraging occurs over short temporal and fine spatial scales (measured in minutes to hours and metres to tens of metres), and is influenced by currents and nutrient availability operating at larger scales (i.e. hours to days and hundreds of metres to kilometres). Similarly, reproductive success is simultaneously (and perhaps paradoxically) tied to fine-scale spatial factors such as foraging, beach substrate and wave exposure, as well as to long-term temporal factors such as seasonal changes and larger-scale spatial factors such as overall rookery locations within - 3 2 - Characteristics of Steller haulouts and rookeries the northern Pacific basin. The scales considered in my analysis ranged from 1- 50 nautical miles over long (annual to decadal) time scales. Hamazaki (2002) used a methodology similar to the one I used, but with a couple of key differences. In his study, the oceanographic data used to generate his model was from a short time period (1990-1996), and the model was tested using an even shorter data set (1997-1998). My model used the longest-term, highest resolution oceanographic data available for this region. Second, Hamazaki's model was dependent upon cetacean sighting data. Such data, as Hamazaki notes, cannot distinguish between when an animal is in its preferred versus its non-preferred habitat. He thus assumed that all areas with sightings of animals represented preferred habitat, whereas all areas with no sightings represented non-preferred habitat. The models I presented do not suffer from these particular assumptions regarding sightings, although other assumptions were made in their place, such as equal (potential) utilization of all waters surrounding a site, and equivalent importance of all selected sites. Logistic models vs. Supervised Classification Models Evaluating model performance revealed that all models developed using logistic regression either underperformed or performed no better than models developed using the supervised classification method. This finding indicates that image classification techniques should be considered as an alternative to logistic regression for generating habitat models, particularly when the behaviour and/or range of a species is well-defined. Logistic regression models have been commonly used for developing habitat models (Morrison et al. 1992), particularly in marine contexts (e.g. Nadeau et al. 1995, Moses & Finn 1997). One of the advantages of logistic regression over supervised classification is their relative ease of interpretation of the results. The use of forward or backward stepwise techniques allows models to be developed using only the variables with the most predictive power. In my analysis, this did not prove to be an important issue because all of the variables were selected for - 3 3 - Characteristics of Steller haulouts and rookeries inclusion in the logistic regression models. Nonetheless, the coefficients associated with each variable in the logistic regression equation also reflected the relative strength or predictive power of those variables. In the supervised classification method I used, all variables were weighted equally, and there was no straightforward way to determine afterwards which of the variables contributed the most to the final prediction, unless models were generated for each of the variables separately and then cross-compared. It is thus more difficult when using a supervised classification model to assess the relative importance and contributions of the input variables in the final model. Rookeries vs Haulouts Models that were first generated by training on rookery sites, and then tested against haulout sites, were as successful or better at predicting where haulouts occur as were the models generated by training on haulout sites, with the exception of the 50 nautical mile distance. This finding indicates that the oceanographic conditions surrounding haulouts were not drastically different from those surrounding rookeries, and that knowing the location of some or all of the rookeries may be sufficient to generate successful haulout prediction models. Comparing the mean values of the oceanographic parameters associated with habitat and non-habitat areas revealed a number of patterns. First, slope was consistently lower in waters near rookeries and haulouts versus other waters. Second, the mixing index was consistently lower. Third, in the eastern population, and in six of eight instances in the western stock, waters were consistently shallower in habitat areas. Finally, tidal speeds were higher in habitat versus non-habitat areas in 13 of 16 cases. This combination of higher average tidal speeds and lower values of the mixing index (indicating better- mixed waters) is consistent with waters that have higher biological productivities. The patterns in oceanographic parameters associated with rookeries and haulouts are consistent with the findings of Call and Loughlin (In press), who found that Steller sea lion rookeries were associated with major oceanic currents - 3 4 - Characteristics of Steller haulouts and rookeries and ecological boundaries (such as oceanic gyres and eddies). They also found that sites with the slowest rate of decline in the eastern population were associated with an area of high tidal flow and mixing (Unimak Pass). Although they also found that rookeries associated with deeper waters had a greater rate of decline, their analysis was confined mainly to rookeries within the western population, and only reflected population trends within that area from 1990-1998. Additionally, they only considered waters within 10 nautical miles of rookeries. The fact that sites are generally situated near shallower waters may indicate that access to benthic prey is important, and that the diving limitations of Steller sea lions may restrict foraging to relatively shallow waters. Common prey species such as Atka mackerel and walleye pollock (Merrick et al. 1997)—particularly smaller individuals—are often found at shallower depths (Wolotira et al. 1993). Loughlin et al. (2003) reported the maximum observed dive depth of a juvenile Steller sea lion was 328 m, with a mean maximum dive depth for all observed dives (young-of-the-year and juveniles) of 62.42 m. In another study, Merrick & Louglin (1997), observed that the maximum dive depth among adult females was >250m, although the exact depth reached was not reported. The maximum dive depth of adult males has not been determined, but the diving abilities of adult females and juvenile animals of both sexes are more likely to be limiting factors if access to benthic habitat is a factor in rookery (if not haulout) site selection. Interestingly, Loughlin et al (2003) found that the diving depths of some animals from Washington State (eastern population) were deeper than animals from the Aleutian Islands (western population). This pattern also appeared in the data presented here, with waters around sites in the eastern stock being generally deeper than those waters in the western stock (with the exception of the waters within 1 nmi of rookeries). This raises the question of whether animals in the eastern stock select sites near deeper waters because they are capable of diving to greater depths, or whether the waters near sites in the eastern stock are generally deeper and thus provided an evolutionary impetus to develop (or retain) greater diving abilities. Additional information about the relationship - 3 5 - Characteristics of Steller haulouts and rookeries between maximum observed dive and maximum available depth is required to resolve this question. The finding that areas near Steller sea lion terrestrial sites had lower bottom slopes seems contrary to the other trends in physical characteristics of surrounding waters. Areas of higher slope tend to be associated with regions of upwelling, at least at large scales (e.g., the continental shelf/slope). At smaller scales, high-slope areas may aggregate prey or make them easier for predators to find (Sjoberg & Ball 2000). However, these smaller-scale effects are also species-dependent, and some benthic or demersal prey species may be associated with lower-slope areas (Dean et al. 2000). Fine-scale changes in slope would also not be detectable with a 600-m cell size, which means that there may be areas of high slope or surface complexity within a cell that would not be apparent. There was no consistent difference between the waters surrounding rookeries and haulouts in average depth, slope, tidal speed, or mixing index across all scales (50, 20, 10, and 1 nautical mile radius) and between both stocks (western and eastern). However, in the western rookeries, tidal speeds were consistently higher, and the mixing index was lower (indicating more well-mixed waters) around rookeries compared to haulouts. The same pattern was observed in three out of the four distances examined in the eastern stock (50, 20, and 10 nautical miles) for tidal speed, and two distances (50 and 10 nmi) for the mixing index. No consistent differences were observed for depth or slope, either across all distances or between eastern and western stocks. While all of the differences were statistically significant, the absolute magnitude of the differences was typically quite small, and may not be biologically significant. However, even if using the uninterpolated, isolated datapoints (which would preclude the calculation of the slope and mixing index), the larger question remains as to what an appropriate "sampling" unit is, and whether it even makes sense to be speaking of "sampling" from a "population" of oceanographic data. In other words, how large of an area of ocean (and at what resolution) would we need to - 3 6 - Characteristics of Steller haulouts and rookeries measure before we are satisfied that we know the characteristics of a given water mass? Relative success of model predictions The highest kappa values were obtained for the smallest (1 nmi) buffer distance, which suggests the 1 nmi models performed better. However, the strength of the model fit may be overestimated because the ratio of non-habitat to habitat pixels increases dramatically as the buffer distance decreases. For example, the ratio of habitat to non-habitat pixels in the reference (true) image of 1 nmi buffered rookeries is over 2 0 0 0 : 1. Thus high K n o scores are largely the result of models more accurately predicting the quantity and location of non-habitat pixels. Correctly identifying the few habitat pixels contributes relatively little towards the overall score. Deciding upon the true "best" model depends on how much weight (or penalty) is given to false negatives (FN) versus false positives (FP). As Fielding & Bell ( 1 9 9 7 ) suggest, models with a higher FP rate may be acceptable for endangered species or other instances where a precautionary approach is desired. On the other hand, if we wish to find only those areas with the highest probability of finding animals, models with a low FP rate would be desirable. Thus, K n o measures of agreement were provided to assess model success assuming that FP and FN errors are given equal weighting, as well as ROC plots, which can be used as a threshold-independent means of assessing overall model success. Thus, models with the highest K n 0 score are not necessarily the "best" model because they are also likely to have a high proportion of false positives as the ratio of true positives to true negatives is reduced. It should be kept in mind that the correct way to interpret the resulting "predictions" of each model is to see each pixel in the map as being indicative of whether there is likely to be a rookery or haulout within x nautical miles of that pixel (where x is the buffer distance used to generate the model). The prediction maps, are, in essence, a prediction of what waters are likely to be near - 3 7 - Characteristics of Steller haulouts and rookeries haulouts/rookeries, not a precise indication of where the sites themselves are likely to be. This is because it is impossible to locate the specific point from which an area (the buffer distance) could have originated. Because K n o considers both the quantity (number of pixels) and quality (spatial location) of model predictions, a threshold that is too high may result in quantification "penalties" in the kappa statistic due to underestimation of pixel counts, whereas too low a threshold may result in a poor score due to overestimation of quantity. Conversely, a lower threshold may increase the K n 0 score of a given model by increasing the chances of obtaining the right location by virtue of number alone. This is certainly a factor at the smaller buffer distances, where the ratio of non-habitat to habitat pixels is high. Implications and future directions Climate change Climate change threatens to alter the structure and function of many ecosystems (Brereton et al. 1995). While my analysis deliberately used oceanographic predictors that were stable over long periods of time, a similar analysis could be conducted using predictors that vary over shorter time scales, or that are otherwise affected by climate change and regime shifts. These variables could include sea surface temperature, salinity, and mean sea level. Forcing the input variables to simulate conditions expected in new climate regimes would allow the models to predict future changes in the sites used by Steller sea lions. Constraints on range Environmental factors other than those I considered, such as ocean temperature, air temperature, or salinity, might have direct physiological impacts on Steller sea lions and might be more useful in determining the theoretical limits of the sea lion's range. The oceanographic predictors I used did not represent a smooth gradient from one end of the animal's range to the other, and there is no reason to suspect that either the depth, slope, tidal speed, or mixing characteristics were sufficiently different at either extreme of the range to inhibit further expansion. More likely, warmer - 3 8 - Characteristics of Steller haulouts and rookeries ocean temperatures, availability of favoured prey species, and competition with other species such California sea lions (Zalophus californianus) (Bartholomew & Boolotian 1960) dictate the maximum range extents of Steller sea lions. Differential weighting of sites If count data were available for sites from a longer time period—such as from archaeological data or historical records—then it would be logical to weight sites in the model according to abundance, to reflect the success or importance of a site. I did not do this because reliable count data were only available for a relatively recent period, and because no count data were available for sites in Washington, Oregon, or California. Additionally, because of the different population trajectories between the eastern and western stocks, it would be difficult to separate the effects of site characteristics from population effects without a reliable baseline count for each site. Conclusions My analysis suggests that there are indeed physiographic differences between the waters surrounding Steller sea lion rookeries and haulouts and those areas not occupied by Steller sea lions. These differences indicate that site selection may be driven by proximity to areas of reliable foraging that are generated by persistent tidal mixing, high tidal speeds, and accessible benthic prey (by virtue of depth). Marine mammals have been shown to make use of prey aggregations associated with these features (Brown & Winn 1989, Suryan & Harvey 1998, Keiper2002), and marine birds have been shown to consistently use tidal rips for foraging (Cairns & Schneider 1990, Hunt & Harrison 1990, Schneider 1991). Haulout sites in other pinnipeds such as harbour seals (Bjorge et al. 2002) and grey seals (McConnell et al. 1999) are also often situated near shallow areas. There were no consistent differences in the physical characteristics of waters surrounding sites in the eastern and western stocks. Therefore, it is unlikely that the particular physical characteristics examined here are related to the different population trends in the eastern and western stocks, especially since these factors are not subject to change over time. There were also no consistent - 3 9 - Characteristics of Steller haulouts and rookeries differences between haulouts and rookeries, indicating that there, does not appear to be anything special about the waters near breeding sites that would have led to their selection. Models developed using waters within 1 nautical mile of rookeries and haulouts tended to have higher K n o values than those using greater search distances. This may mean that waters closer to these sites are either more important for foraging or more characteristic of preferred habitat than waters further from shore. However, because the ROC values for models across distances did not differ greatly, the higher K n 0 scores may be more indicative of biases in this metric rather than actual differences in model fit, as discussed earlier. Nonetheless, the patterns observed in the average depth, slope, tidal speed, and mixing index (Figure 2.5) also support the interpretation that conditions within 1 nautical mile of rookeries and haulouts are different—and perhaps distinct—habitat. The techniques I used can be readily applied to other species, but are particularly relevant to central-place foraging species. For species whose range and distributions are not well-known, similar models could be built to predict likely, but unknown, habitat. The technique of supervised classification proved superior in this instance to logistic regression in terms of output accuracy, but provided only limited information about the relative importance of the input variables. This supervised classification technique warrants further exploration and evaluation in other habitat-modelling applications. Summary Spatial models were constructed to determine which oceanographic factors are associated with the terrestrial sites used by Steller sea lions to rest (at haulouts) and breed (at rookeries). Predictive models were generated to explain the locations of these sites across the North Pacific rim, and how the waters surrounding sites differ from waters elsewhere. The two modelling techniques employed (logistic regression and supervised classification) were evaluated - 4 0 - Characteristics of Steller haulouts and rookeries using the kappa statistics (K n o), and receiver-operating characteristic (ROC) plots. In general, Steller sea lions showed preferences for terrestrial sites associated with waters that were relatively shallow, well-mixed, had higher average tidal speeds and less-steep bottom slopes. Conditions within 1 nautical mile of land were better predictors of haulout and rookery locations than were conditions within 10, 20, and 50 nautical miles. No consistent differences were found in the physical characteristics of waters surrounding sites in the eastern and western populations of Steller sea lions. It is therefore unlikely that the particular physical oceanographic characteristics associated with the rookeries and haulouts examined are related to the different trends of the two populations, especially since these factors (bathymetry, slope, tidal speed, and mixing) are not subject to significant change across years. There were also no consistent differences between the oceanographic features associated with haulouts and rookeries, indicating that there does not appear to be anything special—at least in terms of the variables examined—about the waters near breeding sites that would have led to their selection. Supervised classification was found to produce better-fitting models than logistic regression, and could be readily applied to address habitat questions associated with other central-place foraging animals in the marine environment. -41 - Characteristics of Steller haulouts and rookeries References Anderson RP, Gomez-Laverde M, Peterson AT (2002) Geographical distributions of spiny pocket mice in South America: insights from predictive models. Global Ecology and Biogeography:131-141 Andrefouet S, Kramer P, Torres-Pulliza D, Joyce KE, Hochberg EJ , Garza-Perez R, Mumby P J , Riegl B, Yamano H, White WH, Zubia M, Brock JC, Phinn SR, Naseer A, Hatcher BG, Muller-Karger FE (2003) Multi-site evaluation of IKONOS data for classification of tropical coral reef environments. Remote Sens Environ 88:128-143 Andries AM, Gulinck H, Herremans M (1994) Spatial modelling of the barn owl Tyto alba habitat using landscape characteristics derived from SPOT data. Ecography 17:278-287 Barber RT, Chavez FP (1983) Biological consequences of El-Niho. Science 222:1203-1210 Bartholomew GA, Boolotian RA (1960) Numbers and population structure of the pinnipeds on the California Channel Islands. J Mammal 41:366-375 Baumgartner MF (1997) The distribution of Risso's dolphin (Grampas griseus) with respect to the physiography of the Northern Gulf of Mexico. Mar Mamm Sci 13:614-638 Bigg MA (1985) Status of the Steller Sea Lion Eumetopias jubatus and California Sea Lion Zalophus californianus in British Columbia Canada. Can Spec Publ Fish Aquat Sci 77:1-20 Bigg MA (1988) Status of the Steller sea lion, Eumetopias jubatus, in Canada. Can Field-Nat 102:315-336 Bjorge A, Bekkby T, Bryant EB (2002) Summer home range and habitat selection of harbor seal (Phoca vitulina) pups. Mar Mamm Sci 18:438-454 Brereton R, Bennett S, Mansergh I (1995) Enhanced greenhouse climate change and its potential effect on selected fauna of South-Eastern Australia: A trend analysis. Biol Conserv 72:339-354 - 4 2 - Characteristics of Steller haulouts and rookeries Brown CW, Winn HE (1989) Relationship between the distribution pattern of right whales, Eubalena glacialis, and satellite-derived sea surface thermal structure in teh Great South Channel. Cont Shelf Res 9:247-260 Cairns DK, Schneider DC (1990) Hot spots in cold water: feeding habitat selection by thick-billed murres. Stud avian biol 14:84-92 Call KA, Loughlin TR (In press) An ecological classification of Alaskan Steller sea lion (Eumetopias jubatus) rookeries: A tool for conservation/management. Fish Oceanogr Couto P (2003) Assessing the accuracy of spatial simulation models. Ecol Model 167:181-198 Dean TA, Haldorson L, Laur DR, Jewett SC, Blanchard A (2000) The distribution of nearshore fishes in kelp and eelgrass communities in Prince William Sound, Alaska: associations with vegetation and physical habitat characteristics. Environ Biol Fish 57:271-287 Eastman JR (2002) IDRISI32. Clark University Press, Worcester, MA ESRI (1992-1999) ArcView. Environmental Systems Research Institute, Redlands, California ESRI (1999-2002) ArcGIS. Environmental Systems Research Institute, Redlands, California Fiedler PC, Bernard HJ (1987) Tuna aggregation and feeding near fronts observed in satellite imagery. Cont Shelf Res 7:871-881 Fielding AH, Bell JF (1997) A review of methods for the assessment of prediction errors in conservation presence/absence models. Environ Conserv 24:38- 49 Genin A (2004) Bio-physical coupling in the formation of zooplankton and fish aggregations over abrupt topographies. Journal of Marine Systems 50:3- 20 Hamazaki T (2002) Spatiotemporal prediction models of cetacean habitats in the mid-western North Atlantic Ocean (from Cape Hatteras, North Carolina, U.S.A. to Nova Scotia, Canada). Mar Mamm Sci 18:920-939 - 4 3 - Characteristics of Steller haulouts and rookeries Hirzel AH, Hausser J , Chessel D, Perrin N (2002) Ecological-niche factor analysis: How to compute habitat-suitability maps without absence data? Ecology 83:2027-2036 Hunt GL, Jr., Harrison NM (1990) Foraging habitat and prey taken by least auklets at King Island, Alaska. Mar Ecol Prog Ser 65:141-150 Keiper CA (2002) Marine mammals off central California relative to hydrography: 1986-1994, 1997-1999, Kery M (2002) Inferring the absence of a species: A case study of snakes. J Wildl Manage 66:330-338 Kitron U, Otieno LH, Hungerford LL, Odulaja A, Brigham WU, Okello OO, Joselyn M, Mohamed-Ahmed MM, Cook E (1996) Spatial analysis of the distribution of tsetse flies in the Lambwe Valley, Kenya, using Landsat TM satellite imagery and GIS. Journal of Animal Ecology 65:371-380 Levin S (1992) The problem of pattern and scale in ecology: The Robert H. MacArthur Award lecture. Ecology 73:1943-1967 Loughlin TR, Sterling JT, Merrick RL, Sease JL, York AE (2003) Diving behavior of immature Steller sea lions (Eumetopias jubatus). Fish Bull 101:566-582 Luoto M, Kuussaari M, Toivonen T (2002) Modelling butterfly distribution based on remote sensing data. J Biogeogr 29:1027-1037 Lyman RL (1988) Zoogeography of Oregon Coast marine mammals: The last 3,000 years. Mar Mamm Sci 4:247-264 McConnell BJ, Fedak MA, Lovell P, Hammond PS (1999) Movements and foraging areas of Grey seal in the North Sea. J Appl Ecol 36:573-590 Merrick RL, Chumbley MK, Byrd GV (1997) Diet diversity of Steller sea lions (Eumetopias jubatus) and their population decline in Alaska: A potential relationship. Can J Fish Aquat Sci 54:1342-1348 Merrick RL, Loughlin TR (1997) Foraging behavior of adult female and young-of- the-year Steller sea lions in Alaskan waters. Can J Zool 75:776-786 Milette LL (1998) Reproductive life history, survival, and site fidelity of Steller sea lions (Eumetopias jubatus) in Alaska. Alaska Department of Fish and Game, Wildlife Conservation Division - 4 4 - Characteristics of Steller haulouts and rookeries Mladenoff DJ, Host G E (1994) Ecological perspective: Current and potential applications of remote sensing and GIS to ecosystem analysis. In: Sample VA (ed) Remote sensing and' GIS in ecosystem management. Island Press, Suite 300, 1718 Connecticut Avenue NW, Washington, D.C., USA, p 218-242 Morrison ML, Marcot BG, Mannan RW (1992) Wildlife-habitat relationships. University of Wisconsin Press, Madison Moses E, Finn J (1997) Using geographic information systems to predict North Atlantic right whale (Eubalaena glacialis) habitat. J Northwest Atl Fish Sci 22:37-46 Mumby PJ , Edwards AJ (2002) Mapping marine environments with IKONOS imagery: Enhanced spatial resolution can deliver greater thematic accuracy. Remote Sens Environ 82:248-257 Nadeau S, Decarie R, Lambert D, St-Georges M (1995) Nonlinear modeling of muskrat use of habitat. J Wildl Manage 59:110-117 NMFS (2001a) Endangered Species Act. Section 7 Consultation - Biological opinion on authorization of Gulf of Alaska and Bering Sea/Aleutian Islands groundfish fisheries based on the Fishery Management Plan for the Gulf of Alaska and Bering Sea/Aleutian Islands groundfish as modified by amendments 61 and 70. NMFS - Alaska Region, Protected Resources Division, Silver Spring, Maryland NMFS (2001b) Steller sea lion protection measures. Final supplemental Environmental Impact Statement. NMFS - Alaska Region, Protected Resources Division, Silver Spring, Maryland Osborne PE, Alonso JC, Bryant RG (2001) Modelling landscape-scale habitat use using GIS and remote sensing: A case study with great bustards. J Appl Ecol 38:458-471 Perry Rl, Dilke BR, Parsons TR (1983) Tidal mixing and summer plankton distributions in Hecate Strait, British Columbia. Can J Fish Aquat Sci 40:871-887 - 4 5 - Characteristics of Steller haulouts and rookeries Pontius RGJ (2000) Quantification error versus location error in comparison of categorical maps. Photogramm Eng Remote Sens 66:1011-1016 Pontius RGJ , Schneider LC (2001) Land-cover change model validation by an ROC method for the Ipswich watershed, Massachusetts, USA. Agric Ecosyst Environ 85:239-248 Poulin M, Careau D, Rochefort L, Desrochers A (2002) From satellite imagery to peatland vegetation diversity: How reliable are habitat maps? Conservation Ecology 6:16 Raum-Suryan KL, Kenneth Pitcher, Donald G. Calkins, John L. Sease and Thomas R. Loughlin (2002) Dispersal, rookery fidelity, and metapopulation structure of Steller sea lions (Eumetopias jubatus) in an increasing and a decreasing population in Alaska. Mar Mamm Sci 18:746-764 Schneider DC, Harrison NM, Hunt GL, Jr. (1990) Seabird diet at a front near the Pribilof Islands, Alaska. Stud avian biol 14:61-66 Schneider LC (1991) The role of fluid dynamics in the ecology of marine birds. Oceanography and Marine Biology Annual Review 29:487-521 Shima M, Hollowed AB, VanBlaricom GR (2000) Response of pinniped populations to directed harvest, climate variability, and commercial fishery activity: A comparative analysis. Rev Fish Sci 8:89-124 Simpson JH, Hunter JR (1974) Fronts in the Irish Sea. Nature 250:404-406 Sjoberg M, Ball JP (2000) Grey seal, Halichoerus grypus, habitat selection around haulout sites in the Baltic Sea: bathymetry or central-place foraging? Can J Zool 78:1661-1667 S P S S (1997) S P S S , Chicago, III. Stamps JA (1988) Conspecific attraction and aggregation in territorial species. Am Nat 131:329-347 Stanbury K, Starr R (1999) Applications of Geographic Information Systems (GIS) to habitat assessment and marine resource management. Oceanol Acta 22:699-703 - 4 6 - Characteristics of Steller haulouts and rookeries Suryan RM, Harvey JT (1998) Tracking harbor seals (Phoca vitulina richardsi) to determine dive behavior, foraging activity, and haul-out site use. Mar Mamm Sci 14:361-372 Trites AW, Larkin PA (1996) Changes in the abundance of Steller sea lions {Eumetopias jubatus) in Alaska from 1956 to 1992: How many were there? Aquat Mamm 22:153-166 Walker PL, Kennett DJ, Jones TL, DeLong R (2000) Archaeological investigations at the Point Bennett pinniped rookery on San Miguel Island. In: Browne D, Mitchell K, Chaney H (eds). US Department of the Interior, Minerals Management Service, Pacific OCS Region 770 Paseo Camarillo Camarillo CA 93010-6064 USA Wolotira RJ, Jr., Sample TM, Noel SF, Iten CR (1993) Geographic and bathymetric distributions for many commercially important fishes and shellfishes off the West Coast of North America, based on research survey and commercial catch data, 1912-84. U.S. Dep. Commer., NOAA Tech. Memo. NMFS-AFSC-6, 184p. Zweig MH, Campbell G (1993) Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine. Clinical Chemistry 39:561-577 i - 4 7 - Terrestrial habitat of Steller sea lions Chapter 3 - Quantification of Terrestrial Habitat Preferences of Steller Sea Lions Introduction Steller sea lions inhabit the north Pacific from central California in the eastern Pacific to northern Japan in the west, where they haul-out on land to breed at 88 known rookeries, and rest at nearly 600 non-breeding sites called haulouts. There do not appear to be archaeological records of Steller sea lion breeding or hauling out at sites (at least on the Oregon coast) prior to about 3,000 b.p. (Lyman 1989), although otariids have presumably lived in this region for nearly all of their evolutionary history (~ 3 million years). In the Aleutians, Alaska, and northern British Columbia, the ice edge at the end of the Last Glacial Maximum would have precluded the occupation of present-day sites until at least 20,000 b.p. (Manley & Kaufman 2002). Isolated fossil finds in British Columbia have placed Steller sea lions in this area as early as circa 12,000 b.p. (Harington et al. 2004), when the waters of British Columbia would have been closer in temperature to present-day Cook Inlet or Prince William Sound, Alaska. A number of people have tried to anecdotally describe characteristics of haulouts and rookeries. For example, Bigg (1985) notes that year-round haulouts are usually found in places that are directly exposed to oceanic swells, whereas winter-only haulouts are generally not exposed directly to these swells, and are sheltered to some extent by the surrounding topography. Lyman (1989) quotes several accounts of the habitat preferences of Steller sea lions, noting that they breed almost exclusively on rocky areas of offshore islands and that few mainland rookery or hauling areas are known. He also noted that Steller sea lions breed only on offshore islets and rocks, and do not habitually enter bays, estuaries, or river mouths—showing a preference for outer reefs and large offshore rocks. He also noted that they are rarely found in inland waters and considered them to be a near-shore species. Kastelein & Weltz (1991) studied - 4 8 - Terrestrial habitat of Steller sea lions two sites in Prince William Sound, Alaska, and observed that haulout behaviour was probably influenced by the physical geography of a colony site, particularly regarding variations in the number of animals at a site as tidal height changes. Fiscus (1970) also reported that sea lions preferred rookery beaches composed of sand, clay and small cobblestones or gravel over sections composed of boulders and large rocks. He also believed that they favoured large, fairly level rock ledges over boulder beaches. Rookeries may be selected according to slightly different criteria from haulouts, since protection of pups from exposure may be an important consideration. Edie (1977) and Fiscus (1970) reported the death of many pups due to wave action from storms—thereby suggesting that sites affording some protection from wave exposure may be favoured for rookeries. Considerable research has been undertaken on the haul-out behaviours of phocids (e.g. Pitcher & McAllister 1981, Schneider & Payne 1983, Calambokidis et al. 1987, Brasseur et al. 1996, Watts 1996, Suryan & Harvey 1998, Sjoberg & Ball 2000, Galimberti & Sanvito 2001, Bjorge et al. 2002, Moulton et al. 2002, Nordstrom 2002, Reder et al. 2003). However, research concerning the similar behaviour of otariids in general—and Steller sea lions in particular—with regard to terrestrial factors is sparse and generally qualitative. Apart from anecdotal descriptions of their habitat preferences, there have been no quantitative studies to date of the factors - biotic or abiotic - that may be influencing the selection of Steller sea lion haulouts and rookeries. Site selection by pinnipeds likely involves either an optimization or compromise of two factors: proximity to favourable at-sea foraging areas, and availability or accessibility of terrain that allows both ingress and egress during variable tidal heights. Protection from wind and waves may also play a role, although observations of sites (Fiscus 1970, Bigg 1985, Lyman 1989, Kastelein & Weltz 1991) indicates that this may not be important, and that favoured sites are actually more exposed than other nearby areas that are more sheltered. - 4 9 - Terrestrial habitat of Steller sea lions The fact that the same rookery and haulout sites have been consistently used by Steller sea lions historically, with some sites documented to have been in use for more than four centuries (Lyman 1988, Walker et al. 2000), indicates that the factors driving site selection are also likely to be stable. If sites are being selected solely on the criterion that they are close to productive at-sea foraging areas, then the current distribution of sites would be expected to correspond with the available distribution of shoreline types. While no clear reason has been found for the precipitous decline in the western stock of Steller sea lions (DeMaster & Atkinson 2002, National Research Council 2003, Trites & Donnelly 2003), differences in the terrestrial physical environment could conceivably affect population trajectories. In particular, low population numbers could be further depressed by such influencing factors as unusual pup mortality (due to exposure or trampling), reduced foraging success, or the inability of males to successfully find and defend territories. The following tests two principal hypotheses concerning the terrestrial habitat needs of Steller sea lions. The first was that the distribution of Steller sea lion haulouts and rookeries was not random with respect to the availability of different shoreline types—and second, that rookeries differ from haulouts in the preferred type of shoreline, presumably because pups and young animals might be less agile than mature animals, and thus less able to access steep or rough sites. Finally, the latitude and longitude, and nature of sites were compared against average non-pup counts to determine whether a relationship exists between either of these factors and the success or popularity of a site. Methods The hypotheses were tested using shoreline classification data from the National Oceanographic and Atmospheric Administration (NOAA) and the British Columbia Ministry of Sustainable Resource Management (MSRM). To aid in oil- spill response, NOAA has produced maps of shoreline types for most of the - 50- Terrestrial habitat of Steller sea lions coastal United States (National Oceanic and Atmospheric Administration (NOAA) Office of Response & Restoration (OR&R) Hazardous Materials Response Division 1997). Shorelines are characterized using a ranking system that considers characteristics such as substrate grain size, permeability, slope, exposure, and ease of cleanup, among others. The ranks range from 1-10, where type "1" shorelines are judged least susceptible to oil damage, and 10 are the most vulnerable to oil damage. Each rank also has subtypes that further characterize the shoreline type (Table 3.1). Depending upon whether a shoreline is estuarine, lacustrine, or riverine, the same ESI number may designate a slightly different habitat type with the same approximate oil spill vulnerability. The scale/accuracy of the final data is approximately 1:250 000, with a minimum mapping unit of approximately 100 feet. In British Columbia, a similar coastline atlas exists with a slightly different classification system (Coastal Resource and Oil Spill Response Atlas, Table 3.2). At the time my research was conducted, data for the Straight of Georgia was not publicly available. However, this area constitutes a relatively small portion of the overall British Columbia shoreline, and no rookeries or year-round haulouts exist in this area. Since there is no one-to-one correspondence between the NOAA ESI system and British Columbia's system; descriptions (and photos, where applicable) were used to make shore classifications from the British Columbia system compatible with the ESI system (Table 3.3) so that analyses could be performed on the entire shoreline from southern California to the Aleutian Islands as a single unit. One shoreline type from the B.C. system (channels) had no equivalent in the ESI system, so this was left as a separate class. Steller sea lion rookery and summer haulout locations in Alaska, Washington, Oregon, and California were compiled from the database of Steller sea lion counts conducted by the National Marine Fisheries Service (NMFS), as well as from V. Burkanov (Natural Resources Consultants, Inc, Seattle, Washington, pers. comm), M. Lowry (NMFS, La Jolla, California, pers. comm), - 51 - Terrestrial habitat of Steller sea lions Table 3.1. NOAA ESI Shoreline types. ESI # ESTUARINE LACUSTRINE RIVERINE (large rivers) 1 Exposed rocky shores with or without wave-cut platform (coastal) 1A Exposed rocky shores Exposed rocky shores Exposed rocky banks 1B Exposed, solid man-made structures Exposed, solid man- Exposed, solid man- made structures made structures 2 Exposed high-energy shoreline (unidentified cliffs, platforms, and beaches) 2A Exposed wave-cut platforms in Shelving bedrock Rocky shoals, bedrock, mud, or clay shores bedrock ledges 2B Exposed scarps and steep slopes in clay 3 Fine and medium sand, coarse sand, sand and gravel, gravel 3A Fine to medium-grained sand Eroding scarps in Exposed, eroding beaches unconsolidated banks in sediment unconsolidated sediments 3B Scarps and steep slopes in sand 3C Tundra cliffs 4 Coarse-grained sand beaches Sand beaches Sandy bars and gently sloping banks 5 Mixed sand and gravel beaches Mixed sand and Mixed sand and gravel beaches gravel bars and gently sloping banks 6 Gravel beaches and exposed riprap 6A Gravel beaches Gravel beaches Gravel bars and gently sloping banks 6B Riprap Riprap Riprap 7 Exposed tidal flats Exposed flats 8A Sheltered rocky shores and sheltered Sheltered scarps in scarps in bedrock, mud, or clay bedrock, mud, or clay 8B Sheltered, solid man-made structures Sheltered, solid man- Sheltered, solid man- made structures made structures 8C Sheltered riprap Sheltered riprap Sheltered riprap 8D Vegetated, steeply-sloping bluffs Vegetated, steeply- sloping bluffs 8E Peat shorelines 9A Sheltered tidal flats Sheltered sand/mud flats 9B Vegetated low banks Sheltered, vegetated Vegetated low banks low banks 10A Salt- and brackish-water marshes 10B Freshwater marshes Freshwater marshes Freshwater marshes 10C Swamps Swamps Swamps 10D Scrub-shrub wetlands Scrub-shrub wetlands Scrub-shrub wetlands 10E Inundated Low-lying Tundra U Unranked/Unsurveyed - 52- Terrestrial habitat of Steller sea lions Table 3.2. British Columbia Coastal Response and Oil Spill Atlas Coastal Class Repetitive Shore Type (#) Description 2,5 2 Rock Platform 1,3,4 3 Rock Cliff 6-10 4 Rock with Gravel Beach 11-15 5 Rock, Sand and Gravel Beach 16-20 6 Rock with Sand Beach 22 7 Gravel Beach 25 8 Sand and Gravel Beach 27,30 9 Sand Beach 28 10 Sand Flat 29 11 Mud Flat 31 12 Estuary, Marsh or Lagoon 32,33 13 Man-made 24,26 14 Sand and Gravel Flat 21,23 15 Gravel Flat 34 16 Channel and C. Stinchcomb (NMFS, La Jolla, California, pers. comm). Data for British Columbian sites were obtained from P. Olesiuk (Fisheries and Oceans Canada, Nanaimo, B.C., pers. comm). ESRI ArcView 3.2 (ESRI 1992-1999) and ArcGIS 8.3 (ESRI 1999-2002) were used to spatially analyze, display, and export data. Site locations and the coastline data were converted from latitude-longitude format to a common projection system to ensure that measurements of length were accurate. To allow for slight positional errors and map inaccuracies, only sites within 0.5 nautical miles (approx. 900 metres) of classified shoreline segments were selected for analysis. This yielded 318 haulouts and 38 rookeries out of a total of 594 haulouts and 88 rookeries, or 54% and 43% of all sites across the entire range, respectively. Using a spatial join operation, each site was then assigned to the nearest shoreline segment. In the ESI system, each shoreline segment may have multiple habitat types assigned to it, with numbers ordered from the most landward to the most seaward type. In cases where a site was assigned to a shoreline segment with multiple shoretypes (which only occurred with 14 of the haulouts and none of the rookeries), two separate analyses were conducted— one using the most landward shoretype, and one using the most seaward type. This allowed for possible changes in exposed shoretype as tidal heights rise and fall. - 5 3 - Terrestrial habitat of Steller sea lions Table 3.3. Conversion between BC Shoreline classification and NOAA ESI BC NOAA BC Shorezone description Type ESI 3 1. Exposed Rocky Shores Rock Cliff, narrow 1A. Exposed Rocky Shores (Estuarine) 2 2. Exposed Rocky Platforms (High Rock ramp, narrow energy shoreline) Rock ramp, wide Rock platform, narrow Rock platform, wide 2A. Rocky shoals, Bedrock Ledges (Estuarine) 6 3. Fine-grained sand beaches Sand beach 9 Ramp with sand beach, wide Ramp with sand beach, narrow Platform with sand beach, wide Cliff with sand beach (Rock with sand beach) (Sand beach) 4. Coarse-grained sand beaches 5 5. Mixed sand and gravel beaches Sand & Gravel Beach, narrow 8 Ramp with Gravel & Sand Beach, wide Ramp with Gravel & Sand Beach, narrow Platform with Gravel & Sand Beach, narrow Platform with Gravel & Sand Beach, wide Cliff with Gravel & Sand Beach (Rock with sand & gravel beach) (Sand and gravel beach) 4 6a. Gravel beaches Gravel beach, narrow 7 Ramp with gravel beach, wide Ramp with gravel beach, narrow Platform with gravel beach, wide Platform with gravel beach, narrow Cliff with gravel beach (Rock with gravel beach) (Gravel beach) 13 6b. Riprap structures Man-made, Permeable Man-made, Impermeable 10 7. Exposed Tidal flats Gravel flat, wide 11 Gravel Flat or Fan 14 Sand & Gravel Flat or Fan 15 (Sand Flat) (Mud flat) (Gravel flat) 8a. Sheltered rocky shores Estuaries 8b. Sheltered artificial structures 17 9. Sheltered tidal flats High tide lagoon 12 10a. Salt to brackish marshes Estuaries (Estuary wetland) 10b. Freshwater marshes 1 Undefined Undefined - 5 4 - Terrestrial habitat of Steller sea lions R (The R Foundation for Statistical Computing 2004) was used to perform statistical analyses. To test whether the usage of shoreline habitat types differed from a random distribution, a Monte Carlo implementation of Fisher's exact test with 10,000 iterations was used to randomly seed sites among shoretypes in proportion to their availability by length. The proportion of rookeries and haulouts in each shoreline type were also compared with Fisher's exact test to determine if habitat usage differed between rookeries and haulouts. Where available, average counts of non-pup animals were compared against the latitudinal and longitudinal position of the site to determine whether there was any correlation between location of sites and population. Population surveys spanned 1979-2002 for Alaskan sites (n = 316), and the years 1987, 1992, 1994, and 1998 for British Columbian sites (n = 30). Results Both haulouts and rookeries had a non-random distribution with respect to available shoreline types (p<0.05, Fisher's exact test, Figure 3.1). Steller sea lions heavily favoured Types 1 and 1A (exposed rocky shores), which accounted for over 70% of rookeries and more than 50% of haulouts. Substrate types 2 and 2A (exposed wave-cut platforms) were the second most frequently used, accounting for more than 13% of rookeries and 27% of haulouts. Shoreline types that were used in lower proportion than their availability included Types 3 (Fine to medium-grained sand beaches), 5 (mixed sand and gravel beaches), 6A (gravel beaches), and 8A (sheltered rocky shores). Types used in approximate proportion to their availability were 4 (Coarse-grained sand beaches), 6 (gravel beachesand exposed riprap), and 6B (riprap). No rookeries were found in Types 4, 6B, or 8A. Neither haulouts nor rookeries were found in shoretypes 10, 10D, 10E, 11, 1B, 1C, 2B, 3A, 3B, 3C, 6C, 7, 8, 8B ,8C ,8D, 8E, 9, 9A, or9B, although most of these types (with the exception of 7, 10A, and 9A) represented less than 5% of the available shoreline. - 55- Terrestrial habitat of Steller sea lions 70 60 O 50 • Ava i lab le % 3 Rooker ies % • Haulouts % ® 40 30 20 H 10 A 1A T 2 6A 2A 3 4 5 6 ESI Shore l ine Type 6B 8A u Figure 3.1 Usage of shoreline types by haulouts and rookeries versus available shoreline across entire range. Types 1, 1A, 2, and 2A appear to be heavily favoured, being used in greater proportion than their availability. Comparison of habitat usage between haulouts and rookeries showed that their distributions were also different (p<0.05, Fisher's exact test). Rookeries were located preferentially in Type 1 (exposed rocky shores with or without wave-cut platform) and 6A (gravel beaches) shoretypes, whereas haulouts were preferentially located in Types 1A (exposed rocky shores) and 2A (exposed wave-cut platforms). - 5 6 - Terrestrial habitat of Steller sea lions 3 ^ £ 2 3 O u a I 1 c o c O) fj o AftT-A#* A A A A AAAAVA Akt A A* ! ^ A ^ 4 A 4 ^ * A i > A A * * AAA A MML. A AAA A AjA A W e s t e r n p o p u l a t i o n V E a s t e r n p o p u l a t i o n -200 -190 -180 -170 -160 -150 -140 430 -120 48 Longitude SO 52 54 56 Latitude 56 60 62 Figure 3.2 Latitude and longitude vs. average non- pup count of haulouts and rookeries. No statistically significant relationship existed between longitude and counts. A weak negative correlation existed between latitude and counts. No relationship was found between longitude (^=0.003, p>0.05) and the average number of animals found at each site. Latitude was very weakly correlated (1^=0.070, p<0.001) with non-pup counts (Figure 3.2). Discussion Haulout and rookery sites were located mostly in exposed areas with solid or rocky substrates. Steller sea lions tended to avoid using sheltered areas and beaches with fine-grained substrates (such as mud and sand). This is consistent with Call & Loughlin (In press), who found that 38 of the 40 rookeries in their study had a rock/slab or cobble beach substrate, while only 2 had a sandy beach substrate. - 5 7 - Terrestrial habitat of Steller sea lions The lack of relationship between longitude and average non-pup counts is not surprising, given that the physical environment tends to differ less in an east-west direction than it does moving from north to south. Also, the western population is spread out across a large longitudinal gradient (144°W - 144°E), whereas the eastern population is concentrated within a comparatively narrow range of longitudes (144°W - 122°W). Given the differences in population trends between the two populations, this would tend to obscure any overall trend with longitude. There was a very weak relationship between latitude and average non-pup counts (only about 7% of the variance in counts could be explained by latitude), but this is likely an artifact of the fairly large sample size (n = 215), and of the generally declining population trend in the western population, which is situated in higher latitudes. Additionally, since data were missing from sites in northern Japan and Russia for the western population, and from California, Oregon, and Washington in the eastern population, the relationship between size of a site and latitude may not be strictly linear or monotonic. Many of the largest sites occur in the Gulf of Alaska and British Columbia, while sites in the southern portion of the eastern and southwestern portion of the western stock are generally smaller. Thus, including all sites would probably produce a bell-shaped curve rather than a strictly linear relationship, with the most populous sites occurring in moderate latitudes towards the northern limit of the eastern stock. Although there were differences in the distribution of haulouts and rookeries among shoreline types, there did not appear to be a preference for more sheltered shoretypes among rookeries. This is also consistent with the findings of Call and Loughlin (In press), who found that rookeries tended to be oriented towards the open ocean, rather than on the sheltered sides of islands. In light of reports of pup mortality during rough weather due to drowning (Fiscus 1970, Edie 1977, Cunningham & Stanford 1978), sites affording some degree of protection to exposure would be expected to be favoured. However, a number of reports (Kenyon & Rice 1961, Mathisen & Lopp 1963, Cunningham & Stanford 1978) also indicate that Steller sea lions do not associate land with safety during - 58- Terrestrial habitat of Steller sea lions a storm, and instead raft off-shore during severe weather events. Thus, the degree of shelter from exposure that a site affords may not be a consideration when the site is initially colonized. In addition, there may also be other factors driving the selection of such exposed sites, such as protection from terrestrial predation or proximity to favourable foraging areas. Higher-resolution terrestrial data detailing information such as the slope, aspect, substrate, and wave exposure of individual sites may reveal differences either between haulouts and rookeries or between western and eastern stocks that were not apparent at this scale of analysis. Anecdotal reports (Fiscus 1970, Bigg 1985, Lyman 1989, Kastelein & Weltz 1991) have described the preferences of Steller sea lions with regard to haulout and rookery locations, but no studies to date have quantified this preference across a broad geographic range. My findings confirm the anecdotal reports of habitat preferences. Such information about habitat preferences may prove useful in making management decisions that minimize the impacts of human development and disturbance, and forecast responses to climate change that may drive changes in sea lion distribution. Summary Steller sea lions range across the Pacific rim from Southern California in the east to northern Japan in the west, where they have continuously occupied terrestrial resting sites (haulouts) and breeding sites (rookeries) for hundreds of years, if not longer. Anecdotal accounts describe Steller sea lions as predominantly occupying exposed, rocky shorelines, but this habitat preference has never been quantified. We compared locations of haulouts and rookeries against a coastline type database to identify and quantify the shoreline preferences of Steller sea lions and to look for other spatial trends in site characteristics. Steller sea lions were found to preferentially locate haulouts and rookeries on exposed rocky shorelines and wave-cut platforms. Shoreline types that were used in lower - 59- Terrestrial habitat of Steller sea lions proportion than their availability included fine-to-medium-grained sand beaches, mixed sand and gravel beaches, gravel beaches, and sheltered rocky shores). No relationship was found between either latitude or longitude of a site and its average non-pup count. References Bigg MA (1985) Status of the Steller Sea Lion Eumetopias jubatus and California Sea Lion Zalophus californianus in British Columbia Canada. Can Spec Publ Fish Aquat Sci 77:1-20 Bjorge A, Bekkby T, Bryant EB (2002) Summer home range and habitat selection of harbor seal (Phoca vitulina) pups. Mar Mamm Sci 18:438-454 Brasseur S, Creuwels J , V D Werf B, Reijnders P (1996) Deprivation indicates necessity for haul-out in harbor seals. Mar Mamm Sci 12:619-624 Calambokidis J , Taylor BL, Cater SD, Steiger GH, Dawson PK, Antrim LD (1987) Distribution and haul-out behaviour of harbor seals in Glacier Bay, Alaska. Can J Zool 65:1391-1396 Call KA, Loughlin TR (In press) An ecological classification of Alaskan Steller sea lion (Eumetopias jubatus) rookeries: A tool for conservation/management. Fish Oceanogr Cunningham W, Stanford S (1978) Steller sea lion investigations at Cape St. Elias, March 22 through July 5, 1978. Alaska Department of Fish and Game, Anchorage, Alaska DeMaster D, Atkinson S (2002) Steller sea lion decline: Is it food II. AK-SG-02- 02, University of Alaska Sea Grant College Program Report Edie AG (1977) Distribution and movements of Steller sea lion cows (Eumetopias jubata) on a pupping colony. M.Sc, University of British Columbia ESRI (1992-1999) ArcView. Environmental Systems Research Institute, Redlands, California ESRI (1999-2002) ArcGIS. Environmental Systems Research Institute, Redlands, California - 6 0 - Terrestrial habitat of Steller sea lions Fiscus CH (1970) Steller sea lions at Ugamak Island, Aleutian Islands, Alaska. Department of the Interior, United States Fish and Wildlife Service, Bureau of Commercial Fisheries, National Marine Mammal Laboratory, Seattle, Washington Galimberti F, Sanvito S (2001) Modeling female haul-out in southern elephant seals (Mirounga leonina). Aquat Mamm 27:92-104 Harington CR, Ross RLM, Mathewes RW, Stewart KM, Beattie O (2004) A late Pleistocene Steller sea lion (Eumetopias jubatus) from Courtenay, British Columbia: its death, associated biota, and paleoenvironment. Can J Earth Sci 41:1285-1297 Kastelein RA, Weltz FC (1991) Distribution and behaviour of Steller sea lions (Eumetopias jubatus) in Prince William Sound, Alaska, June 1989. Aquat Mamm 17:91-97 Kenyon KW, Rice DW (1961) Abundance and distribution of the Steller sea lion. J Mammal 42:223-234 Lyman RL (1988) Zoogeography of Oregon Coast marine mammals: The last 3,000 years. Mar Mamm Sci 4:247-264 Lyman RL (1989) Seal and sea lion hunting: A zooarchaeological study from the Southern Northwest coast of North America. J anthrop Archaeol 8:68-99 Manley WF, Kaufman DS (2002) Alaska PaleoGlacier Atlas: Institute of Arctic and Alpine Research (INSTAAR), University of Colorado Mathisen OA, Lopp RJ (1963) Photographic census of the Steller sea lion herds in Alaska, 1956-58. Special Scientific Report on Fisheries 424, U.S. Fish and Wildlife Service Moulton VD, Richardson WJ, McDonald TL, Elliot RE, Williams MT (2002) Factors influencing local abundance and haulout behaviour of ringed seals (Phoca hispida) on landfast ice of the Alaskan Beaufort. Can J Zool 80:1900-1917 National Oceanic and Atmospheric Administration (NOAA) Office of Response & Restoration (OR&R) Hazardous Materials Response Division (1997) Kodiak Island and Shelikof Strait, Alaska; Bristol Bay, Alaska; Southern - 6 1 - Terrestrial habitat of Steller sea lions California; San Francisco Bay, California; Northern California; Western Alaska (Volume 1&2); Kodiak, Alaska; Southeast Alaska (Volume 1&2); Cook Inlet, Alaska; North Slope, Alaska; Aleutian Islands, Alaska; NW Arctic, Alaska (Volume 1&2); Prince William Sound, Alaska (Volume 1&2); Oregon and Washington Environmentally Sensitive Areas. NOAA OR&R, Seattle, WA, p ESI Maps, CD-ROM National Research Council (2003) The decline of the Steller sea lion in Alaskan Waters: Untangling food webs and fishing nets. National Academies Press, Washington, D.C. Nordstrom CA (2002) Haul-out selection by Pacific harbor seals (Phoca vitulina richardii): Isolation and perceived predation risk. Mar Mamm Sci 18:194- 205 Pitcher KW, McAllister DC (1981) Movements and haulout behavior of radio- tagged harbor seals, Phoca vitulina. Canadian Field Naturalist Reder S, Lydersen C, Arnold W, Kovacs KM (2003) Haulout behaviour of High Arctic harbour seals (Phoca vitulina vitulina) in Svalbard, Norway. Polar Biol 27:6-16 Schneider DC, Payne PM (1983) Factors affecting haul-out of harbor seals at a site in southeastern Massachusetts. J Mammal 64:518-520 Sjoberg M, Ball JP (2000) Grey seal, Halichoerus grypus, habitat selection around haulout sites in the Baltic Sea: bathymetry or central-place foraging? Can J Zool 78:1661 -1667 Suryan RM, Harvey JT (1998) Tracking harbor seals (Phoca vitulina richardsi) to determine dive behavior, foraging activity, and haul-out site use. Mar Mamm Sci 14:361-372 The R Foundation for Statistical Computing (2004) R Trites AW, Donnelly LP (2003) The decline of Steller sea lions in Alaska: A review of the nutritional stress hypothesis. Mammal Rev 33:3-28 Walker PL, Kennett DJ, Jones TL, DeLong R (2000) Archaeological investigations at the Point Bennett pinniped rookery on San Miguel Island. In: Browne D, Mitchell K, Chaney H (eds). US Department of the Interior, - 6 2 - Terrestrial habitat of Steller sea lions Minerals Management Service, Pacific OCS Region 770 Paseo Camarillo Camarillo CA 93010-6064 USA Watts P (1996) The diel hauling-out cycle of harbour seals in an open marine environment: correlates and constraints. Journal of Zoological Society London 240:175-200 - 6 3 - Conclusions Chapter 4 - Conclusions Chapter 2 investigated the differences in oceanographic conditions around rookeries and haulouts versus other areas of coastline where these sites did not occur. These differences indicate that site selection may be driven by proximity to areas of reliable foraging that are generated by persistent tidal mixing, high tidal speeds, and shallow bottoms (thus making prey more accessible). Marine mammals have been shown to make use of prey aggregations at these features (Brown & Winn 1989, Suryan & Harvey 1998, Keiper 2002), and marine birds have been shown to consistently use tidal rips for foraging (Cairns & Schneider 1990, Hunt & Harrison 1990, Schneider 1991). Haulout sites in other pinnipeds such as harbour seals (Bjorge et al. 2002) and grey seals (McConnell et al. 1999) are also often situated near shallow areas. Chapter 3 sought to investigate anecdotal reports (Fiscus 1970, Bigg 1985, Lyman 1989, Kastelein & Weltz 1991) that have described the terrestrial preferences of Steller sea lions with regard to haulout and rookery locations. Haulout and rookery sites were located mostly in exposed areas with solid or rocky substrates; sheltered areas and beaches with fine-grained substrates (such as mud and sand) tended to be avoided. This is consistent with (Call & Loughlin In press), who found that 38 of the 40 rookeries in their study had a rock/slab or cobble beach substrate, while only 2 had a sandy beach substrate. The findings of my research confirm the anecdotal reports of habitat preferences—Steller sea lions appear to prefer exposed, rocky shorelines over sheltered beaches. Although there were differences in the distribution of haulouts and rookeries among shoreline types, surprisingly there did not appear to be a preference for more sheltered shoretypes among rookeries. This is also consistent with the findings of Call & Loughlin (In press), who found that rookeries tended to be oriented towards the open ocean, rather than on the sheltered sides of islands. - 64- Conclusions In light of reports of pup mortality during rough weather due to drowning (Fiscus 1970, Edie 1977, Cunningham & Stanford 1978), one would expect that sites affording some degree of protection to exposure would be favoured. However, numerous reports (Kenyon & Rice 1961, Mathisen & Lopp 1963, Cunningham & Stanford 1978) also indicate that Steller sea lions do not associate land with safety during a storm, and instead raft off-shore during severe weather events; thus, the degree of shelter from exposure that a site affords may not be a consideration when the site is initially colonized. Additionally, there may be other factors driving the selection of such exposed sites, such as protection from terrestrial predation or proximity to favourable foraging areas. There was a weak relationship between latitude and average non-pup counts, but this is likely an artifact of the generally declining population trend in the western population, which is situated in higher latitudes. Additionally, since data were missing from sites in northern Japan and Russia for the western population, and from California, Oregon, and Washington in the eastern population, the relationship between size of a site and latitude may not be strictly linear. Many of the largest sites occur in the Gulf of Alaska and British Columbia, while sites in the southern portion of the eastern and southwestern portion of the western population are generally smaller. Thus, including all sites would probably produce a bell-shaped curve rather than a line, with the most populous sites in moderate latitudes towards the northern limit of the eastern population. Summary of findings My analysis suggests that there are indeed physiographic differences between the waters surrounding Steller sea lion rookeries and haulouts and those areas not occupied by Steller sea lions. Steller sea lions also appear to have a strong preferences with regard to the terrestrial characteristics of their breeding and resting sites. There were no consistent differences in the physical characteristics of waters surrounding sites in the eastern and western populations. Therefore, it is unlikely - 6 5 - Conclusions that the particular physical characteristics examined here are related to the different population trends in the eastern and western populations, especially since these factors are not subject to change over time. There were also no consistent differences between haulouts and rookeries, indicating that there does not appear to be anything special about the waters near breeding sites that would have led to their selection. Strengths and weaknesses Some strengths of the approach used in Chapter 2 are that it: (1) applies a modelling technique (supervised classification) in a novel context, (2) evaluates it against another technique (logistic regression) that is commonly used, and (3) did not require detailed information on the movements of individual animals or groups of animals. Although certain assumptions had to be made about the way in which animals utilized their habitat, the models took advantage of the central foraging strategy to obviate the need for information on individual animal movements. Unlike many other models of species-habitat associations, since the location of every haulout and rookery is known across the range of the animal, there was no need to account for areas that were unsurveyed or had low sighting probabilities, and there was no need to differentiate between "suspected" non- habitat areas (i.e., areas that are actually habitat but are categorized as non- habitat because animals were not sighted there) and actual non-habitat areas The weaknesses of this approach are, however: (1) Interpreting tests of statistical significance is difficult because the method artificially inflates the apparent sample size by, using a cell-based (raster) approach involving millions of grid cells, in which the values of each cell were interpolated from an original data set of hundreds of thousands of points. (2) Issues of spatial autocorrelation (the values in one cell are not independent of the values in adjacent or nearby cells). Another issue is that some of the data themselves were the result of a model; however, since the behaviour of tidal currents is well-described by deterministic physical equations, this may not be as problematic as it first seems, even if the - 6 6 - Conclusions resulting data is not a perfect or complete description of reality. (3) Other oceanographic factors besides the ones used (depth, slope, average tidal speed, and mixing index) may be more direct indicators of biological productivity and prey accessibility, or have more physiological relevance. Such variables may include chlorophyll concentration, water temperature, or salinity. (4) Data were not available for the entire range of rookeries and haulouts, so sites in the Commander Islands and northern Japan were not included. These areas could be included if similar oceanographic data become available in these regions. (5) Finally, the validation of the models could have been improved. Ideally, a random subset of each of the sites from each of the populations would be used to generate a model. This series of models could then be tested against the ideal model, providing an estimate of the average fit and error of each training set. This was not possible in this analysis due to the amount of processing time required to develop each model combined with the number of putative models to be evaluated. With regard to the terrestrial (shoreline) data analysis in Chapter 3, the strengths of this approach were: (1) the methods are very straightforward, and can be applied to any other species that occupies the mapped coastline without requiring expensive fieldwork and on-site surveys, as long as the home ranges of the species of interest are well known and/or characterised. (2) it uses a preexisting data set in which the quality of the data were exceptional in terms of both resolution and coverage. This allowed for characterizing nearly half (46%) of all sites according to shoretype. The weaknesses of this analysis were: (1) the data resolution was still insufficient to identify the shoretypes of sites that were located on very small offshore islands, and did not contain enough information to determine other potentially important factors such as slope, aspect, and degree of wave exposure. This data could only realistically be collected by high-resolution aerial photography or surveying of each site individually. (2) Complete count data were not available for conducting a population-weighted analysis, although historic count data (i.e. prior - 6 7 - Conclusions to circa 1950) are either unavailable or very sparse for most locations. Ideally, factors with little or no temporal variation (such as latitude, longitude, aspect, etc) should be correlated only with long-term averages of site populations to avoid biases due to recent changes in population status between populations and short-term shifts in population densities. (3) Shoreline type data were not available for sites in Russia and northern Japan, so these sites were not included in the analysis. Nonetheless, clear patterns were apparent with the remaining sites. If count data were available for sites from a longer time period—such as from archaeological data or historical records—then it would be logical to weight sites in the models according to abundance, to reflect the success or importance of a site. This was not done in this research because reliable count data were only available for a relatively recent period, and because no count data was available for sites in Washington, Oregon, or California. Additionally, because of the different population trajectories between the eastern and western populations, it would be difficult to separate the effects of site characteristics from population effects without a reliable baseline count for each site. Implications Climate change threatens to alter the structure and function of many ecosystems (Brereton et al. 1995). While this analysis deliberately used oceanographic predictors that are stable over long periods of time, a similar analysis could be conducted using predictors that varied over shorter time scales, or that are otherwise affected by climate change and regime shifts. These variables could include sea surface temperature, salinity, and mean sea level. Conducting such an analysis would then allow for predictions of changes in range by forcing of the input variables to simulate conditions expected in new climate regimes. Other environmental factors that have direct physiological impacts, such as ocean temperature, air temperature, or salinity, would be more useful than the factors that were available for this analysis in determining the theoretical limits of - 6 8 - Conclusions the Steller sea lion's range. The oceanographic predictors used here do not represent a smooth gradient from one end of the animal's range to the other, and there is no reason to suspect that either the depth, slope, tidal speed, or mixing characteristics are sufficiently different at either extreme of the range to inhibit further expansion. More likely, warmer ocean temperatures, availability of favoured prey species, and competition with other species such as Zalophus californianus (California sea lions) (Bartholomew & Boolotian 1960) are what dictate the maximum extents of Steller sea lions. The same techniques used in this analysis could also be applied to other species, but are particularly relevant to central-place foraging species. For species whose range and distributions are not well-known, similar models could be built to predict likely, but unknown, habitat. The technique of supervised classification proved superior in this instance to logistic regression in terms of output accuracy, but provides only limited information about the relative importance of the input variables. This technique warrants further exploration and evaluation in other habitat-modelling applications. References Anderson RP, Gomez-Laverde M, Peterson AT (2002) Geographical distributions of spiny pocket mice in South America: insights from predictive models. Global Ecology and Biogeography:131-141 Andrefouet S, Kramer P, Torres-Pulliza D, Joyce KE, Hochberg EJ , Garza-Perez R, Mumby P J , Riegl B, Yamano H, White WH, Zubia M, Brock JC, Phinn SR, Naseer A, Hatcher BG, Muller-Karger FE (2003) Multi-site evaluation of IKONOS data for classification of tropical coral reef environments. Remote Sens Environ 88:128-143 - 69- Conclusions Andries AM, Gulinck H, Herremans M (1994) Spatial modelling of the barn owl Tyto alba habitat using landscape characteristics derived from SPOT data. Ecography 17:278-287 Aydin K (2002) The Eastern Bering Sea. Fisheries Centre Research Reports 10(2), University of British Columbia, Vancouver, B.C. Barber RT, Chavez FP (1983) Biological consequences of El-Nino. Science 222:1203-1210 Bartholomew GA, Boolotian RA (1960) Numbers and population structure of the pinnipeds on the California Channel Islands. J Mammal 41:366-375 Baumgartner MF (1997) The distribution of Risso's dolphin (Grampas griseus) with respect to the physiography of the Northern Gulf of Mexico. Mar Mamm Sci 13:614-638 Beauchamp JJ , Christensen SW, Smith EP (1992) Selection of factors affecting the presence of brook trout (Salvelinus fontinalis) in Adirondack lakes: A case study. Can J Fish Aquat Sci 49:597-608 Bigg MA (1985) Status of the Steller sea lion Eumetopias jubatus and California sea lion Zalophus californianus in British Columbia, Canada. Can Spec Publ Fish Aquat Sci 77:1-20 Bigg MA (1988) Status of the Steller sea lion, Eumetopias jubatus, in Canada. Can Field-Nat 102:315-336 Bjorge A, Bekkby T, Bryant EB (2002) Summer home range and habitat selection of harbor seal (Phoca vitulina) pups. Mar Mamm Sci 18:438-454 Brasseur S, Creuwels J , V D Werf B, Reijnders P (1996) Deprivation indicates necessity for haul-out in harbor seals. Mar Mamm Sci 12:619-624 Brereton R, Bennett S, Mansergh I (1995) Enhanced greenhouse climate change and its potential effect on selected fauna of South-Eastern Australia: A trend analysis. Biol Conserv 72:339-354 Broad TL, Townsend CR, Arbuckle C J , Jellyman DJ (2001) A model to predict the presence of longfin eels in some New Zealand streams, with particular reference to riparian vegetation and elevation. J Fish Biol 58:1098-1112 - 70- Conclusions Brown CW, Winn HE (1989) Relationship between the distribution pattern of right whales, Eubalena glacialis, and satellite-derived sea surface thermal structure in teh Great South Channel. Cont Shelf Res 9:247-260 Cairns DK, Schneider DC (1990) Hot spots in cold water: feeding habitat selection by thick-billed murres. Stud avian biol 14:84-92 Calambokidis J , Taylor BL, Cater SD, Steiger GH, Dawson PK, Antrim LD (1987) Distribution and haul-out behaviour of harbor seals in Glacier Bay, Alaska. Can J Zool 65:1391-1396 Calkins DG, McAllister DC, Pitcher KW (1999) Steller sea lion status and trend in Southeast Alaska: 1979-1997. Mar Mamm Sci 15:462-477 Call KA, Loughlin TR (In press) An ecological classification of Alaskan Steller sea lion (Eumetopias jubatus) rookeries: A tool for conservation/management. Fish Oceanogr Couto P (2003) Assessing the accuracy of spatial simulation models. Ecol Model 167:181-198 Cunningham W, Stanford S (1978) Steller sea lion investigations at Cape St. Elias, March 22 through July 5, 1978. Alaska Department of Fish and Game, Anchorage, Alaska Dean TA, Haldorson L, Laur DR, Jewett SC, Blanchard A (2000) The distribution of nearshore fishes in kelp and eelgrass communities in Prince William Sound, Alaska: associations with vegetation and physical habitat characteristics. Environ Biol Fish 57:271-287 DeMaster D, Atkinson S (2002) Steller sea lion decline: Is it food II. AK-SG-02- 02, University of Alaska Sea Grant College Program Report Dettmers R, Bart J (1999) A GIS modeling method applied to predicting forest songbird habitat. Ecol App 9:152-163 Diller LV, Wallace RL (1999) Distribution and habitat of Ascaphus trueiln streams on managed, young growth forests in north coastal California. J , Herpetol 33:71-79 Eastman JR (2002) IDRISI32. Clark University Press, Worcester, MA - 71 - Conclusions Edie AG (1977) Distribution and movements of Steller sea lion cows (Eumetopias jubata) on a pupping colony. M.Sc, University of British Columbia ESRI (1992-1999) ArcView. Environmental Systems Research Institute, Redlands, California ESRI (1999-2002) ArcGIS. Environmental Systems Research Institute, Redlands, California Fiedler PC, Bernard HJ (1987) Tuna aggregation and feeding near fronts observed in satellite imagery. Cont Shelf Res 7:871-881 Fielding AH, Bell JF (1997) A review of methods for the assessment of prediction errors in conservation presence/absence models. Environ Conserv 24:38- 49 Fiscus CH (1970) Steller sea lions at Ugamak Island, Aleutian Islands, Alaska. Department of the Interior, United States Fish and Wildlife Service, Bureau of Commercial Fisheries, National Marine Mammal Laboratory, Seattle, Washington Galimberti F, Sanvito S (2001) Modeling female haul-out in southern elephant seals (Mirounga leonina). Aquat Mamm 27:92-104 Gregr EJ, Trites AW (2001) Predictions of critical habitat for five whale species in the waters of coastal British Columbia. Can J Fish Aquat Sci 58:1265- 1285 Guay JC, Boisclair D, Rioux D, Leclerc C, Lapointe M, Legendre P (2000) Development and validation of numerical habitat models for juveniles of Atlantic salmon {Salmo salar). Can J Fish Aquat Sci 57:2065-2075 Hamazaki T (2002) Spatiotemporal prediction models of cetacean habitats in the mid-western North Atlantic Ocean (from Cape Hatteras, North Carolina, U.S.A. to Nova Scotia, Canada). Mar Mamm Sci 18:920-939 Harington CR, Ross RLM, Mathewes RW, Stewart KM, Beattie O (2004) A late Pleistocene Steller sea lion (Eumetopias jubatus) from Courtenay, British Columbia: its death, associated biota, and paleoenvironment. Can J Earth Sci 41:1285-1297 - 72- Conclusions Hirzel AH, Hausser J , Chessel D, Perrin N (2002) Ecological-niche factor analysis: How to compute habitat-suitability maps without absence data? Ecology 83:2027-2036 Hunt GL, Jr., Harrison NM (1990) Foraging habitat and prey taken by least auklets at King Island, Alaska. Mar Ecol Prog Ser 65:141-150 Hunter AMJ, Trites AW (2001) An annotated bibliography of scientific literature (1751-2000) pertaining to Steller sea lions (Eumetopias jubatus) in Alaska. Fisheries Centre Research Reports, University of British Columbia, Vancouver, B.C. Jaquet N (1996) How spatial and temporal scales influence understanding of Sperm Whale distribution: A review. Mammal Rev 26:51-65 Kastelein RA, Weltz FC (1991) Distribution and behaviour of Steller sea lions (Eumetopias jubatus) in Prince William Sound, Alaska, June 1989. Aquat Mamm 17:91-97 Keiper CA (2002) Marine mammals off central California relative to hydrography: 1986-1994, 1997-1999, Kenyon KW, Rice DW (1961) Abundance and distribution of the Steller sea lion. J Mammal 42:223-234 Kery M (2002) Inferring the absence of a species: A case study of snakes. J Wildl Manage 66:330-338 Kitron U, Otieno LH, Hungerford LL, Odulaja A, Brigham WU, Okello OO, Joselyn M, Mohamed-Ahmed MM, Cook E (1996) Spatial analysis of the distribution of tsetse flies in the Lambwe Valley, Kenya, using Landsat TM satellite imagery and GIS. Journal of Animal Ecology 65:371-380 Knapp RA, Preisler HK (1999) Is it possible to predict habitat use by spawning salmonids? A test using California golden trout (Oncorhynchus mykiss aguabonita). Can J Fish Aquat Sci 56:1576-1584 Levin S (1992) The problem of pattern and scale in ecology: The Robert H. MacArthur Award lecture. Ecology 73:1943-1967 Littell R (1992) Endangered and other protected species: Federal law and regulations. The Bureau of National Affairs:p.185 - 7 3 - Conclusions Loughlin TR, Perlov AS, Vladimirov VA (1992) Range-wide survey and estimation of total number of Steller sea lions in 1989. Mar Mamm Sci 8:220-239 Loughlin TR, Rush DJ, Fiscus CH (1984) Northern sea lion distribution and abundance: 1956-80. J Wildl Manage 48:729-740 Loughlin TR, Sterling JT, Merrick RL, Sease JL, York AE (2003) Diving behavior of immature Steller sea lions (Eumefop/as jubatus). Fish Bull 101:566-582 Luoto M, Kuussaari M, Toivonen T (2002) Modelling butterfly distribution based on remote sensing data. J Biogeogr 29:1027-1037 Lyman RL (1988) Zoogeography of Oregon Coast marine mammals: The last 3,000 years. Mar Mamm Sci 4:247-264 Lyman RL (1989) Seal and sea lion hunting: A zooarchaeological study from the Southern Northwest coast of North America. J anthrop Archaeol 8:68-99 Manley WF, Kaufman DS (2002) Alaska PaleoGlacier Atlas: Institute of Arctic and Alpine Research (INSTAAR), University of Colorado Mathisen OA, Lopp RJ (1963) Photographic census of the Steller sea lion herds in Alaska, 1956-58. Special Scientific Report on Fisheries 424, U.S. Fish and Wildlife Service Mattingly HT, Galat DL (2002) Distributional patterns of the threatened Niangua darter, Etheostoma nianguae, at three spatial scales, with implications for species conservation. Copeia 2002:573-585 McConnell BJ, Fedak MA, Lovell P, Hammond PS (1999) Movements and foraging areas of Grey seal in the North Sea. J Appl Ecol 36:573-590 Merrick RL, Loughlin TR (1997) Foraging behavior of adult female and young-of- the-year Steller sea lions in Alaskan waters. Can J Zool 75:776-786 Milette LL (1998) Reproductive life history, survival, and site fidelity of Steller sea lions (Eumetopias jubatus) in Alaska. Alaska Department of Fish and Game, Wildlife Conservation Division Mladenoff DJ, Host G E (1994) Ecological perspective: Current and potential applications of remote sensing and GIS to ecosystem analysis. In: Sample VA (ed) Remote sensing and GIS in ecosystem management. Island - 74- Conclusions Press, Suite 300, 1718 Connecticut Avenue NW, Washington, D.C., USA, p 218-242 Mladenoff DJ, Sickley TA (1998) Assessing potential gray wolf restoration in the northeastern United States: A spatial prediction of favorable habitat and potential population levels. J Wildl Manage 62:1-10 Morita K, Yamamoto S (2002) Effects of habitat fragmentation by damming on the persistence of stream-dwelling charr populations. Conserv Biol 16:1318-1323 Morrison ML, Marcot BG, Mannan RW (1992) Wildlife-habitat relationships. University of Wisconsin Press, Madison Moses E, Finn J (1997) Using geographic information systems to predict North Atlantic right whale [Eubalaena glacialis) habitat. J Northwest Atl Fish Sci 22:37-46 Moulton VD, Richardson WJ, McDonald TL, Elliot RE, Williams MT (2002) Factors influencing local abundance and haulout behaviour of ringed seals (Phoca hispida) on landfast ice of the Alaskan Beaufort. Can J Zool 80:1900-1917 Mumby PJ , Edwards AJ (2002) Mapping marine environments with IKONOS imagery: Enhanced spatial resolution can deliver greater thematic accuracy. Remote Sens Environ 82:248-257 Nadeau S, Decarie R, Lambert D, St-Georges M (1995) Nonlinear modeling of muskrat use of habitat. J Wildl Manage 59:110-117 National Oceanic and Atmospheric Administration (NOAA) Office of Response & Restoration (OR&R) Hazardous Materials Response Division (1997) Kodiak Island and Shelikof Strait, Alaska; Bristol Bay, Alaska; Southern California; San Francisco Bay, California; Northern California; Western Alaska (Volume 1&2); Kodiak, Alaska; Southeast Alaska (Volume 1&2); Cook Inlet, Alaska; North Slope, Alaska; Aleutian Islands, Alaska; NW Arctic, Alaska (Volume 1&2); Prince William Sound, Alaska (Volume 1&2); Oregon and Washington Environmentally Sensitive Areas. NOAA OR&R, Seattle, WA, p ESI Maps, CD-ROM - 75- Conclusions National Research Council (2003) The decline of the Steller sea lion in Alaskan Waters: Untangling food webs and fishing nets. National Academies Press, Washington, D.C. NMFS (1992) Recovery plan for the Steller sea lion (Eumetopias jubatus). NOAA, Silver Spring, MD. NMFS (2001a) Endangered Species Act. Section 7 Consultation - Biological opinion on authorization of Gulf of Alaska and Bering Sea/Aleutian Islands groundfish fisheries based on the Fishery Management Plan for the Gulf of Alaska and Bering Sea/Aleutian Islands groundfish as modified by amendments 61 and 70. NMFS - Alaska Region, Protected Resources Division, Silver Spring, Maryland NMFS (2001b) Steller sea lion protection measures. Final supplemental Environmental Impact Statement. NMFS - Alaska Region, Protected Resources Division, Silver Spring, Maryland Nordstrom CA (2002) Haul-out selection by Pacific harbor seals (Phoca vitulina richardii): Isolation and perceived predation risk. Mar Mamm Sci 18:194- 205 Oberdorff T, Pont D, Hugueny B, Chessel D (2001) A probabilistic model characterizing fish assemblages of French rivers: a framework for environmental assessment. Freshw Biol 46:399-415 Osborne PE, Alonso JC , Bryant RG (2001) Modelling landscape-scale habitat use using GIS and remote sensing: A case study with great bustards. J Appl Ecol 38:458-471 Parrish FA, DeMartini EE, Ellis DM (1997) Nursery habitat in relation to production of juvenile pink snapper, Pristipomoides filamentosus, in the Hawaiian Archipelago. Fish Bull 95:137-148 Pascual MA, Adkison MD (1994) The decline of the Steller sea lion in the Northeast Pacific: Demography, harvest or environment? Ecol App 4:393- 403 Perlov AS (1996) Catch as one of the main causes of decline in the abundance of Steller's sea lion. Tinro Tsentr, Vladivostok (Russia) - 76- Conclusions Perry Rl, Dilke BR, Parsons TR (1983) Tidal mixing and summer plankton distributions in Hecate Strait, British Columbia. Can J Fish Aquat Sci 40:871-887 Pitcher KW, McAllister DC (1981) Movements and haulout behavior of radio- tagged harbor seals, Phoca vitulina. Canadian Field Naturalist Pontius RGJ (2000) Quantification error versus location error in comparison of categorical maps. Photogramm Eng Remote Sens 66:1011-1016 Pontius RGJ , Schneider LC (2001) Land-cover change model validation by an ROC method for the Ipswich watershed, Massachusetts, USA. Agric Ecosyst Environ 85:239-248 Porter MS, Rosenfeld J , Parkinson EA (2000) Predictive models offish species distribution in the Blackwater drainage, British Columbia. N Am J Fish Manage 20:349-359 Poulin M, Careau D, Rochefort L, Desrochers A (2002) From satellite imagery to peatland vegetation diversity: How reliable are habitat maps? Conservation Ecology 6:16 Raum-Suryan KL, Kenneth Pitcher, Donald G. Calkins, John L. Sease and Thomas R. Loughlin (2002) Dispersal, rookery fidelity, and metapopulation structure of Steller sea lions (Eumetopias jubatus) in an increasing and a decreasing population in Alaska. Mar Mamm Sci 18:746-764 Reder S, Lydersen C, Arnold W, Kovacs KM (2003) Haulout behaviour of High Arctic harbour seals (Phoca vitulina vitulina) in Svalbard, Norway. Polar Biol 27:6-16 Sato Y, Okabe T, Takebayashi H (2002) Prediction model of occurrence/non- occurrence of stream fishes in the Katsuura River, Tokushima Prefecture. Jpn J Ichthyol 49:41-52 Schneider DC, Harrison NM, Hunt GL, Jr. (1990) Seabird diet at a front near the Pribilof Islands, Alaska. Stud avian biol 14:61-66 Schneider DC, Payne PM (1983) Factors affecting haul-out of harbor seals at a site in southeastern Massachusetts. J Mammal 64:518-520 - 77- Conclusions Schneider LC (1991) The role of fluid dynamics in the ecology of marine birds. Oceanography and Marine Biology Annual Review 29:487-521 Shima M, Hollowed AB, VanBlaricom GR (2000) Response of pinniped populations to directed harvest, climate variability, and commercial fishery activity: A comparative analysis. Rev Fish Sci 8:89-124 Simpson JH, Hunter JR (1974) Fronts in the Irish Sea. Nature 250:404-406 Sjoberg M, Ball JP (2000) Grey seal, Halichoerus grypus, habitat selection around haulout sites in the Baltic Sea: bathymetry or central-place foraging? Can J Zool 78:1661-1667 Springer AM (1992) A review: walleye pollock in the North Pacific - how much difference do they really make? Fish Oceanogr 1:80-96 S P S S (1997) S P S S , Chicago, III. Stamps JA (1988) Conspecific attraction and aggregation in territorial species. Am Nat 131:329-347 Stanbury K, Starr R (1999) Applications of Geographic Information Systems (GIS) to habitat assessment and marine resource management. Oceanol Acta 22:699-703 Su Y, Sheng Y (1999) Visualizing upwelling at Monterey Bay in an integrated environment of GIS and scientific visualization. Mar Geod 22:93-104 Suryan RM, Harvey JT (1998) Tracking harbor seals (Phoca vitulina richardsi) to determine dive behavior, foraging activity, and haul-out site use. Mar Mamm Sci 14:361-372 Sydeman WJ, Allen S G (1999) Pinniped population dynamics in Central California: Correlations with sea surface temperature and upwelling indices. Mar Mamm Sci 15:446-461 The R Foundation for Statistical Computing (2004) R Trites AW, Donnelly LP (2003) The decline of Steller sea lions in Alaska: A review of the nutritional stress hypothesis. Mammal Rev 33:3-28 Trites AW, Larkin PA (1996) Changes in the abundance of Steller sea lions (Eumetopias jubatus) in Alaska from 1956 to 1992: How many were there? Aquat Mamm 22:153-166 - 7 8 - Conclusions Walker PL, Kennett DJ, Jones TL, DeLong R (2000) Archaeological investigations at the Point Bennett pinniped rookery on San Miguel Island. In: Browne D, Mitchell K, Chaney H (eds). US Department of the Interior, Minerals Management Service, Pacific OCS Region 770 Paseo Camarillo Camarillo CA 93010-6064 USA Watling L, Norse EA (1998) Disturbance of the seabed by mobile fishing gear: A comparison to forest clearcutting. Conserv Biol 12:1180-1197 Watts P (1996) The diel hauling-out cycle of harbour seals in an open marine environment: correlates and constraints. Journal of Zoological Society London 240:175-200 Yu S-L, Peters EJ , Stroup WW (1995) Application of logistic regression to develop habitat suitability criteria for sand shiner, Notropis stramineus. Rivers 5:22-34 Zweig MH, Campbell G (1993) Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine. Clinical Chemistry 39:561-577 - 7 9 - Appendix I Appendix I - ROC values and K n o values for models at varying probability thresholds Prediction of Rookery Locations - Mainland & Offshore Classification models Using K n o as an indicator of overall model success at predicting rookery locations, training the model using the western rookeries using a 1 nautical mile buffer, then applying a 99% cut-off produced the best-fitting model (Table A1.1). Generally - but not in all cases - a higher cut-off threshold resulted in a higher kappa value. Whether the model was trained using the western or eastern set of rookeries did not significantly affect the model fit, with the exception of the 50 nautical mile buffer distance, where training using the eastern set of rookeries yielded a better model fit than training on the western set (p<0.001). Logistic models Using the forward stepwise procedure, only the independent variables with the most explanatory power are selected for inclusion in the final model. In all but two of the model runs, all of the variables were selected for inclusion. When using the 1 nautical mile buffer distance, depth was excluded from the final model in both the western and eastern training sets. However, to facilitate a direct comparison with the supervised classification models, in these two instances, the depth variable was added back in to the logistic models. Interestingly, adding the depth variable back into the models markedly improved their performance versus the models without depth as an explanatory variable. Only three of the logistic models performed better than chance (according to ROC plots - Table A1.2)) at predicting rookery locations, compared with all of the classification models—50 nautical miles (west and east), and 20 nautical miles with eastern training sites (Table A1.3). - 80- Appendix I Table A1.1. K n o (Kappa) values for predicting rookery locations using rookeries as training areas for classification model. Values are given for each training set (west and east), buffer distance (1, 10, 20, or 50 nmi), and probability threshold (60-99%). The highest kappa value for each distance and training set is highlighted in bold. Threshold 50 nmi 50 nmi 20 nmi 20 nmi 10 nmi 10 nmi 1 nmi 1 nmi (W) (E) (W) (E) (W) (E) (W) (E) 99% 0.4674 0.7611 0.7580 0.8915 0.8712 0.9853 0.9761 95% 0.4084 0.4441 0.7576 0.5169 0.8862 0.6210 0.9791 0.8344 90% 0.4205 0.3873 0.7452 0.2671 0.8758 0.3403 0.9622 0.6157 80% 0.4039 0.6823 0.8452 0.9065 70% 0.5001 0.7903 0.8454 60% 0.1837 Table A1.2. ROC values predicting rookeries using rookeries as training areas. Bold values are above 0.5, and therefore represent models that perform better than a random or trivial model. ROC 50 50 20 20 10 10 1 1 R-R nmi nmi nmi nmi nmi nmi nmi nmi W E W E W E W E Classification 0.693 0.677 0.672 0.626 0.676 0.668 0.506 0.688 Logistic 0.514 0.663 0.491 0.661 0.448 0.447 0.483 0.478 Table At.3. Logistic model performance (Kno) predicting rookeries using rookeries as training areas at varying probability thresholds. Hatched columns represent models that did not perform better than random in ROC tests, and thus were excluded. Threshold 50 nmi 50 nmi 20 nmi 20 nmi 10 nmi 10 nmi 1 nmi 1 nmi (W) (E) (W) (E) (W) (E) (W) (E) 99% 0.3152 0.4381 0.7585 95% 0.1930 0.3651 0.7088 ' - " ' 90% 0.1513 0.3623 0.7055 80% t 70% 60% , " ' , The 20 nautical mile logistic model performed slightly better than the corresponding classification model, while the 50 nautical mile logistic models did not perform as well as the corresponding supervised classification models (p<0.001)—compare the highest (bolded) values in Tables A1.1 with A1.3. - 8 1 - Appendix I Prediction of Haulout Locations - Mainland & Offshore Classification models Actual haulout locations were compared against models trained using both haulouts and rookeries separately as training areas to test if conditions near rookery locations alone were sufficient to predict haulouts. Tables A1.4 and A1.5 show the K n o values for models generated by training on rookeries and haulouts, respectively. The same trend in model fits is generally seen, with better fits as the buffer distance decreases, although the model fits for the 20 nmi buffer distances were lower than the 50 nmi buffers. No training set (west or east) was consistently better at predicting haulouts across all distances, although the difference between kappa values was usually negligible. In comparing the best model from each distance, training using rookery sites alone did not produce a significantly better-fitting model (p>0.05) than training on haulout sites, except at the 20 nautical mile distance, where training on rookeries produced a better-fitting model than training on haulouts (p<0.001). At 50 nautical miles, training using haulouts produced a better-fitting model (p<0.001). Logistic models Only one of the logistic models trained using haulouts (50 nautical miles east), and two trained on rookeries (50 and 20 nautical miles east) performed better than chance at predicting haulout locations according to ROC plots (Tables A1.6, A1.7). Neither model outperformed the supervised classification models on the K n o score (p<0.001) (Compare Tables A1.4 and A1.5 with Tables A1.8, A1.9). - 8 2 - Appendix I Table A1.4 K n o (Kappa) values for predicting haulout locations using rookeries as training areas. Values are given for each training set (west and east), buffer distance (1, 10, 20, or 50 nmi), and probability threshold (40- 99%). The highest kappa value for each distance and training set is highlighted in bold. Threshold 50 nmi 50 nmi 20 nmi 20 nmi 10 nmi 10 nmi 1 nmi 1 nmi (W) (E) (W) (E) (W) (E) (W) (E) 99% 0.0713 0.1902 0.2765 0.2852 0.5291 0.5242 0.9547 0.9494 95% 0.1510 0.3127 0.2747 0.3023 0.5274 0.4525 0.9481 0.9810 90% 0.2704 0.3727 0.2753 0.2995 0.6108 80% 0.4155 0.3007 0.2773 70% 0.4604 0.3041 60% 0.4654 0.3480 50% 0.4651 0.3489 40% 0.3409 Table A1.5 K n o (Kappa) values for predicting haulout locations using haulouts as training areas. Threshold 50 nmi 50 nmi 20 nmi 20 nmi 10 nmi 10 nmi 1 nmi 1 nmi (W) (E) (W) (E) (W) (E) (W) (E) 99% -0.0734 0.3813 0.2642 0.2663 0.5214 0.5286 0.9562 0.9669 95% -0.1000 0.4436 0.2647 0.2683 0.5129 0.5257 0.9393 0.9393 90% 0.3872 0.4595 0.2599 0.2816 0.4936 0.4931 0.7553 0.9422 80% 0.4911 0.4644 0.1118 70% 0.4997 0.4618 60% 0.4830 Table A1.6 ROC values predicting haulouts using rookeries as training areas. ROC 50 50 20 20 10 10 1 1 R-H nmi nmi nmi nmi nmi nmi nmi nmi W E W E W E W E Classification 0.664 0.735 0.701 0.655 0.667 0.672 0.688 0.685 Logistic 0.474 0.637 0.484 0.670 0.484 0.481 0.421 0.424 Table A1.7 ROC values predicting haulouts using haulouts as training areas. ROC 50 50 20 20 10 10 1 1 H-H nmi nmi nmi nmi nmi nmi nmi nmi W E W E W E W E Classification 0.769 0.755 0.671 0.681 0.644 0.655 0.724 0.711 Logistic 0.458 0.541 0.495 0.482 0.486 0.488 0.424 0.431 - 83 - Appendix I Table A1.8. Logistic model performance (Kno) predicting haulouts using rookeries as training areas. Hatched columns represent models that did not perform better than random in ROC tests, and thus were excluded. Threshold 50 50 nmi 20 nmi 20 nmi . 10 nmi 10 nmi 1 nmi 1 nmi I nmi (W) (E) (W) (E) (W) (E) (W) (E) i 99% • * •, i 95% -0.0275 • 1 90% -0.0206 0.2312 • / ' ".[ 80% -0.0052 0.2328 ' • ••! 70% 0.0047 0.2310 i 60% 0.0118 \,> r\ • • 1 Table A1.9. Logistic model performance (Kno) predicting haulouts using haulouts as training areas. Threshold . 50 nmi 50 nmi 20 nmi 20 nmi 10 nmi 10 nmi 1 nmi T nmi I (Mainland & Islands) • (W) (E) (W) (E) (W) (E) (W) (E) 1 \ 99% lijiRililij^ •i 95% -0.1611 •j 90% -0.1227 * * ,"< i v - " V ' \ 80% 0.1027 70% n 60% Prediction of Rookery Locations - Offshore Only Classification Models Exclusion of the mainland to simulate avoidance of terrestrial predators did not generally improve model performance in predicting rookery locations; in six of eight cases, (50, 20, and 10 nmi, east and western training areas) models that included the mainland outperformed those that excluded it (p<0.001, except for 10 nmi west, p<0.01). There was no significant difference (p>0.05) between the mainland and offshore models at the 1 nmi distance. The same trend in Kno values was apparent with the offshore models: higher values were associated with smaller buffer distances. With the exception of the 50 nmi distance, where the western-trained model had a better fit than the eastern-trained model (p<0.001), there was no significant difference between models with the mainland mask applied. -84- Appendix I Logistic Models Only two of the logistic models exceeded chance when combined with the mainland exclusion on the ROC plots: 50 nautical miles east and west. In comparison with models that included the mainland, these two models did not show any increase in performance (p<0.001). Table A1.10 K n o (Kappa) values for predicting rookery locations using rookeries as training areas. Values are given for each training set (west and east), buffer distance (1, 10, 20, or 50 nmi), and probability threshold (40- 99%). The highest kappa value for each distance and training set is highlighted in bold. Threshold 50 nmi 50 nmi 20 nmi 20 nmi 10 nmi 10 nmi 1 nmi 1 nmi (W) (E) (W) (E) (W) (E) (W) (E) 99% 0.2112 0.2776 0.6660 0.6611 0.7768 0.8240 0.9844 0.9759 95% 0.2055 0.2522 0.6610 0.3311 0.8382 0.4802 0.9764 0.8026 90% 0.2285 0.1977 0.6471 0.0062 0.8220 0.1041 0.9580 0.5521 80% 0.2135 70% Table A1.11 ROC values predicting rookeries using rookeries as training areas. 50 50 20 20 10 10 1 1 nmi nmi nmi nmi nmi nmi nmi nmi W E W E W E W E Classification 0.619 0.595 0.608 0.538 0.637 0.571 0.583 0.638 Logistic 0.449 0.625 0.436 0.585 0.393 0.393 0.330 0.385 Table A1.12. Logistic model performance (Kno) predicting rookeries using rookeries as training areas. Hatched columns represent models that did not perform better than random in ROC tests, and thus were excluded. Threshold 50 nmi 50 nmi 20 nmi 20 nmi .10 nmi 10 nmi 1 nmi 1 nmi (W) (E) (W) (E) (W) (E) (W) (E)- 99% 0.2342 0.6643 95% 0.2367 0.6514 • ; . 90% 0.2417 \ 0.6463 80% 0.2460 ^ ' ' ' " " . 70% 0.2415 - 8 5 - Appendix I Prediction of Haulout Locations - Offshore Only Classification Models Applying the mainland mask to predictions of haulout locations also did not result in any marked improvement in model performance when using rookeries as training areas; again, in six of eight instances, models that included the mainland outperformed the offshore-only models (p<0.001). At the 1 nautical mile distance, there was no difference between the mainland and offshore models (p>0.05) (Compare Table A1.4 with Table A1.11). Table A1.13. Classification model performance (K n o) predicting haulouts using rookeries as training areas. Threshold 50 nmi 50 nmi 20 nmi 20 nmi \ 10 nmi 10 nmi 1 nmi 1 nmi (W) (E) (W) (E) (W) (E) (W) (E) 99% -0.2210 -0.0472 0.0171 0.0278 0.3338 0.3497 0.9724 0.9647 95% -0.0979 0.1354 0.0146 0.0519 0.3591 0.2398 0.9645 0.7940 90% 0.0792 0.2427 0.0171 0.0872 0.3526 0.1698 0.9463 0.5487 80% 0.2969 0.2434 0.0339 0.0209 70% 0.3673 0.2351 0.0947 60% 0.3787 0.1774 50% 0.3818 40% 0.3819 30% 0.3770 Similarly, when haulouts were used as training areas to predict haulout sites, models that included the mainland outperformed offshore-only models in six of eight instances (p<0.001, 50, 20, and 10 nautical miles with eastern and western training areas—compare Table A1.5 with Table A1.12). Again, at the 1 nautical mile distances, there was no difference between the models (p>0.05). When comparing the haulout models on the basis of training area (west vs. east), rookery-trained models outperformed haulout-trained models in three cases (p<0.001,50 nautical miles west, 20 nautical miles west and east), one haulout- trained model outperformed a rookery-trained model (p<0.001,50 nautical miles east), and the remaining models showed no significant difference (p>0.05, 10 and 1 nautical miles west and east). - 86- Appendix I Logistic Models None of the haulout-trained models, and only two (50 nautical miles and 20 nautical miles, both eastern) of the rookery-trained logistic regression models performed better than chance according to their ROC scores when predicting haulout locations (Tables A1.15 & 16). The 50 nmi rookery-trained logistic model with the mainland mask outperformed the corresponding model without a mainland mask, while the 20 nmi model did not perform as well as the unmasked model (compare Table Tables A1.8 and A1.17). Table A1.14. Classification model performance (Kno) predicting haulouts using haulouts as training areas. Threshold 50 nmi 50 nmi 20 nmi 20 nmi 10 nmi 10 nmi 1 nmi 1 nmi (W) (E) (W) (E) (W) (E) (W) (E) 99% 0.0062 0.3662 0.9874 95% 0.2497 0.1760 0.0126 0.0206 0.3440 0.3628 0.9548 0.9840 90% 0.2659 0.2113 0.0444 0.0237 0.3448 0.2712 0.9604 0.4732 80% 0.3766 0.3737 0.0964 0.1614 0.3231 -0.2330 0.9255 -0.1568 70% 0.3493 0.3786 0.1164 0.1913 0.8078 -0.2800 60% 0.3744 0.1354 0.1552 0.2330 -0.3384 50% 0.1392 40% 0.1371 Table A1.15. ROC values predicting haulouts using rookeries as training areas. 50 50 20 20 10 10 1 1 nmi nmi nmi nmi nmi nmi nmi nmi W E W E W E W E Classification 0.733 0.675 0.624 0.534 0.608 0.616 0.558 0.628 Logistic 0.367 0.595 0.411 0.569 0.432 0.444 0.428 0.420 Table A1.16. ROC values predicting haulouts using haulouts as training areas. 50 50 20 20 10 10 1 1 nmi nmi nmi nmi nmi nmi nmi nmi W E W E W E W E Classification 0.734 0.706 0.583 0.567 0.571 0.559 0.605 0.623 Logistic 0.354 0.460 0.392 0.414 0.435 0.447 0.405 0.415 - 8 7 - Appendix I Table A1.17. Logistic model performance (Kno) predicting haulouts using rookeries as training areas. Threshold ' 50 nmi (W) 50 nmi •(E) 20 nmi (W) 20 nmi (E) 10 nmi 10 nmi 1 nmi 1 nmi. (W) (E) (W) (E) 99% -0.2770 0.0172 95% "©IIP -0.2570 0.0104 90% -0.2441 0.0130 80% -0.2235 70% -0.2098 60% -0.1999 50% -0.1841 40% -0.1545 30% -0.1438 20% -0.0536 10% 0.0433 - However, the 50 nmi model required a very low threshold (10%) to achieve a positive score, whereas the unmasked model achieved positive scores at all thresholds below 70%, indicating that the unmasked model is more robust overall. This interpretation is also supported by the higher ROC score of the unmasked model (compare Tables A1.6 and A1.15) - 8 8 - Appendix II - shoretypes Appendix II Environmental Sensitivity Index (ESI) Plate 1. ESI shoreline type 1 - exposed rocky shores. (Picture courtesy OR&R, NOS, NOAA) - 8 9 - Appendix II Appendix II Appendix II Plate 6. ESI shoreline type 6a - gravel beaches. (Picture courtesy OR&R, NOS, NOAA) Plate 7. ESI shoreline type 6b - riprap structures. (Picture courtesy OR&R, NOS, NOAA) - 9 2 - Appendix II Plate 8. ESI shoreline type 7 - exposed tidal flats. (Picture courtesy OR&R, NOS, NOAA) Plate 9. ESI shoreline type 8a - sheltered rocky shores. (Picture courtesy OR&R, NOS, NOAA) - 9 3 - Appendix II Plate 10. ESI shoreline type 8b - Sheltered artificial structures. (Picture courtesy OR&R, NOS, NOAA) Plate 11. ESI shoreline type 9 - sheltered tidal flats. (Picture courtesy OR&R, NOS, NOAA). - 9 4 -

Cite

Citation Scheme:

    

Usage Statistics

Country Views Downloads
United Kingdom 7 0
United States 5 0
China 4 17
France 2 0
Japan 2 0
City Views Downloads
Unknown 11 0
Beijing 4 0
Tokyo 2 0
Mountain View 1 0
Salem 1 0
Ashburn 1 0

{[{ mDataHeader[type] }]} {[{ month[type] }]} {[{ tData[type] }]}

Share

Share to:

Comment

Related Items